NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
pub.dev
pub.dev, #2766 most downloaded on pub.dev
Cross-platform AI agent harness for Dart and Flutter: streaming provider adapters, agent loop with tools, session persistence, context compaction.
Latest stable
1.0.701
Last release 13 days ago
09 Nov 2026
Ships on a steady schedule
a new release about every 8 days
Nearly every release is documented
notes for 57 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
4 months old
404 releases · first in 2026
Drag across the ruler to compare, or tap a release to open it.
fix(gh-2075): the daily-publish macOS legs no longer depend on the package:sqlite3 native-assets hook's single un-retried build-time download. The bui
fix(gh-2075): the daily-publish macOS legs no longer depend on the
package:sqlite3 native-assets hook's single un-retried build-time
download. The build-macos.yml app-build and CLI-bundle steps now
pre-warm the hook's digest-checked cache via the ONE shared tool
(scripts/sqlite3_hook_prewarm.dart, gh-2074 — the same contract
build-mobile.yml + ci.yml already ride) after pub/config resolution and
before the build; a brownout fails there in seconds with a named
::error:: instead of killing a release build ~20 minutes in. The
interim bash seeder (scripts/seed_sqlite3_native_assets.sh + selftest)
is removed — round-2 review flagged it as a second implementation of the
same hook-cache contract; the lane wiring + the single-mechanism
invariant are pinned by test/ci_sqlite3_native_asset_seed_guard_test.dart.
fix(gh-2074): the daily-publish play leg no longer reds on a GitHub
release-download brownout. Three build-mobile.yml Android jobs died in
package:sqlite3's native-assets build hook with Bad state: Hash of downloaded file libsqlite3.x64.android.so is bef140a1… — the hook
(released ≤3.9.1; the lockfile pins 3.7.0, the root floats 3.9.1)
hashes the download WITHOUT honoring the HTTP status, so GitHub's 504
page (a constant hash, upstream simolus3/sqlite3.dart#435) failed the
pin check and the build exited 255. The assets themselves were intact;
ci.yml had solved this exact class for itself on 2026-09-21 by caching
.dart_tool/hooks_runner/shared, but build-mobile.yml never got the
protection. Cure: new scripts/sqlite3_hook_prewarm.dart — derives the
release tag + pinned hashes from the RESOLVED sqlite3 package (no
second pin), gates on the HTTP status before hashing, bounded-retries
with backoff, verifies SHA-256 before any cache write (tmp+rename),
skips idempotently on a verified cache hit — wired as the
Native-assets hook cache + Pre-warm sqlite3 hook assets steps in
the two Android build jobs (both roots, hooks-shared-mobile- keys)
and inside assert-builtin-skills for the root host asset (the step
that failed first). Guard test test/ci_sqlite3_hook_prewarm_guard_ test.dart pins the wiring (AC1-AC4); the script's byte-level behavior
and the live hook-reuse layout are pinned in
test/scripts/sqlite3_hook_prewarm_test.dart.
Rework (2026-11-09, the PR's own CI run): five reds, two root causes,
both fixed. (1) expandTargets accepted its hostOs/hostArch
injection but ignored it — expandTargets('host', hostOs: 'linux', hostArch: 'x64') resolved against the REAL machine and failed on the
ubuntu-24.04-arm Core-tests shard; a fully injected platform now never
consults the host (unit-provable on x64 and arm64 alike). (2) The
pre-warm was wired as plain dart run — which executes the whole
package graph's build hooks BEFORE the script's main(), so on exactly
the gh-2074 conditions (cold cache + brownout) the pre-warm died before
it could pre-warm (proven live; the pty-visual red was this same shape).
Every invocation is now dart --disable-dart-dev (no hook machinery)
and the script is self-contained (no package: imports — SHA-256
inlined) so that mode can load it. The same protection extended to
ci.yml: the lanes whose first hook consumer is a plain dart command
(test-core, integration-mock, hostile-env, fa-aot-build,
pty-integration-linux, binaries-pr, pty-visual — the last had no
hooks cache at all) now restore hooks-shared- + pre-warm the host
asset after pub get and before that consumer; the guard pins it as
AC6/AC7.
feat(gh-2059): browser-extension compaction parity with CLI — over-window
relief, /compact with visible receipts, and the shared under-window
guarantee ride the same CompactionHostWiring resolution as the CLI and
app (resolveExtensionCompactionWiring); receipts are durable across
panel attach/SW restart. Machine-review round 1: exact ladder pass
counts, busy-matrix routing (routeUserInput) for panel text and second
/compact during a pass, window settings from the ACTIVE stored source
with positive validation, string-encoded int merges.
fix(gh-2055): the daily-publish cli leg red-failed in pod install
(Build macOS x86_64, run 41778615815) because the vendored macOS pod's
XCFramework download block guarded on the downloaded-zip MARKER
(if [ ! -f wasm_run-v0.1.0.zip ]) instead of on the artifact the build
consumes. A fresh CI clone never has the marker, so pod install re-ran
the block on every build: it removed the WasmRun.xcframework the
build-macos.yml restore step had already placed (cache hit,
sha256-verified), re-downloaded the 92.7 MB asset with a bare
no---fail curl, and one transient reset (curl 56 at ~3 MB/92.7 MB)
left the Podfile nm -gU symbol check with no
macos-arm64_x86_64/libwasm_run_dart.a to read. The guard now keys on
WasmRun.xcframework/macos-arm64_x86_64/libwasm_run_dart.a — the exact
path the Podfile force_loads and nm-checks — so the block is a no-op
whenever the restore step (or a previous run) put the framework in
place, and the fallback download for a truly cold tree carries
--fail AND aborts the block at the curl step on failure
(|| { echo "wasm_run download failed" >&2; exit 1; } — a swallowed
status fell through to unzip/truncate residue and masked the root cause
behind the Podfile nm diagnostic; review round 1 on this PR), and the
podspec now assigns the backtick result and raises "wasm_run framework
fetch failed" unless $?.success? — Ruby ignores a bare backtick's
exit status, so pod install would otherwise have sailed past the failed
fetch to the downstream nm diagnostic (review round 2). Linted
behaviorally (stubbed curl, no network; POSIX-gated sandbox, plus a
ruby evaluation of the real fetch statement) by
test/wasmrun_macos_podspec_download_guard_test.dart.
fix(gh-2050): the daily-publish cli leg red-failed daily when its
macOS attach arrived after a sibling train had already PUBLISHED the
derived tag (the same-tag convergence race: mobile leg drafts + the tag
train publishes ~25 min after dispatch while the serialized self-hosted
mac builds need ~66 min — v1.0.685: draft 05:22:46, published 05:44:52,
mac leg attached 06:28:53). scripts/release_asset_gate.sh gains a
caller-declared pre-stage policy — RELEASE_GATE_PUBLISHED_SKIP=1 —
that turns the PUBLISHED+unattached-fa-* verdict into a loud whole-call
skip (::warning::, exit 0, nothing uploaded, no twin create, no flip;
all-or-nothing so latest-mac.yml can never land alone on a release
whose zips were skipped). The default stays the gh-2042 AC1 loud
refusal; only build-macos.yml's release-macos desktop attach opts in
(its DMGs/zips/manifest are clobber-refresher pre-stage assets, not
publish blockers), the CLI-tar/extension attaches keep the refusal (they
carry REQUIRED release assets). The stranded desktop feed self-heals:
the next daily leg pre-stages the next version's draft with fresh
assets.
fix(gh-2044): the auto-compact fragile-band escalation could fold the
ENTIRE visible projection when the remaining path fit inside
keepRecentTokens — the keep-recent walk defaulted to folding
everything when its budget never exhausted, swallowing the live tail
(the newest user ask and assistant turn included). Real-world trigger:
resumed subagent runs ended with no assistant message at all
(subagent produced no assistant message — CI cores shards 1/4 and
3/4, hostile-env, both red). _pickCheckpointRange now declines the
fold when the whole foldable path fits the keep-recent region and
still carries live records; all-marker paths keep the lossless
deterministic consolidation (issue #387). The escalation receipt also
reports the REAL summarizer call count (a #729 chunked fold makes N
calls, not 1), and the factory wires its run token into the engine's
cancelToken so a watchdog-shaped cancel reason (any
TimeoutException) can no longer dissolve a user abort into a
"summarizer returned no text" failed receipt while the run reports
success (#1085). Follow-up CI red on the same branch: the decline
moved tiny-transcript runs to the classic residual engine, where an
EMPTY summarizable region folded zero #729 chunks into a blank
summary that still landed as a junk CompactionRecord and a hollow
success: true through the no-work terminal — compactSession now
refuses that empty-region blank fold (gh-1743's failure-verdict
contract kept; a dialed blank answer over a real region still lands
as a record, issue #578 unchanged). Third CI red on the same branch
(flutter-app over-window relief, both shards): the whole-world decline
lived in the shared _pickCheckpointRange, so it also starved the
checkpoint LADDER — a tool-result-heavy transcript whose every ledger
token fits keepRecentTokens (the request estimate is far over the
trigger) lost its only summarizer-based relief, and the hosted app's
over-window guard ended the turn instead of continuing it. The decline
is now escalation-only: ladder passes run over the trigger, where the
fold IS the relief (the escalation still declines — semantic 5 — and
its pass body was extracted below the CRAP ratchet's 12.00).
(gh-2044 rework, UT-ESC-13)
fix(gh-2003): the Copilot picker filter is tier-aware. Individual-tier
/models responses omit model_picker_enabled entirely, so phase 1
dropped every such entry as «picker flag absent» and individual/OAuth
accounts still listed zero usable models. When the response comes from
the individual tier — the individual host (api.githubcopilot.com or
pi-mono's api.individual.githubcopilot.com spelling) or a
no-proxy-ep token (individual by design) — an ABSENT flag is now
selectable: absence ≠ an explicit false, which still drops
everywhere (embeddings, completion-only). Business/enterprise answers
keep the strict picker list unchanged, and the keep is never silent:
the fetch note names it («2 of 2 models kept via the individual-tier
fallback (picker flag absent)») through the existing fetch notes:
surfacing. pi-mono parity: the trigger mirrors allowPolicyFallback = baseUrl === "https://api.individual.githubcopilot.com" in
packages/ai/src/auth/oauth/github-copilot.ts.
fix(gh-1990): the Copilot /models dialect never answers a silent
empty list again. Every silent-empty exit names its branch through the
new fetchModelsForEndpointReport dispatch (ModelsFetchReport +
ModelsFetchDiagnostic): the token-exchange HTTP status (with the
OAuth device-flow / fine-grained PAT remedies for classic ghp_…
keys), the /models HTTP status or timeout (with the seat + per-tier
host hint), proxy-ep-less individual tokens aimed at a business/
enterprise configured host (the tier-misrouting hint), and the picker
filter with per-model drop reasons (picker flag absent, picker flag false, policy disabled, no tool_calls, endpoint unsupported) —
«12 models fetched, 12 filtered (picker flag absent)». A failed
primary credential path retries the other path ONCE before declaring
empty (dual-path fallback, gh-1606): an exchange-401 classic PAT tries
PAT-direct (the provenance note names which path answered), a
fine-grained PAT never falls into the exchange, an OAuth token never
becomes a direct PAT Bearer. Surfacing: /models prints a fetch notes: section per provider, the /provider copilot connect flow
prints the named failure before manual entry, the app's model
quick-select (EndpointModelsController.emptyReason) renders it under
the field, and the Copilot connect sheet shows the reason above the
manual model id. The legacy fetchModelsForEndpoint answers are
unchanged (parity); the filter criteria themselves are unchanged
(pi-mono parity — only the reporting is new). Review round 2: the
fetch notes: wiring covers the unsaved sibling paths too (the
active-catalog and env-keyed Copilot fetches surface the same named
failure instead of a silent seed fallback), the connect flow's
no models: line reuses ModelsFetchReport.reasonText (the last
hand-rolled message — hint join), and the ModelsFetchReport
docs scope the diagnostic contract to the dialects that name their
failure branches.
feat(network): channel payload sender block — app members' channel
messages carry {"text": …, "sender": {"name": …, "id": …}} (public
base64 JSON and keyed fanet1 plaintext alike), so DAP agents can
attribute the HUMAN author behind the fa_network relay's hub id. The
harness unwraps it tolerantly (raw text passes through), rides the
name as AgentMessage.fromName through the fabric, and the CLI
steering line greets people again: from Uladzimir (<relay id>) in #channel: hello instead of from 3cfaee1aa054fe6e…. The app falls
back to the payload sender for bubble names when the envelope carries
none; DMs also name hub peers from the whois cache. Legacy payloads
without the block decode exactly as before.
feat(gh-1956): reply auto-spill — oversized child reply / agent_message
payloads spill to an agent:// artifact automatically; the parent receives a
bounded excerpt + pointer + completeness marker with char AND byte counts.
Manual chunking degrades to a documented fallback, restorable across parent
resume via durable deliverable_artifact records, behind the
subagents.replyAutoSpill kill switch (default on; honored on the CLI and
the app host alike). Rework round: the salvage hook (output-truncated
deliverable calls) restores the durable ids before allocating its own
(no post-resume pointer shadowing), durability losses emit a
[deliverables] diagnostic line, the app host honors the kill switch, the
spill-reach gate is renamed spillAppliesTo (was spillApliesTo), byte
counts are encoded once per spill, and DeliverableRecord.artifact(bytes)
is optional again so external constructors keep compiling.
fix(#1947): schedule_message from the orchestrator no longer misroutes
into a finished subagent's mailbox. A background shell job started by a
CHILD registered its settle continuation inside the child's subagent
zone; when the job settled after the child's run ended, the CLI's
settle-wake (_onShellJobSettled → _startRun) started MAIN's reaction
run still carrying the finished child's activeSubagentId(), and every
identity consumer of that run misattributed it — the reported shape was
a self-reminder scheduled "in <session>/<child>" that auto-resumed the
completed child (duplicate-release risk). Three layers: the
ShellJobRegistry now binds its settle continuation to the
construction-time (boot) zone (_bootZone.run), both hosts'
scheduleSenderMailbox wiring trusts the zone scope only while the
child is genuinely in flight on the session executor
(TaskExecutor.isInFlight), and the tool response names the RECIPIENT
the record carries — an explicit to: used to be shadowed in the reply
by the sender's mailbox, which read as "the override was ignored" even
when it was honored. E2E repro drives the real AgentCli
(test/cli/agent_cli_schedule_recipient_test.dart) plus a registry-level
zone pin (test/tools/shell_job_settle_zone_test.dart).
fix(gh-1948): task_observe works on a RUNNING background subagent
instead of crashing with an uncaught SessionException on the
<parentSessionId>/<agentName> placeholder. Three legs: the executor
flushes each child's transcript incrementally at every turn boundary
(the JSONL exists and grows mid-run — also crash-resilience for
aborted children), task_observe resolves a placeholder handle's real
session file from the sessions tree (the same agent: subagent /
parent: / id: header scan the boot adoption uses) and heals the
handle onto the found path, and every read failure degrades to a
structured "transcript unavailable" / "no transcript available yet"
message instead of an uncaught tool error. The child-transcript reader
rides the transient-ENOENT retry (issue #427) — it reads while the
executor appends.
fix(#1941): the selected CodeMie project now rides EVERY model request.
The header is baked into Model.headers only by the connect-time
switches (gh-1809), so every other rebuild silently dropped
x-codemie-project — and the backend bills a header-less request against
the user's DEFAULT project (codemie-code header-injection.plugin.ts
contract): the reported symptom is limit-exhausted from the exhausted
default project despite picking a budgeted one. Rebuilds now re-resolve
the project from the saved entry through one helper
(projectHeaderForBaseUrl): the boot folder-state restore (JWT entries
previously booted header-less — the SSO cookie patch was the only boot
carrier), the session-resume re-apply (_buildRestoredModel), the
roles-chain failover (ModelRef.project, mirroring the #964 authHeader
leg), the roles-mode /model + settings-flow pins, and the generic
/provider switch. The models list (/llm_models, SSO + JWT + guided
setup + entry dispatch) and the quota adapter (dark until pinned) scope
themselves to the same project, so display and billing can never diverge.
Tests pin the header VALUE on the wire (real adapter over a fake
endpoint) for the restored and failed-over models.
feat(network): private networks created outside the app surface after
sign-in. GET /api/networks/mine (fa_network#16) is synced into the
wallet after OAuth/dev sign-in and both restore paths — best-effort,
silent on older backends (404) and any failure, and never destructive.
Synced memberships keep the server's name/class, preserve local
passwords/join facts, and land without a password (the endpoint never
returns one) until the first join. Entering such a network — sidebar,
⌘K/⌘3-9, or a restored selection — opens the join sheet prefilled with
the network id instead of a resume error or endless spinner; joining by
name now keys the wallet/session by the echoed canonical id and migrates
legacy name-keyed rows (password + channel keys) instead of duplicating
them. A malformed /mine row is skipped, not the whole list.
fix(gh-1931): one render per tool call on the non-TUI console. Line mode
and the headless -p log printed every executed call twice — a start row
(• bash · cmd, no elapsed) followed by the end row (✓/✗ bash · cmd 0s). The duplicate start-row print is gone from _onToolExecutionStart;
the end row (command, status glyph, elapsed, failure news) is the single
render per call. The TUI keeps its running card + settled card, start
bookkeeping stays (_toolStarts feeds the end row's detail/elapsed, the
waiting coordinator still arms per-call liveness), and persisted
artifacts (session JSONL, trajectory records) are untouched —
rendering-only change. Long foreground calls keep the gh-1055 liveness
reminders past waiting.toolLivenessSeconds (default 60s).
fix(gh-1929): esc during the pre-flight auto-compaction no longer crashes
the CLI. The compaction layer's deliberate abort (#1085) used to escape
_runPrompt outside its guarded body and reach the root zone handler via
an unlistened whenComplete chain — fa crashed: CancelledException +
crash.log + a lost session view. The pre-flight window (submit on an
over-window transcript) now translates the cancel into a user-cancel
outcome: neutral dim submission cancelled receipt (never error:), the
TUI composer hands the typed text back (sendInputText) so a resubmit is
one Enter away, headless exits 130 with the same receipt, and the next
prompt works (the funnel's local-trim recovery or a retried compaction).
Cancels DURING the run — in-flight relief, settle funnel — keep their
pinned loud-abort surfaces, and the post-run window keeps its
compaction interrupted receipt via the shared _renderUserCancel
helper.
feat(1923): stall-proof steering. Steering arriving while the loop awaits
an UNPRODUCTIVE provider submit (zero bytes, or silent past the steering
stall window) now cancels the attempt via a SteeringInterruptFire
cancel, injects the pending steering into the rebuilt request, and
resubmits — the #1916-class submit stalls can no longer queue user or
agent mail invisibly. Productive streams keep next-boundary semantics;
the interrupt never consumes a retry attempt, and the run latch re-arms
in place so watchdog replays and user aborts keep working. task_resume
caps its foreground wait (like task_send) and continues in the
background; video polling honors soft-yield checkpoints (remote render
never restarted, honest hand-back text) and media generation honors run
cancellation; SLO-breached undelivered child mail surfaces in
task_status (UNDELIVERED … past the 2s delivery SLO); the TUI prints
steering queued — provider stalled Ns, interrupting to deliver plus the
delivered ack.
feat(1916): image-heavy submits no longer livelock — byte-based
images.maxBytesPerRequest (default 8MB, layered onto the count cap;
/images names both drop reasons), legacy text-embedded data-URL
screenshots in history fold into the once-only registry instead of
replaying raw forever, stall watchdog windows scale with the outbound
payload (capped at 2x; explicit connect clamps untouched), abandoned
connect attempts answering during the retry backoff are adopted, and
repeated identical-payload stalls fire ONE actionable /compact +
image-budget advice. TUI stall banner + docs/images.md carry the
image-heavy playbook.
fix(gh-1912): Add an agent → DAP now turns a retired hub's unknown
enroll_create response into an actionable upgrade message instead of
exposing bad_frame. Independent enrollment never falls back to the
legacy same-name credential rotation endpoint, and existing credentials
remain untouched on failure.
fix(gh-1913): Add an agent is usable on narrow phone screens: scope and format choices wrap without per-character labels, enrollment controls stack at full width, and actions remain reachable by scrolling. Desktop keeps the segmented layout; phone/desktop light/dark goldens guard both.
fix(gh-1825): DAP agent enrollments carry a stable, independent
agentInstance through app/CLI bundles, persistence, direct boot and
reconnect. Independent same-name agents keep distinct identities and stay
online together, while legacy bundles retain their project-scoped identity.
feat(gh-1825): DAP private-channel sends now use self-describing
fanet1 frames (sender X25519 key in the frame, channel-bound AAD) via
fa_hub_client 0.2.11, so fa_network app members decrypt them without
a whois/senderKey round-trip. Inbound channel frames open fanet1
first and fall back to legacy dap/1 payloads. Verified live in both
directions between a headless DAP agent and the Flutter app.
feat(gh-1825): the network offline banner now has a Reconnect now
action which restarts the socket and reloads the channel/member
snapshot. Fresh sockets also clear a stale offline flag (the server
sends no historical drain to a socket that was away during recovery).
AgentNetworkController.start stops after disposal instead of
notifying a dead controller.
feat(gh-1825): the roster sheet shows access-scope badges (whole network vs N channels, decoded additively from Member.access in snapshots and presence.changed) and owner/admins can drop an agent (DELETE /api/networks/{id}/agents/{agentId} with confirmation; the server tombstone keeps it out of future snapshots). Guests and store members are removable too via DELETE /api/networks/{id}/members/{id} (sessions revoked server-side); owner/admin/self rows stay protected.
feat(gh-1825): channel/network management in the app — rename a channel (PATCH), delete a channel or an owned network (REST DELETE with confirmation and state/navigation cleanup), and unnamed CLI/hub-created channels show a short friendly label instead of the raw 64-char id.
feat(gh-1825): pasting a DAP invite no longer demands a passphrase on macOS — importing NEW channel secrets auto-provisions a random passphrase into the per-channel Keychain item (decryption of existing envelopes never auto-generates; env/prompt/headless behavior unchanged), and passphrase warnings now name the concrete recovery (re-paste the invite to re-encrypt, or set DAP_KEY_PASSPHRASE).
feat(gh-1825): the network online counter opens a live-first roster sheet (names, member class, presence), and the channel composer offers roster-filtered @-mention autocomplete — the draft keeps the literal @displayName for DAP agents and the envelope carries the resolved member ids. On the CLI side the DAP channel notice names the agent's own identity (display name + agentId), and inbound channel mail that @-mentions that name (case-insensitive, word-boundary — no email or @name2 false positives) is marked "[mentions you — you are @name]".
fix(gh-1825): the app's network roster no longer reads as permanently empty ("0 online"): fa_network sends roster.snapshot with an OBJECT payload ({members: [...]}) and the app only accepted a bare list, so every snapshot silently decoded to zero members. Both shapes are now accepted, with a production-shape regression test. The DAP paste hint names web alongside CLI/macOS/iOS (plain-web join is supported).
fix(gh-1825): resumed CLI sessions rejoin their DAP channels and receive channel mail again — session replay now re-arms membership synchronization (plugin session-ready listeners on boot/resume and in-process switches) and idle channel probes complete deferred joins before inbox peek/drain. The post-replay sync also dials the membership's OWN hub with its enrolled credential before joining: a resumed session-scoped agent used to stay on the zero-config local URL forever, and because a disconnected fabric probe short-circuits before the sync, the remote dial never happened — the session-ready listener now drives it directly (throttled to one attempt per 10 s). App channel history no longer flashes raw payloads on first open: decoding waits for startup's channel listing instead of caching an undecoded guess until a manual switch.
feat(gh-1825): the fa app consumes the same DAP enrollment block as the
CLI — paste it into the agent chat composer (macOS/iOS/Android; the
smart-paste .txt chip path included) and the channel join happens
without the block ever reaching sendText, analytics, or session
records. Native joins ride the live agent-fabric hub repository:
per-project scoped identity, invite pub/priv import, keyless public
channels, reconnect-persisted membership in a dedicated secure
enrollment store (Keychain/Keystore where available). Harness additions:
HubMessagingRepository.joinChannel/waitForConnection with
rejoin-before-drain on reconnect, keyed channel decrypt
(hubDecryptPayloadWithKeyPair), public showcase payloads, and
HubBearerTransport (native Authorization header or the browser
?token= equivalent; authed dials never stringify transport errors
into logs). Plain web joins through the browser
WebHubTransport (?token= bearer equivalent, IndexedDB identity);
the browser extension rides the service worker's existing DAP client
(hub.join op), never a panel-local second connection. Add agent now defaults to the
DAP tab with a paste-into-chat hint; web gains ai-native provider
sign-in through the existing popup/callback OAuth flow.
fix(fa_network#4): the app inserts hub agents arriving via
presence.changed into the roster (additive displayName/class fields)
instead of dropping them, so the online counter no longer sticks at 0
when an agent comes online after roster.snapshot; legacy frames
without those fields create no phantom members. roster.snapshot/REST
contents are logged for decisive presence debugging.
feat(gh-1825): /network + scoped DAP channel membership — join by
paste (invite url / url+keys / bare channel id / whole fa dap import
command) with session/project/global scopes, live join without
restart, per-scope leave + copy invite, and the composer-paste
shortcut into the same wizard. Session scope rides a
dap_channel_membership custom record (leave tombstones, branch
resume); project scope is .fah/dap.json (committable, dapkey1
PBKDF2+ChaCha20-Poly1305 encrypted keys); global keeps the legacy
~/.dap files. Headless CI: DAP_CHANNELS_JSON (+DAP_HUB_URL,
DAP_CLIENT_SECRET, DAP_KEY_PASSPHRASE) joins at boot with zero
prompts, plaintext or encrypted privs mixed. Passphrase chain:
env → per-channel macOS Keychain item → one prompt with save default.
feat(gh-1844): media slots never fall back to the main connection
implicitly — SDK/CLI, app store, and the JS bridge resolve a slot
without an override as UNCONFIGURED with an actionable
"configure the slot" error; the main connection serves a slot only via
the explicit per-slot useMainConnection: true opt-in (YAML
models.slots.<slot>.useMainConnection / app's "Same as main
connection" pick, preserved through media_models.json). Image and
video generation accept multimodal input (text + reference images:
sandbox paths, http(s) URLs, base64) on capability-declaring
OpenRouter endpoints (input_references + declared/probed
input_modalities gate; unsupported endpoints fail before any network
call); the OpenRouter video job API is a first-class video dialect
(submit → poll → mp4). New decide tool: structured decisions via
OpenRouter's Decisions API (/v1/systemone) — bool/choice/score
questions with calibrated probabilities, structured error results, no
streaming.
feat(gh-1834): the FinalizeGate contract counts a parity/conformance
claim as verified only with negative + combination self-checks — new
numbered step 2.6 in prompts/cli/finalize_gate.md (refused
incompatible/mutually exclusive options, empty/corrupt inputs, boundary
values, must-NOT-work cases, alongside happy paths; a blanket
"everything matches" is not verified). Universality: abstract example
only, no benchmark/task/flag names in the prompt; the step is
conditional, so tasks with no conformance claim see no gate change.
Ledger/verdict wire formats unchanged; validated end-to-end by the
scripted IT-4 replay (real CLI + real bash) whose persisted
task_ledger carries the rejection rows.
fix(gh-1832): release v1.0.635 published without fa-extension.zip — the
landing-page download 404'd. The extension's only producer (the daily
build-macos.yml build-extension leg) derives its tag as "latest+1 at its
own dispatch moment"; on 2026-10-24 the v1.0.635 tag was cut by
auto_release before those legs ran, so the extension (and the DMGs)
landed on an orphaned v1.0.636 draft while the tag train published
v1.0.635 — every leg green, nothing asserted the published asset set.
Now scripts/release_required_assets.txt is the single source of truth,
scripts/assert_release_assets.sh gates the release-provenance publish
flip (a missing asset leaves the release a draft and fails the run
loudly) and a new release-asset-smoke ci.yml job post-publish probes
every landing-page release link + the full required set (red on non-200).
Remediation: the exact-tag fa-extension.zip (built from the v1.0.635
tree by run 40064876732) was uploaded to release v1.0.635 — the landing
link serves 200 again.
fix(gh-1820): the arm64 DMG leg's stale size floor is reseeded onto the
shipped v1.0.634 shape — run 4004758753 (post-#1815 SPM-lane adoption)
passed the min-platform gate and the full build, then failed the
size_manifest.py +5% budget on App.framework/App +5.1%
(25,020,112 → 26,284,144, the accumulated Dart AOT of #1793/#1794/
#1796/#1805/#1806); the committed scripts/size_budget/fa-macos-arm64.tsv
was still the pre-#1096-diet floor (TOTAL 210.5 MB vs the artifact's
116.4 MB), so the documented intentional-growth remediation applies:
the floor is now the exact --reseed output for the failing run's
artifact manifest (26 rows ≥ 64 KiB + TOTAL 116,386,369, from the run
log's printed table — the zip never uploaded past the gate), locking in
Fa 23.4→18.6 MB, quickjs_bridge 950→844 KB; new
test/ci_1820_macos_arm64_size_floor_test.dart gates the shape green
through the REAL size_manifest.py check and pins the committed floor
byte-identical to a --reseed round-trip.
fix(gh-1814): the desktop DMG lane ADOPTS the SwiftPM lane — Flutter
3.47.7 removed the FLUTTER_SWIFT_PACKAGE_MANAGER disable knob
(flutter/flutter#193978), so the gh-180 pure-CocoaPods pin became a no-op
and the min-platform gate (which forbade any SwiftPM state) failed both
DMG legs on every pinned-3.47.7 run (39991663029, 40022900373); the
#1813 escape (pin 3.47.4) is unreachable on the ~60kB/s runner link.
scripts/assert_macos_min_platform.sh now ASSERTS the SPM lane instead:
SwiftPM state expected after flutter build macos --config-only,
generated package floor ≥ the app floor (the silent bump must not skip —
the original gh-180 killer), no SwiftPM product above the app minimum,
vendored inappwebview manifest floored at 14.0 if it ever reappears.
Floors were audited and need no raising: Runner pbxproj 3×14.0, Podfile
:osx, '14.0', generated floor 14.0 on 3.47.7, vendored podspec 14.0
(its Package.swift stays deleted — the plugin rides the CocoaPods half of
the hybrid lane via Podfile.lock). The dead env pin left the build-macos
job env; the SDK pin reverts #1813 back to 3.47.7. The ci.yml macOS debug
leg gets the same treatment (gh-1814 review): its
FLUTTER_SWIFT_PACKAGE_MANAGER=0 pin — a silent no-op on the floating
3.47.x stable it resolves past the knob's removal — is removed and the
lane comment corrected; the leg rides SPM enabled as it already has been
(run 399572012 green).
fix(gh-1805): modifier chords no longer type their base letter into the
composer while the slash menu is open — the vendored dart_tui decoder
puts the base letter into the keystroke's text for control bytes
(Ctrl+S → ctrl+s with text s), and the slash-menu edit cluster was
the one insert surface missing the _isCommandKeystroke guard, so a
chord pressed mid-menu poisoned the typed prefix (/memory landed as
/smemory). Plain letters and Shift+letters still type and filter.
feat(1515): ONE model-limits settings surface — the settings hub's
standalone "Model parameters" row is retired and bare /model-edit
(TUI) aliases into /settings → Model capabilities, pre-focused on the
active model. The caps flow gains the model-edit preset ladders +
Custom… free input for context window / max output (a preset pick
writes models.overrides AND rebuilds the active model on the spot —
the old flow's live-mutation semantics, now durable), every field row
shows the current EFFECTIVE value with its source layer (pinned /
role override / catalog), the thinking row is an omp-style selector
(only levels the model supports, per-rung budget metadata, current
level preselected), and the hub row describes the effective summary
(200K ctx · 16K out · thinking off) instead of a pin count. The
scripted /model-edit <field> <n> form and the line-mode limits print
are unchanged.
feat(gh-1746): opt-in session disk retention — sessions.retention in
~/.fah/config.yaml caps the sessions root and sweeps whole session
units (transcript + sidecars) oldest-last-used-first at boot, exempting
live, named, young, and the current session; nothing is EVER
auto-deleted without the section (byte-identical legacy boot). Adds
fa session clean [--dry-run], the /sessions + fa session list
usage footer under a cap, and header-cwd/path-escape guards so a
planted session header can never steer a delete outside any root.
fix(gh-1763): a bash tool command running broad process kills can no
longer signal its own shell or the harness. Scripts are delivered from a
per-exec private temp file (sh -c '. "$1"' sh <script>) so the wrapper
argv never carries the script body (pkill -f <pattern> cannot
self-match its own sh -c; $0/$#/exit codes/set -e/stdin stay
identical), a PATH shim for pkill/killall plus a kill() prelude
guard exempt the harness pid chain + process group (exported as
FA_PROTECTED_PIDS) and the invoking shell from the child's own kills
while legitimate victims (stuck test runners, background servers) are
still terminated, and headless SIGTERM aborts now log
fa: SIGTERM received … so a child-issued kill never masquerades as a
user Ctrl+C (exit code contract unchanged: 130 graceful, 143 escalate).
fix(gh-1778): the release gate no longer reds a fully-green shard when
GitHub's cache service degrades. v1.0.621's Core tests (shard 2/4) was
killed by its own timeout-minutes: 12 with 2535/2761 tests green — both
cache steps had stalled ~133 s on keys that existed (degraded service,
not eviction), cold pub get + native-asset hook builds delayed the
first test to ~5 min, and the shard needed ~14 min against a budget
sized for the warm ~8.8 min wall. The job budget is now 20 min (sibling
test-leg band), and both test-core cache steps are best-effort
(timeout-minutes: 2 + continue-on-error: true) — a failed restore is
slower setup, never a red shard. test/ci_core_shard_timeout_guard_test.dart
pins the budget and the best-effort cache contract (RED-first on the old
YAML).
feat(#1761): the 32KB project-context budget never cuts silently again.
A file truncated to its head excerpt now carries an extended
<!-- From: path — N bytes, only the first M shown --> marker plus an
explicit MUST-read-the-full-file instruction (offset/limit/ranges,
prompt copy in prompts/context/project_context_read_full.md); a file
dropped entirely leaves a one-line tombstone naming its path instead of
vanishing; and the user gets a dim boot line per truncated/dropped file
(CLI line mode prints directly, TUI re-prints after the banner, the app
logs through its AppLog boot-notice seam). New projectContext: config
section: maxBytes (default 32768) — 0 disables the section entirely
and reads no files (fast boot); unknown keys are tolerated with a note.
Leaf-first allocation is unchanged.
fix(gh-1757): the Android build no longer re-enters the sdkmanager
download roulette for CMake. v1.0.619's build-android leg died at
:jni:configureCMakeDebug[arm64-v8a] — the :jni module (transitive
jni package via path_provider_android) wires externalNativeBuild.cmake
with no version, so AGP auto-installed its default cmake;3.22.1
mid-configuration and one corrupt archive ("Archive is not a ZIP
archive") red'd the leg, the aggregate Quality gate, and the
Advance-vpinned-marker job. The gh-1655 pre-install action now also
pre-installs cmake;3.22.1 with bounded retry + idempotent skip (same
hardening the NDK path already carries); the guard test pins the new
AC6 assertions and the AGP-version coupling (AC7).
feat(#1750): per-model image caps — the outbound request never carries
more unique images than the target model accepts. Model gains an
optional maxImages (catalog-seeded: gemini-4-argon: 10, measured);
images.perModel.<modelId> config overrides it (strict images:
section, user truth wins); the registry resolves
perModel > catalog > images.maxPerRequest (first present wins, the
issue-#1554 shape). A Google-shaped too many images … maximum allowed for model <id> is <n> 400 is classified (image_overflow.dart, the
overflow.ts twin), latches the learned cap for the (provider, model)
pair for the run, and re-sends ONCE through the registry at the
smaller cap — current-first-then-newest drops with the existing
never-silent notices naming the learned cap; a second refusal
surfaces exactly as today. JSONL is never rewritten;
images.registry: false keeps the byte-identical legacy shape
including the recovery-off behavior.
feat(gh-1760): oversized user-message intake spills to a file instead of
blowing the context window. Any incoming user message — headless
-p/positional, TUI first/mid-session, steering/follow-up drain,
external inbox — whose text passes the new relative threshold (a
spills.userFraction of the live window budget minus the compaction
reserve, floored by spills.userThreshold bytes) is redacted and
written verbatim to .fah/spills/<sessionId>/user-<n>.md (the #678
layout) from one choke point (Agent.userMessageGuard); the model
receives a bounded head+tail preview plus a read-in-parts instruction
riding the replacement USER message (copy in
prompts/spill/user_read_instruction.md). Images stay inline, a
refused spill write degrades to today's inline behavior, and
spills.enabled: false / userThreshold: 0 / the new
FAH_USER_SPILL_THRESHOLD env knob (bytes) keep the legacy shapes.
fix(gh-1743): a structured compaction that hits its total-budget
watchdog (the factory's 300s totalBudget) no longer aborts the
live session. AutoCompactorFactory.run() now applies the
mechanical local-trim fallback (keepRecentTokens, pairing-repaired,
[context trimmed locally] marker — the same valve the classic
pass ladder ends on) instead of failing fast; the honest
both-roles-failed verdict is kept only when nothing is droppable.
fix(gh-1636): auto-compaction stops dumping the internal checkpoint
summary into the console — the default report is ONE marker line
(● auto-compacted · engine: N → M tokens (F freed · P%)); the records
line + fenced summary moved behind the new tui.showCompactionDetails
opt-in (default false), and the manual /compact still prints the full
block. The headless log face no longer streams the summarizer deltas by
default (the gh-1433 E4 live stream is part of the same opt-in; the
gh-1430 … reasoning liveness line still covers the window). Nothing
is lost — the summary stays in the session file, reachable via
compact_expand.
fix(test): gh-1732 — the wrap_ansi_line_perf_test sustained-contention
validation leg spikes 3 of the 5 doubled runs per attempt instead of 4
(CI run 39546301132, core shard 2/4: the old shape left exactly ONE
real-wall-clock-timed doubled sample per attempt, and real jitter on
that single sample — the very shape being modeled — inflated all three
attempts' min ratios to a 3.30x best against the 3.2 bound while the
real 500k growth legs on the same runner stayed green). 3-of-5 is the
documented live-red shape ("≥3 of the 5 timed runs"): median-of-5
[c,c,s] is still a spiked value, so the leg still proves the
pre-gh-1532 median harness fails, and the min harness now rides two
clean doubled samples per attempt.
fix(#1705): CLI fresh-install boarding asks the provider choice FIRST —
the same list the /add-provider picker shows (presets, Custom last) —
and only Custom continues into the api-type wizard; a preset opens its
own flow with the api type prefilled. Previously boarding opened the
custom wizard directly, cold-asking "api type" before any provider was
chosen.
feat(providers): gh-1704 — first-class Codex/ChatGPT credential export,
CI refresh action, and Provider Queue integration (Copilot-export
parity). fa provider export <name> [--check] resolves the saved
ChatGPT login's OAuth blob from its secure-store slot, refreshes it when
stale, and prints the shell lines (OPENAI_API_KEY + OPENAI_BASE_URL
= the Codex backend — these tokens never serve api.openai.com); the
refresh token never prints and --check reports credential health with
the SSO caveat. /provider chatgpt import [path] adopts an existing
codex login session (~/.codex/auth.json) into the same slot — no
browser round-trip. .github/actions/chatgpt-codex-auth mints a fresh
access token from a CHATGPT_REFRESH_TOKEN secret in CI (direct curl,
no third-party action); invalid_grant fails fast and loudly.
providersQueue: entries speak chatgpt-codex: an omitted apiKeyEnv
resolves the Keychain slot automatically (in-memory refresh), a new
refreshTokenEnv: field mints the initial access token on slot
activation, and a raw Bearer token in apiKeyEnv is tolerated by
chatgpt_codex.dart (no more FormatException — proactive/reactive
refresh degrade, a 401 surfaces as the auth death the queue fails over
on).
feat(bench): gh-1681 — leaderboard submission readiness. ATIF
trajectories: new exportAtifJson whole-snapshot exporter
(lib/src/trajectory/atif_export.dart, ATIF-v1.7 — the schema harbor
0.23.0's flagship adapters emit) behind a headless fa trajectory export [--format atif] [--output path] verb; the harbor adapter runs
it as a bounded best-effort exec after the agent phase on every
terminal path, so <trial>/agent/trajectory.json reaches the hub with
the other agent logs (the leaderboard static analysis requires one per
rewarded trial) and result.json metadata carries an
atif_trajectory present/missing marker. The golden fixture
(bench/harbor_fa/fixtures/atif_sample.json) is pinned Dart-side and
cross-validated against harbor's pydantic Trajectory +
validate_trajectory in test_atif_contract.py. Workflow: a
submission: true dispatch input turns the submission rules into hard
setup preflights (full task coverage, ≥5 attempts/task, no
agent-timeout override, pinned dataset) and makes the new per-shard
ATIF presence gate fail hard; default runs gain only a warn-only scan
(artifact flow unchanged). Runbook: docs/leaderboard-submission.md
(pinned dataset + @latest content-hash recording, quota-window runner
policy, lb CLI steps).
fix(headless): gh-1589 — the headless session becomes a SEALED artifact:
no session-JSONL writes land after the run's agent phase ends. New
Session.seal() (state in the writer, no file-format change): raised in
runHeadless' finally and the REPL teardown as the LAST session action
(close order: drain → usage fold → seal → exit, one implementation both
paths — REG-2), it makes every writer (_persistMessages,
shell_job_registry snapshots, compaction, any appendCustomEntry,
leaf moves) drop late appends with exactly ONE stderr warning per
writer class — never a throw into an unhandled zone; the exit code
stays the turn's terminal outcome (issue #413 semantics untouched). The
settle→reaction path is sealed with the file (I2): the drain wrap-up
raises a fence (_reactionRunsSealed) so a background-shell settle or
task completion arriving after the final answer's drain — including the
registry's one-turn-deferred listener racing _finishHeadlessDrain's
check-to-return window (the 65gGW7b 10:13:42–10:14:00 burst) — does
board bookkeeping only: no _startRun, no model request, no record.
An honest drain's records (final answer → seal) stay legal (E7); the
seal is per-run state, so resume/replay/session_repair are unchanged.
Adapter side (bench/harbor_fa, I3/AC4+AC7): run()/perform_task
returns only after the fa process exited on EVERY path — the ladder
kill became terminate-and-reap (SIGTERM → bounded exec await → SIGKILL
escalation → pgrep confirm), and the external-cancel paths (harbor's
own wait_for firing first, the no-knobs exec, external-cap-only mode)
now fire the kill + bounded reap before propagating, with a loud
agent-phase exit guarantee degraded stderr note when fa survives both
signals (never a silent proceed, never an unbounded hang).
fix(bench): gh-1643 — a trial killed mid-stream no longer reports 0/None
usage. The fold read only FINALIZED assistant messages, so a run killed
before its first assistant message (the costliest rows) folded to silent
zeros. Both folds now price a dangling model_request_summary at its
KNOWN request-side volume (system prompt + outbound message chars at the
shared chars/4 rate, estimated input; output unknown) and mark the usage
incomplete instead: UsageChainScan.usageIncomplete → usage.json
(usageIncomplete: true, written only when set so completed folds stay
byte-identical), SessionUsage.usage_incomplete → the tb adapter's
usage-guard.json row and harbor's agent_result.metadata, and the
post-mortem stamps usage_incomplete on every results.json row it
patches from a partial fold. Harbor also exports the run outcome class
(agent_result.metadata.failure_mode: agent_timeout self-kill /
external_kill / completed / crashed / never_started) so downstream
classification stops relying on exception.txt archaeology; a summary
superseded by another summary is flushed instead of silently dropped on
both sides. Never-started trials keep the gh-1638/gh-1576 audit-only
precedence (no fake usage).
fix(stuck-calls): gh-1632 — a stuck tool call in an interactive session now
escalates like everywhere else. effectiveStuckTool() defaulted every
non-headless host to StuckFollowUpMode.advisory when no explicit
agent.stuckTool: was configured, so the fa 1.0.580 incident session (a
dead git fetch in the REPL) fired its tool_stuck flag at the 300 s
threshold and then hung for 27 more minutes — advisory mode never cancels,
retries, or converts; only the owner's manual steering broke the deadlock.
