Fixes since the September 16 entries

• Shortcuts sidebar no longer breaks other views (GH #209). With the sidebar open (;), only the list view made room for it. Board, graph, insights, actionable, history, tree, label dashboard, attention, flow matrix and the sprint dashboard were drawn at full width with the sidebar added on, so rows wrapped into each other and the footer was pushed off screen. Every view is now laid out in the space left beside the sidebar, overlays such as help and the pickers no longer get the sidebar next to them, and on a terminal too narrow for it the sidebar is hidden and the status line says why (fix, row tests, live-reload speed).
• Dashboard time travel works again after merges. --export-pages could skip history.json with the warning "timeline commit … missing from source history" once a branch that edited the tracker had been merged, so the exported dashboard had no time travel. The same happened when the tracker had been renamed (for example beads.jsonl to issues.jsonl) and an older branch that edited the old name was merged later. The timeline now orders its commits by their full Git ancestry (merged branches, renamed tracker).
• The right tracker is loaded. When BEADS_DB, BEADS_DIR or --db names a tracker directory, a newer export in an unrelated Git worktree can no longer win. A SQLite tracker whose recent writes are still in its WAL file is treated as fresh instead of losing to an older JSONL export, and when two files are equally fresh issues.jsonl is preferred over legacy names. The cached validation result is dropped whenever the source actually changes (worktrees, WAL freshness, name ties, cache identity).
• Stable history and correlation output. Several beads changed in one commit are reported in the same order on every run, commit references keep the full SHA, a commit seen through several bead histories is counted once, and a deleted bead's history now ends with its deletion (ordering, commit identity, deletion).
• Searching for an issue ID finds it. Typing an exact issue ID now always shows that issue, even with hybrid ranking and more than 75 matches (search).
• Agent-file reads are bounded. Inspecting or updating AGENTS.md refuses files over 16 MiB, rejects a file that changes size while being read, and no longer follows a symlink to an unrelated file (agent files).
• Source installs report the real compiler. install.sh and install.ps1 now print the Go version the build actually used (the module's pinned toolchain), not the version of the go command that launched it (installer).
• Vendored dashboard libraries, the graph WASM glue and frozen test fixtures are byte-identical to their recorded hashes again; a formatting pass had rewritten them (assets, fixtures).
• The exported dashboard of this repository now ships its database twice (as one file and as 1 MiB chunks) because the tracker passed the 5 MiB chunking threshold, so the recorded bundle size in tests/artifacts/perf/pages_load.json was re-measured (9.8 MB to 17.5 MB). The export code is unchanged; v0.25.0 produces the same bundle for the same tracker.

Dashboard graph and timeline

• The force graph now validates exported node positions and blocking edges against the loaded database, then paints matching positions without synchronous layout warmup. Completed PageRank and exact betweenness results can be reused; missing, invalid, or sampled results fall back to browser computation. Live physics and other browser algorithms remain enabled. Detail-pane resizing, cyclic fallback layouts, and graph-state resets are also repaired (layout integration, rendering and metric reuse).
• Timeline playback follows recorded Git lifecycle events, distinguishes removal from closure, preserves the observed baseline at the retained history boundary, and supports reverse scrubbing. Nodes fade, grow, pulse, and shrink; reduced motion skips transitions. Sprint buttons use current definitions. Playback resets on graph replacement and avoids repeated synchronous layout warmup (recorded events, removal, canvas animation). History remains bounded to 500 commits and the current export's issue universe.
• Real Chromium desktop/mobile-viewport journeys verify controls, drawing, fallbacks, detail interaction, and offline updates. A 1,000-node, 21-commit fixture measured playback p99 at 33.3 ms in both viewports. This is not a physical-phone result or the separate CLI/TUI performance qualification. Original workstreams bv-643f and bv-z38b are complete.