The autonomous default (cancel → retry once → background-convert, then a
session-visible escalation) now applies in EVERY mode; advisory remains
an explicit opt-out (agent.stuckTool.followUp: advisory), never an
implicit default.
fix(bindings): gh-1624 — the gh-1570 binding preflight no longer kills
translating-proxy factories. _validateModelEndpoint's vendor-prefix
verdict was fatal on ANY endpoint, so FA_PROVIDER_TYPE=anthropic +
bifrost's /anthropic path + zai/glm-5.3 exit-64'd every boot on
1.0.576 (the prefix is a routing key on a translating proxy, not a
dialect declaration). The vendor-prefix check now fires ONLY on the
endpoints that carry a dialect verdict — the three family-exclusive
canonical vendor endpoints (api.anthropic.com, api.openai.com,
generativelanguage), mirroring the family verdict's scope: those keep
the gh-1570 fail-fast, ROUTERS (OpenRouter/Copilot/DIAL/kimi — the
prefix is a routing key there too; OpenRouter's canonical ids are
anthropic/claude-*) and unknown/private endpoints carry the mismatch
as a ProviderModelBinding.warnings entry, never exit 64. The warning
is rendered on EVERY resolve path (gh-1624 review): the FA_PROVIDER_*
boot notice, the provider queue (per entry at boot, on queue reload,
and in /provider queue test), and the --model/--base-url flag
revalidation — a rider never boots silently.
Breaking-change note for 1.0.576 upgraders behind proxy
factories: the migration is the wire: FA_PROVIDER_CONFIG key —
documented now (README env-preconfig section,
docs/agents/providers.md §Translating proxies, --help) with its
caveat: the override engages only when the declared wire DIFFERS from
the provider record's own kind (record anthropic + wire: anthropic
still runs the checks). REG matrix (AC1/AC2/AC3) pinned in
test/model_roles/provider_model_binding_test.dart; AC4 IT drives a
bifrost-shaped loopback proxy through the production anthropic adapter
(test/model_roles/bifrost_proxy_binding_test.dart).
gh-1596: the first-run wizard (site/setup.sh) no longer crashes or
corrupts input under curl | sh on dash (Ubuntu sh) or busybox ash
(Alpine): scripts/gen_installers.dart now renders the provider/mode/
approval menus from the live provider catalog into plain POSIX sh — the
bashisms that killed dash (printf -v, read -p) and silently corrupted
busybox ash are gone; the hidden API-key prompt uses stty -echo + a
by-name read (dash has no read -s) and the key never rides printf
argv; EOF/Ctrl-C mid-wizard aborts non-zero with no half-written
~/.fah, and non-TTY runs print guidance and exit 1. The wizard's temp
config is mktemp-created (no predictable, symlink-followable fixed
path). A new path-triggered installer-portability CI leg proves the
script under dash + busybox ash (ubuntu) and bash (macOS) on every
site/** or generator change.
fix(cli): gh-1572 — fa boot no longer rewrites ~/.fah/config.yaml
from a lossy config. persistConfig (bin/fah_runapp.dart) built a
fresh CliConfig carrying only some fields and persistConfig() ran
unconditionally at boot, so every launch silently dropped the user's
spills:, subagents:, jobs:, headless:, agent:, quota:,
tui:, trajectory:, waiting:, power.hold and auto_update
settings (first reported on 1.0.547). Three cooperating fixes: the
boot save is now saved.copyWith(<live overlays>) so every parsed
section rides along by construction; CliConfig.copyWith replaces the
drifted hand-rolled withCustomProviders field list (itself dropping
spills/subagents/waiting/jobs/agentMode/misuseBreaker/streamThinking
inside saveCliConfig's merge step); the _diskPreservedSections
disk backstop covers every known top-level key. spills: renders in
toYaml (new SpillsConfig.toYaml) and agent.misuseBreaker: false
persists too; an unchanged save no longer rewrites the file at all.
Rework round: the boot snapshot re-syncs from disk after every
sanctioned mid-session settings write (onUserConfigFileChanged — a
cleared context cap, an auto_update key drop or a global compaction
pick is no longer resurrected/overwritten by the next save in the
same session); validateAgentSection (the strict fa config check /
config set / settings-hub mirror) accepts the emitter's
agent.misuseBreaker and agent.stuckTool keys so fa config check
never false-fails on fa's own output; clearing the context cap drops
only the cap key instead of the whole agent: block (the sibling
mode:/misuseBreaker: keys survive); and the providersQueue:
block (no typed renderer) joins the disk backstop so a save can never
silently delete the user's queue. Rework round 2 (merge + guards):
validateAgentSection's per-key switch became the _agentKeyValidators
table (the two gh-1572 keys pushed the switch form to complexity 13 —
over the repo's CRAP ratchet; the table keeps the mirror complete at
complexity 4), and the last two direct USER-config writers —
_writeTtsrSection and _writeMcpSection — fire the re-sync hook, so
a mid-session ttsr rule toggle or mcp server edit survives the next
persist too (the writer roster in settings_flow_config_sync.dart is
now exhaustive).
gh-1606: the copilot provider accepts fine-grained PATs
(github_pat_…) directly — PAT mode skips the copilot_internal/v2/token
exchange entirely (GitHub 404s every PAT exchange) and sends the raw PAT
as the Bearer against the configured Copilot base URL with the official
Copilot CLI's first-party identity headers
(copilot-integration-id: copilot-developer-cli,
user-agent: copilot/<fa-version> (<os>) term/unknown, installed by the
host at startup). OAuth device-flow tokens keep the exchange +
proactive-refresh flow unchanged; a PAT 401/403 surfaces the error
instead of a doomed refresh retry. Headless COPILOT_GITHUB_TOKEN now
works with a PAT. Connect flow: the paste hint explains PAT mode, the
dead-PAT case surfaces at the model fetch, and /provider export copies
the saved Copilot account's full token to the clipboard (account picker
on several entries; the token never prints to the transcript).
gh-1575: macOS clipboard paste fixed for TIFF-only pasteboards + the
Ctrl+V text fallback. _readMacos probes «class TIFF» after PNGf/JPEG
miss and converts via sips -s format png (a Preview/browser copy used
to dead-end on clipboard holds no image; a failed conversion keeps the
unavailable fall-through and every scratch file is cleaned up); new
readPasteboardText (pbpaste / xclip+wl-paste / PowerShell
Get-Clipboard) behind the same injectable ProcessRunner/platform:
seam, and Ctrl+V with a TEXT clipboard now inserts the clipboard text
into the composer instead of printing the image probe's dead-end reason
— bracketed-paste routing, gesture pairing and silent-failure semantics
unchanged. Review pass: the Windows Get-Clipboard read pins UTF-8
console output and normalizes embedded \r\n (OEM-code-page mojibake),
the Linux text backends pin text/plain;charset=utf-8 (an image-only
clipboard can no longer leak bytes as "text"), the library doc no longer
overclaims the FA_FAKE_PASTEBOARD seam, and FaTuiCallbacks moved to
a fa_tui_callbacks.dart part file keeping fa_tui.dart under the
2800-line gate.
gh-1555 (W1, lazy/windowed-everywhere reads): the legacy full-open call
sites migrate to windowed opens, so a session ledger past the gh-1073
2 GiB full-open backstop stays openable on every user-facing read path.
The TUI session picker, /resume, and the boot name-resolution scan
open windowed and resolve the display name through paged reads; the
subagent transcript preview opens the child windowed (tail only); the
_loadSession marathon fallback no longer degenerates to a refused
full open when the boundary walk exhausts its page cap (it proceeds
with the tail-anchored windowed session; the boot cap owns over-window
residue); and the headless fa trajectory resolver opens windowed,
reading the whole ledger for the cost/inspect views through the new
SessionTree.ensureWholeLedgerResident — a sequential forward scan in
bounded strips (WindowedSessionStorage.loadAllOlder), not the
boundary-walk's doubling strips (whose per-strip record cap measured
~13x read/parse amplification over the whole file). The same change
block-prepends paged chunks in _indexChunk (one insertAll per
list), removing an O(n²) the suspended-eviction walk paid per record.
Trajectory export streaming (AC3a), split-and-retain rotation
activation (W2), and the opt-in write-side caps (W3) remain follow-up
work.
gh-1605: the FinalizeGate gains a "grader parity" named check. A
produced entry point (script/CLI the verifier will run itself) is only
verified when it survives the grader's own invocation: the checklist
must re-run each produced entry point from a different working
directory in a fresh subprocess with a clean env, capture stdout/
stderr/exit code, and diff against the earlier in-session run — any
divergence (different output, relative-path artifacts, locale/
encoding-dependent output) is a fail ledger item, never a pass.
Closes the modernize-scientific-stack__nwbgy7a near-miss (every
in-session checklist item passed, the verifier's clean-cwd subprocess
collapsed Station 102 to Station 101's mean, reward 0.0). REG coverage:
the parity procedure catches cwd- and env-dependent entry points and
does not false-positive cwd-independent ones
(test/agent/finalize_gate_grader_parity_test.dart), and an IT replays
the evidence shape end-to-end — divergence detected and persisted as a
fail ledger item before finalize.
gh-1603: unattended runs can no longer end as a clean success on empty
model responses. Loop: from the second empty-hiccup retry on, the request
carries a continuation nudge user message ("your last reply was
empty/truncated — continue the task"; exact-text constant
emptyContinuationNudge, emitted with MessageStart/End so the transcript
stays truthful) — a cap-truncated reasoning model re-streamed verbatim
repeats the empty outcome, the nudge breaks the symmetry; the classifier
treats that exact message as transparent, so the turn's partial-content
window (gh-1577) survives a nudged retry. Headless: the empty-reply nudge
budget rises to AgentCliConfig.headlessEmptyReplyNudges (default 2 — the
session ends only after 3 consecutive unrecoverable empties; interactive
keeps the #1085 M2b one-nudge semantics), and a run that still settles
empty prints the explicit __FA_INCOMPLETE marker and exits 1 — a
distinguishable failure for the bench adapter, never a silent exit 0.
gh-1577: an empty assistant completion after partial content (assistant
text, tool activity, or a truncation signal earlier in the turn) is a
bounded-retry provider hiccup — up to emptyHiccupRetries (default 3)
re-streams with exponential backoff (emptyHiccupBaseDelay,
emptyHiccupMaxBackoff, both threaded through the Agent facade,
AgentWiringSpec, and the CLI config), and exhaustion ends the run with
a loud StopReason.error (headless exit 1, interactive error + prompt)
instead of the silent exit-0 settle that killed the gh-1577 trials.
Clean empties keep the #1085 M2b one-nudge semantics. The abort always
wins: a cancel racing the final retry settles StopReason.aborted for
any empty shape (the exit-contract race is pinned by tests), and the
backoff sleep is cancellation-aware.
gh-1574: SQLite reads work on installed binaries — bundle libs land where
the binary looks. The AOT binary's native-assets resolver reads
<exe_dir>/../lib/<name> (the dart build cli bundle layout), but every
POSIX installer copied the bundle libs beside the binary
(<install_dir>/lib), so read path=x.db died with a raw FFI dump
(Couldn't resolve native function 'sqlite3_initialize' … Failed to load dynamic library '~/.local/lib/libsqlite3.dylib'); it only worked on machines
with a manual symlink in ~/.local/lib. Fixed the layout in all three
install paths — site/install.sh (incl. the quarantine-xattr pass and the
source-build advice), install_local.sh, and fa update
(bin/self_manage.dart copyBundleLibs, now public and layout-pinned by
tests; also aligns Windows fa update with install.ps1, which was
already correct) — plus clean degradation in the FFI engine:
Sqlite3Engine.openReadOnly classifies the dart:ffi resolver's
library-load failure (isSqliteDynamicLibraryLoadFailure) and throws a
clean SQLite is unavailable — reinstall or run fa update note instead of
the raw resolver dump; a Dart-level system-library fallback is impossible
with package:sqlite3 native assets (build-time asset mapping, no loader
override). Installs already broken by the old layout self-heal: on the
bundled library's first failed load the binary moves the stale
<exe_dir>/libsqlite3.* into <exe_dir>/../lib/ and retries the load
(repairStaleSqliteBundleLayout) — so one fa update suffices (the
successor binary repairs on its first SQLite read; a pre-fix binary cannot
repair itself), and the stale <bin>/lib leftovers disappear as they
migrate. The POSIX installers also report a failed lib copy instead of
swallowing it (site/install.sh warns naming the lib dir; same in
install_local.sh), and site/install.sh.template carries the same
layout so regenerating the installers cannot resurrect the old one.
gh-1556: bench pipeline stops bleeding score + fa deadline awareness.
Bench: the harbor ladder kill uses the [f]a bracket trick and is
wrapped so a kill failure can never replace the TimeoutError (the
pre-fix self-matching pkill classified TB 2.0 ladder kills as agent
crashes — a SOLVED task scored 0); progress-watch (stall-gap 360s per
the gh-1430 watchdog invariant, flat abs ceiling 3600s) is the
bench-harbor.yml default on every shard, with the harbor adapter
picking the gap-aware ProgressWatch for parity with the
terminal-bench adapter; the aggregate counts completeness by
result.json presence (errored trials report as their own breakdown
line; exit 1 only on genuinely lost trials) and renders the full
report — table, errored breakdown, ledger row — to
$GITHUB_STEP_SUMMARY (TB1 parity). Both bench adapters export
FA_DEADLINE_EPOCH_MS (effective cap − 120s wrap-up margin) into the
agent session and audit the exact value. fa core: deadline awareness —
FA_DEADLINE_EPOCH_MS / --timer/--deadline / /timer write one
session clock, the per-turn context carries a time-left: line while
armed, and one-shot system notices steer wrap-up at ~50%, T-3min
(drain background jobs), and T-30s; absent deadline = byte-for-byte
today's behavior.
gh-1522: tag-sourced versions — the git tag is the single source of
truth. The release-bot file bumps are retired: cutting a release = a
tag + GitHub Release (zero commits to main, zero file mutations).
Both pubspecs carry the fixed 0.0.0-dev placeholder (CI guards
hand-bumps), the publish job stamps the staged pubspec + generated
changelog section from the tag (scripts/stamp_staged_release.sh,
staged version == tag guarded pre-upload), store builds stamp
--build-name/--build-number from the derived tag version
(pubspec.lock stops churning across releases), and
install_local.sh stamps version.txt from git describe.
feat(usage): gh-1460 — the fa-tokens: segment-close line carries the
session's model id, so every token row the factory publishes is
attributable to a model and therefore priceable. The fold now captures
each segment's LAST-SEEN model id (the model that served the final
request, the same string settings/UI display — mid-session fallback
take-overs attribute the segment to the model that actually finished it;
zero-filled fallbacks never clobber a seen model) and the pinned line
shape grows an OPTIONAL "model":"<id>" key after "segment" — omitted
when no model was observed, so legacy lines parse unchanged and the
pinned regex (usageTokensLogPattern, gh-1241 AC7) keeps matching both
shapes. usage.json's schema is untouched (the model is rebuildable from
the chain, I6); old consumers using tolerant JSON ignore the new key,
and the repo's single strict consumer (the pinned constant) is updated
in the same change.
fix(test): gh-1460 rework — the wrap_ansi_line_perf_test growth-shape
legs measure the doubling ratio with the MINIMUM over 9 runs instead of
the median of 5 (CI run 38112524699, core shard 0/4: a scheduling pause
inflated the doubled-size median — 12.09ms against a true ~7ms — and
tripped the 3.2 bound at 3.65x on a genuinely linear wrap). Runner
jitter only ever inflates a wall-clock sample, so the min converges on
the true compute time; genuine superlinearity still trips the ratio
(the gh-1496 quadratic candidates measure ~4x by min). The absolute-
budget legs keep the median — their gh-1357 headroom already prices in
jitter. A synthetic regression test pins the pause-proof property.
<!-- release-draft-owner: run/42695820867/job/binaries-linux-arm64 -->
Your coding agent can read these notes before it upgrades. Set up the MCP server →
One column per month.
fix(gh-2075): the daily-publish macOS legs no longer depend on the package:sqlite3 native-assets hook's single un-retried build-time download. The bui
fix(gh-2075): the daily-publish macOS legs no longer depend on the
package:sqlite3 native-assets hook's single un-retried build-time
download. The build-macos.yml app-build and CLI-bundle steps now
pre-warm the hook's digest-checked cache via the ONE shared tool
(scripts/sqlite3_hook_prewarm.dart, gh-2074 — the same contract
build-mobile.yml + ci.yml already ride) after pub/config resolution and
before the build; a brownout fails there in seconds with a named
::error:: instead of killing a release build ~20 minutes in. The
interim bash seeder (scripts/seed_sqlite3_native_assets.sh + selftest)
is removed — round-2 review flagged it as a second implementation of the
same hook-cache contract; the lane wiring + the single-mechanism
invariant are pinned by test/ci_sqlite3_native_asset_seed_guard_test.dart.
fix(gh-2074): the daily-publish play leg no longer reds on a GitHub
release-download brownout. Three build-mobile.yml Android jobs died in
package:sqlite3's native-assets build hook with Bad state: Hash of downloaded file libsqlite3.x64.android.so is bef140a1… — the hook
(released ≤3.9.1; the lockfile pins 3.7.0, the root floats 3.9.1)
hashes the download WITHOUT honoring the HTTP status, so GitHub's 504
page (a constant hash, upstream simolus3/sqlite3.dart#435) failed the
pin check and the build exited 255. The assets themselves were intact;
ci.yml had solved this exact class for itself on 2026-09-21 by caching
.dart_tool/hooks_runner/shared, but build-mobile.yml never got the
protection. Cure: new scripts/sqlite3_hook_prewarm.dart — derives the
release tag + pinned hashes from the RESOLVED sqlite3 package (no
second pin), gates on the HTTP status before hashing, bounded-retries
with backoff, verifies SHA-256 before any cache write (tmp+rename),
skips idempotently on a verified cache hit — wired as the
Native-assets hook cache + Pre-warm sqlite3 hook assets steps in
the two Android build jobs (both roots, hooks-shared-mobile- keys)
and inside assert-builtin-skills for the root host asset (the step
that failed first). Guard test test/ci_sqlite3_hook_prewarm_guard_ test.dart pins the wiring (AC1-AC4); the script's byte-level behavior
and the live hook-reuse layout are pinned in
test/scripts/sqlite3_hook_prewarm_test.dart.
Rework (2026-11-09, the PR's own CI run): five reds, two root causes,
both fixed. (1) expandTargets accepted its hostOs/hostArch
injection but ignored it — expandTargets('host', hostOs: 'linux', hostArch: 'x64') resolved against the REAL machine and failed on the
ubuntu-24.04-arm Core-tests shard; a fully injected platform now never
consults the host (unit-provable on x64 and arm64 alike). (2) The
pre-warm was wired as plain dart run — which executes the whole
package graph's build hooks BEFORE the script's main(), so on exactly
the gh-2074 conditions (cold cache + brownout) the pre-warm died before
it could pre-warm (proven live; the pty-visual red was this same shape).
Every invocation is now dart --disable-dart-dev (no hook machinery)
and the script is self-contained (no package: imports — SHA-256
inlined) so that mode can load it. The same protection extended to
ci.yml: the lanes whose first hook consumer is a plain dart command
(test-core, integration-mock, hostile-env, fa-aot-build,
pty-integration-linux, binaries-pr, pty-visual — the last had no
hooks cache at all) now restore hooks-shared- + pre-warm the host
asset after pub get and before that consumer; the guard pins it as
AC6/AC7.
feat(gh-2059): browser-extension compaction parity with CLI — over-window
relief, /compact with visible receipts, and the shared under-window
guarantee ride the same CompactionHostWiring resolution as the CLI and
app (resolveExtensionCompactionWiring); receipts are durable across
panel attach/SW restart. Machine-review round 1: exact ladder pass
counts, busy-matrix routing (routeUserInput) for panel text and second
/compact during a pass, window settings from the ACTIVE stored source
with positive validation, string-encoded int merges.
fix(gh-2055): the daily-publish cli leg red-failed in pod install
(Build macOS x86_64, run 41778615815) because the vendored macOS pod's
XCFramework download block guarded on the downloaded-zip MARKER
(if [ ! -f wasm_run-v0.1.0.zip ]) instead of on the artifact the build
consumes. A fresh CI clone never has the marker, so pod install re-ran
the block on every build: it removed the WasmRun.xcframework the
build-macos.yml restore step had already placed (cache hit,
sha256-verified), re-downloaded the 92.7 MB asset with a bare
no---fail curl, and one transient reset (curl 56 at ~3 MB/92.7 MB)
left the Podfile nm -gU symbol check with no
macos-arm64_x86_64/libwasm_run_dart.a to read. The guard now keys on
WasmRun.xcframework/macos-arm64_x86_64/libwasm_run_dart.a — the exact
path the Podfile force_loads and nm-checks — so the block is a no-op
whenever the restore step (or a previous run) put the framework in
place, and the fallback download for a truly cold tree carries
--fail AND aborts the block at the curl step on failure
(|| { echo "wasm_run download failed" >&2; exit 1; } — a swallowed
status fell through to unzip/truncate residue and masked the root cause
behind the Podfile nm diagnostic; review round 1 on this PR), and the
podspec now assigns the backtick result and raises "wasm_run framework
fetch failed" unless $?.success? — Ruby ignores a bare backtick's
exit status, so pod install would otherwise have sailed past the failed
fetch to the downstream nm diagnostic (review round 2). Linted
behaviorally (stubbed curl, no network; POSIX-gated sandbox, plus a
ruby evaluation of the real fetch statement) by
test/wasmrun_macos_podspec_download_guard_test.dart.
fix(gh-2050): the daily-publish cli leg red-failed daily when its
macOS attach arrived after a sibling train had already PUBLISHED the
derived tag (the same-tag convergence race: mobile leg drafts + the tag
train publishes ~25 min after dispatch while the serialized self-hosted
mac builds need ~66 min — v1.0.685: draft 05:22:46, published 05:44:52,
mac leg attached 06:28:53). scripts/release_asset_gate.sh gains a
caller-declared pre-stage policy — RELEASE_GATE_PUBLISHED_SKIP=1 —
that turns the PUBLISHED+unattached-fa-* verdict into a loud whole-call
skip (::warning::, exit 0, nothing uploaded, no twin create, no flip;
all-or-nothing so latest-mac.yml can never land alone on a release
whose zips were skipped). The default stays the gh-2042 AC1 loud
refusal; only build-macos.yml's release-macos desktop attach opts in
(its DMGs/zips/manifest are clobber-refresher pre-stage assets, not
publish blockers), the CLI-tar/extension attaches keep the refusal (they
carry REQUIRED release assets). The stranded desktop feed self-heals:
the next daily leg pre-stages the next version's draft with fresh
assets.
fix(gh-2044): the auto-compact fragile-band escalation could fold the
ENTIRE visible projection when the remaining path fit inside
keepRecentTokens — the keep-recent walk defaulted to folding
everything when its budget never exhausted, swallowing the live tail
(the newest user ask and assistant turn included). Real-world trigger:
resumed subagent runs ended with no assistant message at all
(subagent produced no assistant message — CI cores shards 1/4 and
3/4, hostile-env, both red). _pickCheckpointRange now declines the
fold when the whole foldable path fits the keep-recent region and
still carries live records; all-marker paths keep the lossless
deterministic consolidation (issue #387). The escalation receipt also
reports the REAL summarizer call count (a #729 chunked fold makes N
calls, not 1), and the factory wires its run token into the engine's
cancelToken so a watchdog-shaped cancel reason (any
TimeoutException) can no longer dissolve a user abort into a
"summarizer returned no text" failed receipt while the run reports
success (#1085). Follow-up CI red on the same branch: the decline
moved tiny-transcript runs to the classic residual engine, where an
EMPTY summarizable region folded zero #729 chunks into a blank
summary that still landed as a junk CompactionRecord and a hollow
success: true through the no-work terminal — compactSession now
refuses that empty-region blank fold (gh-1743's failure-verdict
contract kept; a dialed blank answer over a real region still lands
as a record, issue #578 unchanged). Third CI red on the same branch
(flutter-app over-window relief, both shards): the whole-world decline
lived in the shared _pickCheckpointRange, so it also starved the
checkpoint LADDER — a tool-result-heavy transcript whose every ledger
token fits keepRecentTokens (the request estimate is far over the
trigger) lost its only summarizer-based relief, and the hosted app's
over-window guard ended the turn instead of continuing it. The decline
is now escalation-only: ladder passes run over the trigger, where the
fold IS the relief (the escalation still declines — semantic 5 — and
its pass body was extracted below the CRAP ratchet's 12.00).
(gh-2044 rework, UT-ESC-13)
fix(gh-2003): the Copilot picker filter is tier-aware. Individual-tier
/models responses omit model_picker_enabled entirely, so phase 1
dropped every such entry as «picker flag absent» and individual/OAuth
accounts still listed zero usable models. When the response comes from
the individual tier — the individual host (api.githubcopilot.com or
pi-mono's api.individual.githubcopilot.com spelling) or a
no-proxy-ep token (individual by design) — an ABSENT flag is now
selectable: absence ≠ an explicit false, which still drops
everywhere (embeddings, completion-only). Business/enterprise answers
keep the strict picker list unchanged, and the keep is never silent:
the fetch note names it («2 of 2 models kept via the individual-tier
fallback (picker flag absent)») through the existing fetch notes:
surfacing. pi-mono parity: the trigger mirrors allowPolicyFallback = baseUrl === "https://api.individual.githubcopilot.com" in
packages/ai/src/auth/oauth/github-copilot.ts.
fix(gh-1990): the Copilot /models dialect never answers a silent
empty list again. Every silent-empty exit names its branch through the
new fetchModelsForEndpointReport dispatch (ModelsFetchReport +
ModelsFetchDiagnostic): the token-exchange HTTP status (with the
OAuth device-flow / fine-grained PAT remedies for classic ghp_…
keys), the /models HTTP status or timeout (with the seat + per-tier
host hint), proxy-ep-less individual tokens aimed at a business/
enterprise configured host (the tier-misrouting hint), and the picker
filter with per-model drop reasons (picker flag absent, picker flag false, policy disabled, no tool_calls, endpoint unsupported) —
«12 models fetched, 12 filtered (picker flag absent)». A failed
primary credential path retries the other path ONCE before declaring
empty (dual-path fallback, gh-1606): an exchange-401 classic PAT tries
PAT-direct (the provenance note names which path answered), a
fine-grained PAT never falls into the exchange, an OAuth token never
becomes a direct PAT Bearer. Surfacing: /models prints a fetch notes: section per provider, the /provider copilot connect flow
prints the named failure before manual entry, the app's model
quick-select (EndpointModelsController.emptyReason) renders it under
the field, and the Copilot connect sheet shows the reason above the
manual model id. The legacy fetchModelsForEndpoint answers are
unchanged (parity); the filter criteria themselves are unchanged
(pi-mono parity — only the reporting is new). Review round 2: the
fetch notes: wiring covers the unsaved sibling paths too (the
active-catalog and env-keyed Copilot fetches surface the same named
failure instead of a silent seed fallback), the connect flow's
no models: line reuses ModelsFetchReport.reasonText (the last
hand-rolled message — hint join), and the ModelsFetchReport
docs scope the diagnostic contract to the dialects that name their
failure branches.
feat(network): channel payload sender block — app members' channel
messages carry {"text": …, "sender": {"name": …, "id": …}} (public
base64 JSON and keyed fanet1 plaintext alike), so DAP agents can
attribute the HUMAN author behind the fa_network relay's hub id. The
harness unwraps it tolerantly (raw text passes through), rides the
name as AgentMessage.fromName through the fabric, and the CLI
steering line greets people again: from Uladzimir (<relay id>) in #channel: hello instead of from 3cfaee1aa054fe6e…. The app falls
back to the payload sender for bubble names when the envelope carries
none; DMs also name hub peers from the whois cache. Legacy payloads
without the block decode exactly as before.
feat(gh-1956): reply auto-spill — oversized child reply / agent_message
payloads spill to an agent:// artifact automatically; the parent receives a
bounded excerpt + pointer + completeness marker with char AND byte counts.
Manual chunking degrades to a documented fallback, restorable across parent
resume via durable deliverable_artifact records, behind the
subagents.replyAutoSpill kill switch (default on; honored on the CLI and
the app host alike). Rework round: the salvage hook (output-truncated
deliverable calls) restores the durable ids before allocating its own
(no post-resume pointer shadowing), durability losses emit a
[deliverables] diagnostic line, the app host honors the kill switch, the
spill-reach gate is renamed spillAppliesTo (was spillApliesTo), byte
counts are encoded once per spill, and DeliverableRecord.artifact(bytes)
is optional again so external constructors keep compiling.
fix(#1947): schedule_message from the orchestrator no longer misroutes
into a finished subagent's mailbox. A background shell job started by a
CHILD registered its settle continuation inside the child's subagent
zone; when the job settled after the child's run ended, the CLI's
settle-wake (_onShellJobSettled → _startRun) started MAIN's reaction
run still carrying the finished child's activeSubagentId(), and every
identity consumer of that run misattributed it — the reported shape was
a self-reminder scheduled "in <session>/<child>" that auto-resumed the
completed child (duplicate-release risk). Three layers: the
ShellJobRegistry now binds its settle continuation to the
construction-time (boot) zone (_bootZone.run), both hosts'
scheduleSenderMailbox wiring trusts the zone scope only while the
child is genuinely in flight on the session executor
(TaskExecutor.isInFlight), and the tool response names the RECIPIENT
the record carries — an explicit to: used to be shadowed in the reply
by the sender's mailbox, which read as "the override was ignored" even
when it was honored. E2E repro drives the real AgentCli
(test/cli/agent_cli_schedule_recipient_test.dart) plus a registry-level
zone pin (test/tools/shell_job_settle_zone_test.dart).
fix(gh-1948): task_observe works on a RUNNING background subagent
instead of crashing with an uncaught SessionException on the
<parentSessionId>/<agentName> placeholder. Three legs: the executor
flushes each child's transcript incrementally at every turn boundary
(the JSONL exists and grows mid-run — also crash-resilience for
aborted children), task_observe resolves a placeholder handle's real
session file from the sessions tree (the same agent: subagent /
parent: / id: header scan the boot adoption uses) and heals the
handle onto the found path, and every read failure degrades to a
structured "transcript unavailable" / "no transcript available yet"
message instead of an uncaught tool error. The child-transcript reader
rides the transient-ENOENT retry (issue #427) — it reads while the
executor appends.
fix(#1941): the selected CodeMie project now rides EVERY model request.
The header is baked into Model.headers only by the connect-time
switches (gh-1809), so every other rebuild silently dropped
x-codemie-project — and the backend bills a header-less request against
the user's DEFAULT project (codemie-code header-injection.plugin.ts
contract): the reported symptom is limit-exhausted from the exhausted
default project despite picking a budgeted one. Rebuilds now re-resolve
the project from the saved entry through one helper
(projectHeaderForBaseUrl): the boot folder-state restore (JWT entries
previously booted header-less — the SSO cookie patch was the only boot
carrier), the session-resume re-apply (_buildRestoredModel), the
roles-chain failover (ModelRef.project, mirroring the #964 authHeader
leg), the roles-mode /model + settings-flow pins, and the generic
/provider switch. The models list (/llm_models, SSO + JWT + guided
setup + entry dispatch) and the quota adapter (dark until pinned) scope
themselves to the same project, so display and billing can never diverge.
Tests pin the header VALUE on the wire (real adapter over a fake
endpoint) for the restored and failed-over models.
feat(network): private networks created outside the app surface after
sign-in. GET /api/networks/mine (fa_network#16) is synced into the
wallet after OAuth/dev sign-in and both restore paths — best-effort,
silent on older backends (404) and any failure, and never destructive.
Synced memberships keep the server's name/class, preserve local
passwords/join facts, and land without a password (the endpoint never
returns one) until the first join. Entering such a network — sidebar,
⌘K/⌘3-9, or a restored selection — opens the join sheet prefilled with
the network id instead of a resume error or endless spinner; joining by
name now keys the wallet/session by the echoed canonical id and migrates
legacy name-keyed rows (password + channel keys) instead of duplicating
them. A malformed /mine row is skipped, not the whole list.
fix(gh-1931): one render per tool call on the non-TUI console. Line mode
and the headless -p log printed every executed call twice — a start row
(• bash · cmd, no elapsed) followed by the end row (✓/✗ bash · cmd 0s). The duplicate start-row print is gone from _onToolExecutionStart;
the end row (command, status glyph, elapsed, failure news) is the single
render per call. The TUI keeps its running card + settled card, start
bookkeeping stays (_toolStarts feeds the end row's detail/elapsed, the
waiting coordinator still arms per-call liveness), and persisted
artifacts (session JSONL, trajectory records) are untouched —
rendering-only change. Long foreground calls keep the gh-1055 liveness
reminders past waiting.toolLivenessSeconds (default 60s).
fix(gh-1929): esc during the pre-flight auto-compaction no longer crashes
the CLI. The compaction layer's deliberate abort (#1085) used to escape
_runPrompt outside its guarded body and reach the root zone handler via
an unlistened whenComplete chain — fa crashed: CancelledException +
crash.log + a lost session view. The pre-flight window (submit on an
over-window transcript) now translates the cancel into a user-cancel
outcome: neutral dim submission cancelled receipt (never error:), the
TUI composer hands the typed text back (sendInputText) so a resubmit is
one Enter away, headless exits 130 with the same receipt, and the next
prompt works (the funnel's local-trim recovery or a retried compaction).
Cancels DURING the run — in-flight relief, settle funnel — keep their
pinned loud-abort surfaces, and the post-run window keeps its
compaction interrupted receipt via the shared _renderUserCancel
helper.
feat(1923): stall-proof steering. Steering arriving while the loop awaits
an UNPRODUCTIVE provider submit (zero bytes, or silent past the steering
stall window) now cancels the attempt via a SteeringInterruptFire
cancel, injects the pending steering into the rebuilt request, and
resubmits — the #1916-class submit stalls can no longer queue user or
agent mail invisibly. Productive streams keep next-boundary semantics;
the interrupt never consumes a retry attempt, and the run latch re-arms
in place so watchdog replays and user aborts keep working. task_resume
caps its foreground wait (like task_send) and continues in the
background; video polling honors soft-yield checkpoints (remote render
never restarted, honest hand-back text) and media generation honors run
cancellation; SLO-breached undelivered child mail surfaces in
task_status (UNDELIVERED … past the 2s delivery SLO); the TUI prints
steering queued — provider stalled Ns, interrupting to deliver plus the
delivered ack.
feat(1916): image-heavy submits no longer livelock — byte-based
images.maxBytesPerRequest (default 8MB, layered onto the count cap;
/images names both drop reasons), legacy text-embedded data-URL
screenshots in history fold into the once-only registry instead of
replaying raw forever, stall watchdog windows scale with the outbound
payload (capped at 2x; explicit connect clamps untouched), abandoned
connect attempts answering during the retry backoff are adopted, and
repeated identical-payload stalls fire ONE actionable /compact +
image-budget advice. TUI stall banner + docs/images.md carry the
image-heavy playbook.
fix(gh-1912): Add an agent → DAP now turns a retired hub's unknown
enroll_create response into an actionable upgrade message instead of
exposing bad_frame. Independent enrollment never falls back to the
legacy same-name credential rotation endpoint, and existing credentials
remain untouched on failure.
fix(gh-1913): Add an agent is usable on narrow phone screens: scope and format choices wrap without per-character labels, enrollment controls stack at full width, and actions remain reachable by scrolling. Desktop keeps the segmented layout; phone/desktop light/dark goldens guard both.
fix(gh-1825): DAP agent enrollments carry a stable, independent
agentInstance through app/CLI bundles, persistence, direct boot and
reconnect. Independent same-name agents keep distinct identities and stay
online together, while legacy bundles retain their project-scoped identity.
feat(gh-1825): DAP private-channel sends now use self-describing
fanet1 frames (sender X25519 key in the frame, channel-bound AAD) via
fa_hub_client 0.2.11, so fa_network app members decrypt them without
a whois/senderKey round-trip. Inbound channel frames open fanet1
first and fall back to legacy dap/1 payloads. Verified live in both
directions between a headless DAP agent and the Flutter app.
feat(gh-1825): the network offline banner now has a Reconnect now
action which restarts the socket and reloads the channel/member
snapshot. Fresh sockets also clear a stale offline flag (the server
sends no historical drain to a socket that was away during recovery).
AgentNetworkController.start stops after disposal instead of
notifying a dead controller.
feat(gh-1825): the roster sheet shows access-scope badges (whole network vs N channels, decoded additively from Member.access in snapshots and presence.changed) and owner/admins can drop an agent (DELETE /api/networks/{id}/agents/{agentId} with confirmation; the server tombstone keeps it out of future snapshots). Guests and store members are removable too via DELETE /api/networks/{id}/members/{id} (sessions revoked server-side); owner/admin/self rows stay protected.
feat(gh-1825): channel/network management in the app — rename a channel (PATCH), delete a channel or an owned network (REST DELETE with confirmation and state/navigation cleanup), and unnamed CLI/hub-created channels show a short friendly label instead of the raw 64-char id.
feat(gh-1825): pasting a DAP invite no longer demands a passphrase on macOS — importing NEW channel secrets auto-provisions a random passphrase into the per-channel Keychain item (decryption of existing envelopes never auto-generates; env/prompt/headless behavior unchanged), and passphrase warnings now name the concrete recovery (re-paste the invite to re-encrypt, or set DAP_KEY_PASSPHRASE).
feat(gh-1825): the network online counter opens a live-first roster sheet (names, member class, presence), and the channel composer offers roster-filtered @-mention autocomplete — the draft keeps the literal @displayName for DAP agents and the envelope carries the resolved member ids. On the CLI side the DAP channel notice names the agent's own identity (display name + agentId), and inbound channel mail that @-mentions that name (case-insensitive, word-boundary — no email or @name2 false positives) is marked "[mentions you — you are @name]".
fix(gh-1825): the app's network roster no longer reads as permanently empty ("0 online"): fa_network sends roster.snapshot with an OBJECT payload ({members: [...]}) and the app only accepted a bare list, so every snapshot silently decoded to zero members. Both shapes are now accepted, with a production-shape regression test. The DAP paste hint names web alongside CLI/macOS/iOS (plain-web join is supported).
fix(gh-1825): resumed CLI sessions rejoin their DAP channels and receive channel mail again — session replay now re-arms membership synchronization (plugin session-ready listeners on boot/resume and in-process switches) and idle channel probes complete deferred joins before inbox peek/drain. The post-replay sync also dials the membership's OWN hub with its enrolled credential before joining: a resumed session-scoped agent used to stay on the zero-config local URL forever, and because a disconnected fabric probe short-circuits before the sync, the remote dial never happened — the session-ready listener now drives it directly (throttled to one attempt per 10 s). App channel history no longer flashes raw payloads on first open: decoding waits for startup's channel listing instead of caching an undecoded guess until a manual switch.
feat(gh-1825): the fa app consumes the same DAP enrollment block as the
CLI — paste it into the agent chat composer (macOS/iOS/Android; the
smart-paste .txt chip path included) and the channel join happens
without the block ever reaching sendText, analytics, or session
records. Native joins ride the live agent-fabric hub repository:
per-project scoped identity, invite pub/priv import, keyless public
channels, reconnect-persisted membership in a dedicated secure
enrollment store (Keychain/Keystore where available). Harness additions:
HubMessagingRepository.joinChannel/waitForConnection with
rejoin-before-drain on reconnect, keyed channel decrypt
(hubDecryptPayloadWithKeyPair), public showcase payloads, and
HubBearerTransport (native Authorization header or the browser
?token= equivalent; authed dials never stringify transport errors
into logs). Plain web joins through the browser
WebHubTransport (?token= bearer equivalent, IndexedDB identity);
the browser extension rides the service worker's existing DAP client
(hub.join op), never a panel-local second connection. Add agent now defaults to the
DAP tab with a paste-into-chat hint; web gains ai-native provider
sign-in through the existing popup/callback OAuth flow.
fix(fa_network#4): the app inserts hub agents arriving via
presence.changed into the roster (additive displayName/class fields)
instead of dropping them, so the online counter no longer sticks at 0
when an agent comes online after roster.snapshot; legacy frames
without those fields create no phantom members. roster.snapshot/REST
contents are logged for decisive presence debugging.
feat(gh-1825): /network + scoped DAP channel membership — join by
paste (invite url / url+keys / bare channel id / whole fa dap import
command) with session/project/global scopes, live join without
restart, per-scope leave + copy invite, and the composer-paste
shortcut into the same wizard. Session scope rides a
dap_channel_membership custom record (leave tombstones, branch
resume); project scope is .fah/dap.json (committable, dapkey1
PBKDF2+ChaCha20-Poly1305 encrypted keys); global keeps the legacy
~/.dap files. Headless CI: DAP_CHANNELS_JSON (+DAP_HUB_URL,
DAP_CLIENT_SECRET, DAP_KEY_PASSPHRASE) joins at boot with zero
prompts, plaintext or encrypted privs mixed. Passphrase chain:
env → per-channel macOS Keychain item → one prompt with save default.
feat(gh-1844): media slots never fall back to the main connection
implicitly — SDK/CLI, app store, and the JS bridge resolve a slot
without an override as UNCONFIGURED with an actionable
"configure the slot" error; the main connection serves a slot only via
the explicit per-slot useMainConnection: true opt-in (YAML
models.slots.<slot>.useMainConnection / app's "Same as main
connection" pick, preserved through media_models.json). Image and
video generation accept multimodal input (text + reference images:
sandbox paths, http(s) URLs, base64) on capability-declaring
OpenRouter endpoints (input_references + declared/probed
input_modalities gate; unsupported endpoints fail before any network
call); the OpenRouter video job API is a first-class video dialect
(submit → poll → mp4). New decide tool: structured decisions via
OpenRouter's Decisions API (/v1/systemone) — bool/choice/score
questions with calibrated probabilities, structured error results, no
streaming.
feat(gh-1834): the FinalizeGate contract counts a parity/conformance
claim as verified only with negative + combination self-checks — new
numbered step 2.6 in prompts/cli/finalize_gate.md (refused
incompatible/mutually exclusive options, empty/corrupt inputs, boundary
values, must-NOT-work cases, alongside happy paths; a blanket
"everything matches" is not verified). Universality: abstract example
only, no benchmark/task/flag names in the prompt; the step is
conditional, so tasks with no conformance claim see no gate change.
Ledger/verdict wire formats unchanged; validated end-to-end by the
scripted IT-4 replay (real CLI + real bash) whose persisted
task_ledger carries the rejection rows.
fix(gh-1832): release v1.0.635 published without fa-extension.zip — the
landing-page download 404'd. The extension's only producer (the daily
build-macos.yml build-extension leg) derives its tag as "latest+1 at its
own dispatch moment"; on 2026-10-24 the v1.0.635 tag was cut by
auto_release before those legs ran, so the extension (and the DMGs)
landed on an orphaned v1.0.636 draft while the tag train published
v1.0.635 — every leg green, nothing asserted the published asset set.
Now scripts/release_required_assets.txt is the single source of truth,
scripts/assert_release_assets.sh gates the release-provenance publish
flip (a missing asset leaves the release a draft and fails the run
loudly) and a new release-asset-smoke ci.yml job post-publish probes
every landing-page release link + the full required set (red on non-200).
Remediation: the exact-tag fa-extension.zip (built from the v1.0.635
tree by run 40064876732) was uploaded to release v1.0.635 — the landing
link serves 200 again.
fix(gh-1820): the arm64 DMG leg's stale size floor is reseeded onto the
shipped v1.0.634 shape — run 4004758753 (post-#1815 SPM-lane adoption)
passed the min-platform gate and the full build, then failed the
size_manifest.py +5% budget on App.framework/App +5.1%
(25,020,112 → 26,284,144, the accumulated Dart AOT of #1793/#1794/
#1796/#1805/#1806); the committed scripts/size_budget/fa-macos-arm64.tsv
was still the pre-#1096-diet floor (TOTAL 210.5 MB vs the artifact's
116.4 MB), so the documented intentional-growth remediation applies:
the floor is now the exact --reseed output for the failing run's
artifact manifest (26 rows ≥ 64 KiB + TOTAL 116,386,369, from the run
log's printed table — the zip never uploaded past the gate), locking in
Fa 23.4→18.6 MB, quickjs_bridge 950→844 KB; new
test/ci_1820_macos_arm64_size_floor_test.dart gates the shape green
through the REAL size_manifest.py check and pins the committed floor
byte-identical to a --reseed round-trip.
fix(gh-1814): the desktop DMG lane ADOPTS the SwiftPM lane — Flutter
3.47.7 removed the FLUTTER_SWIFT_PACKAGE_MANAGER disable knob
(flutter/flutter#193978), so the gh-180 pure-CocoaPods pin became a no-op
and the min-platform gate (which forbade any SwiftPM state) failed both
DMG legs on every pinned-3.47.7 run (39991663029, 40022900373); the
#1813 escape (pin 3.47.4) is unreachable on the ~60kB/s runner link.
scripts/assert_macos_min_platform.sh now ASSERTS the SPM lane instead:
SwiftPM state expected after flutter build macos --config-only,
generated package floor ≥ the app floor (the silent bump must not skip —
the original gh-180 killer), no SwiftPM product above the app minimum,
vendored inappwebview manifest floored at 14.0 if it ever reappears.
Floors were audited and need no raising: Runner pbxproj 3×14.0, Podfile
:osx, '14.0', generated floor 14.0 on 3.47.7, vendored podspec 14.0
(its Package.swift stays deleted — the plugin rides the CocoaPods half of
the hybrid lane via Podfile.lock). The dead env pin left the build-macos
job env; the SDK pin reverts #1813 back to 3.47.7. The ci.yml macOS debug
leg gets the same treatment (gh-1814 review): its
FLUTTER_SWIFT_PACKAGE_MANAGER=0 pin — a silent no-op on the floating
3.47.x stable it resolves past the knob's removal — is removed and the
lane comment corrected; the leg rides SPM enabled as it already has been
(run 399572012 green).
fix(gh-1805): modifier chords no longer type their base letter into the
composer while the slash menu is open — the vendored dart_tui decoder
puts the base letter into the keystroke's text for control bytes
(Ctrl+S → ctrl+s with text s), and the slash-menu edit cluster was
the one insert surface missing the _isCommandKeystroke guard, so a
chord pressed mid-menu poisoned the typed prefix (/memory landed as
/smemory). Plain letters and Shift+letters still type and filter.
feat(1515): ONE model-limits settings surface — the settings hub's
standalone "Model parameters" row is retired and bare /model-edit
(TUI) aliases into /settings → Model capabilities, pre-focused on the
active model. The caps flow gains the model-edit preset ladders +
Custom… free input for context window / max output (a preset pick
writes models.overrides AND rebuilds the active model on the spot —
the old flow's live-mutation semantics, now durable), every field row
shows the current EFFECTIVE value with its source layer (pinned /
role override / catalog), the thinking row is an omp-style selector
(only levels the model supports, per-rung budget metadata, current
level preselected), and the hub row describes the effective summary
(200K ctx · 16K out · thinking off) instead of a pin count. The
scripted /model-edit <field> <n> form and the line-mode limits print
are unchanged.
feat(gh-1746): opt-in session disk retention — sessions.retention in
~/.fah/config.yaml caps the sessions root and sweeps whole session
units (transcript + sidecars) oldest-last-used-first at boot, exempting
live, named, young, and the current session; nothing is EVER
auto-deleted without the section (byte-identical legacy boot). Adds
fa session clean [--dry-run], the /sessions + fa session list
usage footer under a cap, and header-cwd/path-escape guards so a
planted session header can never steer a delete outside any root.
fix(gh-1763): a bash tool command running broad process kills can no
longer signal its own shell or the harness. Scripts are delivered from a
per-exec private temp file (sh -c '. "$1"' sh <script>) so the wrapper
argv never carries the script body (pkill -f <pattern> cannot
self-match its own sh -c; $0/$#/exit codes/set -e/stdin stay
identical), a PATH shim for pkill/killall plus a kill() prelude
guard exempt the harness pid chain + process group (exported as
FA_PROTECTED_PIDS) and the invoking shell from the child's own kills
while legitimate victims (stuck test runners, background servers) are
still terminated, and headless SIGTERM aborts now log
fa: SIGTERM received … so a child-issued kill never masquerades as a
user Ctrl+C (exit code contract unchanged: 130 graceful, 143 escalate).
fix(gh-1778): the release gate no longer reds a fully-green shard when
GitHub's cache service degrades. v1.0.621's Core tests (shard 2/4) was
killed by its own timeout-minutes: 12 with 2535/2761 tests green — both
cache steps had stalled ~133 s on keys that existed (degraded service,
not eviction), cold pub get + native-asset hook builds delayed the
first test to ~5 min, and the shard needed ~14 min against a budget
sized for the warm ~8.8 min wall. The job budget is now 20 min (sibling
test-leg band), and both test-core cache steps are best-effort
(timeout-minutes: 2 + continue-on-error: true) — a failed restore is
slower setup, never a red shard. test/ci_core_shard_timeout_guard_test.dart
pins the budget and the best-effort cache contract (RED-first on the old
YAML).
feat(#1761): the 32KB project-context budget never cuts silently again.
A file truncated to its head excerpt now carries an extended
<!-- From: path — N bytes, only the first M shown --> marker plus an
explicit MUST-read-the-full-file instruction (offset/limit/ranges,
prompt copy in prompts/context/project_context_read_full.md); a file
dropped entirely leaves a one-line tombstone naming its path instead of
vanishing; and the user gets a dim boot line per truncated/dropped file
(CLI line mode prints directly, TUI re-prints after the banner, the app
logs through its AppLog boot-notice seam). New projectContext: config
section: maxBytes (default 32768) — 0 disables the section entirely
and reads no files (fast boot); unknown keys are tolerated with a note.
Leaf-first allocation is unchanged.
fix(gh-1757): the Android build no longer re-enters the sdkmanager
download roulette for CMake. v1.0.619's build-android leg died at
:jni:configureCMakeDebug[arm64-v8a] — the :jni module (transitive
jni package via path_provider_android) wires externalNativeBuild.cmake
with no version, so AGP auto-installed its default cmake;3.22.1
mid-configuration and one corrupt archive ("Archive is not a ZIP
archive") red'd the leg, the aggregate Quality gate, and the
Advance-vpinned-marker job. The gh-1655 pre-install action now also
pre-installs cmake;3.22.1 with bounded retry + idempotent skip (same
hardening the NDK path already carries); the guard test pins the new
AC6 assertions and the AGP-version coupling (AC7).
feat(#1750): per-model image caps — the outbound request never carries
more unique images than the target model accepts. Model gains an
optional maxImages (catalog-seeded: gemini-4-argon: 10, measured);
images.perModel.<modelId> config overrides it (strict images:
section, user truth wins); the registry resolves
perModel > catalog > images.maxPerRequest (first present wins, the
issue-#1554 shape). A Google-shaped too many images … maximum allowed for model <id> is <n> 400 is classified (image_overflow.dart, the
overflow.ts twin), latches the learned cap for the (provider, model)
pair for the run, and re-sends ONCE through the registry at the
smaller cap — current-first-then-newest drops with the existing
never-silent notices naming the learned cap; a second refusal
surfaces exactly as today. JSONL is never rewritten;
images.registry: false keeps the byte-identical legacy shape
including the recovery-off behavior.
feat(gh-1760): oversized user-message intake spills to a file instead of
blowing the context window. Any incoming user message — headless
-p/positional, TUI first/mid-session, steering/follow-up drain,
external inbox — whose text passes the new relative threshold (a
spills.userFraction of the live window budget minus the compaction
reserve, floored by spills.userThreshold bytes) is redacted and
written verbatim to .fah/spills/<sessionId>/user-<n>.md (the #678
layout) from one choke point (Agent.userMessageGuard); the model
receives a bounded head+tail preview plus a read-in-parts instruction
riding the replacement USER message (copy in
prompts/spill/user_read_instruction.md). Images stay inline, a
refused spill write degrades to today's inline behavior, and
spills.enabled: false / userThreshold: 0 / the new
FAH_USER_SPILL_THRESHOLD env knob (bytes) keep the legacy shapes.
fix(gh-1743): a structured compaction that hits its total-budget
watchdog (the factory's 300s totalBudget) no longer aborts the
live session. AutoCompactorFactory.run() now applies the
mechanical local-trim fallback (keepRecentTokens, pairing-repaired,
[context trimmed locally] marker — the same valve the classic
pass ladder ends on) instead of failing fast; the honest
both-roles-failed verdict is kept only when nothing is droppable.
fix(gh-1636): auto-compaction stops dumping the internal checkpoint
summary into the console — the default report is ONE marker line
(● auto-compacted · engine: N → M tokens (F freed · P%)); the records
line + fenced summary moved behind the new tui.showCompactionDetails
opt-in (default false), and the manual /compact still prints the full
block. The headless log face no longer streams the summarizer deltas by
default (the gh-1433 E4 live stream is part of the same opt-in; the
gh-1430 … reasoning liveness line still covers the window). Nothing
is lost — the summary stays in the session file, reachable via
compact_expand.
fix(test): gh-1732 — the wrap_ansi_line_perf_test sustained-contention
validation leg spikes 3 of the 5 doubled runs per attempt instead of 4
(CI run 39546301132, core shard 2/4: the old shape left exactly ONE
real-wall-clock-timed doubled sample per attempt, and real jitter on
that single sample — the very shape being modeled — inflated all three
attempts' min ratios to a 3.30x best against the 3.2 bound while the
real 500k growth legs on the same runner stayed green). 3-of-5 is the
documented live-red shape ("≥3 of the 5 timed runs"): median-of-5
[c,c,s] is still a spiked value, so the leg still proves the
pre-gh-1532 median harness fails, and the min harness now rides two
clean doubled samples per attempt.
fix(#1705): CLI fresh-install boarding asks the provider choice FIRST —
the same list the /add-provider picker shows (presets, Custom last) —
and only Custom continues into the api-type wizard; a preset opens its
own flow with the api type prefilled. Previously boarding opened the
custom wizard directly, cold-asking "api type" before any provider was
chosen.
feat(providers): gh-1704 — first-class Codex/ChatGPT credential export,
CI refresh action, and Provider Queue integration (Copilot-export
parity). fa provider export <name> [--check] resolves the saved
ChatGPT login's OAuth blob from its secure-store slot, refreshes it when
stale, and prints the shell lines (OPENAI_API_KEY + OPENAI_BASE_URL
= the Codex backend — these tokens never serve api.openai.com); the
refresh token never prints and --check reports credential health with
the SSO caveat. /provider chatgpt import [path] adopts an existing
codex login session (~/.codex/auth.json) into the same slot — no
browser round-trip. .github/actions/chatgpt-codex-auth mints a fresh
access token from a CHATGPT_REFRESH_TOKEN secret in CI (direct curl,
no third-party action); invalid_grant fails fast and loudly.
providersQueue: entries speak chatgpt-codex: an omitted apiKeyEnv
resolves the Keychain slot automatically (in-memory refresh), a new
refreshTokenEnv: field mints the initial access token on slot
activation, and a raw Bearer token in apiKeyEnv is tolerated by
chatgpt_codex.dart (no more FormatException — proactive/reactive
refresh degrade, a 401 surfaces as the auth death the queue fails over
on).
feat(bench): gh-1681 — leaderboard submission readiness. ATIF
trajectories: new exportAtifJson whole-snapshot exporter
(lib/src/trajectory/atif_export.dart, ATIF-v1.7 — the schema harbor
0.23.0's flagship adapters emit) behind a headless fa trajectory export [--format atif] [--output path] verb; the harbor adapter runs
it as a bounded best-effort exec after the agent phase on every
terminal path, so <trial>/agent/trajectory.json reaches the hub with
the other agent logs (the leaderboard static analysis requires one per
rewarded trial) and result.json metadata carries an
atif_trajectory present/missing marker. The golden fixture
(bench/harbor_fa/fixtures/atif_sample.json) is pinned Dart-side and
cross-validated against harbor's pydantic Trajectory +
validate_trajectory in test_atif_contract.py. Workflow: a
submission: true dispatch input turns the submission rules into hard
setup preflights (full task coverage, ≥5 attempts/task, no
agent-timeout override, pinned dataset) and makes the new per-shard
ATIF presence gate fail hard; default runs gain only a warn-only scan
(artifact flow unchanged). Runbook: docs/leaderboard-submission.md
(pinned dataset + @latest content-hash recording, quota-window runner
policy, lb CLI steps).
fix(headless): gh-1589 — the headless session becomes a SEALED artifact:
no session-JSONL writes land after the run's agent phase ends. New
Session.seal() (state in the writer, no file-format change): raised in
runHeadless' finally and the REPL teardown as the LAST session action
(close order: drain → usage fold → seal → exit, one implementation both
paths — REG-2), it makes every writer (_persistMessages,
shell_job_registry snapshots, compaction, any appendCustomEntry,
leaf moves) drop late appends with exactly ONE stderr warning per
writer class — never a throw into an unhandled zone; the exit code
stays the turn's terminal outcome (issue #413 semantics untouched). The
settle→reaction path is sealed with the file (I2): the drain wrap-up
raises a fence (_reactionRunsSealed) so a background-shell settle or
task completion arriving after the final answer's drain — including the
registry's one-turn-deferred listener racing _finishHeadlessDrain's
check-to-return window (the 65gGW7b 10:13:42–10:14:00 burst) — does
board bookkeeping only: no _startRun, no model request, no record.
An honest drain's records (final answer → seal) stay legal (E7); the
seal is per-run state, so resume/replay/session_repair are unchanged.
Adapter side (bench/harbor_fa, I3/AC4+AC7): run()/perform_task
returns only after the fa process exited on EVERY path — the ladder
kill became terminate-and-reap (SIGTERM → bounded exec await → SIGKILL
escalation → pgrep confirm), and the external-cancel paths (harbor's
own wait_for firing first, the no-knobs exec, external-cap-only mode)
now fire the kill + bounded reap before propagating, with a loud
agent-phase exit guarantee degraded stderr note when fa survives both
signals (never a silent proceed, never an unbounded hang).
fix(bench): gh-1643 — a trial killed mid-stream no longer reports 0/None
usage. The fold read only FINALIZED assistant messages, so a run killed
before its first assistant message (the costliest rows) folded to silent
zeros. Both folds now price a dangling model_request_summary at its
KNOWN request-side volume (system prompt + outbound message chars at the
shared chars/4 rate, estimated input; output unknown) and mark the usage
incomplete instead: UsageChainScan.usageIncomplete → usage.json
(usageIncomplete: true, written only when set so completed folds stay
byte-identical), SessionUsage.usage_incomplete → the tb adapter's
usage-guard.json row and harbor's agent_result.metadata, and the
post-mortem stamps usage_incomplete on every results.json row it
patches from a partial fold. Harbor also exports the run outcome class
(agent_result.metadata.failure_mode: agent_timeout self-kill /
external_kill / completed / crashed / never_started) so downstream
classification stops relying on exception.txt archaeology; a summary
superseded by another summary is flushed instead of silently dropped on
both sides. Never-started trials keep the gh-1638/gh-1576 audit-only
precedence (no fake usage).
fix(stuck-calls): gh-1632 — a stuck tool call in an interactive session now
escalates like everywhere else. effectiveStuckTool() defaulted every
non-headless host to StuckFollowUpMode.advisory when no explicit
agent.stuckTool: was configured, so the fa 1.0.580 incident session (a
dead git fetch in the REPL) fired its tool_stuck flag at the 300 s
threshold and then hung for 27 more minutes — advisory mode never cancels,
retries, or converts; only the owner's manual steering broke the deadlock.
The autonomous default (cancel → retry once → background-convert, then a
session-visible escalation) now applies in EVERY mode; advisory remains
an explicit opt-out (agent.stuckTool.followUp: advisory), never an
implicit default.
fix(bindings): gh-1624 — the gh-1570 binding preflight no longer kills
translating-proxy factories. _validateModelEndpoint's vendor-prefix
verdict was fatal on ANY endpoint, so FA_PROVIDER_TYPE=anthropic +
bifrost's /anthropic path + zai/glm-5.3 exit-64'd every boot on
1.0.576 (the prefix is a routing key on a translating proxy, not a
dialect declaration). The vendor-prefix check now fires ONLY on the
endpoints that carry a dialect verdict — the three family-exclusive
canonical vendor endpoints (api.anthropic.com, api.openai.com,
generativelanguage), mirroring the family verdict's scope: those keep
the gh-1570 fail-fast, ROUTERS (OpenRouter/Copilot/DIAL/kimi — the
prefix is a routing key there too; OpenRouter's canonical ids are
anthropic/claude-*) and unknown/private endpoints carry the mismatch
as a ProviderModelBinding.warnings entry, never exit 64. The warning
is rendered on EVERY resolve path (gh-1624 review): the FA_PROVIDER_*
boot notice, the provider queue (per entry at boot, on queue reload,
and in /provider queue test), and the --model/--base-url flag
revalidation — a rider never boots silently.
Breaking-change note for 1.0.576 upgraders behind proxy
factories: the migration is the wire: FA_PROVIDER_CONFIG key —
documented now (README env-preconfig section,
docs/agents/providers.md §Translating proxies, --help) with its
caveat: the override engages only when the declared wire DIFFERS from
the provider record's own kind (record anthropic + wire: anthropic
still runs the checks). REG matrix (AC1/AC2/AC3) pinned in
test/model_roles/provider_model_binding_test.dart; AC4 IT drives a
bifrost-shaped loopback proxy through the production anthropic adapter
(test/model_roles/bifrost_proxy_binding_test.dart).
gh-1596: the first-run wizard (site/setup.sh) no longer crashes or
corrupts input under curl | sh on dash (Ubuntu sh) or busybox ash
(Alpine): scripts/gen_installers.dart now renders the provider/mode/
approval menus from the live provider catalog into plain POSIX sh — the
bashisms that killed dash (printf -v, read -p) and silently corrupted
busybox ash are gone; the hidden API-key prompt uses stty -echo + a
by-name read (dash has no read -s) and the key never rides printf
argv; EOF/Ctrl-C mid-wizard aborts non-zero with no half-written
~/.fah, and non-TTY runs print guidance and exit 1. The wizard's temp
config is mktemp-created (no predictable, symlink-followable fixed
path). A new path-triggered installer-portability CI leg proves the
script under dash + busybox ash (ubuntu) and bash (macOS) on every
site/** or generator change.
fix(cli): gh-1572 — fa boot no longer rewrites ~/.fah/config.yaml
from a lossy config. persistConfig (bin/fah_runapp.dart) built a
fresh CliConfig carrying only some fields and persistConfig() ran
unconditionally at boot, so every launch silently dropped the user's
spills:, subagents:, jobs:, headless:, agent:, quota:,
tui:, trajectory:, waiting:, power.hold and auto_update
settings (first reported on 1.0.547). Three cooperating fixes: the
boot save is now saved.copyWith(<live overlays>) so every parsed
section rides along by construction; CliConfig.copyWith replaces the
drifted hand-rolled withCustomProviders field list (itself dropping
spills/subagents/waiting/jobs/agentMode/misuseBreaker/streamThinking
inside saveCliConfig's merge step); the _diskPreservedSections
disk backstop covers every known top-level key. spills: renders in
toYaml (new SpillsConfig.toYaml) and agent.misuseBreaker: false
persists too; an unchanged save no longer rewrites the file at all.
Rework round: the boot snapshot re-syncs from disk after every
sanctioned mid-session settings write (onUserConfigFileChanged — a
cleared context cap, an auto_update key drop or a global compaction
pick is no longer resurrected/overwritten by the next save in the
same session); validateAgentSection (the strict fa config check /
config set / settings-hub mirror) accepts the emitter's
agent.misuseBreaker and agent.stuckTool keys so fa config check
never false-fails on fa's own output; clearing the context cap drops
only the cap key instead of the whole agent: block (the sibling
mode:/misuseBreaker: keys survive); and the providersQueue:
block (no typed renderer) joins the disk backstop so a save can never
silently delete the user's queue. Rework round 2 (merge + guards):
validateAgentSection's per-key switch became the _agentKeyValidators
table (the two gh-1572 keys pushed the switch form to complexity 13 —
over the repo's CRAP ratchet; the table keeps the mirror complete at
complexity 4), and the last two direct USER-config writers —
_writeTtsrSection and _writeMcpSection — fire the re-sync hook, so
a mid-session ttsr rule toggle or mcp server edit survives the next
persist too (the writer roster in settings_flow_config_sync.dart is
now exhaustive).
gh-1606: the copilot provider accepts fine-grained PATs
(github_pat_…) directly — PAT mode skips the copilot_internal/v2/token
exchange entirely (GitHub 404s every PAT exchange) and sends the raw PAT
as the Bearer against the configured Copilot base URL with the official
Copilot CLI's first-party identity headers
(copilot-integration-id: copilot-developer-cli,
user-agent: copilot/<fa-version> (<os>) term/unknown, installed by the
host at startup). OAuth device-flow tokens keep the exchange +
proactive-refresh flow unchanged; a PAT 401/403 surfaces the error
instead of a doomed refresh retry. Headless COPILOT_GITHUB_TOKEN now
works with a PAT. Connect flow: the paste hint explains PAT mode, the
dead-PAT case surfaces at the model fetch, and /provider export copies
the saved Copilot account's full token to the clipboard (account picker
on several entries; the token never prints to the transcript).
gh-1575: macOS clipboard paste fixed for TIFF-only pasteboards + the
Ctrl+V text fallback. _readMacos probes «class TIFF» after PNGf/JPEG
miss and converts via sips -s format png (a Preview/browser copy used
to dead-end on clipboard holds no image; a failed conversion keeps the
unavailable fall-through and every scratch file is cleaned up); new
readPasteboardText (pbpaste / xclip+wl-paste / PowerShell
Get-Clipboard) behind the same injectable ProcessRunner/platform:
seam, and Ctrl+V with a TEXT clipboard now inserts the clipboard text
into the composer instead of printing the image probe's dead-end reason
— bracketed-paste routing, gesture pairing and silent-failure semantics
unchanged. Review pass: the Windows Get-Clipboard read pins UTF-8
console output and normalizes embedded \r\n (OEM-code-page mojibake),
the Linux text backends pin text/plain;charset=utf-8 (an image-only
clipboard can no longer leak bytes as "text"), the library doc no longer
overclaims the FA_FAKE_PASTEBOARD seam, and FaTuiCallbacks moved to
a fa_tui_callbacks.dart part file keeping fa_tui.dart under the
2800-line gate.
gh-1555 (W1, lazy/windowed-everywhere reads): the legacy full-open call
sites migrate to windowed opens, so a session ledger past the gh-1073
2 GiB full-open backstop stays openable on every user-facing read path.
The TUI session picker, /resume, and the boot name-resolution scan
open windowed and resolve the display name through paged reads; the
subagent transcript preview opens the child windowed (tail only); the
_loadSession marathon fallback no longer degenerates to a refused
full open when the boundary walk exhausts its page cap (it proceeds
with the tail-anchored windowed session; the boot cap owns over-window
residue); and the headless fa trajectory resolver opens windowed,
reading the whole ledger for the cost/inspect views through the new
SessionTree.ensureWholeLedgerResident — a sequential forward scan in
bounded strips (WindowedSessionStorage.loadAllOlder), not the
boundary-walk's doubling strips (whose per-strip record cap measured
~13x read/parse amplification over the whole file). The same change
block-prepends paged chunks in _indexChunk (one insertAll per
list), removing an O(n²) the suspended-eviction walk paid per record.
Trajectory export streaming (AC3a), split-and-retain rotation
activation (W2), and the opt-in write-side caps (W3) remain follow-up
work.
gh-1605: the FinalizeGate gains a "grader parity" named check. A
produced entry point (script/CLI the verifier will run itself) is only
verified when it survives the grader's own invocation: the checklist
must re-run each produced entry point from a different working
directory in a fresh subprocess with a clean env, capture stdout/
stderr/exit code, and diff against the earlier in-session run — any
divergence (different output, relative-path artifacts, locale/
encoding-dependent output) is a fail ledger item, never a pass.
Closes the modernize-scientific-stack__nwbgy7a near-miss (every
in-session checklist item passed, the verifier's clean-cwd subprocess
collapsed Station 102 to Station 101's mean, reward 0.0). REG coverage:
the parity procedure catches cwd- and env-dependent entry points and
does not false-positive cwd-independent ones
(test/agent/finalize_gate_grader_parity_test.dart), and an IT replays
the evidence shape end-to-end — divergence detected and persisted as a
fail ledger item before finalize.
gh-1603: unattended runs can no longer end as a clean success on empty
model responses. Loop: from the second empty-hiccup retry on, the request
carries a continuation nudge user message ("your last reply was
empty/truncated — continue the task"; exact-text constant
emptyContinuationNudge, emitted with MessageStart/End so the transcript
stays truthful) — a cap-truncated reasoning model re-streamed verbatim
repeats the empty outcome, the nudge breaks the symmetry; the classifier
treats that exact message as transparent, so the turn's partial-content
window (gh-1577) survives a nudged retry. Headless: the empty-reply nudge
budget rises to AgentCliConfig.headlessEmptyReplyNudges (default 2 — the
session ends only after 3 consecutive unrecoverable empties; interactive
keeps the #1085 M2b one-nudge semantics), and a run that still settles
empty prints the explicit __FA_INCOMPLETE marker and exits 1 — a
distinguishable failure for the bench adapter, never a silent exit 0.
gh-1577: an empty assistant completion after partial content (assistant
text, tool activity, or a truncation signal earlier in the turn) is a
bounded-retry provider hiccup — up to emptyHiccupRetries (default 3)
re-streams with exponential backoff (emptyHiccupBaseDelay,
emptyHiccupMaxBackoff, both threaded through the Agent facade,
AgentWiringSpec, and the CLI config), and exhaustion ends the run with
a loud StopReason.error (headless exit 1, interactive error + prompt)
instead of the silent exit-0 settle that killed the gh-1577 trials.
Clean empties keep the #1085 M2b one-nudge semantics. The abort always
wins: a cancel racing the final retry settles StopReason.aborted for
any empty shape (the exit-contract race is pinned by tests), and the
backoff sleep is cancellation-aware.
gh-1574: SQLite reads work on installed binaries — bundle libs land where
the binary looks. The AOT binary's native-assets resolver reads
<exe_dir>/../lib/<name> (the dart build cli bundle layout), but every
POSIX installer copied the bundle libs beside the binary
(<install_dir>/lib), so read path=x.db died with a raw FFI dump
(Couldn't resolve native function 'sqlite3_initialize' … Failed to load dynamic library '~/.local/lib/libsqlite3.dylib'); it only worked on machines
with a manual symlink in ~/.local/lib. Fixed the layout in all three
install paths — site/install.sh (incl. the quarantine-xattr pass and the
source-build advice), install_local.sh, and fa update
(bin/self_manage.dart copyBundleLibs, now public and layout-pinned by
tests; also aligns Windows fa update with install.ps1, which was
already correct) — plus clean degradation in the FFI engine:
Sqlite3Engine.openReadOnly classifies the dart:ffi resolver's
library-load failure (isSqliteDynamicLibraryLoadFailure) and throws a
clean SQLite is unavailable — reinstall or run fa update note instead of
the raw resolver dump; a Dart-level system-library fallback is impossible
with package:sqlite3 native assets (build-time asset mapping, no loader
override). Installs already broken by the old layout self-heal: on the
bundled library's first failed load the binary moves the stale
<exe_dir>/libsqlite3.* into <exe_dir>/../lib/ and retries the load
(repairStaleSqliteBundleLayout) — so one fa update suffices (the
successor binary repairs on its first SQLite read; a pre-fix binary cannot
repair itself), and the stale <bin>/lib leftovers disappear as they
migrate. The POSIX installers also report a failed lib copy instead of
swallowing it (site/install.sh warns naming the lib dir; same in
install_local.sh), and site/install.sh.template carries the same
layout so regenerating the installers cannot resurrect the old one.
gh-1556: bench pipeline stops bleeding score + fa deadline awareness.
Bench: the harbor ladder kill uses the [f]a bracket trick and is
wrapped so a kill failure can never replace the TimeoutError (the
pre-fix self-matching pkill classified TB 2.0 ladder kills as agent
crashes — a SOLVED task scored 0); progress-watch (stall-gap 360s per
the gh-1430 watchdog invariant, flat abs ceiling 3600s) is the
bench-harbor.yml default on every shard, with the harbor adapter
picking the gap-aware ProgressWatch for parity with the
terminal-bench adapter; the aggregate counts completeness by
result.json presence (errored trials report as their own breakdown
line; exit 1 only on genuinely lost trials) and renders the full
report — table, errored breakdown, ledger row — to
$GITHUB_STEP_SUMMARY (TB1 parity). Both bench adapters export
FA_DEADLINE_EPOCH_MS (effective cap − 120s wrap-up margin) into the
agent session and audit the exact value. fa core: deadline awareness —
FA_DEADLINE_EPOCH_MS / --timer/--deadline / /timer write one
session clock, the per-turn context carries a time-left: line while
armed, and one-shot system notices steer wrap-up at ~50%, T-3min
(drain background jobs), and T-30s; absent deadline = byte-for-byte
today's behavior.
gh-1522: tag-sourced versions — the git tag is the single source of
truth. The release-bot file bumps are retired: cutting a release = a
tag + GitHub Release (zero commits to main, zero file mutations).
Both pubspecs carry the fixed 0.0.0-dev placeholder (CI guards
hand-bumps), the publish job stamps the staged pubspec + generated
changelog section from the tag (scripts/stamp_staged_release.sh,
staged version == tag guarded pre-upload), store builds stamp
--build-name/--build-number from the derived tag version
(pubspec.lock stops churning across releases), and
install_local.sh stamps version.txt from git describe.
feat(usage): gh-1460 — the fa-tokens: segment-close line carries the
session's model id, so every token row the factory publishes is
attributable to a model and therefore priceable. The fold now captures
each segment's LAST-SEEN model id (the model that served the final
request, the same string settings/UI display — mid-session fallback
take-overs attribute the segment to the model that actually finished it;
zero-filled fallbacks never clobber a seen model) and the pinned line
shape grows an OPTIONAL "model":"<id>" key after "segment" — omitted
when no model was observed, so legacy lines parse unchanged and the
pinned regex (usageTokensLogPattern, gh-1241 AC7) keeps matching both
shapes. usage.json's schema is untouched (the model is rebuildable from
the chain, I6); old consumers using tolerant JSON ignore the new key,
and the repo's single strict consumer (the pinned constant) is updated
in the same change.
fix(test): gh-1460 rework — the wrap_ansi_line_perf_test growth-shape
legs measure the doubling ratio with the MINIMUM over 9 runs instead of
the median of 5 (CI run 38112524699, core shard 0/4: a scheduling pause
inflated the doubled-size median — 12.09ms against a true ~7ms — and
tripped the 3.2 bound at 3.65x on a genuinely linear wrap). Runner
jitter only ever inflates a wall-clock sample, so the min converges on
the true compute time; genuine superlinearity still trips the ratio
(the gh-1496 quadratic candidates measure ~4x by min). The absolute-
budget legs keep the median — their gh-1357 headroom already prices in
jitter. A synthetic regression test pins the pause-proof property.
<!-- release-draft-owner: run/42657846711/job/binaries-linux-x64 -->
fix(gh-2075): the daily-publish macOS legs no longer depend on the package:sqlite3 native-assets hook's single un-retried build-time download. The bui
fix(gh-2075): the daily-publish macOS legs no longer depend on the
package:sqlite3 native-assets hook's single un-retried build-time
download. The build-macos.yml app-build and CLI-bundle steps now
pre-warm the hook's digest-checked cache via the ONE shared tool
(scripts/sqlite3_hook_prewarm.dart, gh-2074 — the same contract
build-mobile.yml + ci.yml already ride) after pub/config resolution and
before the build; a brownout fails there in seconds with a named
::error:: instead of killing a release build ~20 minutes in. The
interim bash seeder (scripts/seed_sqlite3_native_assets.sh + selftest)
is removed — round-2 review flagged it as a second implementation of the
same hook-cache contract; the lane wiring + the single-mechanism
invariant are pinned by test/ci_sqlite3_native_asset_seed_guard_test.dart.
fix(gh-2074): the daily-publish play leg no longer reds on a GitHub
release-download brownout. Three build-mobile.yml Android jobs died in
package:sqlite3's native-assets build hook with Bad state: Hash of downloaded file libsqlite3.x64.android.so is bef140a1… — the hook
(released ≤3.9.1; the lockfile pins 3.7.0, the root floats 3.9.1)
hashes the download WITHOUT honoring the HTTP status, so GitHub's 504
page (a constant hash, upstream simolus3/sqlite3.dart#435) failed the
pin check and the build exited 255. The assets themselves were intact;
ci.yml had solved this exact class for itself on 2026-09-21 by caching
.dart_tool/hooks_runner/shared, but build-mobile.yml never got the
protection. Cure: new scripts/sqlite3_hook_prewarm.dart — derives the
release tag + pinned hashes from the RESOLVED sqlite3 package (no
second pin), gates on the HTTP status before hashing, bounded-retries
with backoff, verifies SHA-256 before any cache write (tmp+rename),
skips idempotently on a verified cache hit — wired as the
Native-assets hook cache + Pre-warm sqlite3 hook assets steps in
the two Android build jobs (both roots, hooks-shared-mobile- keys)
and inside assert-builtin-skills for the root host asset (the step
that failed first). Guard test test/ci_sqlite3_hook_prewarm_guard_ test.dart pins the wiring (AC1-AC4); the script's byte-level behavior
and the live hook-reuse layout are pinned in
test/scripts/sqlite3_hook_prewarm_test.dart.
Rework (2026-11-09, the PR's own CI run): five reds, two root causes,
both fixed. (1) expandTargets accepted its hostOs/hostArch
injection but ignored it — expandTargets('host', hostOs: 'linux', hostArch: 'x64') resolved against the REAL machine and failed on the
ubuntu-24.04-arm Core-tests shard; a fully injected platform now never
consults the host (unit-provable on x64 and arm64 alike). (2) The
pre-warm was wired as plain dart run — which executes the whole
package graph's build hooks BEFORE the script's main(), so on exactly
the gh-2074 conditions (cold cache + brownout) the pre-warm died before
it could pre-warm (proven live; the pty-visual red was this same shape).
Every invocation is now dart --disable-dart-dev (no hook machinery)
and the script is self-contained (no package: imports — SHA-256
inlined) so that mode can load it. The same protection extended to
ci.yml: the lanes whose first hook consumer is a plain dart command
(test-core, integration-mock, hostile-env, fa-aot-build,
pty-integration-linux, binaries-pr, pty-visual — the last had no
hooks cache at all) now restore hooks-shared- + pre-warm the host
asset after pub get and before that consumer; the guard pins it as
AC6/AC7.
feat(gh-2059): browser-extension compaction parity with CLI — over-window
relief, /compact with visible receipts, and the shared under-window
guarantee ride the same CompactionHostWiring resolution as the CLI and
app (resolveExtensionCompactionWiring); receipts are durable across
panel attach/SW restart. Machine-review round 1: exact ladder pass
counts, busy-matrix routing (routeUserInput) for panel text and second
/compact during a pass, window settings from the ACTIVE stored source
with positive validation, string-encoded int merges.
fix(gh-2055): the daily-publish cli leg red-failed in pod install
(Build macOS x86_64, run 41778615815) because the vendored macOS pod's
XCFramework download block guarded on the downloaded-zip MARKER
(if [ ! -f wasm_run-v0.1.0.zip ]) instead of on the artifact the build
consumes. A fresh CI clone never has the marker, so pod install re-ran
the block on every build: it removed the WasmRun.xcframework the
build-macos.yml restore step had already placed (cache hit,
sha256-verified), re-downloaded the 92.7 MB asset with a bare
no---fail curl, and one transient reset (curl 56 at ~3 MB/92.7 MB)
left the Podfile nm -gU symbol check with no
macos-arm64_x86_64/libwasm_run_dart.a to read. The guard now keys on
WasmRun.xcframework/macos-arm64_x86_64/libwasm_run_dart.a — the exact
path the Podfile force_loads and nm-checks — so the block is a no-op
whenever the restore step (or a previous run) put the framework in
place, and the fallback download for a truly cold tree carries
--fail AND aborts the block at the curl step on failure
(|| { echo "wasm_run download failed" >&2; exit 1; } — a swallowed
status fell through to unzip/truncate residue and masked the root cause
behind the Podfile nm diagnostic; review round 1 on this PR), and the
podspec now assigns the backtick result and raises "wasm_run framework
fetch failed" unless $?.success? — Ruby ignores a bare backtick's
exit status, so pod install would otherwise have sailed past the failed
fetch to the downstream nm diagnostic (review round 2). Linted
behaviorally (stubbed curl, no network; POSIX-gated sandbox, plus a
ruby evaluation of the real fetch statement) by
test/wasmrun_macos_podspec_download_guard_test.dart.
fix(gh-2050): the daily-publish cli leg red-failed daily when its
macOS attach arrived after a sibling train had already PUBLISHED the
derived tag (the same-tag convergence race: mobile leg drafts + the tag
train publishes ~25 min after dispatch while the serialized self-hosted
mac builds need ~66 min — v1.0.685: draft 05:22:46, published 05:44:52,
mac leg attached 06:28:53). scripts/release_asset_gate.sh gains a
caller-declared pre-stage policy — RELEASE_GATE_PUBLISHED_SKIP=1 —
that turns the PUBLISHED+unattached-fa-* verdict into a loud whole-call
skip (::warning::, exit 0, nothing uploaded, no twin create, no flip;
all-or-nothing so latest-mac.yml can never land alone on a release
whose zips were skipped). The default stays the gh-2042 AC1 loud
refusal; only build-macos.yml's release-macos desktop attach opts in
(its DMGs/zips/manifest are clobber-refresher pre-stage assets, not
publish blockers), the CLI-tar/extension attaches keep the refusal (they
carry REQUIRED release assets). The stranded desktop feed self-heals:
the next daily leg pre-stages the next version's draft with fresh
assets.
fix(gh-2044): the auto-compact fragile-band escalation could fold the
ENTIRE visible projection when the remaining path fit inside
keepRecentTokens — the keep-recent walk defaulted to folding
everything when its budget never exhausted, swallowing the live tail
(the newest user ask and assistant turn included). Real-world trigger:
resumed subagent runs ended with no assistant message at all
(subagent produced no assistant message — CI cores shards 1/4 and
3/4, hostile-env, both red). _pickCheckpointRange now declines the
fold when the whole foldable path fits the keep-recent region and
still carries live records; all-marker paths keep the lossless
deterministic consolidation (issue #387). The escalation receipt also
reports the REAL summarizer call count (a #729 chunked fold makes N
calls, not 1), and the factory wires its run token into the engine's
cancelToken so a watchdog-shaped cancel reason (any
TimeoutException) can no longer dissolve a user abort into a
"summarizer returned no text" failed receipt while the run reports
success (#1085). Follow-up CI red on the same branch: the decline
moved tiny-transcript runs to the classic residual engine, where an
EMPTY summarizable region folded zero #729 chunks into a blank
summary that still landed as a junk CompactionRecord and a hollow
success: true through the no-work terminal — compactSession now
refuses that empty-region blank fold (gh-1743's failure-verdict
contract kept; a dialed blank answer over a real region still lands
as a record, issue #578 unchanged). Third CI red on the same branch
(flutter-app over-window relief, both shards): the whole-world decline
lived in the shared _pickCheckpointRange, so it also starved the
checkpoint LADDER — a tool-result-heavy transcript whose every ledger
token fits keepRecentTokens (the request estimate is far over the
trigger) lost its only summarizer-based relief, and the hosted app's
over-window guard ended the turn instead of continuing it. The decline
is now escalation-only: ladder passes run over the trigger, where the
fold IS the relief (the escalation still declines — semantic 5 — and
its pass body was extracted below the CRAP ratchet's 12.00).
(gh-2044 rework, UT-ESC-13)
fix(gh-2003): the Copilot picker filter is tier-aware. Individual-tier
/models responses omit model_picker_enabled entirely, so phase 1
dropped every such entry as «picker flag absent» and individual/OAuth
accounts still listed zero usable models. When the response comes from
the individual tier — the individual host (api.githubcopilot.com or
pi-mono's api.individual.githubcopilot.com spelling) or a
no-proxy-ep token (individual by design) — an ABSENT flag is now
selectable: absence ≠ an explicit false, which still drops
everywhere (embeddings, completion-only). Business/enterprise answers
keep the strict picker list unchanged, and the keep is never silent:
the fetch note names it («2 of 2 models kept via the individual-tier
fallback (picker flag absent)») through the existing fetch notes:
surfacing. pi-mono parity: the trigger mirrors allowPolicyFallback = baseUrl === "https://api.individual.githubcopilot.com" in
packages/ai/src/auth/oauth/github-copilot.ts.
fix(gh-1990): the Copilot /models dialect never answers a silent
empty list again. Every silent-empty exit names its branch through the
new fetchModelsForEndpointReport dispatch (ModelsFetchReport +
ModelsFetchDiagnostic): the token-exchange HTTP status (with the
OAuth device-flow / fine-grained PAT remedies for classic ghp_…
keys), the /models HTTP status or timeout (with the seat + per-tier
host hint), proxy-ep-less individual tokens aimed at a business/
enterprise configured host (the tier-misrouting hint), and the picker
filter with per-model drop reasons (picker flag absent, picker flag false, policy disabled, no tool_calls, endpoint unsupported) —
«12 models fetched, 12 filtered (picker flag absent)». A failed
primary credential path retries the other path ONCE before declaring
empty (dual-path fallback, gh-1606): an exchange-401 classic PAT tries
PAT-direct (the provenance note names which path answered), a
fine-grained PAT never falls into the exchange, an OAuth token never
becomes a direct PAT Bearer. Surfacing: /models prints a fetch notes: section per provider, the /provider copilot connect flow
prints the named failure before manual entry, the app's model
quick-select (EndpointModelsController.emptyReason) renders it under
the field, and the Copilot connect sheet shows the reason above the
manual model id. The legacy fetchModelsForEndpoint answers are
unchanged (parity); the filter criteria themselves are unchanged
(pi-mono parity — only the reporting is new). Review round 2: the
fetch notes: wiring covers the unsaved sibling paths too (the
active-catalog and env-keyed Copilot fetches surface the same named
failure instead of a silent seed fallback), the connect flow's
no models: line reuses ModelsFetchReport.reasonText (the last
hand-rolled message — hint join), and the ModelsFetchReport
docs scope the diagnostic contract to the dialects that name their
failure branches.
feat(network): channel payload sender block — app members' channel
messages carry {"text": …, "sender": {"name": …, "id": …}} (public
base64 JSON and keyed fanet1 plaintext alike), so DAP agents can
attribute the HUMAN author behind the fa_network relay's hub id. The
harness unwraps it tolerantly (raw text passes through), rides the
name as AgentMessage.fromName through the fabric, and the CLI
steering line greets people again: from Uladzimir (<relay id>) in #channel: hello instead of from 3cfaee1aa054fe6e…. The app falls
back to the payload sender for bubble names when the envelope carries
none; DMs also name hub peers from the whois cache. Legacy payloads
without the block decode exactly as before.
feat(gh-1956): reply auto-spill — oversized child reply / agent_message
payloads spill to an agent:// artifact automatically; the parent receives a
bounded excerpt + pointer + completeness marker with char AND byte counts.