Tree visibility, terminal integration, and history

• Tree view keeps rootless parent-child cycles visible and preserves repeated cycle/diamond occurrences when replacing analysis snapshots (display roots, snapshot copying).
• Clipboard copy can fall back to OSC 52 over SSH when local helpers are absent or fail. tmux receives the raw sequence; GNU screen retains its passthrough wrapping (clipboard repair).
• Windows automatic theme selection can query Windows Terminal's background; BV_NO_BG_QUERY disables the query. Parser and cross-build checks pass, but actual Windows Terminal console I/O remains unverified (background query).
• Filtered history requests can reuse fresh full-history event caches when both Git blob identities match. Rebuildable event caches no longer require a durability flush; corruption still triggers reconstruction. Real Git regression tests cover freshness, identity, filtering, and interrupted writes (cache and documentation checks).
• Public guidance now describes actual search, history, routing, and forecasting behavior. Copied recipe and robot queries execute against real fixtures, including lifecycle duration units. Overall documentation qualification still depends on the open performance and native-platform work (guidance corrections).
• --robot-docs no longer tells agents that TOON "saves ~30-50% tokens" and no longer offers bv robot-triage --toon as the token-saving example. The measured artifact records TOON at 0.93x JSON for --robot-graph and 1.09x to 1.15x for the nested payloads, and the guidance now says so and points at --stats. When no tru encoder is discoverable, --format toon also stops declaring output_format: "toon" over JSON bytes; the envelope now reports json, matching the fallback warning already printed on stderr. The --robot-schema entry for Cores and Slack is corrected from object to array, which is what --robot-insights emits.
• A full README and root SKILL.md audit against the current code corrected the fuzzy-filter field set, the HTML export's search fields and its 2-character / 8-result limits, the board age band at exactly 30 days, the untyped dependency as a blocking type, the informational status of related and discovered-from, the correlation feedback stats fields and the unwritable ignore decision, the Cass session modal mock and its relative timestamps, the eight-hour workday used by --robot-capacity and --robot-forecast summary days, the hybrid-only candidate widening, the impossible "ms": 0 on a skipped metric, the --robot-diff cycle field names, the --robot-recipes example descriptions, and SKILL.md's claim that .status is top-level on --robot-triage. The in-app tutorial no longer advertises fuzzy search as typo-tolerant or the hashed-keyword embedder as meaning-based.

Duplicate detection and rendering

• Duplicate detection now retains only the requested best candidates for unique issue IDs instead of storing and sorting every matching pair. It preserves score/ID ordering and source-specific keyword explanations; repeated IDs retain the existing full-sort behavior. Earlier changes defer explanations until selection, compact candidate storage, and reuse overlap counts. In the dense 500-issue benchmark requesting 20 suggestions, the final bounded-selection step reduced median allocated bytes from 42.8 MB to 2.18 MB and median time from 232 ms to 31 ms across five runs on the same Linux worker with Go 1.26.8 and GOMAXPROCS=4. Overlap enumeration can still be quadratic (bounded selection; source-pair regression).
• Reusable theme styles precompute ANSI palette conversions while retaining live light/dark and terminal-profile selection. Exact rendering comparisons cover all four profiles; out-of-range numeric colors preserve their original lazy behavior. The focused ANSI256 label benchmark uses nine allocations instead of ten; it does not establish an application-wide latency gain (palette conversion).
• Panel borders now reuse that palette conversion instead of repeating it for each border character. Exact comparisons preserve output across profiles, light/dark settings, and narrow layouts. In five runs rendering two 65×40 panels with ANSI256 dark colors, median time fell from 0.780 ms to 0.443 ms and allocations from 1,658 to 1,494 per iteration. This measures panel rendering, not complete UI latency (panel borders).
• A priority-confidence test now uses the calculator's fixed clock when selecting its fixture, preventing calendar-dependent staleness from invalidating the test. Its confidence thresholds and assertions are unchanged (fixture clock).

These changes advance bv-apal.1; its full performance qualification remains open. All 36 current-only UI cohorts passed the unchanged 50 ms limits on hz4, with 36,000 samples and a worst p99 of 34.47 ms. This does not replace the original paired UI and CLI matrix.

Nix usage and verification

• The README now explains that another flake's nixpkgs configuration does not configure bv's own import. Consumers of its exported package must keep the documented explicit unfree allowance (corrected instructions).
• Native Linux amd64 Nix build and executable checks were completed against the v0.25.0 source. The subsequent RCH source-verification timeout and the remaining native-platform gaps are recorded separately in release verification; this does not close bv-oonu.9.

Latest

Latest

Compare

Published from 87cee258 after all ten release-gate stages passed through RCH with no skips. Five platform archives, checksums, the sealed gate receipt and a packaged-binary SBOM were uploaded with DSR; all 14 assets were downloaded and verified before publication. Homebrew and Scoop manifests were updated. Native Linux amd64, macOS arm64 and Windows amd64 installer/update checks passed, as did the packaged binary's desktop/mobile/offline browser journeys. Native source-build acceptance and the incomplete P1 performance matrix remain open.