Manual chunking degrades to a documented fallback, restorable across parent
resume via durable deliverable_artifact records, behind the
subagents.replyAutoSpill kill switch (default on; honored on the CLI and
the app host alike). Rework round: the salvage hook (output-truncated
deliverable calls) restores the durable ids before allocating its own
(no post-resume pointer shadowing), durability losses emit a
[deliverables] diagnostic line, the app host honors the kill switch, the
spill-reach gate is renamed spillAppliesTo (was spillApliesTo), byte
counts are encoded once per spill, and DeliverableRecord.artifact(bytes)
is optional again so external constructors keep compiling.
fix(#1947): schedule_message from the orchestrator no longer misroutes
into a finished subagent's mailbox. A background shell job started by a
CHILD registered its settle continuation inside the child's subagent
zone; when the job settled after the child's run ended, the CLI's
settle-wake (_onShellJobSettled → _startRun) started MAIN's reaction
run still carrying the finished child's activeSubagentId(), and every
identity consumer of that run misattributed it — the reported shape was
a self-reminder scheduled "in <session>/<child>" that auto-resumed the
completed child (duplicate-release risk). Three layers: the
ShellJobRegistry now binds its settle continuation to the
construction-time (boot) zone (_bootZone.run), both hosts'
scheduleSenderMailbox wiring trusts the zone scope only while the
child is genuinely in flight on the session executor
(TaskExecutor.isInFlight), and the tool response names the RECIPIENT
the record carries — an explicit to: used to be shadowed in the reply
by the sender's mailbox, which read as "the override was ignored" even
when it was honored. E2E repro drives the real AgentCli
(test/cli/agent_cli_schedule_recipient_test.dart) plus a registry-level
zone pin (test/tools/shell_job_settle_zone_test.dart).
fix(gh-1948): task_observe works on a RUNNING background subagent
instead of crashing with an uncaught SessionException on the
<parentSessionId>/<agentName> placeholder. Three legs: the executor
flushes each child's transcript incrementally at every turn boundary
(the JSONL exists and grows mid-run — also crash-resilience for
aborted children), task_observe resolves a placeholder handle's real
session file from the sessions tree (the same agent: subagent /
parent: / id: header scan the boot adoption uses) and heals the
handle onto the found path, and every read failure degrades to a
structured "transcript unavailable" / "no transcript available yet"
message instead of an uncaught tool error. The child-transcript reader
rides the transient-ENOENT retry (issue #427) — it reads while the
executor appends.
fix(#1941): the selected CodeMie project now rides EVERY model request.
The header is baked into Model.headers only by the connect-time
switches (gh-1809), so every other rebuild silently dropped
x-codemie-project — and the backend bills a header-less request against
the user's DEFAULT project (codemie-code header-injection.plugin.ts
contract): the reported symptom is limit-exhausted from the exhausted
default project despite picking a budgeted one. Rebuilds now re-resolve
the project from the saved entry through one helper
(projectHeaderForBaseUrl): the boot folder-state restore (JWT entries
previously booted header-less — the SSO cookie patch was the only boot
carrier), the session-resume re-apply (_buildRestoredModel), the
roles-chain failover (ModelRef.project, mirroring the #964 authHeader
leg), the roles-mode /model + settings-flow pins, and the generic
/provider switch. The models list (/llm_models, SSO + JWT + guided
setup + entry dispatch) and the quota adapter (dark until pinned) scope
themselves to the same project, so display and billing can never diverge.
Tests pin the header VALUE on the wire (real adapter over a fake
endpoint) for the restored and failed-over models.
feat(network): private networks created outside the app surface after
sign-in. GET /api/networks/mine (fa_network#16) is synced into the
wallet after OAuth/dev sign-in and both restore paths — best-effort,
silent on older backends (404) and any failure, and never destructive.
Synced memberships keep the server's name/class, preserve local
passwords/join facts, and land without a password (the endpoint never
returns one) until the first join. Entering such a network — sidebar,
⌘K/⌘3-9, or a restored selection — opens the join sheet prefilled with
the network id instead of a resume error or endless spinner; joining by
name now keys the wallet/session by the echoed canonical id and migrates
legacy name-keyed rows (password + channel keys) instead of duplicating
them. A malformed /mine row is skipped, not the whole list.
fix(gh-1931): one render per tool call on the non-TUI console. Line mode
and the headless -p log printed every executed call twice — a start row
(• bash · cmd, no elapsed) followed by the end row (✓/✗ bash · cmd 0s). The duplicate start-row print is gone from _onToolExecutionStart;
the end row (command, status glyph, elapsed, failure news) is the single
render per call. The TUI keeps its running card + settled card, start
bookkeeping stays (_toolStarts feeds the end row's detail/elapsed, the
waiting coordinator still arms per-call liveness), and persisted
artifacts (session JSONL, trajectory records) are untouched —
rendering-only change. Long foreground calls keep the gh-1055 liveness
reminders past waiting.toolLivenessSeconds (default 60s).
fix(gh-1929): esc during the pre-flight auto-compaction no longer crashes
the CLI. The compaction layer's deliberate abort (#1085) used to escape
_runPrompt outside its guarded body and reach the root zone handler via
an unlistened whenComplete chain — fa crashed: CancelledException +
crash.log + a lost session view. The pre-flight window (submit on an
over-window transcript) now translates the cancel into a user-cancel
outcome: neutral dim submission cancelled receipt (never error:), the
TUI composer hands the typed text back (sendInputText) so a resubmit is
one Enter away, headless exits 130 with the same receipt, and the next
prompt works (the funnel's local-trim recovery or a retried compaction).
Cancels DURING the run — in-flight relief, settle funnel — keep their
pinned loud-abort surfaces, and the post-run window keeps its
compaction interrupted receipt via the shared _renderUserCancel
helper.
feat(1923): stall-proof steering. Steering arriving while the loop awaits
an UNPRODUCTIVE provider submit (zero bytes, or silent past the steering
stall window) now cancels the attempt via a SteeringInterruptFire
cancel, injects the pending steering into the rebuilt request, and
resubmits — the #1916-class submit stalls can no longer queue user or
agent mail invisibly. Productive streams keep next-boundary semantics;
the interrupt never consumes a retry attempt, and the run latch re-arms
in place so watchdog replays and user aborts keep working. task_resume
caps its foreground wait (like task_send) and continues in the
background; video polling honors soft-yield checkpoints (remote render
never restarted, honest hand-back text) and media generation honors run
cancellation; SLO-breached undelivered child mail surfaces in
task_status (UNDELIVERED … past the 2s delivery SLO); the TUI prints
steering queued — provider stalled Ns, interrupting to deliver plus the
delivered ack.
feat(1916): image-heavy submits no longer livelock — byte-based
images.maxBytesPerRequest (default 8MB, layered onto the count cap;
/images names both drop reasons), legacy text-embedded data-URL
screenshots in history fold into the once-only registry instead of
replaying raw forever, stall watchdog windows scale with the outbound
payload (capped at 2x; explicit connect clamps untouched), abandoned
connect attempts answering during the retry backoff are adopted, and
repeated identical-payload stalls fire ONE actionable /compact +
image-budget advice. TUI stall banner + docs/images.md carry the
image-heavy playbook.
fix(gh-1912): Add an agent → DAP now turns a retired hub's unknown
enroll_create response into an actionable upgrade message instead of
exposing bad_frame. Independent enrollment never falls back to the
legacy same-name credential rotation endpoint, and existing credentials
remain untouched on failure.
fix(gh-1913): Add an agent is usable on narrow phone screens: scope and format choices wrap without per-character labels, enrollment controls stack at full width, and actions remain reachable by scrolling. Desktop keeps the segmented layout; phone/desktop light/dark goldens guard both.
fix(gh-1825): DAP agent enrollments carry a stable, independent
agentInstance through app/CLI bundles, persistence, direct boot and
reconnect. Independent same-name agents keep distinct identities and stay
online together, while legacy bundles retain their project-scoped identity.
feat(gh-1825): DAP private-channel sends now use self-describing
fanet1 frames (sender X25519 key in the frame, channel-bound AAD) via
fa_hub_client 0.2.11, so fa_network app members decrypt them without
a whois/senderKey round-trip. Inbound channel frames open fanet1
first and fall back to legacy dap/1 payloads. Verified live in both
directions between a headless DAP agent and the Flutter app.
feat(gh-1825): the network offline banner now has a Reconnect now
action which restarts the socket and reloads the channel/member
snapshot. Fresh sockets also clear a stale offline flag (the server
sends no historical drain to a socket that was away during recovery).
AgentNetworkController.start stops after disposal instead of
notifying a dead controller.
feat(gh-1825): the roster sheet shows access-scope badges (whole network vs N channels, decoded additively from Member.access in snapshots and presence.changed) and owner/admins can drop an agent (DELETE /api/networks/{id}/agents/{agentId} with confirmation; the server tombstone keeps it out of future snapshots). Guests and store members are removable too via DELETE /api/networks/{id}/members/{id} (sessions revoked server-side); owner/admin/self rows stay protected.
feat(gh-1825): channel/network management in the app — rename a channel (PATCH), delete a channel or an owned network (REST DELETE with confirmation and state/navigation cleanup), and unnamed CLI/hub-created channels show a short friendly label instead of the raw 64-char id.
feat(gh-1825): pasting a DAP invite no longer demands a passphrase on macOS — importing NEW channel secrets auto-provisions a random passphrase into the per-channel Keychain item (decryption of existing envelopes never auto-generates; env/prompt/headless behavior unchanged), and passphrase warnings now name the concrete recovery (re-paste the invite to re-encrypt, or set DAP_KEY_PASSPHRASE).
feat(gh-1825): the network online counter opens a live-first roster sheet (names, member class, presence), and the channel composer offers roster-filtered @-mention autocomplete — the draft keeps the literal @displayName for DAP agents and the envelope carries the resolved member ids. On the CLI side the DAP channel notice names the agent's own identity (display name + agentId), and inbound channel mail that @-mentions that name (case-insensitive, word-boundary — no email or @name2 false positives) is marked "[mentions you — you are @name]".
fix(gh-1825): the app's network roster no longer reads as permanently empty ("0 online"): fa_network sends roster.snapshot with an OBJECT payload ({members: [...]}) and the app only accepted a bare list, so every snapshot silently decoded to zero members. Both shapes are now accepted, with a production-shape regression test. The DAP paste hint names web alongside CLI/macOS/iOS (plain-web join is supported).
fix(gh-1825): resumed CLI sessions rejoin their DAP channels and receive channel mail again — session replay now re-arms membership synchronization (plugin session-ready listeners on boot/resume and in-process switches) and idle channel probes complete deferred joins before inbox peek/drain. The post-replay sync also dials the membership's OWN hub with its enrolled credential before joining: a resumed session-scoped agent used to stay on the zero-config local URL forever, and because a disconnected fabric probe short-circuits before the sync, the remote dial never happened — the session-ready listener now drives it directly (throttled to one attempt per 10 s). App channel history no longer flashes raw payloads on first open: decoding waits for startup's channel listing instead of caching an undecoded guess until a manual switch.
feat(gh-1825): the fa app consumes the same DAP enrollment block as the
CLI — paste it into the agent chat composer (macOS/iOS/Android; the
smart-paste .txt chip path included) and the channel join happens
without the block ever reaching sendText, analytics, or session
records. Native joins ride the live agent-fabric hub repository:
per-project scoped identity, invite pub/priv import, keyless public
channels, reconnect-persisted membership in a dedicated secure
enrollment store (Keychain/Keystore where available). Harness additions:
HubMessagingRepository.joinChannel/waitForConnection with
rejoin-before-drain on reconnect, keyed channel decrypt
(hubDecryptPayloadWithKeyPair), public showcase payloads, and
HubBearerTransport (native Authorization header or the browser
?token= equivalent; authed dials never stringify transport errors
into logs). Plain web joins through the browser
WebHubTransport (?token= bearer equivalent, IndexedDB identity);
the browser extension rides the service worker's existing DAP client
(hub.join op), never a panel-local second connection. Add agent now defaults to the
DAP tab with a paste-into-chat hint; web gains ai-native provider
sign-in through the existing popup/callback OAuth flow.
fix(fa_network#4): the app inserts hub agents arriving via
presence.changed into the roster (additive displayName/class fields)
instead of dropping them, so the online counter no longer sticks at 0
when an agent comes online after roster.snapshot; legacy frames
without those fields create no phantom members. roster.snapshot/REST
contents are logged for decisive presence debugging.
feat(gh-1825): /network + scoped DAP channel membership — join by
paste (invite url / url+keys / bare channel id / whole fa dap import
command) with session/project/global scopes, live join without
restart, per-scope leave + copy invite, and the composer-paste
shortcut into the same wizard. Session scope rides a
dap_channel_membership custom record (leave tombstones, branch
resume); project scope is .fah/dap.json (committable, dapkey1
PBKDF2+ChaCha20-Poly1305 encrypted keys); global keeps the legacy
~/.dap files. Headless CI: DAP_CHANNELS_JSON (+DAP_HUB_URL,
DAP_CLIENT_SECRET, DAP_KEY_PASSPHRASE) joins at boot with zero
prompts, plaintext or encrypted privs mixed. Passphrase chain:
env → per-channel macOS Keychain item → one prompt with save default.
feat(gh-1844): media slots never fall back to the main connection
implicitly — SDK/CLI, app store, and the JS bridge resolve a slot
without an override as UNCONFIGURED with an actionable
"configure the slot" error; the main connection serves a slot only via
the explicit per-slot useMainConnection: true opt-in (YAML
models.slots.<slot>.useMainConnection / app's "Same as main
connection" pick, preserved through media_models.json). Image and
video generation accept multimodal input (text + reference images:
sandbox paths, http(s) URLs, base64) on capability-declaring
OpenRouter endpoints (input_references + declared/probed
input_modalities gate; unsupported endpoints fail before any network
call); the OpenRouter video job API is a first-class video dialect
(submit → poll → mp4). New decide tool: structured decisions via
OpenRouter's Decisions API (/v1/systemone) — bool/choice/score
questions with calibrated probabilities, structured error results, no
streaming.
feat(gh-1834): the FinalizeGate contract counts a parity/conformance
claim as verified only with negative + combination self-checks — new
numbered step 2.6 in prompts/cli/finalize_gate.md (refused
incompatible/mutually exclusive options, empty/corrupt inputs, boundary
values, must-NOT-work cases, alongside happy paths; a blanket
"everything matches" is not verified). Universality: abstract example
only, no benchmark/task/flag names in the prompt; the step is
conditional, so tasks with no conformance claim see no gate change.
Ledger/verdict wire formats unchanged; validated end-to-end by the
scripted IT-4 replay (real CLI + real bash) whose persisted
task_ledger carries the rejection rows.
fix(gh-1832): release v1.0.635 published without fa-extension.zip — the
landing-page download 404'd. The extension's only producer (the daily
build-macos.yml build-extension leg) derives its tag as "latest+1 at its
own dispatch moment"; on 2026-10-24 the v1.0.635 tag was cut by
auto_release before those legs ran, so the extension (and the DMGs)
landed on an orphaned v1.0.636 draft while the tag train published
v1.0.635 — every leg green, nothing asserted the published asset set.
Now scripts/release_required_assets.txt is the single source of truth,
scripts/assert_release_assets.sh gates the release-provenance publish
flip (a missing asset leaves the release a draft and fails the run
loudly) and a new release-asset-smoke ci.yml job post-publish probes
every landing-page release link + the full required set (red on non-200).
Remediation: the exact-tag fa-extension.zip (built from the v1.0.635
tree by run 40064876732) was uploaded to release v1.0.635 — the landing
link serves 200 again.
fix(gh-1820): the arm64 DMG leg's stale size floor is reseeded onto the
shipped v1.0.634 shape — run 4004758753 (post-#1815 SPM-lane adoption)
passed the min-platform gate and the full build, then failed the
size_manifest.py +5% budget on App.framework/App +5.1%
(25,020,112 → 26,284,144, the accumulated Dart AOT of #1793/#1794/
#1796/#1805/#1806); the committed scripts/size_budget/fa-macos-arm64.tsv
was still the pre-#1096-diet floor (TOTAL 210.5 MB vs the artifact's
116.4 MB), so the documented intentional-growth remediation applies:
the floor is now the exact --reseed output for the failing run's
artifact manifest (26 rows ≥ 64 KiB + TOTAL 116,386,369, from the run
log's printed table — the zip never uploaded past the gate), locking in
Fa 23.4→18.6 MB, quickjs_bridge 950→844 KB; new
test/ci_1820_macos_arm64_size_floor_test.dart gates the shape green
through the REAL size_manifest.py check and pins the committed floor
byte-identical to a --reseed round-trip.
fix(gh-1814): the desktop DMG lane ADOPTS the SwiftPM lane — Flutter
3.47.7 removed the FLUTTER_SWIFT_PACKAGE_MANAGER disable knob
(flutter/flutter#193978), so the gh-180 pure-CocoaPods pin became a no-op
and the min-platform gate (which forbade any SwiftPM state) failed both
DMG legs on every pinned-3.47.7 run (39991663029, 40022900373); the
#1813 escape (pin 3.47.4) is unreachable on the ~60kB/s runner link.
scripts/assert_macos_min_platform.sh now ASSERTS the SPM lane instead:
SwiftPM state expected after flutter build macos --config-only,
generated package floor ≥ the app floor (the silent bump must not skip —
the original gh-180 killer), no SwiftPM product above the app minimum,
vendored inappwebview manifest floored at 14.0 if it ever reappears.
Floors were audited and need no raising: Runner pbxproj 3×14.0, Podfile
:osx, '14.0', generated floor 14.0 on 3.47.7, vendored podspec 14.0
(its Package.swift stays deleted — the plugin rides the CocoaPods half of
the hybrid lane via Podfile.lock). The dead env pin left the build-macos
job env; the SDK pin reverts #1813 back to 3.47.7. The ci.yml macOS debug
leg gets the same treatment (gh-1814 review): its
FLUTTER_SWIFT_PACKAGE_MANAGER=0 pin — a silent no-op on the floating
3.47.x stable it resolves past the knob's removal — is removed and the
lane comment corrected; the leg rides SPM enabled as it already has been
(run 399572012 green).
fix(gh-1805): modifier chords no longer type their base letter into the
composer while the slash menu is open — the vendored dart_tui decoder
puts the base letter into the keystroke's text for control bytes
(Ctrl+S → ctrl+s with text s), and the slash-menu edit cluster was
the one insert surface missing the _isCommandKeystroke guard, so a
chord pressed mid-menu poisoned the typed prefix (/memory landed as
/smemory). Plain letters and Shift+letters still type and filter.
feat(1515): ONE model-limits settings surface — the settings hub's
standalone "Model parameters" row is retired and bare /model-edit
(TUI) aliases into /settings → Model capabilities, pre-focused on the
active model. The caps flow gains the model-edit preset ladders +
Custom… free input for context window / max output (a preset pick
writes models.overrides AND rebuilds the active model on the spot —
the old flow's live-mutation semantics, now durable), every field row
shows the current EFFECTIVE value with its source layer (pinned /
role override / catalog), the thinking row is an omp-style selector
(only levels the model supports, per-rung budget metadata, current
level preselected), and the hub row describes the effective summary
(200K ctx · 16K out · thinking off) instead of a pin count. The
scripted /model-edit <field> <n> form and the line-mode limits print
are unchanged.
feat(gh-1746): opt-in session disk retention — sessions.retention in
~/.fah/config.yaml caps the sessions root and sweeps whole session
units (transcript + sidecars) oldest-last-used-first at boot, exempting
live, named, young, and the current session; nothing is EVER
auto-deleted without the section (byte-identical legacy boot). Adds
fa session clean [--dry-run], the /sessions + fa session list
usage footer under a cap, and header-cwd/path-escape guards so a
planted session header can never steer a delete outside any root.
fix(gh-1763): a bash tool command running broad process kills can no
longer signal its own shell or the harness. Scripts are delivered from a
per-exec private temp file (sh -c '. "$1"' sh <script>) so the wrapper
argv never carries the script body (pkill -f <pattern> cannot
self-match its own sh -c; $0/$#/exit codes/set -e/stdin stay
identical), a PATH shim for pkill/killall plus a kill() prelude
guard exempt the harness pid chain + process group (exported as
FA_PROTECTED_PIDS) and the invoking shell from the child's own kills
while legitimate victims (stuck test runners, background servers) are
still terminated, and headless SIGTERM aborts now log
fa: SIGTERM received … so a child-issued kill never masquerades as a
user Ctrl+C (exit code contract unchanged: 130 graceful, 143 escalate).
fix(gh-1778): the release gate no longer reds a fully-green shard when
GitHub's cache service degrades. v1.0.621's Core tests (shard 2/4) was
killed by its own timeout-minutes: 12 with 2535/2761 tests green — both
cache steps had stalled ~133 s on keys that existed (degraded service,
not eviction), cold pub get + native-asset hook builds delayed the
first test to ~5 min, and the shard needed ~14 min against a budget
sized for the warm ~8.8 min wall. The job budget is now 20 min (sibling
test-leg band), and both test-core cache steps are best-effort
(timeout-minutes: 2 + continue-on-error: true) — a failed restore is
slower setup, never a red shard. test/ci_core_shard_timeout_guard_test.dart
pins the budget and the best-effort cache contract (RED-first on the old
YAML).
feat(#1761): the 32KB project-context budget never cuts silently again.
A file truncated to its head excerpt now carries an extended
<!-- From: path — N bytes, only the first M shown --> marker plus an
explicit MUST-read-the-full-file instruction (offset/limit/ranges,
prompt copy in prompts/context/project_context_read_full.md); a file
dropped entirely leaves a one-line tombstone naming its path instead of
vanishing; and the user gets a dim boot line per truncated/dropped file
(CLI line mode prints directly, TUI re-prints after the banner, the app
logs through its AppLog boot-notice seam). New projectContext: config
section: maxBytes (default 32768) — 0 disables the section entirely
and reads no files (fast boot); unknown keys are tolerated with a note.
Leaf-first allocation is unchanged.
fix(gh-1757): the Android build no longer re-enters the sdkmanager
download roulette for CMake. v1.0.619's build-android leg died at
:jni:configureCMakeDebug[arm64-v8a] — the :jni module (transitive
jni package via path_provider_android) wires externalNativeBuild.cmake
with no version, so AGP auto-installed its default cmake;3.22.1
mid-configuration and one corrupt archive ("Archive is not a ZIP
archive") red'd the leg, the aggregate Quality gate, and the
Advance-vpinned-marker job. The gh-1655 pre-install action now also
pre-installs cmake;3.22.1 with bounded retry + idempotent skip (same
hardening the NDK path already carries); the guard test pins the new
AC6 assertions and the AGP-version coupling (AC7).
feat(#1750): per-model image caps — the outbound request never carries
more unique images than the target model accepts. Model gains an
optional maxImages (catalog-seeded: gemini-4-argon: 10, measured);
images.perModel.<modelId> config overrides it (strict images:
section, user truth wins); the registry resolves
perModel > catalog > images.maxPerRequest (first present wins, the
issue-#1554 shape). A Google-shaped too many images … maximum allowed for model <id> is <n> 400 is classified (image_overflow.dart, the
overflow.ts twin), latches the learned cap for the (provider, model)
pair for the run, and re-sends ONCE through the registry at the
smaller cap — current-first-then-newest drops with the existing
never-silent notices naming the learned cap; a second refusal
surfaces exactly as today. JSONL is never rewritten;
images.registry: false keeps the byte-identical legacy shape
including the recovery-off behavior.
feat(gh-1760): oversized user-message intake spills to a file instead of
blowing the context window. Any incoming user message — headless
-p/positional, TUI first/mid-session, steering/follow-up drain,
external inbox — whose text passes the new relative threshold (a
spills.userFraction of the live window budget minus the compaction
reserve, floored by spills.userThreshold bytes) is redacted and
written verbatim to .fah/spills/<sessionId>/user-<n>.md (the #678
layout) from one choke point (Agent.userMessageGuard); the model
receives a bounded head+tail preview plus a read-in-parts instruction
riding the replacement USER message (copy in
prompts/spill/user_read_instruction.md). Images stay inline, a
refused spill write degrades to today's inline behavior, and
spills.enabled: false / userThreshold: 0 / the new
FAH_USER_SPILL_THRESHOLD env knob (bytes) keep the legacy shapes.
fix(gh-1743): a structured compaction that hits its total-budget
watchdog (the factory's 300s totalBudget) no longer aborts the
live session. AutoCompactorFactory.run() now applies the
mechanical local-trim fallback (keepRecentTokens, pairing-repaired,
[context trimmed locally] marker — the same valve the classic
pass ladder ends on) instead of failing fast; the honest
both-roles-failed verdict is kept only when nothing is droppable.
fix(gh-1636): auto-compaction stops dumping the internal checkpoint
summary into the console — the default report is ONE marker line
(● auto-compacted · engine: N → M tokens (F freed · P%)); the records
line + fenced summary moved behind the new tui.showCompactionDetails
opt-in (default false), and the manual /compact still prints the full
block. The headless log face no longer streams the summarizer deltas by
default (the gh-1433 E4 live stream is part of the same opt-in; the
gh-1430 … reasoning liveness line still covers the window). Nothing
is lost — the summary stays in the session file, reachable via
compact_expand.
fix(test): gh-1732 — the wrap_ansi_line_perf_test sustained-contention
validation leg spikes 3 of the 5 doubled runs per attempt instead of 4
(CI run 39546301132, core shard 2/4: the old shape left exactly ONE
real-wall-clock-timed doubled sample per attempt, and real jitter on
that single sample — the very shape being modeled — inflated all three
attempts' min ratios to a 3.30x best against the 3.2 bound while the
real 500k growth legs on the same runner stayed green). 3-of-5 is the
documented live-red shape ("≥3 of the 5 timed runs"): median-of-5
[c,c,s] is still a spiked value, so the leg still proves the
pre-gh-1532 median harness fails, and the min harness now rides two
clean doubled samples per attempt.
fix(#1705): CLI fresh-install boarding asks the provider choice FIRST —
the same list the /add-provider picker shows (presets, Custom last) —
and only Custom continues into the api-type wizard; a preset opens its
own flow with the api type prefilled. Previously boarding opened the
custom wizard directly, cold-asking "api type" before any provider was
chosen.
feat(providers): gh-1704 — first-class Codex/ChatGPT credential export,
CI refresh action, and Provider Queue integration (Copilot-export
parity). fa provider export <name> [--check] resolves the saved
ChatGPT login's OAuth blob from its secure-store slot, refreshes it when
stale, and prints the shell lines (OPENAI_API_KEY + OPENAI_BASE_URL
= the Codex backend — these tokens never serve api.openai.com); the
refresh token never prints and --check reports credential health with
the SSO caveat. /provider chatgpt import [path] adopts an existing
codex login session (~/.codex/auth.json) into the same slot — no
browser round-trip. .github/actions/chatgpt-codex-auth mints a fresh
access token from a CHATGPT_REFRESH_TOKEN secret in CI (direct curl,
no third-party action); invalid_grant fails fast and loudly.
providersQueue: entries speak chatgpt-codex: an omitted apiKeyEnv
resolves the Keychain slot automatically (in-memory refresh), a new
refreshTokenEnv: field mints the initial access token on slot
activation, and a raw Bearer token in apiKeyEnv is tolerated by
chatgpt_codex.dart (no more FormatException — proactive/reactive
refresh degrade, a 401 surfaces as the auth death the queue fails over
on).
feat(bench): gh-1681 — leaderboard submission readiness. ATIF
trajectories: new exportAtifJson whole-snapshot exporter
(lib/src/trajectory/atif_export.dart, ATIF-v1.7 — the schema harbor
0.23.0's flagship adapters emit) behind a headless fa trajectory export [--format atif] [--output path] verb; the harbor adapter runs
it as a bounded best-effort exec after the agent phase on every
terminal path, so <trial>/agent/trajectory.json reaches the hub with
the other agent logs (the leaderboard static analysis requires one per
rewarded trial) and result.json metadata carries an
atif_trajectory present/missing marker. The golden fixture
(bench/harbor_fa/fixtures/atif_sample.json) is pinned Dart-side and
cross-validated against harbor's pydantic Trajectory +
validate_trajectory in test_atif_contract.py. Workflow: a
submission: true dispatch input turns the submission rules into hard
setup preflights (full task coverage, ≥5 attempts/task, no
agent-timeout override, pinned dataset) and makes the new per-shard
ATIF presence gate fail hard; default runs gain only a warn-only scan
(artifact flow unchanged). Runbook: docs/leaderboard-submission.md
(pinned dataset + @latest content-hash recording, quota-window runner
policy, lb CLI steps).
fix(headless): gh-1589 — the headless session becomes a SEALED artifact:
no session-JSONL writes land after the run's agent phase ends. New
Session.seal() (state in the writer, no file-format change): raised in
runHeadless' finally and the REPL teardown as the LAST session action
(close order: drain → usage fold → seal → exit, one implementation both
paths — REG-2), it makes every writer (_persistMessages,
shell_job_registry snapshots, compaction, any appendCustomEntry,
leaf moves) drop late appends with exactly ONE stderr warning per
writer class — never a throw into an unhandled zone; the exit code
stays the turn's terminal outcome (issue #413 semantics untouched). The
settle→reaction path is sealed with the file (I2): the drain wrap-up
raises a fence (_reactionRunsSealed) so a background-shell settle or
task completion arriving after the final answer's drain — including the
registry's one-turn-deferred listener racing _finishHeadlessDrain's
check-to-return window (the 65gGW7b 10:13:42–10:14:00 burst) — does
board bookkeeping only: no _startRun, no model request, no record.
An honest drain's records (final answer → seal) stay legal (E7); the
seal is per-run state, so resume/replay/session_repair are unchanged.
Adapter side (bench/harbor_fa, I3/AC4+AC7): run()/perform_task
returns only after the fa process exited on EVERY path — the ladder
kill became terminate-and-reap (SIGTERM → bounded exec await → SIGKILL
escalation → pgrep confirm), and the external-cancel paths (harbor's
own wait_for firing first, the no-knobs exec, external-cap-only mode)
now fire the kill + bounded reap before propagating, with a loud
agent-phase exit guarantee degraded stderr note when fa survives both
signals (never a silent proceed, never an unbounded hang).
fix(bench): gh-1643 — a trial killed mid-stream no longer reports 0/None
usage. The fold read only FINALIZED assistant messages, so a run killed
before its first assistant message (the costliest rows) folded to silent
zeros. Both folds now price a dangling model_request_summary at its
KNOWN request-side volume (system prompt + outbound message chars at the
shared chars/4 rate, estimated input; output unknown) and mark the usage
incomplete instead: UsageChainScan.usageIncomplete → usage.json
(usageIncomplete: true, written only when set so completed folds stay
byte-identical), SessionUsage.usage_incomplete → the tb adapter's
usage-guard.json row and harbor's agent_result.metadata, and the
post-mortem stamps usage_incomplete on every results.json row it
patches from a partial fold. Harbor also exports the run outcome class
(agent_result.metadata.failure_mode: agent_timeout self-kill /
external_kill / completed / crashed / never_started) so downstream
classification stops relying on exception.txt archaeology; a summary
superseded by another summary is flushed instead of silently dropped on
both sides. Never-started trials keep the gh-1638/gh-1576 audit-only
precedence (no fake usage).
fix(stuck-calls): gh-1632 — a stuck tool call in an interactive session now
escalates like everywhere else. effectiveStuckTool() defaulted every
non-headless host to StuckFollowUpMode.advisory when no explicit
agent.stuckTool: was configured, so the fa 1.0.580 incident session (a
dead git fetch in the REPL) fired its tool_stuck flag at the 300 s
threshold and then hung for 27 more minutes — advisory mode never cancels,
retries, or converts; only the owner's manual steering broke the deadlock.
The autonomous default (cancel → retry once → background-convert, then a
session-visible escalation) now applies in EVERY mode; advisory remains
an explicit opt-out (agent.stuckTool.followUp: advisory), never an
implicit default.
fix(bindings): gh-1624 — the gh-1570 binding preflight no longer kills
translating-proxy factories. _validateModelEndpoint's vendor-prefix
verdict was fatal on ANY endpoint, so FA_PROVIDER_TYPE=anthropic +
bifrost's /anthropic path + zai/glm-5.3 exit-64'd every boot on
1.0.576 (the prefix is a routing key on a translating proxy, not a
dialect declaration). The vendor-prefix check now fires ONLY on the
endpoints that carry a dialect verdict — the three family-exclusive
canonical vendor endpoints (api.anthropic.com, api.openai.com,
generativelanguage), mirroring the family verdict's scope: those keep
the gh-1570 fail-fast, ROUTERS (OpenRouter/Copilot/DIAL/kimi — the
prefix is a routing key there too; OpenRouter's canonical ids are
anthropic/claude-*) and unknown/private endpoints carry the mismatch
as a ProviderModelBinding.warnings entry, never exit 64. The warning
is rendered on EVERY resolve path (gh-1624 review): the FA_PROVIDER_*
boot notice, the provider queue (per entry at boot, on queue reload,
and in /provider queue test), and the --model/--base-url flag
revalidation — a rider never boots silently.
Breaking-change note for 1.0.576 upgraders behind proxy
factories: the migration is the wire: FA_PROVIDER_CONFIG key —
documented now (README env-preconfig section,
docs/agents/providers.md §Translating proxies, --help) with its
caveat: the override engages only when the declared wire DIFFERS from
the provider record's own kind (record anthropic + wire: anthropic
still runs the checks). REG matrix (AC1/AC2/AC3) pinned in
test/model_roles/provider_model_binding_test.dart; AC4 IT drives a
bifrost-shaped loopback proxy through the production anthropic adapter
(test/model_roles/bifrost_proxy_binding_test.dart).
gh-1596: the first-run wizard (site/setup.sh) no longer crashes or
corrupts input under curl | sh on dash (Ubuntu sh) or busybox ash
(Alpine): scripts/gen_installers.dart now renders the provider/mode/
approval menus from the live provider catalog into plain POSIX sh — the
bashisms that killed dash (printf -v, read -p) and silently corrupted
busybox ash are gone; the hidden API-key prompt uses stty -echo + a
by-name read (dash has no read -s) and the key never rides printf
argv; EOF/Ctrl-C mid-wizard aborts non-zero with no half-written
~/.fah, and non-TTY runs print guidance and exit 1. The wizard's temp
config is mktemp-created (no predictable, symlink-followable fixed
path). A new path-triggered installer-portability CI leg proves the
script under dash + busybox ash (ubuntu) and bash (macOS) on every
site/** or generator change.
fix(cli): gh-1572 — fa boot no longer rewrites ~/.fah/config.yaml
from a lossy config. persistConfig (bin/fah_runapp.dart) built a
fresh CliConfig carrying only some fields and persistConfig() ran
unconditionally at boot, so every launch silently dropped the user's
spills:, subagents:, jobs:, headless:, agent:, quota:,
tui:, trajectory:, waiting:, power.hold and auto_update
settings (first reported on 1.0.547). Three cooperating fixes: the
boot save is now saved.copyWith(<live overlays>) so every parsed
section rides along by construction; CliConfig.copyWith replaces the
drifted hand-rolled withCustomProviders field list (itself dropping
spills/subagents/waiting/jobs/agentMode/misuseBreaker/streamThinking
inside saveCliConfig's merge step); the _diskPreservedSections
disk backstop covers every known top-level key. spills: renders in
toYaml (new SpillsConfig.toYaml) and agent.misuseBreaker: false
persists too; an unchanged save no longer rewrites the file at all.
Rework round: the boot snapshot re-syncs from disk after every
sanctioned mid-session settings write (onUserConfigFileChanged — a
cleared context cap, an auto_update key drop or a global compaction
pick is no longer resurrected/overwritten by the next save in the
same session); validateAgentSection (the strict fa config check /
config set / settings-hub mirror) accepts the emitter's
agent.misuseBreaker and agent.stuckTool keys so fa config check
never false-fails on fa's own output; clearing the context cap drops
only the cap key instead of the whole agent: block (the sibling
mode:/misuseBreaker: keys survive); and the providersQueue:
block (no typed renderer) joins the disk backstop so a save can never
silently delete the user's queue. Rework round 2 (merge + guards):
validateAgentSection's per-key switch became the _agentKeyValidators
table (the two gh-1572 keys pushed the switch form to complexity 13 —
over the repo's CRAP ratchet; the table keeps the mirror complete at
complexity 4), and the last two direct USER-config writers —
_writeTtsrSection and _writeMcpSection — fire the re-sync hook, so
a mid-session ttsr rule toggle or mcp server edit survives the next
persist too (the writer roster in settings_flow_config_sync.dart is
now exhaustive).
gh-1606: the copilot provider accepts fine-grained PATs
(github_pat_…) directly — PAT mode skips the copilot_internal/v2/token
exchange entirely (GitHub 404s every PAT exchange) and sends the raw PAT
as the Bearer against the configured Copilot base URL with the official
Copilot CLI's first-party identity headers
(copilot-integration-id: copilot-developer-cli,
user-agent: copilot/<fa-version> (<os>) term/unknown, installed by the
host at startup). OAuth device-flow tokens keep the exchange +
proactive-refresh flow unchanged; a PAT 401/403 surfaces the error
instead of a doomed refresh retry. Headless COPILOT_GITHUB_TOKEN now
works with a PAT. Connect flow: the paste hint explains PAT mode, the
dead-PAT case surfaces at the model fetch, and /provider export copies
the saved Copilot account's full token to the clipboard (account picker
on several entries; the token never prints to the transcript).
gh-1575: macOS clipboard paste fixed for TIFF-only pasteboards + the
Ctrl+V text fallback. _readMacos probes «class TIFF» after PNGf/JPEG
miss and converts via sips -s format png (a Preview/browser copy used
to dead-end on clipboard holds no image; a failed conversion keeps the
unavailable fall-through and every scratch file is cleaned up); new
readPasteboardText (pbpaste / xclip+wl-paste / PowerShell
Get-Clipboard) behind the same injectable ProcessRunner/platform:
seam, and Ctrl+V with a TEXT clipboard now inserts the clipboard text
into the composer instead of printing the image probe's dead-end reason
— bracketed-paste routing, gesture pairing and silent-failure semantics
unchanged. Review pass: the Windows Get-Clipboard read pins UTF-8
console output and normalizes embedded \r\n (OEM-code-page mojibake),
the Linux text backends pin text/plain;charset=utf-8 (an image-only
clipboard can no longer leak bytes as "text"), the library doc no longer
overclaims the FA_FAKE_PASTEBOARD seam, and FaTuiCallbacks moved to
a fa_tui_callbacks.dart part file keeping fa_tui.dart under the
2800-line gate.
gh-1555 (W1, lazy/windowed-everywhere reads): the legacy full-open call
sites migrate to windowed opens, so a session ledger past the gh-1073
2 GiB full-open backstop stays openable on every user-facing read path.
The TUI session picker, /resume, and the boot name-resolution scan
open windowed and resolve the display name through paged reads; the
subagent transcript preview opens the child windowed (tail only); the
_loadSession marathon fallback no longer degenerates to a refused
full open when the boundary walk exhausts its page cap (it proceeds
with the tail-anchored windowed session; the boot cap owns over-window
residue); and the headless fa trajectory resolver opens windowed,
reading the whole ledger for the cost/inspect views through the new
SessionTree.ensureWholeLedgerResident — a sequential forward scan in
bounded strips (WindowedSessionStorage.loadAllOlder), not the
boundary-walk's doubling strips (whose per-strip record cap measured
~13x read/parse amplification over the whole file). The same change
block-prepends paged chunks in _indexChunk (one insertAll per
list), removing an O(n²) the suspended-eviction walk paid per record.
Trajectory export streaming (AC3a), split-and-retain rotation
activation (W2), and the opt-in write-side caps (W3) remain follow-up
work.
gh-1605: the FinalizeGate gains a "grader parity" named check. A
produced entry point (script/CLI the verifier will run itself) is only
verified when it survives the grader's own invocation: the checklist
must re-run each produced entry point from a different working
directory in a fresh subprocess with a clean env, capture stdout/
stderr/exit code, and diff against the earlier in-session run — any
divergence (different output, relative-path artifacts, locale/
encoding-dependent output) is a fail ledger item, never a pass.
Closes the modernize-scientific-stack__nwbgy7a near-miss (every
in-session checklist item passed, the verifier's clean-cwd subprocess
collapsed Station 102 to Station 101's mean, reward 0.0). REG coverage:
the parity procedure catches cwd- and env-dependent entry points and
does not false-positive cwd-independent ones
(test/agent/finalize_gate_grader_parity_test.dart), and an IT replays
the evidence shape end-to-end — divergence detected and persisted as a
fail ledger item before finalize.
gh-1603: unattended runs can no longer end as a clean success on empty
model responses. Loop: from the second empty-hiccup retry on, the request
carries a continuation nudge user message ("your last reply was
empty/truncated — continue the task"; exact-text constant
emptyContinuationNudge, emitted with MessageStart/End so the transcript
stays truthful) — a cap-truncated reasoning model re-streamed verbatim
repeats the empty outcome, the nudge breaks the symmetry; the classifier
treats that exact message as transparent, so the turn's partial-content
window (gh-1577) survives a nudged retry. Headless: the empty-reply nudge
budget rises to AgentCliConfig.headlessEmptyReplyNudges (default 2 — the
session ends only after 3 consecutive unrecoverable empties; interactive
keeps the #1085 M2b one-nudge semantics), and a run that still settles
empty prints the explicit __FA_INCOMPLETE marker and exits 1 — a
distinguishable failure for the bench adapter, never a silent exit 0.
gh-1577: an empty assistant completion after partial content (assistant
text, tool activity, or a truncation signal earlier in the turn) is a
bounded-retry provider hiccup — up to emptyHiccupRetries (default 3)
re-streams with exponential backoff (emptyHiccupBaseDelay,
emptyHiccupMaxBackoff, both threaded through the Agent facade,
AgentWiringSpec, and the CLI config), and exhaustion ends the run with
a loud StopReason.error (headless exit 1, interactive error + prompt)
instead of the silent exit-0 settle that killed the gh-1577 trials.
Clean empties keep the #1085 M2b one-nudge semantics. The abort always
wins: a cancel racing the final retry settles StopReason.aborted for
any empty shape (the exit-contract race is pinned by tests), and the
backoff sleep is cancellation-aware.
gh-1574: SQLite reads work on installed binaries — bundle libs land where
the binary looks. The AOT binary's native-assets resolver reads
<exe_dir>/../lib/<name> (the dart build cli bundle layout), but every
POSIX installer copied the bundle libs beside the binary
(<install_dir>/lib), so read path=x.db died with a raw FFI dump
(Couldn't resolve native function 'sqlite3_initialize' … Failed to load dynamic library '~/.local/lib/libsqlite3.dylib'); it only worked on machines
with a manual symlink in ~/.local/lib. Fixed the layout in all three
install paths — site/install.sh (incl. the quarantine-xattr pass and the
source-build advice), install_local.sh, and fa update
(bin/self_manage.dart copyBundleLibs, now public and layout-pinned by
tests; also aligns Windows fa update with install.ps1, which was
already correct) — plus clean degradation in the FFI engine:
Sqlite3Engine.openReadOnly classifies the dart:ffi resolver's
library-load failure (isSqliteDynamicLibraryLoadFailure) and throws a
clean SQLite is unavailable — reinstall or run fa update note instead of
the raw resolver dump; a Dart-level system-library fallback is impossible
with package:sqlite3 native assets (build-time asset mapping, no loader
override). Installs already broken by the old layout self-heal: on the
bundled library's first failed load the binary moves the stale
<exe_dir>/libsqlite3.* into <exe_dir>/../lib/ and retries the load
(repairStaleSqliteBundleLayout) — so one fa update suffices (the
successor binary repairs on its first SQLite read; a pre-fix binary cannot
repair itself), and the stale <bin>/lib leftovers disappear as they
migrate. The POSIX installers also report a failed lib copy instead of
swallowing it (site/install.sh warns naming the lib dir; same in
install_local.sh), and site/install.sh.template carries the same
layout so regenerating the installers cannot resurrect the old one.
gh-1556: bench pipeline stops bleeding score + fa deadline awareness.
Bench: the harbor ladder kill uses the [f]a bracket trick and is
wrapped so a kill failure can never replace the TimeoutError (the
pre-fix self-matching pkill classified TB 2.0 ladder kills as agent
crashes — a SOLVED task scored 0); progress-watch (stall-gap 360s per
the gh-1430 watchdog invariant, flat abs ceiling 3600s) is the
bench-harbor.yml default on every shard, with the harbor adapter
picking the gap-aware ProgressWatch for parity with the
terminal-bench adapter; the aggregate counts completeness by
result.json presence (errored trials report as their own breakdown
line; exit 1 only on genuinely lost trials) and renders the full
report — table, errored breakdown, ledger row — to
$GITHUB_STEP_SUMMARY (TB1 parity). Both bench adapters export
FA_DEADLINE_EPOCH_MS (effective cap − 120s wrap-up margin) into the
agent session and audit the exact value. fa core: deadline awareness —
FA_DEADLINE_EPOCH_MS / --timer/--deadline / /timer write one
session clock, the per-turn context carries a time-left: line while
armed, and one-shot system notices steer wrap-up at ~50%, T-3min
(drain background jobs), and T-30s; absent deadline = byte-for-byte
today's behavior.
gh-1522: tag-sourced versions — the git tag is the single source of
truth. The release-bot file bumps are retired: cutting a release = a
tag + GitHub Release (zero commits to main, zero file mutations).
Both pubspecs carry the fixed 0.0.0-dev placeholder (CI guards
hand-bumps), the publish job stamps the staged pubspec + generated
changelog section from the tag (scripts/stamp_staged_release.sh,
staged version == tag guarded pre-upload), store builds stamp
--build-name/--build-number from the derived tag version
(pubspec.lock stops churning across releases), and
install_local.sh stamps version.txt from git describe.
feat(usage): gh-1460 — the fa-tokens: segment-close line carries the
session's model id, so every token row the factory publishes is
attributable to a model and therefore priceable. The fold now captures
each segment's LAST-SEEN model id (the model that served the final
request, the same string settings/UI display — mid-session fallback
take-overs attribute the segment to the model that actually finished it;
zero-filled fallbacks never clobber a seen model) and the pinned line
shape grows an OPTIONAL "model":"<id>" key after "segment" — omitted
when no model was observed, so legacy lines parse unchanged and the
pinned regex (usageTokensLogPattern, gh-1241 AC7) keeps matching both
shapes. usage.json's schema is untouched (the model is rebuildable from
the chain, I6); old consumers using tolerant JSON ignore the new key,
and the repo's single strict consumer (the pinned constant) is updated
in the same change.
fix(test): gh-1460 rework — the wrap_ansi_line_perf_test growth-shape
legs measure the doubling ratio with the MINIMUM over 9 runs instead of
the median of 5 (CI run 38112524699, core shard 0/4: a scheduling pause
inflated the doubled-size median — 12.09ms against a true ~7ms — and
tripped the 3.2 bound at 3.65x on a genuinely linear wrap). Runner
jitter only ever inflates a wall-clock sample, so the min converges on
the true compute time; genuine superlinearity still trips the ratio
(the gh-1496 quadratic candidates measure ~4x by min). The absolute-
budget legs keep the median — their gh-1357 headroom already prices in
jitter. A synthetic regression test pins the pause-proof property.
<!-- release-draft-owner: run/42620276095/job/binaries-linux-arm64 -->
feat(gh-2059): browser-extension compaction parity with CLI — over-window relief, /compact with visible receipts, and the shared under-window guarante
feat(gh-2059): browser-extension compaction parity with CLI — over-window
relief, /compact with visible receipts, and the shared under-window
guarantee ride the same CompactionHostWiring resolution as the CLI and
app (resolveExtensionCompactionWiring); receipts are durable across
panel attach/SW restart. Machine-review round 1: exact ladder pass
counts, busy-matrix routing (routeUserInput) for panel text and second
/compact during a pass, window settings from the ACTIVE stored source
with positive validation, string-encoded int merges.
fix(gh-2055): the daily-publish cli leg red-failed in pod install
(Build macOS x86_64, run 41778615815) because the vendored macOS pod's
XCFramework download block guarded on the downloaded-zip MARKER
(if [ ! -f wasm_run-v0.1.0.zip ]) instead of on the artifact the build
consumes. A fresh CI clone never has the marker, so pod install re-ran
the block on every build: it removed the WasmRun.xcframework the
build-macos.yml restore step had already placed (cache hit,
sha256-verified), re-downloaded the 92.7 MB asset with a bare
no---fail curl, and one transient reset (curl 56 at ~3 MB/92.7 MB)
left the Podfile nm -gU symbol check with no
macos-arm64_x86_64/libwasm_run_dart.a to read. The guard now keys on
WasmRun.xcframework/macos-arm64_x86_64/libwasm_run_dart.a — the exact
path the Podfile force_loads and nm-checks — so the block is a no-op
whenever the restore step (or a previous run) put the framework in
place, and the fallback download for a truly cold tree carries
--fail AND aborts the block at the curl step on failure
(|| { echo "wasm_run download failed" >&2; exit 1; } — a swallowed
status fell through to unzip/truncate residue and masked the root cause
behind the Podfile nm diagnostic; review round 1 on this PR), and the
podspec now assigns the backtick result and raises "wasm_run framework
fetch failed" unless $?.success? — Ruby ignores a bare backtick's
exit status, so pod install would otherwise have sailed past the failed
fetch to the downstream nm diagnostic (review round 2). Linted
behaviorally (stubbed curl, no network; POSIX-gated sandbox, plus a
ruby evaluation of the real fetch statement) by
test/wasmrun_macos_podspec_download_guard_test.dart.
fix(gh-2050): the daily-publish cli leg red-failed daily when its
macOS attach arrived after a sibling train had already PUBLISHED the
derived tag (the same-tag convergence race: mobile leg drafts + the tag
train publishes ~25 min after dispatch while the serialized self-hosted
mac builds need ~66 min — v1.0.685: draft 05:22:46, published 05:44:52,
mac leg attached 06:28:53). scripts/release_asset_gate.sh gains a
caller-declared pre-stage policy — RELEASE_GATE_PUBLISHED_SKIP=1 —
that turns the PUBLISHED+unattached-fa-* verdict into a loud whole-call
skip (::warning::, exit 0, nothing uploaded, no twin create, no flip;
all-or-nothing so latest-mac.yml can never land alone on a release
whose zips were skipped). The default stays the gh-2042 AC1 loud
refusal; only build-macos.yml's release-macos desktop attach opts in
(its DMGs/zips/manifest are clobber-refresher pre-stage assets, not
publish blockers), the CLI-tar/extension attaches keep the refusal (they
carry REQUIRED release assets). The stranded desktop feed self-heals:
the next daily leg pre-stages the next version's draft with fresh
assets.
fix(gh-2044): the auto-compact fragile-band escalation could fold the
ENTIRE visible projection when the remaining path fit inside
keepRecentTokens — the keep-recent walk defaulted to folding
everything when its budget never exhausted, swallowing the live tail
(the newest user ask and assistant turn included). Real-world trigger:
resumed subagent runs ended with no assistant message at all
(subagent produced no assistant message — CI cores shards 1/4 and
3/4, hostile-env, both red). _pickCheckpointRange now declines the
fold when the whole foldable path fits the keep-recent region and
still carries live records; all-marker paths keep the lossless
deterministic consolidation (issue #387). The escalation receipt also
reports the REAL summarizer call count (a #729 chunked fold makes N
calls, not 1), and the factory wires its run token into the engine's
cancelToken so a watchdog-shaped cancel reason (any
TimeoutException) can no longer dissolve a user abort into a
"summarizer returned no text" failed receipt while the run reports
success (#1085). Follow-up CI red on the same branch: the decline
moved tiny-transcript runs to the classic residual engine, where an
EMPTY summarizable region folded zero #729 chunks into a blank
summary that still landed as a junk CompactionRecord and a hollow
success: true through the no-work terminal — compactSession now
refuses that empty-region blank fold (gh-1743's failure-verdict
contract kept; a dialed blank answer over a real region still lands
as a record, issue #578 unchanged). Third CI red on the same branch
(flutter-app over-window relief, both shards): the whole-world decline
lived in the shared _pickCheckpointRange, so it also starved the
checkpoint LADDER — a tool-result-heavy transcript whose every ledger
token fits keepRecentTokens (the request estimate is far over the
trigger) lost its only summarizer-based relief, and the hosted app's
over-window guard ended the turn instead of continuing it. The decline
is now escalation-only: ladder passes run over the trigger, where the
fold IS the relief (the escalation still declines — semantic 5 — and
its pass body was extracted below the CRAP ratchet's 12.00).
(gh-2044 rework, UT-ESC-13)
fix(gh-2003): the Copilot picker filter is tier-aware. Individual-tier
/models responses omit model_picker_enabled entirely, so phase 1
dropped every such entry as «picker flag absent» and individual/OAuth
accounts still listed zero usable models. When the response comes from
the individual tier — the individual host (api.githubcopilot.com or
pi-mono's api.individual.githubcopilot.com spelling) or a
no-proxy-ep token (individual by design) — an ABSENT flag is now
selectable: absence ≠ an explicit false, which still drops
everywhere (embeddings, completion-only). Business/enterprise answers
keep the strict picker list unchanged, and the keep is never silent:
the fetch note names it («2 of 2 models kept via the individual-tier
fallback (picker flag absent)») through the existing fetch notes:
surfacing. pi-mono parity: the trigger mirrors allowPolicyFallback = baseUrl === "https://api.individual.githubcopilot.com" in
packages/ai/src/auth/oauth/github-copilot.ts.
fix(gh-1990): the Copilot /models dialect never answers a silent
empty list again. Every silent-empty exit names its branch through the
new fetchModelsForEndpointReport dispatch (ModelsFetchReport +
ModelsFetchDiagnostic): the token-exchange HTTP status (with the
OAuth device-flow / fine-grained PAT remedies for classic ghp_…
keys), the /models HTTP status or timeout (with the seat + per-tier
host hint), proxy-ep-less individual tokens aimed at a business/
enterprise configured host (the tier-misrouting hint), and the picker
filter with per-model drop reasons (picker flag absent, picker flag false, policy disabled, no tool_calls, endpoint unsupported) —
«12 models fetched, 12 filtered (picker flag absent)». A failed
primary credential path retries the other path ONCE before declaring
empty (dual-path fallback, gh-1606): an exchange-401 classic PAT tries
PAT-direct (the provenance note names which path answered), a
fine-grained PAT never falls into the exchange, an OAuth token never
becomes a direct PAT Bearer. Surfacing: /models prints a fetch notes: section per provider, the /provider copilot connect flow
prints the named failure before manual entry, the app's model
quick-select (EndpointModelsController.emptyReason) renders it under
the field, and the Copilot connect sheet shows the reason above the
manual model id. The legacy fetchModelsForEndpoint answers are
unchanged (parity); the filter criteria themselves are unchanged
(pi-mono parity — only the reporting is new). Review round 2: the
fetch notes: wiring covers the unsaved sibling paths too (the
active-catalog and env-keyed Copilot fetches surface the same named
failure instead of a silent seed fallback), the connect flow's
no models: line reuses ModelsFetchReport.reasonText (the last
hand-rolled message — hint join), and the ModelsFetchReport
docs scope the diagnostic contract to the dialects that name their
failure branches.
feat(network): channel payload sender block — app members' channel
messages carry {"text": …, "sender": {"name": …, "id": …}} (public
base64 JSON and keyed fanet1 plaintext alike), so DAP agents can
attribute the HUMAN author behind the fa_network relay's hub id. The
harness unwraps it tolerantly (raw text passes through), rides the
name as AgentMessage.fromName through the fabric, and the CLI
steering line greets people again: from Uladzimir (<relay id>) in #channel: hello instead of from 3cfaee1aa054fe6e…. The app falls
back to the payload sender for bubble names when the envelope carries
none; DMs also name hub peers from the whois cache. Legacy payloads
without the block decode exactly as before.
feat(gh-1956): reply auto-spill — oversized child reply / agent_message
payloads spill to an agent:// artifact automatically; the parent receives a
bounded excerpt + pointer + completeness marker with char AND byte counts.
Manual chunking degrades to a documented fallback, restorable across parent
resume via durable deliverable_artifact records, behind the
subagents.replyAutoSpill kill switch (default on; honored on the CLI and
the app host alike). Rework round: the salvage hook (output-truncated
deliverable calls) restores the durable ids before allocating its own
(no post-resume pointer shadowing), durability losses emit a
[deliverables] diagnostic line, the app host honors the kill switch, the
spill-reach gate is renamed spillAppliesTo (was spillApliesTo), byte
counts are encoded once per spill, and DeliverableRecord.artifact(bytes)
is optional again so external constructors keep compiling.
fix(#1947): schedule_message from the orchestrator no longer misroutes
into a finished subagent's mailbox. A background shell job started by a
CHILD registered its settle continuation inside the child's subagent
zone; when the job settled after the child's run ended, the CLI's
settle-wake (_onShellJobSettled → _startRun) started MAIN's reaction
run still carrying the finished child's activeSubagentId(), and every
identity consumer of that run misattributed it — the reported shape was
a self-reminder scheduled "in <session>/<child>" that auto-resumed the
completed child (duplicate-release risk). Three layers: the
ShellJobRegistry now binds its settle continuation to the
construction-time (boot) zone (_bootZone.run), both hosts'
scheduleSenderMailbox wiring trusts the zone scope only while the
child is genuinely in flight on the session executor
(TaskExecutor.isInFlight), and the tool response names the RECIPIENT
the record carries — an explicit to: used to be shadowed in the reply
by the sender's mailbox, which read as "the override was ignored" even
when it was honored. E2E repro drives the real AgentCli
(test/cli/agent_cli_schedule_recipient_test.dart) plus a registry-level
zone pin (test/tools/shell_job_settle_zone_test.dart).
fix(gh-1948): task_observe works on a RUNNING background subagent
instead of crashing with an uncaught SessionException on the
<parentSessionId>/<agentName> placeholder. Three legs: the executor
flushes each child's transcript incrementally at every turn boundary
(the JSONL exists and grows mid-run — also crash-resilience for
aborted children), task_observe resolves a placeholder handle's real
session file from the sessions tree (the same agent: subagent /
parent: / id: header scan the boot adoption uses) and heals the
handle onto the found path, and every read failure degrades to a
structured "transcript unavailable" / "no transcript available yet"
message instead of an uncaught tool error. The child-transcript reader
rides the transient-ENOENT retry (issue #427) — it reads while the
executor appends.
fix(#1941): the selected CodeMie project now rides EVERY model request.
The header is baked into Model.headers only by the connect-time
switches (gh-1809), so every other rebuild silently dropped
x-codemie-project — and the backend bills a header-less request against
the user's DEFAULT project (codemie-code header-injection.plugin.ts
contract): the reported symptom is limit-exhausted from the exhausted
default project despite picking a budgeted one. Rebuilds now re-resolve
the project from the saved entry through one helper
(projectHeaderForBaseUrl): the boot folder-state restore (JWT entries
previously booted header-less — the SSO cookie patch was the only boot
carrier), the session-resume re-apply (_buildRestoredModel), the
roles-chain failover (ModelRef.project, mirroring the #964 authHeader
leg), the roles-mode /model + settings-flow pins, and the generic
/provider switch. The models list (/llm_models, SSO + JWT + guided
setup + entry dispatch) and the quota adapter (dark until pinned) scope
themselves to the same project, so display and billing can never diverge.
Tests pin the header VALUE on the wire (real adapter over a fake
endpoint) for the restored and failed-over models.
feat(network): private networks created outside the app surface after
sign-in. GET /api/networks/mine (fa_network#16) is synced into the
wallet after OAuth/dev sign-in and both restore paths — best-effort,
silent on older backends (404) and any failure, and never destructive.
Synced memberships keep the server's name/class, preserve local
passwords/join facts, and land without a password (the endpoint never
returns one) until the first join. Entering such a network — sidebar,
⌘K/⌘3-9, or a restored selection — opens the join sheet prefilled with
the network id instead of a resume error or endless spinner; joining by
name now keys the wallet/session by the echoed canonical id and migrates
legacy name-keyed rows (password + channel keys) instead of duplicating
them. A malformed /mine row is skipped, not the whole list.
fix(gh-1931): one render per tool call on the non-TUI console. Line mode
and the headless -p log printed every executed call twice — a start row
(• bash · cmd, no elapsed) followed by the end row (✓/✗ bash · cmd 0s). The duplicate start-row print is gone from _onToolExecutionStart;
the end row (command, status glyph, elapsed, failure news) is the single
render per call. The TUI keeps its running card + settled card, start
bookkeeping stays (_toolStarts feeds the end row's detail/elapsed, the
waiting coordinator still arms per-call liveness), and persisted
artifacts (session JSONL, trajectory records) are untouched —
rendering-only change. Long foreground calls keep the gh-1055 liveness
reminders past waiting.toolLivenessSeconds (default 60s).
fix(gh-1929): esc during the pre-flight auto-compaction no longer crashes
the CLI. The compaction layer's deliberate abort (#1085) used to escape
_runPrompt outside its guarded body and reach the root zone handler via
an unlistened whenComplete chain — fa crashed: CancelledException +
crash.log + a lost session view. The pre-flight window (submit on an
over-window transcript) now translates the cancel into a user-cancel
outcome: neutral dim submission cancelled receipt (never error:), the
TUI composer hands the typed text back (sendInputText) so a resubmit is
one Enter away, headless exits 130 with the same receipt, and the next
prompt works (the funnel's local-trim recovery or a retried compaction).
Cancels DURING the run — in-flight relief, settle funnel — keep their
pinned loud-abort surfaces, and the post-run window keeps its
compaction interrupted receipt via the shared _renderUserCancel
helper.
feat(1923): stall-proof steering. Steering arriving while the loop awaits
an UNPRODUCTIVE provider submit (zero bytes, or silent past the steering
stall window) now cancels the attempt via a SteeringInterruptFire
cancel, injects the pending steering into the rebuilt request, and
resubmits — the #1916-class submit stalls can no longer queue user or
agent mail invisibly. Productive streams keep next-boundary semantics;
the interrupt never consumes a retry attempt, and the run latch re-arms
in place so watchdog replays and user aborts keep working. task_resume
caps its foreground wait (like task_send) and continues in the
background; video polling honors soft-yield checkpoints (remote render
never restarted, honest hand-back text) and media generation honors run
cancellation; SLO-breached undelivered child mail surfaces in
task_status (UNDELIVERED … past the 2s delivery SLO); the TUI prints
steering queued — provider stalled Ns, interrupting to deliver plus the
delivered ack.
feat(1916): image-heavy submits no longer livelock — byte-based
images.maxBytesPerRequest (default 8MB, layered onto the count cap;
/images names both drop reasons), legacy text-embedded data-URL
screenshots in history fold into the once-only registry instead of
replaying raw forever, stall watchdog windows scale with the outbound
payload (capped at 2x; explicit connect clamps untouched), abandoned
connect attempts answering during the retry backoff are adopted, and
repeated identical-payload stalls fire ONE actionable /compact +
image-budget advice. TUI stall banner + docs/images.md carry the
image-heavy playbook.
fix(gh-1912): Add an agent → DAP now turns a retired hub's unknown
enroll_create response into an actionable upgrade message instead of
exposing bad_frame. Independent enrollment never falls back to the
legacy same-name credential rotation endpoint, and existing credentials
remain untouched on failure.
fix(gh-1913): Add an agent is usable on narrow phone screens: scope and format choices wrap without per-character labels, enrollment controls stack at full width, and actions remain reachable by scrolling. Desktop keeps the segmented layout; phone/desktop light/dark goldens guard both.
fix(gh-1825): DAP agent enrollments carry a stable, independent
agentInstance through app/CLI bundles, persistence, direct boot and
reconnect. Independent same-name agents keep distinct identities and stay
online together, while legacy bundles retain their project-scoped identity.
feat(gh-1825): DAP private-channel sends now use self-describing
fanet1 frames (sender X25519 key in the frame, channel-bound AAD) via
fa_hub_client 0.2.11, so fa_network app members decrypt them without
a whois/senderKey round-trip. Inbound channel frames open fanet1
first and fall back to legacy dap/1 payloads. Verified live in both
directions between a headless DAP agent and the Flutter app.
feat(gh-1825): the network offline banner now has a Reconnect now
action which restarts the socket and reloads the channel/member
snapshot. Fresh sockets also clear a stale offline flag (the server
sends no historical drain to a socket that was away during recovery).
AgentNetworkController.start stops after disposal instead of
notifying a dead controller.
feat(gh-1825): the roster sheet shows access-scope badges (whole network vs N channels, decoded additively from Member.access in snapshots and presence.changed) and owner/admins can drop an agent (DELETE /api/networks/{id}/agents/{agentId} with confirmation; the server tombstone keeps it out of future snapshots). Guests and store members are removable too via DELETE /api/networks/{id}/members/{id} (sessions revoked server-side); owner/admin/self rows stay protected.
feat(gh-1825): channel/network management in the app — rename a channel (PATCH), delete a channel or an owned network (REST DELETE with confirmation and state/navigation cleanup), and unnamed CLI/hub-created channels show a short friendly label instead of the raw 64-char id.
feat(gh-1825): pasting a DAP invite no longer demands a passphrase on macOS — importing NEW channel secrets auto-provisions a random passphrase into the per-channel Keychain item (decryption of existing envelopes never auto-generates; env/prompt/headless behavior unchanged), and passphrase warnings now name the concrete recovery (re-paste the invite to re-encrypt, or set DAP_KEY_PASSPHRASE).
feat(gh-1825): the network online counter opens a live-first roster sheet (names, member class, presence), and the channel composer offers roster-filtered @-mention autocomplete — the draft keeps the literal @displayName for DAP agents and the envelope carries the resolved member ids. On the CLI side the DAP channel notice names the agent's own identity (display name + agentId), and inbound channel mail that @-mentions that name (case-insensitive, word-boundary — no email or @name2 false positives) is marked "[mentions you — you are @name]".
fix(gh-1825): the app's network roster no longer reads as permanently empty ("0 online"): fa_network sends roster.snapshot with an OBJECT payload ({members: [...]}) and the app only accepted a bare list, so every snapshot silently decoded to zero members. Both shapes are now accepted, with a production-shape regression test. The DAP paste hint names web alongside CLI/macOS/iOS (plain-web join is supported).
fix(gh-1825): resumed CLI sessions rejoin their DAP channels and receive channel mail again — session replay now re-arms membership synchronization (plugin session-ready listeners on boot/resume and in-process switches) and idle channel probes complete deferred joins before inbox peek/drain. The post-replay sync also dials the membership's OWN hub with its enrolled credential before joining: a resumed session-scoped agent used to stay on the zero-config local URL forever, and because a disconnected fabric probe short-circuits before the sync, the remote dial never happened — the session-ready listener now drives it directly (throttled to one attempt per 10 s). App channel history no longer flashes raw payloads on first open: decoding waits for startup's channel listing instead of caching an undecoded guess until a manual switch.
feat(gh-1825): the fa app consumes the same DAP enrollment block as the
CLI — paste it into the agent chat composer (macOS/iOS/Android; the
smart-paste .txt chip path included) and the channel join happens
without the block ever reaching sendText, analytics, or session
records. Native joins ride the live agent-fabric hub repository:
per-project scoped identity, invite pub/priv import, keyless public
channels, reconnect-persisted membership in a dedicated secure
enrollment store (Keychain/Keystore where available). Harness additions:
HubMessagingRepository.joinChannel/waitForConnection with
rejoin-before-drain on reconnect, keyed channel decrypt
(hubDecryptPayloadWithKeyPair), public showcase payloads, and
HubBearerTransport (native Authorization header or the browser
?token= equivalent; authed dials never stringify transport errors
into logs). Plain web joins through the browser
WebHubTransport (?token= bearer equivalent, IndexedDB identity);
the browser extension rides the service worker's existing DAP client
(hub.join op), never a panel-local second connection. Add agent now defaults to the
DAP tab with a paste-into-chat hint; web gains ai-native provider
sign-in through the existing popup/callback OAuth flow.
fix(fa_network#4): the app inserts hub agents arriving via
presence.changed into the roster (additive displayName/class fields)
instead of dropping them, so the online counter no longer sticks at 0
when an agent comes online after roster.snapshot; legacy frames
without those fields create no phantom members. roster.snapshot/REST
contents are logged for decisive presence debugging.
feat(gh-1825): /network + scoped DAP channel membership — join by
paste (invite url / url+keys / bare channel id / whole fa dap import
command) with session/project/global scopes, live join without
restart, per-scope leave + copy invite, and the composer-paste
shortcut into the same wizard. Session scope rides a
dap_channel_membership custom record (leave tombstones, branch
resume); project scope is .fah/dap.json (committable, dapkey1
PBKDF2+ChaCha20-Poly1305 encrypted keys); global keeps the legacy
~/.dap files. Headless CI: DAP_CHANNELS_JSON (+DAP_HUB_URL,
DAP_CLIENT_SECRET, DAP_KEY_PASSPHRASE) joins at boot with zero
prompts, plaintext or encrypted privs mixed. Passphrase chain:
env → per-channel macOS Keychain item → one prompt with save default.
feat(gh-1844): media slots never fall back to the main connection
implicitly — SDK/CLI, app store, and the JS bridge resolve a slot
without an override as UNCONFIGURED with an actionable
"configure the slot" error; the main connection serves a slot only via
the explicit per-slot useMainConnection: true opt-in (YAML
models.slots.<slot>.useMainConnection / app's "Same as main
connection" pick, preserved through media_models.json). Image and
video generation accept multimodal input (text + reference images:
sandbox paths, http(s) URLs, base64) on capability-declaring
OpenRouter endpoints (input_references + declared/probed
input_modalities gate; unsupported endpoints fail before any network
call); the OpenRouter video job API is a first-class video dialect
(submit → poll → mp4). New decide tool: structured decisions via
OpenRouter's Decisions API (/v1/systemone) — bool/choice/score
questions with calibrated probabilities, structured error results, no
streaming.
feat(gh-1834): the FinalizeGate contract counts a parity/conformance
claim as verified only with negative + combination self-checks — new
numbered step 2.6 in prompts/cli/finalize_gate.md (refused
incompatible/mutually exclusive options, empty/corrupt inputs, boundary
values, must-NOT-work cases, alongside happy paths; a blanket
"everything matches" is not verified). Universality: abstract example
only, no benchmark/task/flag names in the prompt; the step is
conditional, so tasks with no conformance claim see no gate change.
Ledger/verdict wire formats unchanged; validated end-to-end by the
scripted IT-4 replay (real CLI + real bash) whose persisted
task_ledger carries the rejection rows.
fix(gh-1832): release v1.0.635 published without fa-extension.zip — the
landing-page download 404'd. The extension's only producer (the daily
build-macos.yml build-extension leg) derives its tag as "latest+1 at its
own dispatch moment"; on 2026-10-24 the v1.0.635 tag was cut by
auto_release before those legs ran, so the extension (and the DMGs)
landed on an orphaned v1.0.636 draft while the tag train published
v1.0.635 — every leg green, nothing asserted the published asset set.
Now scripts/release_required_assets.txt is the single source of truth,
scripts/assert_release_assets.sh gates the release-provenance publish
flip (a missing asset leaves the release a draft and fails the run
loudly) and a new release-asset-smoke ci.yml job post-publish probes
every landing-page release link + the full required set (red on non-200).
Remediation: the exact-tag fa-extension.zip (built from the v1.0.635
tree by run 40064876732) was uploaded to release v1.0.635 — the landing
link serves 200 again.
fix(gh-1820): the arm64 DMG leg's stale size floor is reseeded onto the
shipped v1.0.634 shape — run 4004758753 (post-#1815 SPM-lane adoption)
passed the min-platform gate and the full build, then failed the
size_manifest.py +5% budget on App.framework/App +5.1%
(25,020,112 → 26,284,144, the accumulated Dart AOT of #1793/#1794/
#1796/#1805/#1806); the committed scripts/size_budget/fa-macos-arm64.tsv
was still the pre-#1096-diet floor (TOTAL 210.5 MB vs the artifact's
116.4 MB), so the documented intentional-growth remediation applies:
the floor is now the exact --reseed output for the failing run's
artifact manifest (26 rows ≥ 64 KiB + TOTAL 116,386,369, from the run
log's printed table — the zip never uploaded past the gate), locking in
Fa 23.4→18.6 MB, quickjs_bridge 950→844 KB; new
test/ci_1820_macos_arm64_size_floor_test.dart gates the shape green
through the REAL size_manifest.py check and pins the committed floor
byte-identical to a --reseed round-trip.
fix(gh-1814): the desktop DMG lane ADOPTS the SwiftPM lane — Flutter
3.47.7 removed the FLUTTER_SWIFT_PACKAGE_MANAGER disable knob
(flutter/flutter#193978), so the gh-180 pure-CocoaPods pin became a no-op
and the min-platform gate (which forbade any SwiftPM state) failed both
DMG legs on every pinned-3.47.7 run (39991663029, 40022900373); the
#1813 escape (pin 3.47.4) is unreachable on the ~60kB/s runner link.
scripts/assert_macos_min_platform.sh now ASSERTS the SPM lane instead:
SwiftPM state expected after flutter build macos --config-only,
generated package floor ≥ the app floor (the silent bump must not skip —
the original gh-180 killer), no SwiftPM product above the app minimum,
vendored inappwebview manifest floored at 14.0 if it ever reappears.
Floors were audited and need no raising: Runner pbxproj 3×14.0, Podfile
:osx, '14.0', generated floor 14.0 on 3.47.7, vendored podspec 14.0
(its Package.swift stays deleted — the plugin rides the CocoaPods half of
the hybrid lane via Podfile.lock). The dead env pin left the build-macos
job env; the SDK pin reverts #1813 back to 3.47.7. The ci.yml macOS debug
leg gets the same treatment (gh-1814 review): its
FLUTTER_SWIFT_PACKAGE_MANAGER=0 pin — a silent no-op on the floating
3.47.x stable it resolves past the knob's removal — is removed and the
lane comment corrected; the leg rides SPM enabled as it already has been
(run 399572012 green).
fix(gh-1805): modifier chords no longer type their base letter into the
composer while the slash menu is open — the vendored dart_tui decoder
puts the base letter into the keystroke's text for control bytes
(Ctrl+S → ctrl+s with text s), and the slash-menu edit cluster was
the one insert surface missing the _isCommandKeystroke guard, so a
chord pressed mid-menu poisoned the typed prefix (/memory landed as
/smemory). Plain letters and Shift+letters still type and filter.
feat(1515): ONE model-limits settings surface — the settings hub's
standalone "Model parameters" row is retired and bare /model-edit
(TUI) aliases into /settings → Model capabilities, pre-focused on the
active model. The caps flow gains the model-edit preset ladders +
Custom… free input for context window / max output (a preset pick
writes models.overrides AND rebuilds the active model on the spot —
the old flow's live-mutation semantics, now durable), every field row
shows the current EFFECTIVE value with its source layer (pinned /
role override / catalog), the thinking row is an omp-style selector
(only levels the model supports, per-rung budget metadata, current
level preselected), and the hub row describes the effective summary
(200K ctx · 16K out · thinking off) instead of a pin count. The
scripted /model-edit <field> <n> form and the line-mode limits print
are unchanged.
feat(gh-1746): opt-in session disk retention — sessions.retention in
~/.fah/config.yaml caps the sessions root and sweeps whole session
units (transcript + sidecars) oldest-last-used-first at boot, exempting
live, named, young, and the current session; nothing is EVER
auto-deleted without the section (byte-identical legacy boot). Adds
fa session clean [--dry-run], the /sessions + fa session list
usage footer under a cap, and header-cwd/path-escape guards so a
planted session header can never steer a delete outside any root.
fix(gh-1763): a bash tool command running broad process kills can no
longer signal its own shell or the harness. Scripts are delivered from a
per-exec private temp file (sh -c '. "$1"' sh <script>) so the wrapper
argv never carries the script body (pkill -f <pattern> cannot
self-match its own sh -c; $0/$#/exit codes/set -e/stdin stay
identical), a PATH shim for pkill/killall plus a kill() prelude
guard exempt the harness pid chain + process group (exported as
FA_PROTECTED_PIDS) and the invoking shell from the child's own kills
while legitimate victims (stuck test runners, background servers) are
still terminated, and headless SIGTERM aborts now log
fa: SIGTERM received … so a child-issued kill never masquerades as a
user Ctrl+C (exit code contract unchanged: 130 graceful, 143 escalate).
fix(gh-1778): the release gate no longer reds a fully-green shard when
GitHub's cache service degrades. v1.0.621's Core tests (shard 2/4) was
killed by its own timeout-minutes: 12 with 2535/2761 tests green — both
cache steps had stalled ~133 s on keys that existed (degraded service,
not eviction), cold pub get + native-asset hook builds delayed the
first test to ~5 min, and the shard needed ~14 min against a budget
sized for the warm ~8.8 min wall. The job budget is now 20 min (sibling
test-leg band), and both test-core cache steps are best-effort
(timeout-minutes: 2 + continue-on-error: true) — a failed restore is
slower setup, never a red shard. test/ci_core_shard_timeout_guard_test.dart
pins the budget and the best-effort cache contract (RED-first on the old
YAML).
feat(#1761): the 32KB project-context budget never cuts silently again.
A file truncated to its head excerpt now carries an extended
<!-- From: path — N bytes, only the first M shown --> marker plus an
explicit MUST-read-the-full-file instruction (offset/limit/ranges,
prompt copy in prompts/context/project_context_read_full.md); a file
dropped entirely leaves a one-line tombstone naming its path instead of
vanishing; and the user gets a dim boot line per truncated/dropped file
(CLI line mode prints directly, TUI re-prints after the banner, the app
logs through its AppLog boot-notice seam). New projectContext: config
section: maxBytes (default 32768) — 0 disables the section entirely
and reads no files (fast boot); unknown keys are tolerated with a note.
Leaf-first allocation is unchanged.
fix(gh-1757): the Android build no longer re-enters the sdkmanager
download roulette for CMake. v1.0.619's build-android leg died at
:jni:configureCMakeDebug[arm64-v8a] — the :jni module (transitive
jni package via path_provider_android) wires externalNativeBuild.cmake
with no version, so AGP auto-installed its default cmake;3.22.1
mid-configuration and one corrupt archive ("Archive is not a ZIP
archive") red'd the leg, the aggregate Quality gate, and the
Advance-vpinned-marker job. The gh-1655 pre-install action now also
pre-installs cmake;3.22.1 with bounded retry + idempotent skip (same
hardening the NDK path already carries); the guard test pins the new
AC6 assertions and the AGP-version coupling (AC7).
feat(#1750): per-model image caps — the outbound request never carries
more unique images than the target model accepts. Model gains an
optional maxImages (catalog-seeded: gemini-4-argon: 10, measured);
images.perModel.<modelId> config overrides it (strict images:
section, user truth wins); the registry resolves
perModel > catalog > images.maxPerRequest (first present wins, the
issue-#1554 shape). A Google-shaped too many images … maximum allowed for model <id> is <n> 400 is classified (image_overflow.dart, the
overflow.ts twin), latches the learned cap for the (provider, model)
pair for the run, and re-sends ONCE through the registry at the
smaller cap — current-first-then-newest drops with the existing
never-silent notices naming the learned cap; a second refusal
surfaces exactly as today. JSONL is never rewritten;
images.registry: false keeps the byte-identical legacy shape
including the recovery-off behavior.
feat(gh-1760): oversized user-message intake spills to a file instead of
blowing the context window. Any incoming user message — headless
-p/positional, TUI first/mid-session, steering/follow-up drain,
external inbox — whose text passes the new relative threshold (a
spills.userFraction of the live window budget minus the compaction
reserve, floored by spills.userThreshold bytes) is redacted and
written verbatim to .fah/spills/<sessionId>/user-<n>.md (the #678
layout) from one choke point (Agent.userMessageGuard); the model
receives a bounded head+tail preview plus a read-in-parts instruction
riding the replacement USER message (copy in
prompts/spill/user_read_instruction.md). Images stay inline, a
refused spill write degrades to today's inline behavior, and
spills.enabled: false / userThreshold: 0 / the new
FAH_USER_SPILL_THRESHOLD env knob (bytes) keep the legacy shapes.
fix(gh-1743): a structured compaction that hits its total-budget
watchdog (the factory's 300s totalBudget) no longer aborts the
live session. AutoCompactorFactory.run() now applies the
mechanical local-trim fallback (keepRecentTokens, pairing-repaired,
[context trimmed locally] marker — the same valve the classic
pass ladder ends on) instead of failing fast; the honest
both-roles-failed verdict is kept only when nothing is droppable.
fix(gh-1636): auto-compaction stops dumping the internal checkpoint
summary into the console — the default report is ONE marker line
(● auto-compacted · engine: N → M tokens (F freed · P%)); the records
line + fenced summary moved behind the new tui.showCompactionDetails
opt-in (default false), and the manual /compact still prints the full
block. The headless log face no longer streams the summarizer deltas by
default (the gh-1433 E4 live stream is part of the same opt-in; the
gh-1430 … reasoning liveness line still covers the window). Nothing
is lost — the summary stays in the session file, reachable via
compact_expand.
fix(test): gh-1732 — the wrap_ansi_line_perf_test sustained-contention
validation leg spikes 3 of the 5 doubled runs per attempt instead of 4
(CI run 39546301132, core shard 2/4: the old shape left exactly ONE
real-wall-clock-timed doubled sample per attempt, and real jitter on
that single sample — the very shape being modeled — inflated all three
attempts' min ratios to a 3.30x best against the 3.2 bound while the
real 500k growth legs on the same runner stayed green). 3-of-5 is the
documented live-red shape ("≥3 of the 5 timed runs"): median-of-5
[c,c,s] is still a spiked value, so the leg still proves the
pre-gh-1532 median harness fails, and the min harness now rides two
clean doubled samples per attempt.
fix(#1705): CLI fresh-install boarding asks the provider choice FIRST —
the same list the /add-provider picker shows (presets, Custom last) —
and only Custom continues into the api-type wizard; a preset opens its
own flow with the api type prefilled. Previously boarding opened the
custom wizard directly, cold-asking "api type" before any provider was
chosen.
feat(providers): gh-1704 — first-class Codex/ChatGPT credential export,
CI refresh action, and Provider Queue integration (Copilot-export
parity). fa provider export <name> [--check] resolves the saved
ChatGPT login's OAuth blob from its secure-store slot, refreshes it when
stale, and prints the shell lines (OPENAI_API_KEY + OPENAI_BASE_URL
= the Codex backend — these tokens never serve api.openai.com); the
refresh token never prints and --check reports credential health with
the SSO caveat. /provider chatgpt import [path] adopts an existing
codex login session (~/.codex/auth.json) into the same slot — no
browser round-trip. .github/actions/chatgpt-codex-auth mints a fresh
access token from a CHATGPT_REFRESH_TOKEN secret in CI (direct curl,
no third-party action); invalid_grant fails fast and loudly.
providersQueue: entries speak chatgpt-codex: an omitted apiKeyEnv
resolves the Keychain slot automatically (in-memory refresh), a new
refreshTokenEnv: field mints the initial access token on slot
activation, and a raw Bearer token in apiKeyEnv is tolerated by
chatgpt_codex.dart (no more FormatException — proactive/reactive
refresh degrade, a 401 surfaces as the auth death the queue fails over
on).
feat(bench): gh-1681 — leaderboard submission readiness. ATIF
trajectories: new exportAtifJson whole-snapshot exporter
(lib/src/trajectory/atif_export.dart, ATIF-v1.7 — the schema harbor
0.23.0's flagship adapters emit) behind a headless fa trajectory export [--format atif] [--output path] verb; the harbor adapter runs
it as a bounded best-effort exec after the agent phase on every
terminal path, so <trial>/agent/trajectory.json reaches the hub with
the other agent logs (the leaderboard static analysis requires one per
rewarded trial) and result.json metadata carries an
atif_trajectory present/missing marker. The golden fixture
(bench/harbor_fa/fixtures/atif_sample.json) is pinned Dart-side and
cross-validated against harbor's pydantic Trajectory +
validate_trajectory in test_atif_contract.py. Workflow: a
submission: true dispatch input turns the submission rules into hard
setup preflights (full task coverage, ≥5 attempts/task, no
agent-timeout override, pinned dataset) and makes the new per-shard
ATIF presence gate fail hard; default runs gain only a warn-only scan
(artifact flow unchanged). Runbook: docs/leaderboard-submission.md
(pinned dataset + @latest content-hash recording, quota-window runner
policy, lb CLI steps).
fix(headless): gh-1589 — the headless session becomes a SEALED artifact:
no session-JSONL writes land after the run's agent phase ends. New
Session.seal() (state in the writer, no file-format change): raised in
runHeadless' finally and the REPL teardown as the LAST session action
(close order: drain → usage fold → seal → exit, one implementation both
paths — REG-2), it makes every writer (_persistMessages,
shell_job_registry snapshots, compaction, any appendCustomEntry,
leaf moves) drop late appends with exactly ONE stderr warning per
writer class — never a throw into an unhandled zone; the exit code
stays the turn's terminal outcome (issue #413 semantics untouched). The
settle→reaction path is sealed with the file (I2): the drain wrap-up
raises a fence (_reactionRunsSealed) so a background-shell settle or
task completion arriving after the final answer's drain — including the
registry's one-turn-deferred listener racing _finishHeadlessDrain's
check-to-return window (the 65gGW7b 10:13:42–10:14:00 burst) — does
board bookkeeping only: no _startRun, no model request, no record.