Priority recommendation performance

• Cascade simulations now reuse sorted dependency and hierarchy adjacency instead of rebuilding it at every step. Completion state and readiness checks remain specific to each query (frontier reuse).
• Enhanced priority recommendations, used by --robot-priority, now reuse one completed analysis snapshot across the batch instead of rereading the analysis cache for every issue. Scoring, explanation fields, ordering, and output caps are preserved (snapshot reuse; full-output regression).

Performance verification and documentation

• Failed CLI measurements now report sample identity, elapsed command time, and cancellation state captured before diagnostic files are written. Slow artifact writes cannot retroactively make a command failure look like a timeout (diagnostics). The original incomplete P1 matrix remains open.
• The September 11 source measurement passed all 144 current UI cohorts at the original 50 ms p99 and delivered-handler limits; the worst cohort p99 was 46.472 ms. The overall matrix failed when a baseline process was killed in the last timed CLI cohort, leaving 70 of 72 records complete. The README and performance guide distinguish those outcomes and keep the P1 work open. This adds measurement evidence, not a runtime optimization or release qualification.
• Documentation now distinguishes settled navigation from startup and background preparation, and makes dependent readiness conditional on remaining prerequisites and eligibility. The v0.22.0 entry names the 50 ms interaction target while preserving its original measurements.

Workflow data and static dashboards

• SQLite dashboard rebuilds construct a private database before publishing it. A failed construction preserves the last good database, and readers no longer lock a partially populated export. On Windows, an open reader may still prevent replacement; that failure preserves the prior database.
• FTS indexing, materialized-view creation and export metadata now each use one transaction. This reduces repeated disk syncs and rolls back the related writes if a later statement fails, without disabling SQLite durability.
• Source discovery now honors issues.jsonl, then beads.jsonl, then beads.base.jsonl before comparing the selected export with SQLite and worktree sources. A newer sync snapshot cannot replace current issue state, and an empty export stays empty. Sidecar-only directories no longer bypass the filename allowlist through the fallback loader; explicit file overrides remain available. Human commands now report ignored left/right merge artifacts while robot stderr stays clean (bv-mvvu, bv-uoyj.1).
• SQLite live refresh now detects committed WAL updates in event and polling modes. TUI and watched Pages exports update while the writer remains open; checkpoint removal of the WAL is handled without reporting the database itself as removed (bv-oonu.21; WAL refresh).
• Live TUI refreshes and single-repository watched exports now use the source that successfully loaded at startup. A corrupt newer file can no longer redirect the watcher away from its valid fallback. Explicit JSONL and SQLite selection remains supported. Historical --as-of exports reject --watch-export before writing files (bv-oonu.20; source repair).
• SQLite and JSONL loading now retain nonblank custom workflow statuses and relationship types. conditional-blocks and waits-for affect blocking analysis; custom statuses do not automatically become claimable (loader change).
• Recipe validation now rejects a blank filters.status entry with a message that says so, instead of calling it an unknown status. Since the loader change above the check has accepted custom workflow states such as done; only the error text and the Validate documentation still described a closed vocabulary.
• Static dashboards now include both blocking variants and legacy untyped dependencies in their relationship lists and browser graphs. Active ID lists agree with blocker counts, including after either endpoint closes, so blocked work no longer appears ready because its relationship type was omitted.
• In issue details, h now navigates to a prerequisite and l to a dependent, with stable ordering. The export and real Chromium regressions are tracked by bv-oonu.12 (dashboard repair).
• “Simulate Close” now follows dependencies in the correct direction and shows the engine's actual direct and downstream counts. Priority picks and cascade cards show their calculated gains. Deleted prerequisites and closed rows excluded from the dashboard remain resolved in both browser graph engines.
• Graph simulation and reset use the bundled renderer's redraw API; cancelling, reloading or cleaning up a simulation cancels its pending animation callbacks. The rebuilt graph WASM preserves unrelated graph metrics. Real Chromium regressions and source-rebuild evidence are tracked in bv-oonu.13 (simulation repair).
• HITS hub and authority panels now display the bundled engine's scores and open their ranked issues. Their JavaScript consumer previously read field names that the engine does not return (bv-oonu.14).
• Dashboard ready counts, quick wins and Ready/Blocked filters now use the full-source readiness snapshot. Missing or filtered prerequisites, inherited parent gates and deferral no longer disappear from eligibility checks; resolved prerequisites still satisfy them when omitted from the display. Ready cards include eligible in-progress work. Active-node counts include each unresolved issue once and remain numeric when a status group is absent (bv-oonu.15; rankings and readiness repair).
• Direct dashboard exports now apply recipes, including sorted max_items selection. Watched exports reapply repository, label and recipe filters to each reload instead of expanding to the full dataset. Hidden prerequisites still govern readiness, newly matching issues enter the selection, and an empty selection clears old rows (bv-oonu.16; export scope repair).
• Missing or filtered dependency endpoints no longer crowd real issues out of WASM-backed ranking panels. HITS, k-core, slack and metric fallback lists apply their limits to exported issue rows while preserving full-graph scores (bv-oonu.17; ranking limit repair).
• Interactive HTML and static SVG/PNG graph exports now honor recipes, including custom files, sorted limits and label/repository intersections. Actionable selection retains full-source prerequisite checks; an empty recipe selection reports an error without creating a graph (bv-oonu.18; graph recipe repair).
• Dashboard Priority Picks now restrict candidates before optimizing gains. Missing issue IDs cannot be selected or implicitly completed to inflate another issue's gain. Cascade suggestions and actionable lookup use the exported full-source readiness snapshot. The rebuilt WASM preserves results for callers without candidate restrictions (bv-oonu.19; candidate selection repair).
• Label-scoped insights rank hypothetical completions only for selected issues; neighboring context can no longer take a result slot. Top-k candidate counts use the same selection, while graph metrics and unresolved prerequisites retain their context (bv-xbvo.11; scoped ranking repair).
• Capacity reports now use full-source readiness and intersect global selection with --capacity-label. Missing prerequisites, inherited parent gates and parked/deferred work no longer produce false actionable counts. Direct bottlenecks exclude non-blocking relationships, resolved endpoints and duplicate pairs; result ordering is stable. The duration calculation remains a heuristic (bv-xbvo.12; capacity repair).
• Capacity path calculation now shares suffix results on acyclic graphs, avoiding exponential enumeration of overlapping paths. It preserves the chosen path, ties and estimates; reachable cycles retain the original exhaustive search. A dense 26-issue fixture fell from a 3.55-second median to 64 ms across ten measured runs per binary (bv-xbvo.14; capacity performance repair).
• Forecasts now honor global selection intersected with the forecast label and sprint. A single requested issue must also pass these filters; an excluded ID returns an error. Selected estimates retain their loaded dependency and closure context (bv-xbvo.13; forecast scope repair).