An honest drain's records (final answer → seal) stay legal (E7); the
seal is per-run state, so resume/replay/session_repair are unchanged.
Adapter side (bench/harbor_fa, I3/AC4+AC7): run()/perform_task
returns only after the fa process exited on EVERY path — the ladder
kill became terminate-and-reap (SIGTERM → bounded exec await → SIGKILL
escalation → pgrep confirm), and the external-cancel paths (harbor's
own wait_for firing first, the no-knobs exec, external-cap-only mode)
now fire the kill + bounded reap before propagating, with a loud
agent-phase exit guarantee degraded stderr note when fa survives both
signals (never a silent proceed, never an unbounded hang).
fix(bench): gh-1643 — a trial killed mid-stream no longer reports 0/None
usage. The fold read only FINALIZED assistant messages, so a run killed
before its first assistant message (the costliest rows) folded to silent
zeros. Both folds now price a dangling model_request_summary at its
KNOWN request-side volume (system prompt + outbound message chars at the
shared chars/4 rate, estimated input; output unknown) and mark the usage
incomplete instead: UsageChainScan.usageIncomplete → usage.json
(usageIncomplete: true, written only when set so completed folds stay
byte-identical), SessionUsage.usage_incomplete → the tb adapter's
usage-guard.json row and harbor's agent_result.metadata, and the
post-mortem stamps usage_incomplete on every results.json row it
patches from a partial fold. Harbor also exports the run outcome class
(agent_result.metadata.failure_mode: agent_timeout self-kill /
external_kill / completed / crashed / never_started) so downstream
classification stops relying on exception.txt archaeology; a summary
superseded by another summary is flushed instead of silently dropped on
both sides. Never-started trials keep the gh-1638/gh-1576 audit-only
precedence (no fake usage).
fix(stuck-calls): gh-1632 — a stuck tool call in an interactive session now
escalates like everywhere else. effectiveStuckTool() defaulted every
non-headless host to StuckFollowUpMode.advisory when no explicit
agent.stuckTool: was configured, so the fa 1.0.580 incident session (a
dead git fetch in the REPL) fired its tool_stuck flag at the 300 s
threshold and then hung for 27 more minutes — advisory mode never cancels,
retries, or converts; only the owner's manual steering broke the deadlock.
The autonomous default (cancel → retry once → background-convert, then a
session-visible escalation) now applies in EVERY mode; advisory remains
an explicit opt-out (agent.stuckTool.followUp: advisory), never an
implicit default.
fix(bindings): gh-1624 — the gh-1570 binding preflight no longer kills
translating-proxy factories. _validateModelEndpoint's vendor-prefix
verdict was fatal on ANY endpoint, so FA_PROVIDER_TYPE=anthropic +
bifrost's /anthropic path + zai/glm-5.3 exit-64'd every boot on
1.0.576 (the prefix is a routing key on a translating proxy, not a
dialect declaration). The vendor-prefix check now fires ONLY on the
endpoints that carry a dialect verdict — the three family-exclusive
canonical vendor endpoints (api.anthropic.com, api.openai.com,
generativelanguage), mirroring the family verdict's scope: those keep
the gh-1570 fail-fast, ROUTERS (OpenRouter/Copilot/DIAL/kimi — the
prefix is a routing key there too; OpenRouter's canonical ids are
anthropic/claude-*) and unknown/private endpoints carry the mismatch
as a ProviderModelBinding.warnings entry, never exit 64. The warning
is rendered on EVERY resolve path (gh-1624 review): the FA_PROVIDER_*
boot notice, the provider queue (per entry at boot, on queue reload,
and in /provider queue test), and the --model/--base-url flag
revalidation — a rider never boots silently.
Breaking-change note for 1.0.576 upgraders behind proxy
factories: the migration is the wire: FA_PROVIDER_CONFIG key —
documented now (README env-preconfig section,
docs/agents/providers.md §Translating proxies, --help) with its
caveat: the override engages only when the declared wire DIFFERS from
the provider record's own kind (record anthropic + wire: anthropic
still runs the checks). REG matrix (AC1/AC2/AC3) pinned in
test/model_roles/provider_model_binding_test.dart; AC4 IT drives a
bifrost-shaped loopback proxy through the production anthropic adapter
(test/model_roles/bifrost_proxy_binding_test.dart).
gh-1596: the first-run wizard (site/setup.sh) no longer crashes or
corrupts input under curl | sh on dash (Ubuntu sh) or busybox ash
(Alpine): scripts/gen_installers.dart now renders the provider/mode/
approval menus from the live provider catalog into plain POSIX sh — the
bashisms that killed dash (printf -v, read -p) and silently corrupted
busybox ash are gone; the hidden API-key prompt uses stty -echo + a
by-name read (dash has no read -s) and the key never rides printf
argv; EOF/Ctrl-C mid-wizard aborts non-zero with no half-written
~/.fah, and non-TTY runs print guidance and exit 1. The wizard's temp
config is mktemp-created (no predictable, symlink-followable fixed
path). A new path-triggered installer-portability CI leg proves the
script under dash + busybox ash (ubuntu) and bash (macOS) on every
site/** or generator change.
fix(cli): gh-1572 — fa boot no longer rewrites ~/.fah/config.yaml
from a lossy config. persistConfig (bin/fah_runapp.dart) built a
fresh CliConfig carrying only some fields and persistConfig() ran
unconditionally at boot, so every launch silently dropped the user's
spills:, subagents:, jobs:, headless:, agent:, quota:,
tui:, trajectory:, waiting:, power.hold and auto_update
settings (first reported on 1.0.547). Three cooperating fixes: the
boot save is now saved.copyWith(<live overlays>) so every parsed
section rides along by construction; CliConfig.copyWith replaces the
drifted hand-rolled withCustomProviders field list (itself dropping
spills/subagents/waiting/jobs/agentMode/misuseBreaker/streamThinking
inside saveCliConfig's merge step); the _diskPreservedSections
disk backstop covers every known top-level key. spills: renders in
toYaml (new SpillsConfig.toYaml) and agent.misuseBreaker: false
persists too; an unchanged save no longer rewrites the file at all.
Rework round: the boot snapshot re-syncs from disk after every
sanctioned mid-session settings write (onUserConfigFileChanged — a
cleared context cap, an auto_update key drop or a global compaction
pick is no longer resurrected/overwritten by the next save in the
same session); validateAgentSection (the strict fa config check /
config set / settings-hub mirror) accepts the emitter's
agent.misuseBreaker and agent.stuckTool keys so fa config check
never false-fails on fa's own output; clearing the context cap drops
only the cap key instead of the whole agent: block (the sibling
mode:/misuseBreaker: keys survive); and the providersQueue:
block (no typed renderer) joins the disk backstop so a save can never
silently delete the user's queue. Rework round 2 (merge + guards):
validateAgentSection's per-key switch became the _agentKeyValidators
table (the two gh-1572 keys pushed the switch form to complexity 13 —
over the repo's CRAP ratchet; the table keeps the mirror complete at
complexity 4), and the last two direct USER-config writers —
_writeTtsrSection and _writeMcpSection — fire the re-sync hook, so
a mid-session ttsr rule toggle or mcp server edit survives the next
persist too (the writer roster in settings_flow_config_sync.dart is
now exhaustive).
gh-1606: the copilot provider accepts fine-grained PATs
(github_pat_…) directly — PAT mode skips the copilot_internal/v2/token
exchange entirely (GitHub 404s every PAT exchange) and sends the raw PAT
as the Bearer against the configured Copilot base URL with the official
Copilot CLI's first-party identity headers
(copilot-integration-id: copilot-developer-cli,
user-agent: copilot/<fa-version> (<os>) term/unknown, installed by the
host at startup). OAuth device-flow tokens keep the exchange +
proactive-refresh flow unchanged; a PAT 401/403 surfaces the error
instead of a doomed refresh retry. Headless COPILOT_GITHUB_TOKEN now
works with a PAT. Connect flow: the paste hint explains PAT mode, the
dead-PAT case surfaces at the model fetch, and /provider export copies
the saved Copilot account's full token to the clipboard (account picker
on several entries; the token never prints to the transcript).
gh-1575: macOS clipboard paste fixed for TIFF-only pasteboards + the
Ctrl+V text fallback. _readMacos probes «class TIFF» after PNGf/JPEG
miss and converts via sips -s format png (a Preview/browser copy used
to dead-end on clipboard holds no image; a failed conversion keeps the
unavailable fall-through and every scratch file is cleaned up); new
readPasteboardText (pbpaste / xclip+wl-paste / PowerShell
Get-Clipboard) behind the same injectable ProcessRunner/platform:
seam, and Ctrl+V with a TEXT clipboard now inserts the clipboard text
into the composer instead of printing the image probe's dead-end reason
— bracketed-paste routing, gesture pairing and silent-failure semantics
unchanged. Review pass: the Windows Get-Clipboard read pins UTF-8
console output and normalizes embedded \r\n (OEM-code-page mojibake),
the Linux text backends pin text/plain;charset=utf-8 (an image-only
clipboard can no longer leak bytes as "text"), the library doc no longer
overclaims the FA_FAKE_PASTEBOARD seam, and FaTuiCallbacks moved to
a fa_tui_callbacks.dart part file keeping fa_tui.dart under the
2800-line gate.
gh-1555 (W1, lazy/windowed-everywhere reads): the legacy full-open call
sites migrate to windowed opens, so a session ledger past the gh-1073
2 GiB full-open backstop stays openable on every user-facing read path.
The TUI session picker, /resume, and the boot name-resolution scan
open windowed and resolve the display name through paged reads; the
subagent transcript preview opens the child windowed (tail only); the
_loadSession marathon fallback no longer degenerates to a refused
full open when the boundary walk exhausts its page cap (it proceeds
with the tail-anchored windowed session; the boot cap owns over-window
residue); and the headless fa trajectory resolver opens windowed,
reading the whole ledger for the cost/inspect views through the new
SessionTree.ensureWholeLedgerResident — a sequential forward scan in
bounded strips (WindowedSessionStorage.loadAllOlder), not the
boundary-walk's doubling strips (whose per-strip record cap measured
~13x read/parse amplification over the whole file). The same change
block-prepends paged chunks in _indexChunk (one insertAll per
list), removing an O(n²) the suspended-eviction walk paid per record.
Trajectory export streaming (AC3a), split-and-retain rotation
activation (W2), and the opt-in write-side caps (W3) remain follow-up
work.
gh-1605: the FinalizeGate gains a "grader parity" named check. A
produced entry point (script/CLI the verifier will run itself) is only
verified when it survives the grader's own invocation: the checklist
must re-run each produced entry point from a different working
directory in a fresh subprocess with a clean env, capture stdout/
stderr/exit code, and diff against the earlier in-session run — any
divergence (different output, relative-path artifacts, locale/
encoding-dependent output) is a fail ledger item, never a pass.
Closes the modernize-scientific-stack__nwbgy7a near-miss (every
in-session checklist item passed, the verifier's clean-cwd subprocess
collapsed Station 102 to Station 101's mean, reward 0.0). REG coverage:
the parity procedure catches cwd- and env-dependent entry points and
does not false-positive cwd-independent ones
(test/agent/finalize_gate_grader_parity_test.dart), and an IT replays
the evidence shape end-to-end — divergence detected and persisted as a
fail ledger item before finalize.
gh-1603: unattended runs can no longer end as a clean success on empty
model responses. Loop: from the second empty-hiccup retry on, the request
carries a continuation nudge user message ("your last reply was
empty/truncated — continue the task"; exact-text constant
emptyContinuationNudge, emitted with MessageStart/End so the transcript
stays truthful) — a cap-truncated reasoning model re-streamed verbatim
repeats the empty outcome, the nudge breaks the symmetry; the classifier
treats that exact message as transparent, so the turn's partial-content
window (gh-1577) survives a nudged retry. Headless: the empty-reply nudge
budget rises to AgentCliConfig.headlessEmptyReplyNudges (default 2 — the
session ends only after 3 consecutive unrecoverable empties; interactive
keeps the #1085 M2b one-nudge semantics), and a run that still settles
empty prints the explicit __FA_INCOMPLETE marker and exits 1 — a
distinguishable failure for the bench adapter, never a silent exit 0.
gh-1577: an empty assistant completion after partial content (assistant
text, tool activity, or a truncation signal earlier in the turn) is a
bounded-retry provider hiccup — up to emptyHiccupRetries (default 3)
re-streams with exponential backoff (emptyHiccupBaseDelay,
emptyHiccupMaxBackoff, both threaded through the Agent facade,
AgentWiringSpec, and the CLI config), and exhaustion ends the run with
a loud StopReason.error (headless exit 1, interactive error + prompt)
instead of the silent exit-0 settle that killed the gh-1577 trials.
Clean empties keep the #1085 M2b one-nudge semantics. The abort always
wins: a cancel racing the final retry settles StopReason.aborted for
any empty shape (the exit-contract race is pinned by tests), and the
backoff sleep is cancellation-aware.
gh-1574: SQLite reads work on installed binaries — bundle libs land where
the binary looks. The AOT binary's native-assets resolver reads
<exe_dir>/../lib/<name> (the dart build cli bundle layout), but every
POSIX installer copied the bundle libs beside the binary
(<install_dir>/lib), so read path=x.db died with a raw FFI dump
(Couldn't resolve native function 'sqlite3_initialize' … Failed to load dynamic library '~/.local/lib/libsqlite3.dylib'); it only worked on machines
with a manual symlink in ~/.local/lib. Fixed the layout in all three
install paths — site/install.sh (incl. the quarantine-xattr pass and the
source-build advice), install_local.sh, and fa update
(bin/self_manage.dart copyBundleLibs, now public and layout-pinned by
tests; also aligns Windows fa update with install.ps1, which was
already correct) — plus clean degradation in the FFI engine:
Sqlite3Engine.openReadOnly classifies the dart:ffi resolver's
library-load failure (isSqliteDynamicLibraryLoadFailure) and throws a
clean SQLite is unavailable — reinstall or run fa update note instead of
the raw resolver dump; a Dart-level system-library fallback is impossible
with package:sqlite3 native assets (build-time asset mapping, no loader
override). Installs already broken by the old layout self-heal: on the
bundled library's first failed load the binary moves the stale
<exe_dir>/libsqlite3.* into <exe_dir>/../lib/ and retries the load
(repairStaleSqliteBundleLayout) — so one fa update suffices (the
successor binary repairs on its first SQLite read; a pre-fix binary cannot
repair itself), and the stale <bin>/lib leftovers disappear as they
migrate. The POSIX installers also report a failed lib copy instead of
swallowing it (site/install.sh warns naming the lib dir; same in
install_local.sh), and site/install.sh.template carries the same
layout so regenerating the installers cannot resurrect the old one.
gh-1556: bench pipeline stops bleeding score + fa deadline awareness.
Bench: the harbor ladder kill uses the [f]a bracket trick and is
wrapped so a kill failure can never replace the TimeoutError (the
pre-fix self-matching pkill classified TB 2.0 ladder kills as agent
crashes — a SOLVED task scored 0); progress-watch (stall-gap 360s per
the gh-1430 watchdog invariant, flat abs ceiling 3600s) is the
bench-harbor.yml default on every shard, with the harbor adapter
picking the gap-aware ProgressWatch for parity with the
terminal-bench adapter; the aggregate counts completeness by
result.json presence (errored trials report as their own breakdown
line; exit 1 only on genuinely lost trials) and renders the full
report — table, errored breakdown, ledger row — to
$GITHUB_STEP_SUMMARY (TB1 parity). Both bench adapters export
FA_DEADLINE_EPOCH_MS (effective cap − 120s wrap-up margin) into the
agent session and audit the exact value. fa core: deadline awareness —
FA_DEADLINE_EPOCH_MS / --timer/--deadline / /timer write one
session clock, the per-turn context carries a time-left: line while
armed, and one-shot system notices steer wrap-up at ~50%, T-3min
(drain background jobs), and T-30s; absent deadline = byte-for-byte
today's behavior.
gh-1522: tag-sourced versions — the git tag is the single source of
truth. The release-bot file bumps are retired: cutting a release = a
tag + GitHub Release (zero commits to main, zero file mutations).
Both pubspecs carry the fixed 0.0.0-dev placeholder (CI guards
hand-bumps), the publish job stamps the staged pubspec + generated
changelog section from the tag (scripts/stamp_staged_release.sh,
staged version == tag guarded pre-upload), store builds stamp
--build-name/--build-number from the derived tag version
(pubspec.lock stops churning across releases), and
install_local.sh stamps version.txt from git describe.
feat(usage): gh-1460 — the fa-tokens: segment-close line carries the
session's model id, so every token row the factory publishes is
attributable to a model and therefore priceable. The fold now captures
each segment's LAST-SEEN model id (the model that served the final
request, the same string settings/UI display — mid-session fallback
take-overs attribute the segment to the model that actually finished it;
zero-filled fallbacks never clobber a seen model) and the pinned line
shape grows an OPTIONAL "model":"<id>" key after "segment" — omitted
when no model was observed, so legacy lines parse unchanged and the
pinned regex (usageTokensLogPattern, gh-1241 AC7) keeps matching both
shapes. usage.json's schema is untouched (the model is rebuildable from
the chain, I6); old consumers using tolerant JSON ignore the new key,
and the repo's single strict consumer (the pinned constant) is updated
in the same change.
fix(test): gh-1460 rework — the wrap_ansi_line_perf_test growth-shape
legs measure the doubling ratio with the MINIMUM over 9 runs instead of
the median of 5 (CI run 38112524699, core shard 0/4: a scheduling pause
inflated the doubled-size median — 12.09ms against a true ~7ms — and
tripped the 3.2 bound at 3.65x on a genuinely linear wrap). Runner
jitter only ever inflates a wall-clock sample, so the min converges on
the true compute time; genuine superlinearity still trips the ratio
(the gh-1496 quadratic candidates measure ~4x by min). The absolute-
budget legs keep the median — their gh-1357 headroom already prices in
jitter. A synthetic regression test pins the pause-proof property.
<!-- release-draft-owner: run/42588687548/job/binaries-linux-arm64 -->
feat(gh-2059): browser-extension compaction parity with CLI — over-window relief, /compact with visible receipts, and the shared under-window guarante
feat(gh-2059): browser-extension compaction parity with CLI — over-window
relief, /compact with visible receipts, and the shared under-window
guarantee ride the same CompactionHostWiring resolution as the CLI and
app (resolveExtensionCompactionWiring); receipts are durable across
panel attach/SW restart. Machine-review round 1: exact ladder pass
counts, busy-matrix routing (routeUserInput) for panel text and second
/compact during a pass, window settings from the ACTIVE stored source
with positive validation, string-encoded int merges.
fix(gh-2055): the daily-publish cli leg red-failed in pod install
(Build macOS x86_64, run 41778615815) because the vendored macOS pod's
XCFramework download block guarded on the downloaded-zip MARKER
(if [ ! -f wasm_run-v0.1.0.zip ]) instead of on the artifact the build
consumes. A fresh CI clone never has the marker, so pod install re-ran
the block on every build: it removed the WasmRun.xcframework the
build-macos.yml restore step had already placed (cache hit,
sha256-verified), re-downloaded the 92.7 MB asset with a bare
no---fail curl, and one transient reset (curl 56 at ~3 MB/92.7 MB)
left the Podfile nm -gU symbol check with no
macos-arm64_x86_64/libwasm_run_dart.a to read. The guard now keys on
WasmRun.xcframework/macos-arm64_x86_64/libwasm_run_dart.a — the exact
path the Podfile force_loads and nm-checks — so the block is a no-op
whenever the restore step (or a previous run) put the framework in
place, and the fallback download for a truly cold tree carries
--fail AND aborts the block at the curl step on failure
(|| { echo "wasm_run download failed" >&2; exit 1; } — a swallowed
status fell through to unzip/truncate residue and masked the root cause
behind the Podfile nm diagnostic; review round 1 on this PR), and the
podspec now assigns the backtick result and raises "wasm_run framework
fetch failed" unless $?.success? — Ruby ignores a bare backtick's
exit status, so pod install would otherwise have sailed past the failed
fetch to the downstream nm diagnostic (review round 2). Linted
behaviorally (stubbed curl, no network; POSIX-gated sandbox, plus a
ruby evaluation of the real fetch statement) by
test/wasmrun_macos_podspec_download_guard_test.dart.
fix(gh-2050): the daily-publish cli leg red-failed daily when its
macOS attach arrived after a sibling train had already PUBLISHED the
derived tag (the same-tag convergence race: mobile leg drafts + the tag
train publishes ~25 min after dispatch while the serialized self-hosted
mac builds need ~66 min — v1.0.685: draft 05:22:46, published 05:44:52,
mac leg attached 06:28:53). scripts/release_asset_gate.sh gains a
caller-declared pre-stage policy — RELEASE_GATE_PUBLISHED_SKIP=1 —
that turns the PUBLISHED+unattached-fa-* verdict into a loud whole-call
skip (::warning::, exit 0, nothing uploaded, no twin create, no flip;
all-or-nothing so latest-mac.yml can never land alone on a release
whose zips were skipped). The default stays the gh-2042 AC1 loud
refusal; only build-macos.yml's release-macos desktop attach opts in
(its DMGs/zips/manifest are clobber-refresher pre-stage assets, not
publish blockers), the CLI-tar/extension attaches keep the refusal (they
carry REQUIRED release assets). The stranded desktop feed self-heals:
the next daily leg pre-stages the next version's draft with fresh
assets.
fix(gh-2044): the auto-compact fragile-band escalation could fold the
ENTIRE visible projection when the remaining path fit inside
keepRecentTokens — the keep-recent walk defaulted to folding
everything when its budget never exhausted, swallowing the live tail
(the newest user ask and assistant turn included). Real-world trigger:
resumed subagent runs ended with no assistant message at all
(subagent produced no assistant message — CI cores shards 1/4 and
3/4, hostile-env, both red). _pickCheckpointRange now declines the
fold when the whole foldable path fits the keep-recent region and
still carries live records; all-marker paths keep the lossless
deterministic consolidation (issue #387). The escalation receipt also
reports the REAL summarizer call count (a #729 chunked fold makes N
calls, not 1), and the factory wires its run token into the engine's
cancelToken so a watchdog-shaped cancel reason (any
TimeoutException) can no longer dissolve a user abort into a
"summarizer returned no text" failed receipt while the run reports
success (#1085). Follow-up CI red on the same branch: the decline
moved tiny-transcript runs to the classic residual engine, where an
EMPTY summarizable region folded zero #729 chunks into a blank
summary that still landed as a junk CompactionRecord and a hollow
success: true through the no-work terminal — compactSession now
refuses that empty-region blank fold (gh-1743's failure-verdict
contract kept; a dialed blank answer over a real region still lands
as a record, issue #578 unchanged). Third CI red on the same branch
(flutter-app over-window relief, both shards): the whole-world decline
lived in the shared _pickCheckpointRange, so it also starved the
checkpoint LADDER — a tool-result-heavy transcript whose every ledger
token fits keepRecentTokens (the request estimate is far over the
trigger) lost its only summarizer-based relief, and the hosted app's
over-window guard ended the turn instead of continuing it. The decline
is now escalation-only: ladder passes run over the trigger, where the
fold IS the relief (the escalation still declines — semantic 5 — and
its pass body was extracted below the CRAP ratchet's 12.00).
(gh-2044 rework, UT-ESC-13)
fix(gh-2003): the Copilot picker filter is tier-aware. Individual-tier
/models responses omit model_picker_enabled entirely, so phase 1
dropped every such entry as «picker flag absent» and individual/OAuth
accounts still listed zero usable models. When the response comes from
the individual tier — the individual host (api.githubcopilot.com or
pi-mono's api.individual.githubcopilot.com spelling) or a
no-proxy-ep token (individual by design) — an ABSENT flag is now
selectable: absence ≠ an explicit false, which still drops
everywhere (embeddings, completion-only). Business/enterprise answers
keep the strict picker list unchanged, and the keep is never silent:
the fetch note names it («2 of 2 models kept via the individual-tier
fallback (picker flag absent)») through the existing fetch notes:
surfacing. pi-mono parity: the trigger mirrors allowPolicyFallback = baseUrl === "https://api.individual.githubcopilot.com" in
packages/ai/src/auth/oauth/github-copilot.ts.
fix(gh-1990): the Copilot /models dialect never answers a silent
empty list again. Every silent-empty exit names its branch through the
new fetchModelsForEndpointReport dispatch (ModelsFetchReport +
ModelsFetchDiagnostic): the token-exchange HTTP status (with the
OAuth device-flow / fine-grained PAT remedies for classic ghp_…
keys), the /models HTTP status or timeout (with the seat + per-tier
host hint), proxy-ep-less individual tokens aimed at a business/
enterprise configured host (the tier-misrouting hint), and the picker
filter with per-model drop reasons (picker flag absent, picker flag false, policy disabled, no tool_calls, endpoint unsupported) —
«12 models fetched, 12 filtered (picker flag absent)». A failed
primary credential path retries the other path ONCE before declaring
empty (dual-path fallback, gh-1606): an exchange-401 classic PAT tries
PAT-direct (the provenance note names which path answered), a
fine-grained PAT never falls into the exchange, an OAuth token never
becomes a direct PAT Bearer. Surfacing: /models prints a fetch notes: section per provider, the /provider copilot connect flow
prints the named failure before manual entry, the app's model
quick-select (EndpointModelsController.emptyReason) renders it under
the field, and the Copilot connect sheet shows the reason above the
manual model id. The legacy fetchModelsForEndpoint answers are
unchanged (parity); the filter criteria themselves are unchanged
(pi-mono parity — only the reporting is new). Review round 2: the
fetch notes: wiring covers the unsaved sibling paths too (the
active-catalog and env-keyed Copilot fetches surface the same named
failure instead of a silent seed fallback), the connect flow's
no models: line reuses ModelsFetchReport.reasonText (the last
hand-rolled message — hint join), and the ModelsFetchReport
docs scope the diagnostic contract to the dialects that name their
failure branches.
feat(network): channel payload sender block — app members' channel
messages carry {"text": …, "sender": {"name": …, "id": …}} (public
base64 JSON and keyed fanet1 plaintext alike), so DAP agents can
attribute the HUMAN author behind the fa_network relay's hub id. The
harness unwraps it tolerantly (raw text passes through), rides the
name as AgentMessage.fromName through the fabric, and the CLI
steering line greets people again: from Uladzimir (<relay id>) in #channel: hello instead of from 3cfaee1aa054fe6e…. The app falls
back to the payload sender for bubble names when the envelope carries
none; DMs also name hub peers from the whois cache. Legacy payloads
without the block decode exactly as before.
feat(gh-1956): reply auto-spill — oversized child reply / agent_message
payloads spill to an agent:// artifact automatically; the parent receives a
bounded excerpt + pointer + completeness marker with char AND byte counts.
Manual chunking degrades to a documented fallback, restorable across parent
resume via durable deliverable_artifact records, behind the
subagents.replyAutoSpill kill switch (default on; honored on the CLI and
the app host alike). Rework round: the salvage hook (output-truncated
deliverable calls) restores the durable ids before allocating its own
(no post-resume pointer shadowing), durability losses emit a
[deliverables] diagnostic line, the app host honors the kill switch, the
spill-reach gate is renamed spillAppliesTo (was spillApliesTo), byte
counts are encoded once per spill, and DeliverableRecord.artifact(bytes)
is optional again so external constructors keep compiling.
fix(#1947): schedule_message from the orchestrator no longer misroutes
into a finished subagent's mailbox. A background shell job started by a
CHILD registered its settle continuation inside the child's subagent
zone; when the job settled after the child's run ended, the CLI's
settle-wake (_onShellJobSettled → _startRun) started MAIN's reaction
run still carrying the finished child's activeSubagentId(), and every
identity consumer of that run misattributed it — the reported shape was
a self-reminder scheduled "in <session>/<child>" that auto-resumed the
completed child (duplicate-release risk). Three layers: the
ShellJobRegistry now binds its settle continuation to the
construction-time (boot) zone (_bootZone.run), both hosts'
scheduleSenderMailbox wiring trusts the zone scope only while the
child is genuinely in flight on the session executor
(TaskExecutor.isInFlight), and the tool response names the RECIPIENT
the record carries — an explicit to: used to be shadowed in the reply
by the sender's mailbox, which read as "the override was ignored" even
when it was honored. E2E repro drives the real AgentCli
(test/cli/agent_cli_schedule_recipient_test.dart) plus a registry-level
zone pin (test/tools/shell_job_settle_zone_test.dart).
fix(gh-1948): task_observe works on a RUNNING background subagent
instead of crashing with an uncaught SessionException on the
<parentSessionId>/<agentName> placeholder. Three legs: the executor
flushes each child's transcript incrementally at every turn boundary
(the JSONL exists and grows mid-run — also crash-resilience for
aborted children), task_observe resolves a placeholder handle's real
session file from the sessions tree (the same agent: subagent /
parent: / id: header scan the boot adoption uses) and heals the
handle onto the found path, and every read failure degrades to a
structured "transcript unavailable" / "no transcript available yet"
message instead of an uncaught tool error. The child-transcript reader
rides the transient-ENOENT retry (issue #427) — it reads while the
executor appends.
fix(#1941): the selected CodeMie project now rides EVERY model request.
The header is baked into Model.headers only by the connect-time
switches (gh-1809), so every other rebuild silently dropped
x-codemie-project — and the backend bills a header-less request against
the user's DEFAULT project (codemie-code header-injection.plugin.ts
contract): the reported symptom is limit-exhausted from the exhausted
default project despite picking a budgeted one. Rebuilds now re-resolve
the project from the saved entry through one helper
(projectHeaderForBaseUrl): the boot folder-state restore (JWT entries
previously booted header-less — the SSO cookie patch was the only boot
carrier), the session-resume re-apply (_buildRestoredModel), the
roles-chain failover (ModelRef.project, mirroring the #964 authHeader
leg), the roles-mode /model + settings-flow pins, and the generic
/provider switch. The models list (/llm_models, SSO + JWT + guided
setup + entry dispatch) and the quota adapter (dark until pinned) scope
themselves to the same project, so display and billing can never diverge.
Tests pin the header VALUE on the wire (real adapter over a fake
endpoint) for the restored and failed-over models.
feat(network): private networks created outside the app surface after
sign-in. GET /api/networks/mine (fa_network#16) is synced into the
wallet after OAuth/dev sign-in and both restore paths — best-effort,
silent on older backends (404) and any failure, and never destructive.
Synced memberships keep the server's name/class, preserve local
passwords/join facts, and land without a password (the endpoint never
returns one) until the first join. Entering such a network — sidebar,
⌘K/⌘3-9, or a restored selection — opens the join sheet prefilled with
the network id instead of a resume error or endless spinner; joining by
name now keys the wallet/session by the echoed canonical id and migrates
legacy name-keyed rows (password + channel keys) instead of duplicating
them. A malformed /mine row is skipped, not the whole list.
fix(gh-1931): one render per tool call on the non-TUI console. Line mode
and the headless -p log printed every executed call twice — a start row
(• bash · cmd, no elapsed) followed by the end row (✓/✗ bash · cmd 0s). The duplicate start-row print is gone from _onToolExecutionStart;
the end row (command, status glyph, elapsed, failure news) is the single
render per call. The TUI keeps its running card + settled card, start
bookkeeping stays (_toolStarts feeds the end row's detail/elapsed, the
waiting coordinator still arms per-call liveness), and persisted
artifacts (session JSONL, trajectory records) are untouched —
rendering-only change. Long foreground calls keep the gh-1055 liveness
reminders past waiting.toolLivenessSeconds (default 60s).
fix(gh-1929): esc during the pre-flight auto-compaction no longer crashes
the CLI. The compaction layer's deliberate abort (#1085) used to escape
_runPrompt outside its guarded body and reach the root zone handler via
an unlistened whenComplete chain — fa crashed: CancelledException +
crash.log + a lost session view. The pre-flight window (submit on an
over-window transcript) now translates the cancel into a user-cancel
outcome: neutral dim submission cancelled receipt (never error:), the
TUI composer hands the typed text back (sendInputText) so a resubmit is
one Enter away, headless exits 130 with the same receipt, and the next
prompt works (the funnel's local-trim recovery or a retried compaction).
Cancels DURING the run — in-flight relief, settle funnel — keep their
pinned loud-abort surfaces, and the post-run window keeps its
compaction interrupted receipt via the shared _renderUserCancel
helper.
feat(1923): stall-proof steering. Steering arriving while the loop awaits
an UNPRODUCTIVE provider submit (zero bytes, or silent past the steering
stall window) now cancels the attempt via a SteeringInterruptFire
cancel, injects the pending steering into the rebuilt request, and
resubmits — the #1916-class submit stalls can no longer queue user or
agent mail invisibly. Productive streams keep next-boundary semantics;
the interrupt never consumes a retry attempt, and the run latch re-arms
in place so watchdog replays and user aborts keep working. task_resume
caps its foreground wait (like task_send) and continues in the
background; video polling honors soft-yield checkpoints (remote render
never restarted, honest hand-back text) and media generation honors run
cancellation; SLO-breached undelivered child mail surfaces in
task_status (UNDELIVERED … past the 2s delivery SLO); the TUI prints
steering queued — provider stalled Ns, interrupting to deliver plus the
delivered ack.
feat(1916): image-heavy submits no longer livelock — byte-based
images.maxBytesPerRequest (default 8MB, layered onto the count cap;
/images names both drop reasons), legacy text-embedded data-URL
screenshots in history fold into the once-only registry instead of
replaying raw forever, stall watchdog windows scale with the outbound
payload (capped at 2x; explicit connect clamps untouched), abandoned
connect attempts answering during the retry backoff are adopted, and
repeated identical-payload stalls fire ONE actionable /compact +
image-budget advice. TUI stall banner + docs/images.md carry the
image-heavy playbook.
fix(gh-1912): Add an agent → DAP now turns a retired hub's unknown
enroll_create response into an actionable upgrade message instead of
exposing bad_frame. Independent enrollment never falls back to the
legacy same-name credential rotation endpoint, and existing credentials
remain untouched on failure.
fix(gh-1913): Add an agent is usable on narrow phone screens: scope and format choices wrap without per-character labels, enrollment controls stack at full width, and actions remain reachable by scrolling. Desktop keeps the segmented layout; phone/desktop light/dark goldens guard both.
fix(gh-1825): DAP agent enrollments carry a stable, independent
agentInstance through app/CLI bundles, persistence, direct boot and
reconnect. Independent same-name agents keep distinct identities and stay
online together, while legacy bundles retain their project-scoped identity.
feat(gh-1825): DAP private-channel sends now use self-describing
fanet1 frames (sender X25519 key in the frame, channel-bound AAD) via
fa_hub_client 0.2.11, so fa_network app members decrypt them without
a whois/senderKey round-trip. Inbound channel frames open fanet1
first and fall back to legacy dap/1 payloads. Verified live in both
directions between a headless DAP agent and the Flutter app.
feat(gh-1825): the network offline banner now has a Reconnect now
action which restarts the socket and reloads the channel/member
snapshot. Fresh sockets also clear a stale offline flag (the server
sends no historical drain to a socket that was away during recovery).
AgentNetworkController.start stops after disposal instead of
notifying a dead controller.
feat(gh-1825): the roster sheet shows access-scope badges (whole network vs N channels, decoded additively from Member.access in snapshots and presence.changed) and owner/admins can drop an agent (DELETE /api/networks/{id}/agents/{agentId} with confirmation; the server tombstone keeps it out of future snapshots). Guests and store members are removable too via DELETE /api/networks/{id}/members/{id} (sessions revoked server-side); owner/admin/self rows stay protected.
feat(gh-1825): channel/network management in the app — rename a channel (PATCH), delete a channel or an owned network (REST DELETE with confirmation and state/navigation cleanup), and unnamed CLI/hub-created channels show a short friendly label instead of the raw 64-char id.
feat(gh-1825): pasting a DAP invite no longer demands a passphrase on macOS — importing NEW channel secrets auto-provisions a random passphrase into the per-channel Keychain item (decryption of existing envelopes never auto-generates; env/prompt/headless behavior unchanged), and passphrase warnings now name the concrete recovery (re-paste the invite to re-encrypt, or set DAP_KEY_PASSPHRASE).
feat(gh-1825): the network online counter opens a live-first roster sheet (names, member class, presence), and the channel composer offers roster-filtered @-mention autocomplete — the draft keeps the literal @displayName for DAP agents and the envelope carries the resolved member ids. On the CLI side the DAP channel notice names the agent's own identity (display name + agentId), and inbound channel mail that @-mentions that name (case-insensitive, word-boundary — no email or @name2 false positives) is marked "[mentions you — you are @name]".
fix(gh-1825): the app's network roster no longer reads as permanently empty ("0 online"): fa_network sends roster.snapshot with an OBJECT payload ({members: [...]}) and the app only accepted a bare list, so every snapshot silently decoded to zero members. Both shapes are now accepted, with a production-shape regression test. The DAP paste hint names web alongside CLI/macOS/iOS (plain-web join is supported).
fix(gh-1825): resumed CLI sessions rejoin their DAP channels and receive channel mail again — session replay now re-arms membership synchronization (plugin session-ready listeners on boot/resume and in-process switches) and idle channel probes complete deferred joins before inbox peek/drain. The post-replay sync also dials the membership's OWN hub with its enrolled credential before joining: a resumed session-scoped agent used to stay on the zero-config local URL forever, and because a disconnected fabric probe short-circuits before the sync, the remote dial never happened — the session-ready listener now drives it directly (throttled to one attempt per 10 s). App channel history no longer flashes raw payloads on first open: decoding waits for startup's channel listing instead of caching an undecoded guess until a manual switch.
feat(gh-1825): the fa app consumes the same DAP enrollment block as the
CLI — paste it into the agent chat composer (macOS/iOS/Android; the
smart-paste .txt chip path included) and the channel join happens
without the block ever reaching sendText, analytics, or session
records. Native joins ride the live agent-fabric hub repository:
per-project scoped identity, invite pub/priv import, keyless public
channels, reconnect-persisted membership in a dedicated secure
enrollment store (Keychain/Keystore where available). Harness additions:
HubMessagingRepository.joinChannel/waitForConnection with
rejoin-before-drain on reconnect, keyed channel decrypt
(hubDecryptPayloadWithKeyPair), public showcase payloads, and
HubBearerTransport (native Authorization header or the browser
?token= equivalent; authed dials never stringify transport errors
into logs). Plain web joins through the browser
WebHubTransport (?token= bearer equivalent, IndexedDB identity);
the browser extension rides the service worker's existing DAP client
(hub.join op), never a panel-local second connection. Add agent now defaults to the
DAP tab with a paste-into-chat hint; web gains ai-native provider
sign-in through the existing popup/callback OAuth flow.
fix(fa_network#4): the app inserts hub agents arriving via
presence.changed into the roster (additive displayName/class fields)
instead of dropping them, so the online counter no longer sticks at 0
when an agent comes online after roster.snapshot; legacy frames
without those fields create no phantom members. roster.snapshot/REST
contents are logged for decisive presence debugging.
feat(gh-1825): /network + scoped DAP channel membership — join by
paste (invite url / url+keys / bare channel id / whole fa dap import
command) with session/project/global scopes, live join without
restart, per-scope leave + copy invite, and the composer-paste
shortcut into the same wizard. Session scope rides a
dap_channel_membership custom record (leave tombstones, branch
resume); project scope is .fah/dap.json (committable, dapkey1
PBKDF2+ChaCha20-Poly1305 encrypted keys); global keeps the legacy
~/.dap files. Headless CI: DAP_CHANNELS_JSON (+DAP_HUB_URL,
DAP_CLIENT_SECRET, DAP_KEY_PASSPHRASE) joins at boot with zero
prompts, plaintext or encrypted privs mixed. Passphrase chain:
env → per-channel macOS Keychain item → one prompt with save default.
feat(gh-1844): media slots never fall back to the main connection
implicitly — SDK/CLI, app store, and the JS bridge resolve a slot
without an override as UNCONFIGURED with an actionable
"configure the slot" error; the main connection serves a slot only via
the explicit per-slot useMainConnection: true opt-in (YAML
models.slots.<slot>.useMainConnection / app's "Same as main
connection" pick, preserved through media_models.json). Image and
video generation accept multimodal input (text + reference images:
sandbox paths, http(s) URLs, base64) on capability-declaring
OpenRouter endpoints (input_references + declared/probed
input_modalities gate; unsupported endpoints fail before any network
call); the OpenRouter video job API is a first-class video dialect
(submit → poll → mp4). New decide tool: structured decisions via
OpenRouter's Decisions API (/v1/systemone) — bool/choice/score
questions with calibrated probabilities, structured error results, no
streaming.
feat(gh-1834): the FinalizeGate contract counts a parity/conformance
claim as verified only with negative + combination self-checks — new
numbered step 2.6 in prompts/cli/finalize_gate.md (refused
incompatible/mutually exclusive options, empty/corrupt inputs, boundary
values, must-NOT-work cases, alongside happy paths; a blanket
"everything matches" is not verified). Universality: abstract example
only, no benchmark/task/flag names in the prompt; the step is
conditional, so tasks with no conformance claim see no gate change.
Ledger/verdict wire formats unchanged; validated end-to-end by the
scripted IT-4 replay (real CLI + real bash) whose persisted
task_ledger carries the rejection rows.
fix(gh-1832): release v1.0.635 published without fa-extension.zip — the
landing-page download 404'd. The extension's only producer (the daily
build-macos.yml build-extension leg) derives its tag as "latest+1 at its
own dispatch moment"; on 2026-10-24 the v1.0.635 tag was cut by
auto_release before those legs ran, so the extension (and the DMGs)
landed on an orphaned v1.0.636 draft while the tag train published
v1.0.635 — every leg green, nothing asserted the published asset set.
Now scripts/release_required_assets.txt is the single source of truth,
scripts/assert_release_assets.sh gates the release-provenance publish
flip (a missing asset leaves the release a draft and fails the run
loudly) and a new release-asset-smoke ci.yml job post-publish probes
every landing-page release link + the full required set (red on non-200).
Remediation: the exact-tag fa-extension.zip (built from the v1.0.635
tree by run 40064876732) was uploaded to release v1.0.635 — the landing
link serves 200 again.
fix(gh-1820): the arm64 DMG leg's stale size floor is reseeded onto the
shipped v1.0.634 shape — run 4004758753 (post-#1815 SPM-lane adoption)
passed the min-platform gate and the full build, then failed the
size_manifest.py +5% budget on App.framework/App +5.1%
(25,020,112 → 26,284,144, the accumulated Dart AOT of #1793/#1794/
#1796/#1805/#1806); the committed scripts/size_budget/fa-macos-arm64.tsv
was still the pre-#1096-diet floor (TOTAL 210.5 MB vs the artifact's
116.4 MB), so the documented intentional-growth remediation applies:
the floor is now the exact --reseed output for the failing run's
artifact manifest (26 rows ≥ 64 KiB + TOTAL 116,386,369, from the run
log's printed table — the zip never uploaded past the gate), locking in
Fa 23.4→18.6 MB, quickjs_bridge 950→844 KB; new
test/ci_1820_macos_arm64_size_floor_test.dart gates the shape green
through the REAL size_manifest.py check and pins the committed floor
byte-identical to a --reseed round-trip.
fix(gh-1814): the desktop DMG lane ADOPTS the SwiftPM lane — Flutter
3.47.7 removed the FLUTTER_SWIFT_PACKAGE_MANAGER disable knob
(flutter/flutter#193978), so the gh-180 pure-CocoaPods pin became a no-op
and the min-platform gate (which forbade any SwiftPM state) failed both
DMG legs on every pinned-3.47.7 run (39991663029, 40022900373); the
#1813 escape (pin 3.47.4) is unreachable on the ~60kB/s runner link.
scripts/assert_macos_min_platform.sh now ASSERTS the SPM lane instead:
SwiftPM state expected after flutter build macos --config-only,
generated package floor ≥ the app floor (the silent bump must not skip —
the original gh-180 killer), no SwiftPM product above the app minimum,
vendored inappwebview manifest floored at 14.0 if it ever reappears.
Floors were audited and need no raising: Runner pbxproj 3×14.0, Podfile
:osx, '14.0', generated floor 14.0 on 3.47.7, vendored podspec 14.0
(its Package.swift stays deleted — the plugin rides the CocoaPods half of
the hybrid lane via Podfile.lock). The dead env pin left the build-macos
job env; the SDK pin reverts #1813 back to 3.47.7. The ci.yml macOS debug
leg gets the same treatment (gh-1814 review): its
FLUTTER_SWIFT_PACKAGE_MANAGER=0 pin — a silent no-op on the floating
3.47.x stable it resolves past the knob's removal — is removed and the
lane comment corrected; the leg rides SPM enabled as it already has been
(run 399572012 green).
fix(gh-1805): modifier chords no longer type their base letter into the
composer while the slash menu is open — the vendored dart_tui decoder
puts the base letter into the keystroke's text for control bytes
(Ctrl+S → ctrl+s with text s), and the slash-menu edit cluster was
the one insert surface missing the _isCommandKeystroke guard, so a
chord pressed mid-menu poisoned the typed prefix (/memory landed as
/smemory). Plain letters and Shift+letters still type and filter.
feat(1515): ONE model-limits settings surface — the settings hub's
standalone "Model parameters" row is retired and bare /model-edit
(TUI) aliases into /settings → Model capabilities, pre-focused on the
active model. The caps flow gains the model-edit preset ladders +
Custom… free input for context window / max output (a preset pick
writes models.overrides AND rebuilds the active model on the spot —
the old flow's live-mutation semantics, now durable), every field row
shows the current EFFECTIVE value with its source layer (pinned /
role override / catalog), the thinking row is an omp-style selector
(only levels the model supports, per-rung budget metadata, current
level preselected), and the hub row describes the effective summary
(200K ctx · 16K out · thinking off) instead of a pin count. The
scripted /model-edit <field> <n> form and the line-mode limits print
are unchanged.
feat(gh-1746): opt-in session disk retention — sessions.retention in
~/.fah/config.yaml caps the sessions root and sweeps whole session
units (transcript + sidecars) oldest-last-used-first at boot, exempting
live, named, young, and the current session; nothing is EVER
auto-deleted without the section (byte-identical legacy boot). Adds
fa session clean [--dry-run], the /sessions + fa session list
usage footer under a cap, and header-cwd/path-escape guards so a
planted session header can never steer a delete outside any root.
fix(gh-1763): a bash tool command running broad process kills can no
longer signal its own shell or the harness. Scripts are delivered from a
per-exec private temp file (sh -c '. "$1"' sh <script>) so the wrapper
argv never carries the script body (pkill -f <pattern> cannot
self-match its own sh -c; $0/$#/exit codes/set -e/stdin stay
identical), a PATH shim for pkill/killall plus a kill() prelude
guard exempt the harness pid chain + process group (exported as
FA_PROTECTED_PIDS) and the invoking shell from the child's own kills
while legitimate victims (stuck test runners, background servers) are
still terminated, and headless SIGTERM aborts now log
fa: SIGTERM received … so a child-issued kill never masquerades as a
user Ctrl+C (exit code contract unchanged: 130 graceful, 143 escalate).
fix(gh-1778): the release gate no longer reds a fully-green shard when
GitHub's cache service degrades. v1.0.621's Core tests (shard 2/4) was
killed by its own timeout-minutes: 12 with 2535/2761 tests green — both
cache steps had stalled ~133 s on keys that existed (degraded service,
not eviction), cold pub get + native-asset hook builds delayed the
first test to ~5 min, and the shard needed ~14 min against a budget
sized for the warm ~8.8 min wall. The job budget is now 20 min (sibling
test-leg band), and both test-core cache steps are best-effort
(timeout-minutes: 2 + continue-on-error: true) — a failed restore is
slower setup, never a red shard. test/ci_core_shard_timeout_guard_test.dart
pins the budget and the best-effort cache contract (RED-first on the old
YAML).
feat(#1761): the 32KB project-context budget never cuts silently again.
A file truncated to its head excerpt now carries an extended
<!-- From: path — N bytes, only the first M shown --> marker plus an
explicit MUST-read-the-full-file instruction (offset/limit/ranges,
prompt copy in prompts/context/project_context_read_full.md); a file
dropped entirely leaves a one-line tombstone naming its path instead of
vanishing; and the user gets a dim boot line per truncated/dropped file
(CLI line mode prints directly, TUI re-prints after the banner, the app
logs through its AppLog boot-notice seam). New projectContext: config
section: maxBytes (default 32768) — 0 disables the section entirely
and reads no files (fast boot); unknown keys are tolerated with a note.
Leaf-first allocation is unchanged.
fix(gh-1757): the Android build no longer re-enters the sdkmanager
download roulette for CMake. v1.0.619's build-android leg died at
:jni:configureCMakeDebug[arm64-v8a] — the :jni module (transitive
jni package via path_provider_android) wires externalNativeBuild.cmake
with no version, so AGP auto-installed its default cmake;3.22.1
mid-configuration and one corrupt archive ("Archive is not a ZIP
archive") red'd the leg, the aggregate Quality gate, and the
Advance-vpinned-marker job. The gh-1655 pre-install action now also
pre-installs cmake;3.22.1 with bounded retry + idempotent skip (same
hardening the NDK path already carries); the guard test pins the new
AC6 assertions and the AGP-version coupling (AC7).
feat(#1750): per-model image caps — the outbound request never carries
more unique images than the target model accepts. Model gains an
optional maxImages (catalog-seeded: gemini-4-argon: 10, measured);
images.perModel.<modelId> config overrides it (strict images:
section, user truth wins); the registry resolves
perModel > catalog > images.maxPerRequest (first present wins, the
issue-#1554 shape). A Google-shaped too many images … maximum allowed for model <id> is <n> 400 is classified (image_overflow.dart, the
overflow.ts twin), latches the learned cap for the (provider, model)
pair for the run, and re-sends ONCE through the registry at the
smaller cap — current-first-then-newest drops with the existing
never-silent notices naming the learned cap; a second refusal
surfaces exactly as today. JSONL is never rewritten;
images.registry: false keeps the byte-identical legacy shape
including the recovery-off behavior.
feat(gh-1760): oversized user-message intake spills to a file instead of
blowing the context window. Any incoming user message — headless
-p/positional, TUI first/mid-session, steering/follow-up drain,
external inbox — whose text passes the new relative threshold (a
spills.userFraction of the live window budget minus the compaction
reserve, floored by spills.userThreshold bytes) is redacted and
written verbatim to .fah/spills/<sessionId>/user-<n>.md (the #678
layout) from one choke point (Agent.userMessageGuard); the model
receives a bounded head+tail preview plus a read-in-parts instruction
riding the replacement USER message (copy in
prompts/spill/user_read_instruction.md). Images stay inline, a
refused spill write degrades to today's inline behavior, and
spills.enabled: false / userThreshold: 0 / the new
FAH_USER_SPILL_THRESHOLD env knob (bytes) keep the legacy shapes.
fix(gh-1743): a structured compaction that hits its total-budget
watchdog (the factory's 300s totalBudget) no longer aborts the
live session. AutoCompactorFactory.run() now applies the
mechanical local-trim fallback (keepRecentTokens, pairing-repaired,
[context trimmed locally] marker — the same valve the classic
pass ladder ends on) instead of failing fast; the honest
both-roles-failed verdict is kept only when nothing is droppable.
fix(gh-1636): auto-compaction stops dumping the internal checkpoint
summary into the console — the default report is ONE marker line
(● auto-compacted · engine: N → M tokens (F freed · P%)); the records
line + fenced summary moved behind the new tui.showCompactionDetails
opt-in (default false), and the manual /compact still prints the full
block. The headless log face no longer streams the summarizer deltas by
default (the gh-1433 E4 live stream is part of the same opt-in; the
gh-1430 … reasoning liveness line still covers the window). Nothing
is lost — the summary stays in the session file, reachable via
compact_expand.
fix(test): gh-1732 — the wrap_ansi_line_perf_test sustained-contention
validation leg spikes 3 of the 5 doubled runs per attempt instead of 4
(CI run 39546301132, core shard 2/4: the old shape left exactly ONE
real-wall-clock-timed doubled sample per attempt, and real jitter on
that single sample — the very shape being modeled — inflated all three
attempts' min ratios to a 3.30x best against the 3.2 bound while the
real 500k growth legs on the same runner stayed green). 3-of-5 is the
documented live-red shape ("≥3 of the 5 timed runs"): median-of-5
[c,c,s] is still a spiked value, so the leg still proves the
pre-gh-1532 median harness fails, and the min harness now rides two
clean doubled samples per attempt.
fix(#1705): CLI fresh-install boarding asks the provider choice FIRST —
the same list the /add-provider picker shows (presets, Custom last) —
and only Custom continues into the api-type wizard; a preset opens its
own flow with the api type prefilled. Previously boarding opened the
custom wizard directly, cold-asking "api type" before any provider was
chosen.
feat(providers): gh-1704 — first-class Codex/ChatGPT credential export,
CI refresh action, and Provider Queue integration (Copilot-export
parity). fa provider export <name> [--check] resolves the saved
ChatGPT login's OAuth blob from its secure-store slot, refreshes it when
stale, and prints the shell lines (OPENAI_API_KEY + OPENAI_BASE_URL
= the Codex backend — these tokens never serve api.openai.com); the
refresh token never prints and --check reports credential health with
the SSO caveat. /provider chatgpt import [path] adopts an existing
codex login session (~/.codex/auth.json) into the same slot — no
browser round-trip. .github/actions/chatgpt-codex-auth mints a fresh
access token from a CHATGPT_REFRESH_TOKEN secret in CI (direct curl,
no third-party action); invalid_grant fails fast and loudly.
providersQueue: entries speak chatgpt-codex: an omitted apiKeyEnv
resolves the Keychain slot automatically (in-memory refresh), a new
refreshTokenEnv: field mints the initial access token on slot
activation, and a raw Bearer token in apiKeyEnv is tolerated by
chatgpt_codex.dart (no more FormatException — proactive/reactive
refresh degrade, a 401 surfaces as the auth death the queue fails over
on).
feat(bench): gh-1681 — leaderboard submission readiness. ATIF
trajectories: new exportAtifJson whole-snapshot exporter
(lib/src/trajectory/atif_export.dart, ATIF-v1.7 — the schema harbor
0.23.0's flagship adapters emit) behind a headless fa trajectory export [--format atif] [--output path] verb; the harbor adapter runs
it as a bounded best-effort exec after the agent phase on every
terminal path, so <trial>/agent/trajectory.json reaches the hub with
the other agent logs (the leaderboard static analysis requires one per
rewarded trial) and result.json metadata carries an
atif_trajectory present/missing marker. The golden fixture
(bench/harbor_fa/fixtures/atif_sample.json) is pinned Dart-side and
cross-validated against harbor's pydantic Trajectory +
validate_trajectory in test_atif_contract.py. Workflow: a
submission: true dispatch input turns the submission rules into hard
setup preflights (full task coverage, ≥5 attempts/task, no
agent-timeout override, pinned dataset) and makes the new per-shard
ATIF presence gate fail hard; default runs gain only a warn-only scan
(artifact flow unchanged). Runbook: docs/leaderboard-submission.md
(pinned dataset + @latest content-hash recording, quota-window runner
policy, lb CLI steps).
fix(headless): gh-1589 — the headless session becomes a SEALED artifact:
no session-JSONL writes land after the run's agent phase ends. New
Session.seal() (state in the writer, no file-format change): raised in
runHeadless' finally and the REPL teardown as the LAST session action
(close order: drain → usage fold → seal → exit, one implementation both
paths — REG-2), it makes every writer (_persistMessages,
shell_job_registry snapshots, compaction, any appendCustomEntry,
leaf moves) drop late appends with exactly ONE stderr warning per
writer class — never a throw into an unhandled zone; the exit code
stays the turn's terminal outcome (issue #413 semantics untouched). The
settle→reaction path is sealed with the file (I2): the drain wrap-up
raises a fence (_reactionRunsSealed) so a background-shell settle or
task completion arriving after the final answer's drain — including the
registry's one-turn-deferred listener racing _finishHeadlessDrain's
check-to-return window (the 65gGW7b 10:13:42–10:14:00 burst) — does
board bookkeeping only: no _startRun, no model request, no record.
An honest drain's records (final answer → seal) stay legal (E7); the
seal is per-run state, so resume/replay/session_repair are unchanged.
Adapter side (bench/harbor_fa, I3/AC4+AC7): run()/perform_task
returns only after the fa process exited on EVERY path — the ladder
kill became terminate-and-reap (SIGTERM → bounded exec await → SIGKILL
escalation → pgrep confirm), and the external-cancel paths (harbor's
own wait_for firing first, the no-knobs exec, external-cap-only mode)
now fire the kill + bounded reap before propagating, with a loud
agent-phase exit guarantee degraded stderr note when fa survives both
signals (never a silent proceed, never an unbounded hang).
fix(bench): gh-1643 — a trial killed mid-stream no longer reports 0/None
usage. The fold read only FINALIZED assistant messages, so a run killed
before its first assistant message (the costliest rows) folded to silent
zeros. Both folds now price a dangling model_request_summary at its
KNOWN request-side volume (system prompt + outbound message chars at the
shared chars/4 rate, estimated input; output unknown) and mark the usage
incomplete instead: UsageChainScan.usageIncomplete → usage.json
(usageIncomplete: true, written only when set so completed folds stay
byte-identical), SessionUsage.usage_incomplete → the tb adapter's
usage-guard.json row and harbor's agent_result.metadata, and the
post-mortem stamps usage_incomplete on every results.json row it
patches from a partial fold. Harbor also exports the run outcome class
(agent_result.metadata.failure_mode: agent_timeout self-kill /
external_kill / completed / crashed / never_started) so downstream
classification stops relying on exception.txt archaeology; a summary
superseded by another summary is flushed instead of silently dropped on
both sides. Never-started trials keep the gh-1638/gh-1576 audit-only
precedence (no fake usage).
fix(stuck-calls): gh-1632 — a stuck tool call in an interactive session now
escalates like everywhere else. effectiveStuckTool() defaulted every
non-headless host to StuckFollowUpMode.advisory when no explicit
agent.stuckTool: was configured, so the fa 1.0.580 incident session (a
dead git fetch in the REPL) fired its tool_stuck flag at the 300 s
threshold and then hung for 27 more minutes — advisory mode never cancels,
retries, or converts; only the owner's manual steering broke the deadlock.
The autonomous default (cancel → retry once → background-convert, then a
session-visible escalation) now applies in EVERY mode; advisory remains
an explicit opt-out (agent.stuckTool.followUp: advisory), never an
implicit default.
fix(bindings): gh-1624 — the gh-1570 binding preflight no longer kills
translating-proxy factories. _validateModelEndpoint's vendor-prefix
verdict was fatal on ANY endpoint, so FA_PROVIDER_TYPE=anthropic +
bifrost's /anthropic path + zai/glm-5.3 exit-64'd every boot on
1.0.576 (the prefix is a routing key on a translating proxy, not a
dialect declaration). The vendor-prefix check now fires ONLY on the
endpoints that carry a dialect verdict — the three family-exclusive
canonical vendor endpoints (api.anthropic.com, api.openai.com,
generativelanguage), mirroring the family verdict's scope: those keep
the gh-1570 fail-fast, ROUTERS (OpenRouter/Copilot/DIAL/kimi — the
prefix is a routing key there too; OpenRouter's canonical ids are
anthropic/claude-*) and unknown/private endpoints carry the mismatch
as a ProviderModelBinding.warnings entry, never exit 64. The warning
is rendered on EVERY resolve path (gh-1624 review): the FA_PROVIDER_*
boot notice, the provider queue (per entry at boot, on queue reload,
and in /provider queue test), and the --model/--base-url flag
revalidation — a rider never boots silently.
Breaking-change note for 1.0.576 upgraders behind proxy
factories: the migration is the wire: FA_PROVIDER_CONFIG key —
documented now (README env-preconfig section,
docs/agents/providers.md §Translating proxies, --help) with its
caveat: the override engages only when the declared wire DIFFERS from
the provider record's own kind (record anthropic + wire: anthropic
still runs the checks). REG matrix (AC1/AC2/AC3) pinned in
test/model_roles/provider_model_binding_test.dart; AC4 IT drives a
bifrost-shaped loopback proxy through the production anthropic adapter
(test/model_roles/bifrost_proxy_binding_test.dart).
gh-1596: the first-run wizard (site/setup.sh) no longer crashes or
corrupts input under curl | sh on dash (Ubuntu sh) or busybox ash
(Alpine): scripts/gen_installers.dart now renders the provider/mode/
approval menus from the live provider catalog into plain POSIX sh — the
bashisms that killed dash (printf -v, read -p) and silently corrupted
busybox ash are gone; the hidden API-key prompt uses stty -echo + a
by-name read (dash has no read -s) and the key never rides printf
argv; EOF/Ctrl-C mid-wizard aborts non-zero with no half-written
~/.fah, and non-TTY runs print guidance and exit 1. The wizard's temp
config is mktemp-created (no predictable, symlink-followable fixed
path). A new path-triggered installer-portability CI leg proves the
script under dash + busybox ash (ubuntu) and bash (macOS) on every
site/** or generator change.
fix(cli): gh-1572 — fa boot no longer rewrites ~/.fah/config.yaml
from a lossy config. persistConfig (bin/fah_runapp.dart) built a
fresh CliConfig carrying only some fields and persistConfig() ran
unconditionally at boot, so every launch silently dropped the user's
spills:, subagents:, jobs:, headless:, agent:, quota:,
tui:, trajectory:, waiting:, power.hold and auto_update
settings (first reported on 1.0.547). Three cooperating fixes: the
boot save is now saved.copyWith(<live overlays>) so every parsed
section rides along by construction; CliConfig.copyWith replaces the
drifted hand-rolled withCustomProviders field list (itself dropping
spills/subagents/waiting/jobs/agentMode/misuseBreaker/streamThinking
inside saveCliConfig's merge step); the _diskPreservedSections
disk backstop covers every known top-level key. spills: renders in
toYaml (new SpillsConfig.toYaml) and agent.misuseBreaker: false
persists too; an unchanged save no longer rewrites the file at all.
Rework round: the boot snapshot re-syncs from disk after every
sanctioned mid-session settings write (onUserConfigFileChanged — a
cleared context cap, an auto_update key drop or a global compaction
pick is no longer resurrected/overwritten by the next save in the
same session); validateAgentSection (the strict fa config check /
config set / settings-hub mirror) accepts the emitter's
agent.misuseBreaker and agent.stuckTool keys so fa config check
never false-fails on fa's own output; clearing the context cap drops
only the cap key instead of the whole agent: block (the sibling
mode:/misuseBreaker: keys survive); and the providersQueue:
block (no typed renderer) joins the disk backstop so a save can never
silently delete the user's queue. Rework round 2 (merge + guards):
validateAgentSection's per-key switch became the _agentKeyValidators
table (the two gh-1572 keys pushed the switch form to complexity 13 —
over the repo's CRAP ratchet; the table keeps the mirror complete at
complexity 4), and the last two direct USER-config writers —
_writeTtsrSection and _writeMcpSection — fire the re-sync hook, so
a mid-session ttsr rule toggle or mcp server edit survives the next
persist too (the writer roster in settings_flow_config_sync.dart is
now exhaustive).
gh-1606: the copilot provider accepts fine-grained PATs
(github_pat_…) directly — PAT mode skips the copilot_internal/v2/token
exchange entirely (GitHub 404s every PAT exchange) and sends the raw PAT
as the Bearer against the configured Copilot base URL with the official
Copilot CLI's first-party identity headers
(copilot-integration-id: copilot-developer-cli,
user-agent: copilot/<fa-version> (<os>) term/unknown, installed by the
host at startup). OAuth device-flow tokens keep the exchange +
proactive-refresh flow unchanged; a PAT 401/403 surfaces the error
instead of a doomed refresh retry. Headless COPILOT_GITHUB_TOKEN now
works with a PAT. Connect flow: the paste hint explains PAT mode, the
dead-PAT case surfaces at the model fetch, and /provider export copies
the saved Copilot account's full token to the clipboard (account picker
on several entries; the token never prints to the transcript).
gh-1575: macOS clipboard paste fixed for TIFF-only pasteboards + the
Ctrl+V text fallback. _readMacos probes «class TIFF» after PNGf/JPEG
miss and converts via sips -s format png (a Preview/browser copy used
to dead-end on clipboard holds no image; a failed conversion keeps the
unavailable fall-through and every scratch file is cleaned up); new
readPasteboardText (pbpaste / xclip+wl-paste / PowerShell
Get-Clipboard) behind the same injectable ProcessRunner/platform:
seam, and Ctrl+V with a TEXT clipboard now inserts the clipboard text
into the composer instead of printing the image probe's dead-end reason
— bracketed-paste routing, gesture pairing and silent-failure semantics
unchanged. Review pass: the Windows Get-Clipboard read pins UTF-8
console output and normalizes embedded \r\n (OEM-code-page mojibake),
the Linux text backends pin text/plain;charset=utf-8 (an image-only
clipboard can no longer leak bytes as "text"), the library doc no longer
overclaims the FA_FAKE_PASTEBOARD seam, and FaTuiCallbacks moved to
a fa_tui_callbacks.dart part file keeping fa_tui.dart under the
2800-line gate.
gh-1555 (W1, lazy/windowed-everywhere reads): the legacy full-open call
sites migrate to windowed opens, so a session ledger past the gh-1073
2 GiB full-open backstop stays openable on every user-facing read path.
The TUI session picker, /resume, and the boot name-resolution scan
open windowed and resolve the display name through paged reads; the
subagent transcript preview opens the child windowed (tail only); the
_loadSession marathon fallback no longer degenerates to a refused
full open when the boundary walk exhausts its page cap (it proceeds
with the tail-anchored windowed session; the boot cap owns over-window
residue); and the headless fa trajectory resolver opens windowed,
reading the whole ledger for the cost/inspect views through the new
SessionTree.ensureWholeLedgerResident — a sequential forward scan in
bounded strips (WindowedSessionStorage.loadAllOlder), not the
boundary-walk's doubling strips (whose per-strip record cap measured
~13x read/parse amplification over the whole file). The same change
block-prepends paged chunks in _indexChunk (one insertAll per
list), removing an O(n²) the suspended-eviction walk paid per record.
Trajectory export streaming (AC3a), split-and-retain rotation
activation (W2), and the opt-in write-side caps (W3) remain follow-up
work.
gh-1605: the FinalizeGate gains a "grader parity" named check. A
produced entry point (script/CLI the verifier will run itself) is only
verified when it survives the grader's own invocation: the checklist
must re-run each produced entry point from a different working
directory in a fresh subprocess with a clean env, capture stdout/
stderr/exit code, and diff against the earlier in-session run — any
divergence (different output, relative-path artifacts, locale/
encoding-dependent output) is a fail ledger item, never a pass.
Closes the modernize-scientific-stack__nwbgy7a near-miss (every
in-session checklist item passed, the verifier's clean-cwd subprocess
collapsed Station 102 to Station 101's mean, reward 0.0). REG coverage:
the parity procedure catches cwd- and env-dependent entry points and
does not false-positive cwd-independent ones
(test/agent/finalize_gate_grader_parity_test.dart), and an IT replays
the evidence shape end-to-end — divergence detected and persisted as a
fail ledger item before finalize.
gh-1603: unattended runs can no longer end as a clean success on empty
model responses. Loop: from the second empty-hiccup retry on, the request
carries a continuation nudge user message ("your last reply was
empty/truncated — continue the task"; exact-text constant
emptyContinuationNudge, emitted with MessageStart/End so the transcript
stays truthful) — a cap-truncated reasoning model re-streamed verbatim
repeats the empty outcome, the nudge breaks the symmetry; the classifier
treats that exact message as transparent, so the turn's partial-content
window (gh-1577) survives a nudged retry. Headless: the empty-reply nudge
budget rises to AgentCliConfig.headlessEmptyReplyNudges (default 2 — the
session ends only after 3 consecutive unrecoverable empties; interactive
keeps the #1085 M2b one-nudge semantics), and a run that still settles
empty prints the explicit __FA_INCOMPLETE marker and exits 1 — a
distinguishable failure for the bench adapter, never a silent exit 0.
gh-1577: an empty assistant completion after partial content (assistant
text, tool activity, or a truncation signal earlier in the turn) is a
bounded-retry provider hiccup — up to emptyHiccupRetries (default 3)
re-streams with exponential backoff (emptyHiccupBaseDelay,
emptyHiccupMaxBackoff, both threaded through the Agent facade,
AgentWiringSpec, and the CLI config), and exhaustion ends the run with
a loud StopReason.error (headless exit 1, interactive error + prompt)
instead of the silent exit-0 settle that killed the gh-1577 trials.
Clean empties keep the #1085 M2b one-nudge semantics. The abort always
wins: a cancel racing the final retry settles StopReason.aborted for
any empty shape (the exit-contract race is pinned by tests), and the
backoff sleep is cancellation-aware.
gh-1574: SQLite reads work on installed binaries — bundle libs land where
the binary looks. The AOT binary's native-assets resolver reads
<exe_dir>/../lib/<name> (the dart build cli bundle layout), but every
POSIX installer copied the bundle libs beside the binary
(<install_dir>/lib), so read path=x.db died with a raw FFI dump
(Couldn't resolve native function 'sqlite3_initialize' … Failed to load dynamic library '~/.local/lib/libsqlite3.dylib'); it only worked on machines
with a manual symlink in ~/.local/lib. Fixed the layout in all three
install paths — site/install.sh (incl. the quarantine-xattr pass and the
source-build advice), install_local.sh, and fa update
(bin/self_manage.dart copyBundleLibs, now public and layout-pinned by
tests; also aligns Windows fa update with install.ps1, which was
already correct) — plus clean degradation in the FFI engine:
Sqlite3Engine.openReadOnly classifies the dart:ffi resolver's
library-load failure (isSqliteDynamicLibraryLoadFailure) and throws a
clean SQLite is unavailable — reinstall or run fa update note instead of
the raw resolver dump; a Dart-level system-library fallback is impossible
with package:sqlite3 native assets (build-time asset mapping, no loader
override). Installs already broken by the old layout self-heal: on the
bundled library's first failed load the binary moves the stale
<exe_dir>/libsqlite3.* into <exe_dir>/../lib/ and retries the load
(repairStaleSqliteBundleLayout) — so one fa update suffices (the
successor binary repairs on its first SQLite read; a pre-fix binary cannot
repair itself), and the stale <bin>/lib leftovers disappear as they
migrate. The POSIX installers also report a failed lib copy instead of
swallowing it (site/install.sh warns naming the lib dir; same in
install_local.sh), and site/install.sh.template carries the same
layout so regenerating the installers cannot resurrect the old one.
gh-1556: bench pipeline stops bleeding score + fa deadline awareness.
Bench: the harbor ladder kill uses the [f]a bracket trick and is
wrapped so a kill failure can never replace the TimeoutError (the
pre-fix self-matching pkill classified TB 2.0 ladder kills as agent
crashes — a SOLVED task scored 0); progress-watch (stall-gap 360s per
the gh-1430 watchdog invariant, flat abs ceiling 3600s) is the
bench-harbor.yml default on every shard, with the harbor adapter
picking the gap-aware ProgressWatch for parity with the
terminal-bench adapter; the aggregate counts completeness by
result.json presence (errored trials report as their own breakdown
line; exit 1 only on genuinely lost trials) and renders the full
report — table, errored breakdown, ledger row — to
$GITHUB_STEP_SUMMARY (TB1 parity). Both bench adapters export
FA_DEADLINE_EPOCH_MS (effective cap − 120s wrap-up margin) into the
agent session and audit the exact value. fa core: deadline awareness —
FA_DEADLINE_EPOCH_MS / --timer/--deadline / /timer write one
session clock, the per-turn context carries a time-left: line while
armed, and one-shot system notices steer wrap-up at ~50%, T-3min
(drain background jobs), and T-30s; absent deadline = byte-for-byte
today's behavior.
gh-1522: tag-sourced versions — the git tag is the single source of
truth. The release-bot file bumps are retired: cutting a release = a
tag + GitHub Release (zero commits to main, zero file mutations).
Both pubspecs carry the fixed 0.0.0-dev placeholder (CI guards
hand-bumps), the publish job stamps the staged pubspec + generated
changelog section from the tag (scripts/stamp_staged_release.sh,
staged version == tag guarded pre-upload), store builds stamp
--build-name/--build-number from the derived tag version
(pubspec.lock stops churning across releases), and
install_local.sh stamps version.txt from git describe.
feat(usage): gh-1460 — the fa-tokens: segment-close line carries the
session's model id, so every token row the factory publishes is
attributable to a model and therefore priceable. The fold now captures
each segment's LAST-SEEN model id (the model that served the final
request, the same string settings/UI display — mid-session fallback
take-overs attribute the segment to the model that actually finished it;
zero-filled fallbacks never clobber a seen model) and the pinned line
shape grows an OPTIONAL "model":"<id>" key after "segment" — omitted
when no model was observed, so legacy lines parse unchanged and the
pinned regex (usageTokensLogPattern, gh-1241 AC7) keeps matching both
shapes. usage.json's schema is untouched (the model is rebuildable from
the chain, I6); old consumers using tolerant JSON ignore the new key,
and the repo's single strict consumer (the pinned constant) is updated
in the same change.
fix(test): gh-1460 rework — the wrap_ansi_line_perf_test growth-shape
legs measure the doubling ratio with the MINIMUM over 9 runs instead of
the median of 5 (CI run 38112524699, core shard 0/4: a scheduling pause
inflated the doubled-size median — 12.09ms against a true ~7ms — and
tripped the 3.2 bound at 3.65x on a genuinely linear wrap). Runner
jitter only ever inflates a wall-clock sample, so the min converges on
the true compute time; genuine superlinearity still trips the ratio
(the gh-1496 quadratic candidates measure ~4x by min). The absolute-
budget legs keep the median — their gh-1357 headroom already prices in
jitter. A synthetic regression test pins the pause-proof property.
<!-- release-draft-owner: run/42521681812/job/binaries-linux-arm64 -->
Nothing published for this version
feat(gh-2059): browser-extension compaction parity with CLI — over-window relief, /compact with visible receipts, and the shared under-window guarante
feat(gh-2059): browser-extension compaction parity with CLI — over-window
relief, /compact with visible receipts, and the shared under-window
guarantee ride the same CompactionHostWiring resolution as the CLI and
app (resolveExtensionCompactionWiring); receipts are durable across
panel attach/SW restart. Machine-review round 1: exact ladder pass
counts, busy-matrix routing (routeUserInput) for panel text and second
/compact during a pass, window settings from the ACTIVE stored source
with positive validation, string-encoded int merges.
fix(gh-2055): the daily-publish cli leg red-failed in pod install
(Build macOS x86_64, run 41778615815) because the vendored macOS pod's
XCFramework download block guarded on the downloaded-zip MARKER
(if [ ! -f wasm_run-v0.1.0.zip ]) instead of on the artifact the build
consumes. A fresh CI clone never has the marker, so pod install re-ran
the block on every build: it removed the WasmRun.xcframework the
build-macos.yml restore step had already placed (cache hit,
sha256-verified), re-downloaded the 92.7 MB asset with a bare
no---fail curl, and one transient reset (curl 56 at ~3 MB/92.7 MB)
left the Podfile nm -gU symbol check with no
macos-arm64_x86_64/libwasm_run_dart.a to read. The guard now keys on
WasmRun.xcframework/macos-arm64_x86_64/libwasm_run_dart.a — the exact
path the Podfile force_loads and nm-checks — so the block is a no-op
whenever the restore step (or a previous run) put the framework in
place, and the fallback download for a truly cold tree carries
--fail AND aborts the block at the curl step on failure
(|| { echo "wasm_run download failed" >&2; exit 1; } — a swallowed
status fell through to unzip/truncate residue and masked the root cause
behind the Podfile nm diagnostic; review round 1 on this PR), and the
podspec now assigns the backtick result and raises "wasm_run framework
fetch failed" unless $?.success? — Ruby ignores a bare backtick's
exit status, so pod install would otherwise have sailed past the failed
fetch to the downstream nm diagnostic (review round 2). Linted
behaviorally (stubbed curl, no network; POSIX-gated sandbox, plus a
ruby evaluation of the real fetch statement) by
test/wasmrun_macos_podspec_download_guard_test.dart.
fix(gh-2050): the daily-publish cli leg red-failed daily when its
macOS attach arrived after a sibling train had already PUBLISHED the
derived tag (the same-tag convergence race: mobile leg drafts + the tag
train publishes ~25 min after dispatch while the serialized self-hosted
mac builds need ~66 min — v1.0.685: draft 05:22:46, published 05:44:52,
mac leg attached 06:28:53). scripts/release_asset_gate.sh gains a
caller-declared pre-stage policy — RELEASE_GATE_PUBLISHED_SKIP=1 —
that turns the PUBLISHED+unattached-fa-* verdict into a loud whole-call
skip (::warning::, exit 0, nothing uploaded, no twin create, no flip;
all-or-nothing so latest-mac.yml can never land alone on a release
whose zips were skipped). The default stays the gh-2042 AC1 loud
refusal; only build-macos.yml's release-macos desktop attach opts in
(its DMGs/zips/manifest are clobber-refresher pre-stage assets, not
publish blockers), the CLI-tar/extension attaches keep the refusal (they
carry REQUIRED release assets). The stranded desktop feed self-heals:
the next daily leg pre-stages the next version's draft with fresh
assets.
fix(gh-2044): the auto-compact fragile-band escalation could fold the
ENTIRE visible projection when the remaining path fit inside
keepRecentTokens — the keep-recent walk defaulted to folding
everything when its budget never exhausted, swallowing the live tail
(the newest user ask and assistant turn included). Real-world trigger:
resumed subagent runs ended with no assistant message at all
(subagent produced no assistant message — CI cores shards 1/4 and
3/4, hostile-env, both red). _pickCheckpointRange now declines the
fold when the whole foldable path fits the keep-recent region and
still carries live records; all-marker paths keep the lossless
deterministic consolidation (issue #387). The escalation receipt also
reports the REAL summarizer call count (a #729 chunked fold makes N
calls, not 1), and the factory wires its run token into the engine's
cancelToken so a watchdog-shaped cancel reason (any
TimeoutException) can no longer dissolve a user abort into a
"summarizer returned no text" failed receipt while the run reports
success (#1085). Follow-up CI red on the same branch: the decline
moved tiny-transcript runs to the classic residual engine, where an
EMPTY summarizable region folded zero #729 chunks into a blank
summary that still landed as a junk CompactionRecord and a hollow
success: true through the no-work terminal — compactSession now
refuses that empty-region blank fold (gh-1743's failure-verdict
contract kept; a dialed blank answer over a real region still lands
as a record, issue #578 unchanged). Third CI red on the same branch
(flutter-app over-window relief, both shards): the whole-world decline
lived in the shared _pickCheckpointRange, so it also starved the
checkpoint LADDER — a tool-result-heavy transcript whose every ledger
token fits keepRecentTokens (the request estimate is far over the
trigger) lost its only summarizer-based relief, and the hosted app's
over-window guard ended the turn instead of continuing it. The decline
is now escalation-only: ladder passes run over the trigger, where the
fold IS the relief (the escalation still declines — semantic 5 — and
its pass body was extracted below the CRAP ratchet's 12.00).
(gh-2044 rework, UT-ESC-13)
fix(gh-2003): the Copilot picker filter is tier-aware. Individual-tier
/models responses omit model_picker_enabled entirely, so phase 1
dropped every such entry as «picker flag absent» and individual/OAuth
accounts still listed zero usable models. When the response comes from
the individual tier — the individual host (api.githubcopilot.com or
pi-mono's api.individual.githubcopilot.com spelling) or a
no-proxy-ep token (individual by design) — an ABSENT flag is now
selectable: absence ≠ an explicit false, which still drops
everywhere (embeddings, completion-only). Business/enterprise answers
keep the strict picker list unchanged, and the keep is never silent:
the fetch note names it («2 of 2 models kept via the individual-tier
fallback (picker flag absent)») through the existing fetch notes:
surfacing. pi-mono parity: the trigger mirrors allowPolicyFallback = baseUrl === "https://api.individual.githubcopilot.com" in
packages/ai/src/auth/oauth/github-copilot.ts.
fix(gh-1990): the Copilot /models dialect never answers a silent
empty list again. Every silent-empty exit names its branch through the
new fetchModelsForEndpointReport dispatch (ModelsFetchReport +
ModelsFetchDiagnostic): the token-exchange HTTP status (with the
OAuth device-flow / fine-grained PAT remedies for classic ghp_…
keys), the /models HTTP status or timeout (with the seat + per-tier
host hint), proxy-ep-less individual tokens aimed at a business/
enterprise configured host (the tier-misrouting hint), and the picker
filter with per-model drop reasons (picker flag absent, picker flag false, policy disabled, no tool_calls, endpoint unsupported) —
«12 models fetched, 12 filtered (picker flag absent)». A failed
primary credential path retries the other path ONCE before declaring
empty (dual-path fallback, gh-1606): an exchange-401 classic PAT tries
PAT-direct (the provenance note names which path answered), a
fine-grained PAT never falls into the exchange, an OAuth token never
becomes a direct PAT Bearer. Surfacing: /models prints a fetch notes: section per provider, the /provider copilot connect flow
prints the named failure before manual entry, the app's model
quick-select (EndpointModelsController.emptyReason) renders it under
the field, and the Copilot connect sheet shows the reason above the
manual model id. The legacy fetchModelsForEndpoint answers are
unchanged (parity); the filter criteria themselves are unchanged
(pi-mono parity — only the reporting is new). Review round 2: the
fetch notes: wiring covers the unsaved sibling paths too (the
active-catalog and env-keyed Copilot fetches surface the same named
failure instead of a silent seed fallback), the connect flow's
no models: line reuses ModelsFetchReport.reasonText (the last
hand-rolled message — hint join), and the ModelsFetchReport
docs scope the diagnostic contract to the dialects that name their
failure branches.
feat(network): channel payload sender block — app members' channel
messages carry {"text": …, "sender": {"name": …, "id": …}} (public
base64 JSON and keyed fanet1 plaintext alike), so DAP agents can
attribute the HUMAN author behind the fa_network relay's hub id. The
harness unwraps it tolerantly (raw text passes through), rides the
name as AgentMessage.fromName through the fabric, and the CLI
steering line greets people again: from Uladzimir (<relay id>) in #channel: hello instead of from 3cfaee1aa054fe6e…. The app falls
back to the payload sender for bubble names when the envelope carries
none; DMs also name hub peers from the whois cache. Legacy payloads
without the block decode exactly as before.
feat(gh-1956): reply auto-spill — oversized child reply / agent_message
payloads spill to an agent:// artifact automatically; the parent receives a
bounded excerpt + pointer + completeness marker with char AND byte counts.
Manual chunking degrades to a documented fallback, restorable across parent
resume via durable deliverable_artifact records, behind the
subagents.replyAutoSpill kill switch (default on; honored on the CLI and
the app host alike). Rework round: the salvage hook (output-truncated
deliverable calls) restores the durable ids before allocating its own
(no post-resume pointer shadowing), durability losses emit a
[deliverables] diagnostic line, the app host honors the kill switch, the
spill-reach gate is renamed spillAppliesTo (was spillApliesTo), byte
counts are encoded once per spill, and DeliverableRecord.artifact(bytes)
is optional again so external constructors keep compiling.
fix(#1947): schedule_message from the orchestrator no longer misroutes
into a finished subagent's mailbox. A background shell job started by a
CHILD registered its settle continuation inside the child's subagent
zone; when the job settled after the child's run ended, the CLI's
settle-wake (_onShellJobSettled → _startRun) started MAIN's reaction
run still carrying the finished child's activeSubagentId(), and every
identity consumer of that run misattributed it — the reported shape was
a self-reminder scheduled "in <session>/<child>" that auto-resumed the
completed child (duplicate-release risk). Three layers: the
ShellJobRegistry now binds its settle continuation to the
construction-time (boot) zone (_bootZone.run), both hosts'
scheduleSenderMailbox wiring trusts the zone scope only while the
child is genuinely in flight on the session executor
(TaskExecutor.isInFlight), and the tool response names the RECIPIENT
the record carries — an explicit to: used to be shadowed in the reply
by the sender's mailbox, which read as "the override was ignored" even
when it was honored. E2E repro drives the real AgentCli
(test/cli/agent_cli_schedule_recipient_test.dart) plus a registry-level
zone pin (test/tools/shell_job_settle_zone_test.dart).
fix(gh-1948): task_observe works on a RUNNING background subagent
instead of crashing with an uncaught SessionException on the
<parentSessionId>/<agentName> placeholder. Three legs: the executor
flushes each child's transcript incrementally at every turn boundary
(the JSONL exists and grows mid-run — also crash-resilience for
aborted children), task_observe resolves a placeholder handle's real
session file from the sessions tree (the same agent: subagent /
parent: / id: header scan the boot adoption uses) and heals the
handle onto the found path, and every read failure degrades to a
structured "transcript unavailable" / "no transcript available yet"
message instead of an uncaught tool error. The child-transcript reader
rides the transient-ENOENT retry (issue #427) — it reads while the
executor appends.
fix(#1941): the selected CodeMie project now rides EVERY model request.
The header is baked into Model.headers only by the connect-time
switches (gh-1809), so every other rebuild silently dropped
x-codemie-project — and the backend bills a header-less request against
the user's DEFAULT project (codemie-code header-injection.plugin.ts
contract): the reported symptom is limit-exhausted from the exhausted
default project despite picking a budgeted one. Rebuilds now re-resolve
the project from the saved entry through one helper
(projectHeaderForBaseUrl): the boot folder-state restore (JWT entries
previously booted header-less — the SSO cookie patch was the only boot
carrier), the session-resume re-apply (_buildRestoredModel), the
roles-chain failover (ModelRef.project, mirroring the #964 authHeader
leg), the roles-mode /model + settings-flow pins, and the generic
/provider switch. The models list (/llm_models, SSO + JWT + guided
setup + entry dispatch) and the quota adapter (dark until pinned) scope
themselves to the same project, so display and billing can never diverge.
Tests pin the header VALUE on the wire (real adapter over a fake
endpoint) for the restored and failed-over models.
feat(network): private networks created outside the app surface after
sign-in. GET /api/networks/mine (fa_network#16) is synced into the
wallet after OAuth/dev sign-in and both restore paths — best-effort,
silent on older backends (404) and any failure, and never destructive.