Build requirements

• Source builds now require Go 1.26 or newer, with Go 1.26.8 selected by go.mod. Both source installers enforce the same minimum. The Nix flake moves to the 26.05 package set, which supplies Go 1.26.7.
• Terminal rendering, Unicode normalization, HTML parsing and graph-image dependencies are updated. The four local dependency patches remain in place; upgrade notes record the individual transitions and tests.
• The SQLite module moves from 1.52.0 to 1.58.0, updating the embedded engine from SQLite 3.53.2 to 3.53.4 with the module's required libc version.
• The bundled graph WASM and JavaScript glue are rebuilt with wasm-bindgen 0.2.128 and refreshed Rust dependencies. Independent builds from two compiler locations produce identical assets; the graph fixtures, viewer HITS adapter and corruption controls pass in Node WebAssembly execution.

Vendored dependency patches

• The four locally patched dependencies (chroma, glamour, reflow, go-json; see docs/PROVENANCE.md) now live under third_party/ as complete Go modules reached through replace directives in go.mod, so go mod vendor copies the patched sources instead of silently reverting them. Previously the patches existed only as hand edits inside vendor/, and a routine go mod vendor reverted all four, including the go-json decoder-cache race repair that no -race test exercises. A new e2e check fails when vendor/ and third_party/ disagree, and docs/RELEASING.md documents editing, upgrading and retiring a patch. The vendored package sources remain byte-identical to the previously committed patched files.