Synced memberships keep the server's name/class, preserve local
passwords/join facts, and land without a password (the endpoint never
returns one) until the first join. Entering such a network — sidebar,
⌘K/⌘3-9, or a restored selection — opens the join sheet prefilled with
the network id instead of a resume error or endless spinner; joining by
name now keys the wallet/session by the echoed canonical id and migrates
legacy name-keyed rows (password + channel keys) instead of duplicating
them. A malformed /mine row is skipped, not the whole list.
fix(gh-1931): one render per tool call on the non-TUI console. Line mode
and the headless -p log printed every executed call twice — a start row
(• bash · cmd, no elapsed) followed by the end row (✓/✗ bash · cmd 0s). The duplicate start-row print is gone from _onToolExecutionStart;
the end row (command, status glyph, elapsed, failure news) is the single
render per call. The TUI keeps its running card + settled card, start
bookkeeping stays (_toolStarts feeds the end row's detail/elapsed, the
waiting coordinator still arms per-call liveness), and persisted
artifacts (session JSONL, trajectory records) are untouched —
rendering-only change. Long foreground calls keep the gh-1055 liveness
reminders past waiting.toolLivenessSeconds (default 60s).
fix(gh-1929): esc during the pre-flight auto-compaction no longer crashes
the CLI. The compaction layer's deliberate abort (#1085) used to escape
_runPrompt outside its guarded body and reach the root zone handler via
an unlistened whenComplete chain — fa crashed: CancelledException +
crash.log + a lost session view. The pre-flight window (submit on an
over-window transcript) now translates the cancel into a user-cancel
outcome: neutral dim submission cancelled receipt (never error:), the
TUI composer hands the typed text back (sendInputText) so a resubmit is
one Enter away, headless exits 130 with the same receipt, and the next
prompt works (the funnel's local-trim recovery or a retried compaction).
Cancels DURING the run — in-flight relief, settle funnel — keep their
pinned loud-abort surfaces, and the post-run window keeps its
compaction interrupted receipt via the shared _renderUserCancel
helper.
feat(1923): stall-proof steering. Steering arriving while the loop awaits
an UNPRODUCTIVE provider submit (zero bytes, or silent past the steering
stall window) now cancels the attempt via a SteeringInterruptFire
cancel, injects the pending steering into the rebuilt request, and
resubmits — the #1916-class submit stalls can no longer queue user or
agent mail invisibly. Productive streams keep next-boundary semantics;
the interrupt never consumes a retry attempt, and the run latch re-arms
in place so watchdog replays and user aborts keep working. task_resume
caps its foreground wait (like task_send) and continues in the
background; video polling honors soft-yield checkpoints (remote render
never restarted, honest hand-back text) and media generation honors run
cancellation; SLO-breached undelivered child mail surfaces in
task_status (UNDELIVERED … past the 2s delivery SLO); the TUI prints
steering queued — provider stalled Ns, interrupting to deliver plus the
delivered ack.
feat(1916): image-heavy submits no longer livelock — byte-based
images.maxBytesPerRequest (default 8MB, layered onto the count cap;
/images names both drop reasons), legacy text-embedded data-URL
screenshots in history fold into the once-only registry instead of
replaying raw forever, stall watchdog windows scale with the outbound
payload (capped at 2x; explicit connect clamps untouched), abandoned
connect attempts answering during the retry backoff are adopted, and
repeated identical-payload stalls fire ONE actionable /compact +
image-budget advice. TUI stall banner + docs/images.md carry the
image-heavy playbook.
fix(gh-1912): Add an agent → DAP now turns a retired hub's unknown
enroll_create response into an actionable upgrade message instead of
exposing bad_frame. Independent enrollment never falls back to the
legacy same-name credential rotation endpoint, and existing credentials
remain untouched on failure.
fix(gh-1913): Add an agent is usable on narrow phone screens: scope and format choices wrap without per-character labels, enrollment controls stack at full width, and actions remain reachable by scrolling. Desktop keeps the segmented layout; phone/desktop light/dark goldens guard both.
fix(gh-1825): DAP agent enrollments carry a stable, independent
agentInstance through app/CLI bundles, persistence, direct boot and
reconnect. Independent same-name agents keep distinct identities and stay
online together, while legacy bundles retain their project-scoped identity.
feat(gh-1825): DAP private-channel sends now use self-describing
fanet1 frames (sender X25519 key in the frame, channel-bound AAD) via
fa_hub_client 0.2.11, so fa_network app members decrypt them without
a whois/senderKey round-trip. Inbound channel frames open fanet1
first and fall back to legacy dap/1 payloads. Verified live in both
directions between a headless DAP agent and the Flutter app.
feat(gh-1825): the network offline banner now has a Reconnect now
action which restarts the socket and reloads the channel/member
snapshot. Fresh sockets also clear a stale offline flag (the server
sends no historical drain to a socket that was away during recovery).
AgentNetworkController.start stops after disposal instead of
notifying a dead controller.
feat(gh-1825): the roster sheet shows access-scope badges (whole network vs N channels, decoded additively from Member.access in snapshots and presence.changed) and owner/admins can drop an agent (DELETE /api/networks/{id}/agents/{agentId} with confirmation; the server tombstone keeps it out of future snapshots). Guests and store members are removable too via DELETE /api/networks/{id}/members/{id} (sessions revoked server-side); owner/admin/self rows stay protected.
feat(gh-1825): channel/network management in the app — rename a channel (PATCH), delete a channel or an owned network (REST DELETE with confirmation and state/navigation cleanup), and unnamed CLI/hub-created channels show a short friendly label instead of the raw 64-char id.
feat(gh-1825): pasting a DAP invite no longer demands a passphrase on macOS — importing NEW channel secrets auto-provisions a random passphrase into the per-channel Keychain item (decryption of existing envelopes never auto-generates; env/prompt/headless behavior unchanged), and passphrase warnings now name the concrete recovery (re-paste the invite to re-encrypt, or set DAP_KEY_PASSPHRASE).
feat(gh-1825): the network online counter opens a live-first roster sheet (names, member class, presence), and the channel composer offers roster-filtered @-mention autocomplete — the draft keeps the literal @displayName for DAP agents and the envelope carries the resolved member ids. On the CLI side the DAP channel notice names the agent's own identity (display name + agentId), and inbound channel mail that @-mentions that name (case-insensitive, word-boundary — no email or @name2 false positives) is marked "[mentions you — you are @name]".
fix(gh-1825): the app's network roster no longer reads as permanently empty ("0 online"): fa_network sends roster.snapshot with an OBJECT payload ({members: [...]}) and the app only accepted a bare list, so every snapshot silently decoded to zero members. Both shapes are now accepted, with a production-shape regression test. The DAP paste hint names web alongside CLI/macOS/iOS (plain-web join is supported).
fix(gh-1825): resumed CLI sessions rejoin their DAP channels and receive channel mail again — session replay now re-arms membership synchronization (plugin session-ready listeners on boot/resume and in-process switches) and idle channel probes complete deferred joins before inbox peek/drain. The post-replay sync also dials the membership's OWN hub with its enrolled credential before joining: a resumed session-scoped agent used to stay on the zero-config local URL forever, and because a disconnected fabric probe short-circuits before the sync, the remote dial never happened — the session-ready listener now drives it directly (throttled to one attempt per 10 s). App channel history no longer flashes raw payloads on first open: decoding waits for startup's channel listing instead of caching an undecoded guess until a manual switch.
feat(gh-1825): the fa app consumes the same DAP enrollment block as the
CLI — paste it into the agent chat composer (macOS/iOS/Android; the
smart-paste .txt chip path included) and the channel join happens
without the block ever reaching sendText, analytics, or session
records. Native joins ride the live agent-fabric hub repository:
per-project scoped identity, invite pub/priv import, keyless public
channels, reconnect-persisted membership in a dedicated secure
enrollment store (Keychain/Keystore where available). Harness additions:
HubMessagingRepository.joinChannel/waitForConnection with
rejoin-before-drain on reconnect, keyed channel decrypt
(hubDecryptPayloadWithKeyPair), public showcase payloads, and
HubBearerTransport (native Authorization header or the browser
?token= equivalent; authed dials never stringify transport errors
into logs). Plain web joins through the browser
WebHubTransport (?token= bearer equivalent, IndexedDB identity);
the browser extension rides the service worker's existing DAP client
(hub.join op), never a panel-local second connection. Add agent now defaults to the
DAP tab with a paste-into-chat hint; web gains ai-native provider
sign-in through the existing popup/callback OAuth flow.
fix(fa_network#4): the app inserts hub agents arriving via
presence.changed into the roster (additive displayName/class fields)
instead of dropping them, so the online counter no longer sticks at 0
when an agent comes online after roster.snapshot; legacy frames
without those fields create no phantom members. roster.snapshot/REST
contents are logged for decisive presence debugging.
feat(gh-1825): /network + scoped DAP channel membership — join by
paste (invite url / url+keys / bare channel id / whole fa dap import
command) with session/project/global scopes, live join without
restart, per-scope leave + copy invite, and the composer-paste
shortcut into the same wizard. Session scope rides a
dap_channel_membership custom record (leave tombstones, branch
resume); project scope is .fah/dap.json (committable, dapkey1
PBKDF2+ChaCha20-Poly1305 encrypted keys); global keeps the legacy
~/.dap files. Headless CI: DAP_CHANNELS_JSON (+DAP_HUB_URL,
DAP_CLIENT_SECRET, DAP_KEY_PASSPHRASE) joins at boot with zero
prompts, plaintext or encrypted privs mixed. Passphrase chain:
env → per-channel macOS Keychain item → one prompt with save default.
feat(gh-1844): media slots never fall back to the main connection
implicitly — SDK/CLI, app store, and the JS bridge resolve a slot
without an override as UNCONFIGURED with an actionable
"configure the slot" error; the main connection serves a slot only via
the explicit per-slot useMainConnection: true opt-in (YAML
models.slots.<slot>.useMainConnection / app's "Same as main
connection" pick, preserved through media_models.json). Image and
video generation accept multimodal input (text + reference images:
sandbox paths, http(s) URLs, base64) on capability-declaring
OpenRouter endpoints (input_references + declared/probed
input_modalities gate; unsupported endpoints fail before any network
call); the OpenRouter video job API is a first-class video dialect
(submit → poll → mp4). New decide tool: structured decisions via
OpenRouter's Decisions API (/v1/systemone) — bool/choice/score
questions with calibrated probabilities, structured error results, no
streaming.
feat(gh-1834): the FinalizeGate contract counts a parity/conformance
claim as verified only with negative + combination self-checks — new
numbered step 2.6 in prompts/cli/finalize_gate.md (refused
incompatible/mutually exclusive options, empty/corrupt inputs, boundary
values, must-NOT-work cases, alongside happy paths; a blanket
"everything matches" is not verified). Universality: abstract example
only, no benchmark/task/flag names in the prompt; the step is
conditional, so tasks with no conformance claim see no gate change.
Ledger/verdict wire formats unchanged; validated end-to-end by the
scripted IT-4 replay (real CLI + real bash) whose persisted
task_ledger carries the rejection rows.
fix(gh-1832): release v1.0.635 published without fa-extension.zip — the
landing-page download 404'd. The extension's only producer (the daily
build-macos.yml build-extension leg) derives its tag as "latest+1 at its
own dispatch moment"; on 2026-10-24 the v1.0.635 tag was cut by
auto_release before those legs ran, so the extension (and the DMGs)
landed on an orphaned v1.0.636 draft while the tag train published
v1.0.635 — every leg green, nothing asserted the published asset set.
Now scripts/release_required_assets.txt is the single source of truth,
scripts/assert_release_assets.sh gates the release-provenance publish
flip (a missing asset leaves the release a draft and fails the run
loudly) and a new release-asset-smoke ci.yml job post-publish probes
every landing-page release link + the full required set (red on non-200).
Remediation: the exact-tag fa-extension.zip (built from the v1.0.635
tree by run 40064876732) was uploaded to release v1.0.635 — the landing
link serves 200 again.
fix(gh-1820): the arm64 DMG leg's stale size floor is reseeded onto the
shipped v1.0.634 shape — run 4004758753 (post-#1815 SPM-lane adoption)
passed the min-platform gate and the full build, then failed the
size_manifest.py +5% budget on App.framework/App +5.1%
(25,020,112 → 26,284,144, the accumulated Dart AOT of #1793/#1794/
#1796/#1805/#1806); the committed scripts/size_budget/fa-macos-arm64.tsv
was still the pre-#1096-diet floor (TOTAL 210.5 MB vs the artifact's
116.4 MB), so the documented intentional-growth remediation applies:
the floor is now the exact --reseed output for the failing run's
artifact manifest (26 rows ≥ 64 KiB + TOTAL 116,386,369, from the run
log's printed table — the zip never uploaded past the gate), locking in
Fa 23.4→18.6 MB, quickjs_bridge 950→844 KB; new
test/ci_1820_macos_arm64_size_floor_test.dart gates the shape green
through the REAL size_manifest.py check and pins the committed floor
byte-identical to a --reseed round-trip.
fix(gh-1814): the desktop DMG lane ADOPTS the SwiftPM lane — Flutter
3.47.7 removed the FLUTTER_SWIFT_PACKAGE_MANAGER disable knob
(flutter/flutter#193978), so the gh-180 pure-CocoaPods pin became a no-op
and the min-platform gate (which forbade any SwiftPM state) failed both
DMG legs on every pinned-3.47.7 run (39991663029, 40022900373); the
#1813 escape (pin 3.47.4) is unreachable on the ~60kB/s runner link.
scripts/assert_macos_min_platform.sh now ASSERTS the SPM lane instead:
SwiftPM state expected after flutter build macos --config-only,
generated package floor ≥ the app floor (the silent bump must not skip —
the original gh-180 killer), no SwiftPM product above the app minimum,
vendored inappwebview manifest floored at 14.0 if it ever reappears.
Floors were audited and need no raising: Runner pbxproj 3×14.0, Podfile
:osx, '14.0', generated floor 14.0 on 3.47.7, vendored podspec 14.0
(its Package.swift stays deleted — the plugin rides the CocoaPods half of
the hybrid lane via Podfile.lock). The dead env pin left the build-macos
job env; the SDK pin reverts #1813 back to 3.47.7. The ci.yml macOS debug
leg gets the same treatment (gh-1814 review): its
FLUTTER_SWIFT_PACKAGE_MANAGER=0 pin — a silent no-op on the floating
3.47.x stable it resolves past the knob's removal — is removed and the
lane comment corrected; the leg rides SPM enabled as it already has been
(run 399572012 green).
fix(gh-1805): modifier chords no longer type their base letter into the
composer while the slash menu is open — the vendored dart_tui decoder
puts the base letter into the keystroke's text for control bytes
(Ctrl+S → ctrl+s with text s), and the slash-menu edit cluster was
the one insert surface missing the _isCommandKeystroke guard, so a
chord pressed mid-menu poisoned the typed prefix (/memory landed as
/smemory). Plain letters and Shift+letters still type and filter.
feat(1515): ONE model-limits settings surface — the settings hub's
standalone "Model parameters" row is retired and bare /model-edit
(TUI) aliases into /settings → Model capabilities, pre-focused on the
active model. The caps flow gains the model-edit preset ladders +
Custom… free input for context window / max output (a preset pick
writes models.overrides AND rebuilds the active model on the spot —
the old flow's live-mutation semantics, now durable), every field row
shows the current EFFECTIVE value with its source layer (pinned /
role override / catalog), the thinking row is an omp-style selector
(only levels the model supports, per-rung budget metadata, current
level preselected), and the hub row describes the effective summary
(200K ctx · 16K out · thinking off) instead of a pin count. The
scripted /model-edit <field> <n> form and the line-mode limits print
are unchanged.
feat(gh-1746): opt-in session disk retention — sessions.retention in
~/.fah/config.yaml caps the sessions root and sweeps whole session
units (transcript + sidecars) oldest-last-used-first at boot, exempting
live, named, young, and the current session; nothing is EVER
auto-deleted without the section (byte-identical legacy boot). Adds
fa session clean [--dry-run], the /sessions + fa session list
usage footer under a cap, and header-cwd/path-escape guards so a
planted session header can never steer a delete outside any root.
fix(gh-1763): a bash tool command running broad process kills can no
longer signal its own shell or the harness. Scripts are delivered from a
per-exec private temp file (sh -c '. "$1"' sh <script>) so the wrapper
argv never carries the script body (pkill -f <pattern> cannot
self-match its own sh -c; $0/$#/exit codes/set -e/stdin stay
identical), a PATH shim for pkill/killall plus a kill() prelude
guard exempt the harness pid chain + process group (exported as
FA_PROTECTED_PIDS) and the invoking shell from the child's own kills
while legitimate victims (stuck test runners, background servers) are
still terminated, and headless SIGTERM aborts now log
fa: SIGTERM received … so a child-issued kill never masquerades as a
user Ctrl+C (exit code contract unchanged: 130 graceful, 143 escalate).
fix(gh-1778): the release gate no longer reds a fully-green shard when
GitHub's cache service degrades. v1.0.621's Core tests (shard 2/4) was
killed by its own timeout-minutes: 12 with 2535/2761 tests green — both
cache steps had stalled ~133 s on keys that existed (degraded service,
not eviction), cold pub get + native-asset hook builds delayed the
first test to ~5 min, and the shard needed ~14 min against a budget
sized for the warm ~8.8 min wall. The job budget is now 20 min (sibling
test-leg band), and both test-core cache steps are best-effort
(timeout-minutes: 2 + continue-on-error: true) — a failed restore is
slower setup, never a red shard. test/ci_core_shard_timeout_guard_test.dart
pins the budget and the best-effort cache contract (RED-first on the old
YAML).
feat(#1761): the 32KB project-context budget never cuts silently again.
A file truncated to its head excerpt now carries an extended
<!-- From: path — N bytes, only the first M shown --> marker plus an
explicit MUST-read-the-full-file instruction (offset/limit/ranges,
prompt copy in prompts/context/project_context_read_full.md); a file
dropped entirely leaves a one-line tombstone naming its path instead of
vanishing; and the user gets a dim boot line per truncated/dropped file
(CLI line mode prints directly, TUI re-prints after the banner, the app
logs through its AppLog boot-notice seam). New projectContext: config
section: maxBytes (default 32768) — 0 disables the section entirely
and reads no files (fast boot); unknown keys are tolerated with a note.
Leaf-first allocation is unchanged.
fix(gh-1757): the Android build no longer re-enters the sdkmanager
download roulette for CMake. v1.0.619's build-android leg died at
:jni:configureCMakeDebug[arm64-v8a] — the :jni module (transitive
jni package via path_provider_android) wires externalNativeBuild.cmake
with no version, so AGP auto-installed its default cmake;3.22.1
mid-configuration and one corrupt archive ("Archive is not a ZIP
archive") red'd the leg, the aggregate Quality gate, and the
Advance-vpinned-marker job. The gh-1655 pre-install action now also
pre-installs cmake;3.22.1 with bounded retry + idempotent skip (same
hardening the NDK path already carries); the guard test pins the new
AC6 assertions and the AGP-version coupling (AC7).
feat(#1750): per-model image caps — the outbound request never carries
more unique images than the target model accepts. Model gains an
optional maxImages (catalog-seeded: gemini-4-argon: 10, measured);
images.perModel.<modelId> config overrides it (strict images:
section, user truth wins); the registry resolves
perModel > catalog > images.maxPerRequest (first present wins, the
issue-#1554 shape). A Google-shaped too many images … maximum allowed for model <id> is <n> 400 is classified (image_overflow.dart, the
overflow.ts twin), latches the learned cap for the (provider, model)
pair for the run, and re-sends ONCE through the registry at the
smaller cap — current-first-then-newest drops with the existing
never-silent notices naming the learned cap; a second refusal
surfaces exactly as today. JSONL is never rewritten;
images.registry: false keeps the byte-identical legacy shape
including the recovery-off behavior.
feat(gh-1760): oversized user-message intake spills to a file instead of
blowing the context window. Any incoming user message — headless
-p/positional, TUI first/mid-session, steering/follow-up drain,
external inbox — whose text passes the new relative threshold (a
spills.userFraction of the live window budget minus the compaction
reserve, floored by spills.userThreshold bytes) is redacted and
written verbatim to .fah/spills/<sessionId>/user-<n>.md (the #678
layout) from one choke point (Agent.userMessageGuard); the model
receives a bounded head+tail preview plus a read-in-parts instruction
riding the replacement USER message (copy in
prompts/spill/user_read_instruction.md). Images stay inline, a
refused spill write degrades to today's inline behavior, and
spills.enabled: false / userThreshold: 0 / the new
FAH_USER_SPILL_THRESHOLD env knob (bytes) keep the legacy shapes.
fix(gh-1743): a structured compaction that hits its total-budget
watchdog (the factory's 300s totalBudget) no longer aborts the
live session. AutoCompactorFactory.run() now applies the
mechanical local-trim fallback (keepRecentTokens, pairing-repaired,
[context trimmed locally] marker — the same valve the classic
pass ladder ends on) instead of failing fast; the honest
both-roles-failed verdict is kept only when nothing is droppable.
fix(gh-1636): auto-compaction stops dumping the internal checkpoint
summary into the console — the default report is ONE marker line
(● auto-compacted · engine: N → M tokens (F freed · P%)); the records
line + fenced summary moved behind the new tui.showCompactionDetails
opt-in (default false), and the manual /compact still prints the full
block. The headless log face no longer streams the summarizer deltas by
default (the gh-1433 E4 live stream is part of the same opt-in; the
gh-1430 … reasoning liveness line still covers the window). Nothing
is lost — the summary stays in the session file, reachable via
compact_expand.
fix(test): gh-1732 — the wrap_ansi_line_perf_test sustained-contention
validation leg spikes 3 of the 5 doubled runs per attempt instead of 4
(CI run 39546301132, core shard 2/4: the old shape left exactly ONE
real-wall-clock-timed doubled sample per attempt, and real jitter on
that single sample — the very shape being modeled — inflated all three
attempts' min ratios to a 3.30x best against the 3.2 bound while the
real 500k growth legs on the same runner stayed green). 3-of-5 is the
documented live-red shape ("≥3 of the 5 timed runs"): median-of-5
[c,c,s] is still a spiked value, so the leg still proves the
pre-gh-1532 median harness fails, and the min harness now rides two
clean doubled samples per attempt.
fix(#1705): CLI fresh-install boarding asks the provider choice FIRST —
the same list the /add-provider picker shows (presets, Custom last) —
and only Custom continues into the api-type wizard; a preset opens its
own flow with the api type prefilled. Previously boarding opened the
custom wizard directly, cold-asking "api type" before any provider was
chosen.
feat(providers): gh-1704 — first-class Codex/ChatGPT credential export,
CI refresh action, and Provider Queue integration (Copilot-export
parity). fa provider export <name> [--check] resolves the saved
ChatGPT login's OAuth blob from its secure-store slot, refreshes it when
stale, and prints the shell lines (OPENAI_API_KEY + OPENAI_BASE_URL
= the Codex backend — these tokens never serve api.openai.com); the
refresh token never prints and --check reports credential health with
the SSO caveat. /provider chatgpt import [path] adopts an existing
codex login session (~/.codex/auth.json) into the same slot — no
browser round-trip. .github/actions/chatgpt-codex-auth mints a fresh
access token from a CHATGPT_REFRESH_TOKEN secret in CI (direct curl,
no third-party action); invalid_grant fails fast and loudly.
providersQueue: entries speak chatgpt-codex: an omitted apiKeyEnv
resolves the Keychain slot automatically (in-memory refresh), a new
refreshTokenEnv: field mints the initial access token on slot
activation, and a raw Bearer token in apiKeyEnv is tolerated by
chatgpt_codex.dart (no more FormatException — proactive/reactive
refresh degrade, a 401 surfaces as the auth death the queue fails over
on).
feat(bench): gh-1681 — leaderboard submission readiness. ATIF
trajectories: new exportAtifJson whole-snapshot exporter
(lib/src/trajectory/atif_export.dart, ATIF-v1.7 — the schema harbor
0.23.0's flagship adapters emit) behind a headless fa trajectory export [--format atif] [--output path] verb; the harbor adapter runs
it as a bounded best-effort exec after the agent phase on every
terminal path, so <trial>/agent/trajectory.json reaches the hub with
the other agent logs (the leaderboard static analysis requires one per
rewarded trial) and result.json metadata carries an
atif_trajectory present/missing marker. The golden fixture
(bench/harbor_fa/fixtures/atif_sample.json) is pinned Dart-side and
cross-validated against harbor's pydantic Trajectory +
validate_trajectory in test_atif_contract.py. Workflow: a
submission: true dispatch input turns the submission rules into hard
setup preflights (full task coverage, ≥5 attempts/task, no
agent-timeout override, pinned dataset) and makes the new per-shard
ATIF presence gate fail hard; default runs gain only a warn-only scan
(artifact flow unchanged). Runbook: docs/leaderboard-submission.md
(pinned dataset + @latest content-hash recording, quota-window runner
policy, lb CLI steps).
fix(headless): gh-1589 — the headless session becomes a SEALED artifact:
no session-JSONL writes land after the run's agent phase ends. New
Session.seal() (state in the writer, no file-format change): raised in
runHeadless' finally and the REPL teardown as the LAST session action
(close order: drain → usage fold → seal → exit, one implementation both
paths — REG-2), it makes every writer (_persistMessages,
shell_job_registry snapshots, compaction, any appendCustomEntry,
leaf moves) drop late appends with exactly ONE stderr warning per
writer class — never a throw into an unhandled zone; the exit code
stays the turn's terminal outcome (issue #413 semantics untouched). The
settle→reaction path is sealed with the file (I2): the drain wrap-up
raises a fence (_reactionRunsSealed) so a background-shell settle or
task completion arriving after the final answer's drain — including the
registry's one-turn-deferred listener racing _finishHeadlessDrain's
check-to-return window (the 65gGW7b 10:13:42–10:14:00 burst) — does
board bookkeeping only: no _startRun, no model request, no record.
An honest drain's records (final answer → seal) stay legal (E7); the
seal is per-run state, so resume/replay/session_repair are unchanged.
Adapter side (bench/harbor_fa, I3/AC4+AC7): run()/perform_task
returns only after the fa process exited on EVERY path — the ladder
kill became terminate-and-reap (SIGTERM → bounded exec await → SIGKILL
escalation → pgrep confirm), and the external-cancel paths (harbor's
own wait_for firing first, the no-knobs exec, external-cap-only mode)
now fire the kill + bounded reap before propagating, with a loud
agent-phase exit guarantee degraded stderr note when fa survives both
signals (never a silent proceed, never an unbounded hang).
fix(bench): gh-1643 — a trial killed mid-stream no longer reports 0/None
usage. The fold read only FINALIZED assistant messages, so a run killed
before its first assistant message (the costliest rows) folded to silent
zeros. Both folds now price a dangling model_request_summary at its
KNOWN request-side volume (system prompt + outbound message chars at the
shared chars/4 rate, estimated input; output unknown) and mark the usage
incomplete instead: UsageChainScan.usageIncomplete → usage.json
(usageIncomplete: true, written only when set so completed folds stay
byte-identical), SessionUsage.usage_incomplete → the tb adapter's
usage-guard.json row and harbor's agent_result.metadata, and the
post-mortem stamps usage_incomplete on every results.json row it
patches from a partial fold. Harbor also exports the run outcome class
(agent_result.metadata.failure_mode: agent_timeout self-kill /
external_kill / completed / crashed / never_started) so downstream
classification stops relying on exception.txt archaeology; a summary
superseded by another summary is flushed instead of silently dropped on
both sides. Never-started trials keep the gh-1638/gh-1576 audit-only
precedence (no fake usage).
fix(stuck-calls): gh-1632 — a stuck tool call in an interactive session now
escalates like everywhere else. effectiveStuckTool() defaulted every
non-headless host to StuckFollowUpMode.advisory when no explicit
agent.stuckTool: was configured, so the fa 1.0.580 incident session (a
dead git fetch in the REPL) fired its tool_stuck flag at the 300 s
threshold and then hung for 27 more minutes — advisory mode never cancels,
retries, or converts; only the owner's manual steering broke the deadlock.
The autonomous default (cancel → retry once → background-convert, then a
session-visible escalation) now applies in EVERY mode; advisory remains
an explicit opt-out (agent.stuckTool.followUp: advisory), never an
implicit default.
fix(bindings): gh-1624 — the gh-1570 binding preflight no longer kills
translating-proxy factories. _validateModelEndpoint's vendor-prefix
verdict was fatal on ANY endpoint, so FA_PROVIDER_TYPE=anthropic +
bifrost's /anthropic path + zai/glm-5.3 exit-64'd every boot on
1.0.576 (the prefix is a routing key on a translating proxy, not a
dialect declaration). The vendor-prefix check now fires ONLY on the
endpoints that carry a dialect verdict — the three family-exclusive
canonical vendor endpoints (api.anthropic.com, api.openai.com,
generativelanguage), mirroring the family verdict's scope: those keep
the gh-1570 fail-fast, ROUTERS (OpenRouter/Copilot/DIAL/kimi — the
prefix is a routing key there too; OpenRouter's canonical ids are
anthropic/claude-*) and unknown/private endpoints carry the mismatch
as a ProviderModelBinding.warnings entry, never exit 64. The warning
is rendered on EVERY resolve path (gh-1624 review): the FA_PROVIDER_*
boot notice, the provider queue (per entry at boot, on queue reload,
and in /provider queue test), and the --model/--base-url flag
revalidation — a rider never boots silently.
Breaking-change note for 1.0.576 upgraders behind proxy
factories: the migration is the wire: FA_PROVIDER_CONFIG key —
documented now (README env-preconfig section,
docs/agents/providers.md §Translating proxies, --help) with its
caveat: the override engages only when the declared wire DIFFERS from
the provider record's own kind (record anthropic + wire: anthropic
still runs the checks). REG matrix (AC1/AC2/AC3) pinned in
test/model_roles/provider_model_binding_test.dart; AC4 IT drives a
bifrost-shaped loopback proxy through the production anthropic adapter
(test/model_roles/bifrost_proxy_binding_test.dart).
gh-1596: the first-run wizard (site/setup.sh) no longer crashes or
corrupts input under curl | sh on dash (Ubuntu sh) or busybox ash
(Alpine): scripts/gen_installers.dart now renders the provider/mode/
approval menus from the live provider catalog into plain POSIX sh — the
bashisms that killed dash (printf -v, read -p) and silently corrupted
busybox ash are gone; the hidden API-key prompt uses stty -echo + a
by-name read (dash has no read -s) and the key never rides printf
argv; EOF/Ctrl-C mid-wizard aborts non-zero with no half-written
~/.fah, and non-TTY runs print guidance and exit 1. The wizard's temp
config is mktemp-created (no predictable, symlink-followable fixed
path). A new path-triggered installer-portability CI leg proves the
script under dash + busybox ash (ubuntu) and bash (macOS) on every
site/** or generator change.
fix(cli): gh-1572 — fa boot no longer rewrites ~/.fah/config.yaml
from a lossy config. persistConfig (bin/fah_runapp.dart) built a
fresh CliConfig carrying only some fields and persistConfig() ran
unconditionally at boot, so every launch silently dropped the user's
spills:, subagents:, jobs:, headless:, agent:, quota:,
tui:, trajectory:, waiting:, power.hold and auto_update
settings (first reported on 1.0.547). Three cooperating fixes: the
boot save is now saved.copyWith(<live overlays>) so every parsed
section rides along by construction; CliConfig.copyWith replaces the
drifted hand-rolled withCustomProviders field list (itself dropping
spills/subagents/waiting/jobs/agentMode/misuseBreaker/streamThinking
inside saveCliConfig's merge step); the _diskPreservedSections
disk backstop covers every known top-level key. spills: renders in
toYaml (new SpillsConfig.toYaml) and agent.misuseBreaker: false
persists too; an unchanged save no longer rewrites the file at all.
Rework round: the boot snapshot re-syncs from disk after every
sanctioned mid-session settings write (onUserConfigFileChanged — a
cleared context cap, an auto_update key drop or a global compaction
pick is no longer resurrected/overwritten by the next save in the
same session); validateAgentSection (the strict fa config check /
config set / settings-hub mirror) accepts the emitter's
agent.misuseBreaker and agent.stuckTool keys so fa config check
never false-fails on fa's own output; clearing the context cap drops
only the cap key instead of the whole agent: block (the sibling
mode:/misuseBreaker: keys survive); and the providersQueue:
block (no typed renderer) joins the disk backstop so a save can never
silently delete the user's queue. Rework round 2 (merge + guards):
validateAgentSection's per-key switch became the _agentKeyValidators
table (the two gh-1572 keys pushed the switch form to complexity 13 —
over the repo's CRAP ratchet; the table keeps the mirror complete at
complexity 4), and the last two direct USER-config writers —
_writeTtsrSection and _writeMcpSection — fire the re-sync hook, so
a mid-session ttsr rule toggle or mcp server edit survives the next
persist too (the writer roster in settings_flow_config_sync.dart is
now exhaustive).
gh-1606: the copilot provider accepts fine-grained PATs
(github_pat_…) directly — PAT mode skips the copilot_internal/v2/token
exchange entirely (GitHub 404s every PAT exchange) and sends the raw PAT
as the Bearer against the configured Copilot base URL with the official
Copilot CLI's first-party identity headers
(copilot-integration-id: copilot-developer-cli,
user-agent: copilot/<fa-version> (<os>) term/unknown, installed by the
host at startup). OAuth device-flow tokens keep the exchange +
proactive-refresh flow unchanged; a PAT 401/403 surfaces the error
instead of a doomed refresh retry. Headless COPILOT_GITHUB_TOKEN now
works with a PAT. Connect flow: the paste hint explains PAT mode, the
dead-PAT case surfaces at the model fetch, and /provider export copies
the saved Copilot account's full token to the clipboard (account picker
on several entries; the token never prints to the transcript).
gh-1575: macOS clipboard paste fixed for TIFF-only pasteboards + the
Ctrl+V text fallback. _readMacos probes «class TIFF» after PNGf/JPEG
miss and converts via sips -s format png (a Preview/browser copy used
to dead-end on clipboard holds no image; a failed conversion keeps the
unavailable fall-through and every scratch file is cleaned up); new
readPasteboardText (pbpaste / xclip+wl-paste / PowerShell
Get-Clipboard) behind the same injectable ProcessRunner/platform:
seam, and Ctrl+V with a TEXT clipboard now inserts the clipboard text
into the composer instead of printing the image probe's dead-end reason
— bracketed-paste routing, gesture pairing and silent-failure semantics
unchanged. Review pass: the Windows Get-Clipboard read pins UTF-8
console output and normalizes embedded \r\n (OEM-code-page mojibake),
the Linux text backends pin text/plain;charset=utf-8 (an image-only
clipboard can no longer leak bytes as "text"), the library doc no longer
overclaims the FA_FAKE_PASTEBOARD seam, and FaTuiCallbacks moved to
a fa_tui_callbacks.dart part file keeping fa_tui.dart under the
2800-line gate.
gh-1555 (W1, lazy/windowed-everywhere reads): the legacy full-open call
sites migrate to windowed opens, so a session ledger past the gh-1073
2 GiB full-open backstop stays openable on every user-facing read path.
The TUI session picker, /resume, and the boot name-resolution scan
open windowed and resolve the display name through paged reads; the
subagent transcript preview opens the child windowed (tail only); the
_loadSession marathon fallback no longer degenerates to a refused
full open when the boundary walk exhausts its page cap (it proceeds
with the tail-anchored windowed session; the boot cap owns over-window
residue); and the headless fa trajectory resolver opens windowed,
reading the whole ledger for the cost/inspect views through the new
SessionTree.ensureWholeLedgerResident — a sequential forward scan in
bounded strips (WindowedSessionStorage.loadAllOlder), not the
boundary-walk's doubling strips (whose per-strip record cap measured
~13x read/parse amplification over the whole file). The same change
block-prepends paged chunks in _indexChunk (one insertAll per
list), removing an O(n²) the suspended-eviction walk paid per record.
Trajectory export streaming (AC3a), split-and-retain rotation
activation (W2), and the opt-in write-side caps (W3) remain follow-up
work.
gh-1605: the FinalizeGate gains a "grader parity" named check. A
produced entry point (script/CLI the verifier will run itself) is only
verified when it survives the grader's own invocation: the checklist
must re-run each produced entry point from a different working
directory in a fresh subprocess with a clean env, capture stdout/
stderr/exit code, and diff against the earlier in-session run — any
divergence (different output, relative-path artifacts, locale/
encoding-dependent output) is a fail ledger item, never a pass.
Closes the modernize-scientific-stack__nwbgy7a near-miss (every
in-session checklist item passed, the verifier's clean-cwd subprocess
collapsed Station 102 to Station 101's mean, reward 0.0). REG coverage:
the parity procedure catches cwd- and env-dependent entry points and
does not false-positive cwd-independent ones
(test/agent/finalize_gate_grader_parity_test.dart), and an IT replays
the evidence shape end-to-end — divergence detected and persisted as a
fail ledger item before finalize.
gh-1603: unattended runs can no longer end as a clean success on empty
model responses. Loop: from the second empty-hiccup retry on, the request
carries a continuation nudge user message ("your last reply was
empty/truncated — continue the task"; exact-text constant
emptyContinuationNudge, emitted with MessageStart/End so the transcript
stays truthful) — a cap-truncated reasoning model re-streamed verbatim
repeats the empty outcome, the nudge breaks the symmetry; the classifier
treats that exact message as transparent, so the turn's partial-content
window (gh-1577) survives a nudged retry. Headless: the empty-reply nudge
budget rises to AgentCliConfig.headlessEmptyReplyNudges (default 2 — the
session ends only after 3 consecutive unrecoverable empties; interactive
keeps the #1085 M2b one-nudge semantics), and a run that still settles
empty prints the explicit __FA_INCOMPLETE marker and exits 1 — a
distinguishable failure for the bench adapter, never a silent exit 0.
gh-1577: an empty assistant completion after partial content (assistant
text, tool activity, or a truncation signal earlier in the turn) is a
bounded-retry provider hiccup — up to emptyHiccupRetries (default 3)
re-streams with exponential backoff (emptyHiccupBaseDelay,
emptyHiccupMaxBackoff, both threaded through the Agent facade,
AgentWiringSpec, and the CLI config), and exhaustion ends the run with
a loud StopReason.error (headless exit 1, interactive error + prompt)
instead of the silent exit-0 settle that killed the gh-1577 trials.
Clean empties keep the #1085 M2b one-nudge semantics. The abort always
wins: a cancel racing the final retry settles StopReason.aborted for
any empty shape (the exit-contract race is pinned by tests), and the
backoff sleep is cancellation-aware.
gh-1574: SQLite reads work on installed binaries — bundle libs land where
the binary looks. The AOT binary's native-assets resolver reads
<exe_dir>/../lib/<name> (the dart build cli bundle layout), but every
POSIX installer copied the bundle libs beside the binary
(<install_dir>/lib), so read path=x.db died with a raw FFI dump
(Couldn't resolve native function 'sqlite3_initialize' … Failed to load dynamic library '~/.local/lib/libsqlite3.dylib'); it only worked on machines
with a manual symlink in ~/.local/lib. Fixed the layout in all three
install paths — site/install.sh (incl. the quarantine-xattr pass and the
source-build advice), install_local.sh, and fa update
(bin/self_manage.dart copyBundleLibs, now public and layout-pinned by
tests; also aligns Windows fa update with install.ps1, which was
already correct) — plus clean degradation in the FFI engine:
Sqlite3Engine.openReadOnly classifies the dart:ffi resolver's
library-load failure (isSqliteDynamicLibraryLoadFailure) and throws a
clean SQLite is unavailable — reinstall or run fa update note instead of
the raw resolver dump; a Dart-level system-library fallback is impossible
with package:sqlite3 native assets (build-time asset mapping, no loader
override). Installs already broken by the old layout self-heal: on the
bundled library's first failed load the binary moves the stale
<exe_dir>/libsqlite3.* into <exe_dir>/../lib/ and retries the load
(repairStaleSqliteBundleLayout) — so one fa update suffices (the
successor binary repairs on its first SQLite read; a pre-fix binary cannot
repair itself), and the stale <bin>/lib leftovers disappear as they
migrate. The POSIX installers also report a failed lib copy instead of
swallowing it (site/install.sh warns naming the lib dir; same in
install_local.sh), and site/install.sh.template carries the same
layout so regenerating the installers cannot resurrect the old one.
gh-1556: bench pipeline stops bleeding score + fa deadline awareness.
Bench: the harbor ladder kill uses the [f]a bracket trick and is
wrapped so a kill failure can never replace the TimeoutError (the
pre-fix self-matching pkill classified TB 2.0 ladder kills as agent
crashes — a SOLVED task scored 0); progress-watch (stall-gap 360s per
the gh-1430 watchdog invariant, flat abs ceiling 3600s) is the
bench-harbor.yml default on every shard, with the harbor adapter
picking the gap-aware ProgressWatch for parity with the
terminal-bench adapter; the aggregate counts completeness by
result.json presence (errored trials report as their own breakdown
line; exit 1 only on genuinely lost trials) and renders the full
report — table, errored breakdown, ledger row — to
$GITHUB_STEP_SUMMARY (TB1 parity). Both bench adapters export
FA_DEADLINE_EPOCH_MS (effective cap − 120s wrap-up margin) into the
agent session and audit the exact value. fa core: deadline awareness —
FA_DEADLINE_EPOCH_MS / --timer/--deadline / /timer write one
session clock, the per-turn context carries a time-left: line while
armed, and one-shot system notices steer wrap-up at ~50%, T-3min
(drain background jobs), and T-30s; absent deadline = byte-for-byte
today's behavior.
gh-1522: tag-sourced versions — the git tag is the single source of
truth. The release-bot file bumps are retired: cutting a release = a
tag + GitHub Release (zero commits to main, zero file mutations).
Both pubspecs carry the fixed 0.0.0-dev placeholder (CI guards
hand-bumps), the publish job stamps the staged pubspec + generated
changelog section from the tag (scripts/stamp_staged_release.sh,
staged version == tag guarded pre-upload), store builds stamp
--build-name/--build-number from the derived tag version
(pubspec.lock stops churning across releases), and
install_local.sh stamps version.txt from git describe.
feat(usage): gh-1460 — the fa-tokens: segment-close line carries the
session's model id, so every token row the factory publishes is
attributable to a model and therefore priceable. The fold now captures
each segment's LAST-SEEN model id (the model that served the final
request, the same string settings/UI display — mid-session fallback
take-overs attribute the segment to the model that actually finished it;
zero-filled fallbacks never clobber a seen model) and the pinned line
shape grows an OPTIONAL "model":"<id>" key after "segment" — omitted
when no model was observed, so legacy lines parse unchanged and the
pinned regex (usageTokensLogPattern, gh-1241 AC7) keeps matching both
shapes. usage.json's schema is untouched (the model is rebuildable from
the chain, I6); old consumers using tolerant JSON ignore the new key,
and the repo's single strict consumer (the pinned constant) is updated
in the same change.
fix(test): gh-1460 rework — the wrap_ansi_line_perf_test growth-shape
legs measure the doubling ratio with the MINIMUM over 9 runs instead of
the median of 5 (CI run 38112524699, core shard 0/4: a scheduling pause
inflated the doubled-size median — 12.09ms against a true ~7ms — and
tripped the 3.2 bound at 3.65x on a genuinely linear wrap). Runner
jitter only ever inflates a wall-clock sample, so the min converges on
the true compute time; genuine superlinearity still trips the ratio
(the gh-1496 quadratic candidates measure ~4x by min). The absolute-
budget legs keep the median — their gh-1357 headroom already prices in
jitter. A synthetic regression test pins the pause-proof property.
<!-- release-draft-owner: run/42496495862/job/build-macos-cli-arm64 -->