Dependency inspection and documentation

• Cass session lookups now run in the background, keeping navigation and resizing available while Cass responds. V or Esc cancels a pending lookup; results from an older selection or dataset cannot open a stale modal. Refreshed data receives a fresh correlation cache. V also works from the detail pane, and closing the modal restores its originating view (bv-xiyd; responsiveness repair).
• Cass session lookup now reads the actual hits response and requests the title, preview, workspace and millisecond timestamp fields used by the UI. Pressing V can show sessions from a searchable stale or rebuilding index while retaining its health warning. Cached sessions keep their computed scores and match reasons; failed or partial lookups remain retryable instead of becoming cached empty results (bv-8phk; search adapter, cache repair and live verification).
• README examples now place impact-network fields under network, show priority reasoning as an array and use the actual alert fields. Flow Matrix rows are documented as blockers and columns as dependents; saved-baseline drift checks are distinguished from Git-revision comparisons. Cass scoring and modal controls now describe the implemented behavior (documentation corrections).
• Generated agent instructions now use br update <id> --claim --json, matching the README and assigning the current actor as work starts. Instruction version 6 makes existing version 5 blocks eligible for refresh through --agents-add; surrounding user instructions remain intact.
• Flow Matrix drilldowns now show the actual blocker and dependent for each relationship, exclude unrelated issues sharing a label, and deduplicate pairs spanning multiple labels. Enter inspects either endpoint without changing the active recipe or selected work; Escape returns to the relationship.
• Open relationships and endpoint details update after snapshots and file reloads, retain surviving selection by ID, and remove obsolete relationships. Unicode and long IDs fit narrow drilldown rows. The implementation and Linux terminal journey are tracked by bv-apal.12/bv-apal.13 (implementation and regression tests).
• README now matches the recipe-picker key, cass health/count indicators and separate history modal, plan/history JSON fields and duration units, dependency direction, and binary versus source requirements. BV_INSIGHTS_MAP_LIMIT documentation now gives its existing default of 200; zero and invalid values use that default. A real CLI regression checks all five documented limit cases. These corrections are part of the still-open bv-apal.3/bv-apal.4 workstream; they do not establish the remaining performance or native-platform claims.

Windows installation

• Version-check failures now retain stdout and stderr, capped at 4,096 characters per stream. The original ten-second execution deadline remains; pipe draining is bounded to one additional second, including when a child inherits the pipes. Truncated version output is rejected before installation (bv-oonu.9; diagnostic repair). Both README install commands now pin this reviewed script. Portable process regressions and native PowerShell 5.1 version/mismatch checks pass; the original native source first-start failure remains open.
• Suppress download progress locally inside Install-FromRelease, avoiding the redirected-download stalls observed with Windows PowerShell 5.1 without changing the caller's preference. The native harness now checks the existing diagnostic_top_pick and metadata-free claim refusal instead of an obsolete top-level ID (3ca2176f). This fixes installer behavior and its test; it does not change robot output.
• README's Windows commands initially pinned that installer revision (87756815). The pinned source option uses a verified tagged checkout and vendored dependencies; the stale reference to its older go install path is corrected. The installer fix follows v0.24.1's immutable tag and is already usable with its published binaries.

Verification and remaining limits

• With installer 3ca2176f, the complete default native Windows suite passed against the public v0.24.0 and v0.24.1 archives: installation, readiness, update/no-update, and preservation of the installed executable and user PATH on failure. Readback covers 28 command logs, eight capability results, and five specific rejection cases; it does not count a transport timeout as a successful rejection.
• An initial native Windows source installation passed, but a later optional source run exceeded the unchanged 10-second first-start guard. Its retained executable succeeded on a second diagnostic invocation; the original failure remains unresolved. Native macOS amd64/arm64 and Linux arm64 execution, native Homebrew installation, and a supported Nix build remain unverified. The broader installation workstream bv-oonu.10 stays open. See the retained release findings.

Latest

Latest

Compare

This patch removes avoidable hashing and cache-lock waiting from robot commands while preserving their data-hash scope and computed analysis results.

Fixed

• Optional analysis-cache publication no longer waits for another process's writer lock. A contended writer leaves the existing cache entry intact and returns the computed result; a later request can publish successfully (f719c41a). Linux and native Windows regression tests exercise the real lock, preservation of the existing entry, and a successful later publication.

Performance

• Robot commands reuse the hash computed while loading an unchanged single source, including historical snapshots, avoiding a second fingerprint pass. Workspace aggregates, tombstone-bearing sources, and --repo filtering still trigger recomputation. The envelope hash remains scoped after --repo and before --label or --recipe; 15 scenarios exercise both robot plan and triage (6c8a4474).

Distribution

• Published through DSR using locally packaged GoReleaser artifacts, without GitHub Actions or repository dispatch. Linux amd64/arm64, macOS amd64/arm64, and Windows amd64 archives identify the clean tagged revision 3e4e61c9, Go 1.25.5, and disabled CGO. The 14 assets include checksums, a sealed gate receipt, and an SPDX SBOM for the Linux amd64 binary. No minisign signature was produced.
• Homebrew and Scoop were advanced from v0.22.0 to v0.24.1 using the verified archive hashes. Go module publication and the tagged Nix flake were checked against the same source revision. These publication checks do not establish native Homebrew installation or a supported Nix build.

Verification

• All ten release-gate stages passed on the clean tagged commit; no gate stage was skipped. All 14 uploaded assets were downloaded and matched by size and SHA-256 before publication. Native Linux installation, readiness, update/no-update, and failed-install preservation checks passed. Windows installer results and remaining native-platform limits are recorded under Unreleased because the installer repair followed the tag.
• The separate pre-release performance campaign retained 288 UI records, 72 timed CLI records and 36 fixed-clock comparisons. Its original exact-output check failed because the baseline reports v0.23.0 and the candidate v0.24.0; readback of all 144 outputs found only those 72 version-field differences. This accounts for the failure without changing the original result or claiming the broader performance campaign complete. The latency workstream bv-apal.1 remains in progress; the distribution workstream bv-l76l is complete. Release verification details preserve the original failures and execution limits.

Performance (latency campaign bv-apal.1)

• Analysis: canonical graph hashing streams through successors instead of boxing and sorting every edge, so allocations scale with nodes (fbc41526); ComputeDataHash / ComputeIssueDiff reuse one fingerprint writer per call, 1,289 -> 777 allocations per 256-issue aggregate (c48bd53c); the ready set is reused across parallel-gain candidates and marginal unlock IDs are selected without copying issues (9a72543d, 187f7525); blocker IDs sort without reflection (347134f1); the regenerable analysis cache no longer forces durable flushes (4b018e1b); the drift Calculator gains ReuseAnalyzer to share precomputed readiness and topology across runs (1713989c); Phase 2 analysis is size-tiered with a deterministic betweenness approximation that reduces in sample order (341a0005).
• TUI: the visible critical chain is computed once, at snapshot construction, instead of on the event loop (735f246a, 30417526); full list rows are initialised in place and incremental row copies are avoided (934da756, e3dcc7ce); owned readiness data is compacted (84cf799b); background Phase 2 preparation no longer mutates the active model (1c5d55f9); snapshot diffing and markdown element layout allocate less (f2bad78d).
• Loader / search: one default 10 MiB JSONL reader is retained between parses while caller-owned buffers keep their semantics (7f708334); streaming readers and concurrency pooling in internal/datasource and pkg/loader (3b4df94d); metrics vector-search caching layers (19323918).
• Vendored renderers: strings.Builder replaces quadratic concatenation in chroma coalesce.go and glamour ansi/elements.go, and reflow padding.go uses runewidth.RuneWidth (5e16bff3). These are hand patches under vendor/; a bare go mod vendor reverts them.

Fixed

• Graph navigation and historical causality restored: pan, scroll and expansion follow bounded visible dependency paths with a deterministic highlighted chain; historical transition and dependency authority survive extraction and caching, observed waits carry explicit uncertainty, and causal history is bound to the selected revision and source path (93b90959).
• Decoder cache synchronised and pseudo-versions filtered (40644bd9); XFetch cache refresh keyed on the actual expiry (8285b6f6).
• SQLite export creates its schema atomically (a4f8245b).
• Robot parallel-gain output omits elapsed timing when the clock is pinned, keeping envelopes reproducible (5814bed8).
• Graph WASM build remaps compiler source paths (8cd63299); the frozen 1,000-issue benchmark input is tracked so a clean checkout can run the release gate (09fed4e9).

Changed

• Robot registry outputs, defer_until flags and CLI fixtures harmonised (54db481b); the interactive HTML graph export embeds the robot envelope (19323918); readiness scopes modelled with projection and authority boundaries, richer triage recommendations, cycle detection and ETA prediction (341a0005); graph-analysis invariants, vector indexing and SQLite export schema validation hardened (3db9dd04); ingestion pipelines, defer_until parsing and workspace path resolution hardened (3b4df94d).
• Release engineering: goreleaser dist output isolated to /tmp/bv-dist with CGO_ENABLED=0 and GOWORK=off (a90029b8, 0aca294f); install.sh / install.ps1 handle native packaging with checksum verification and isolated release verification; scripts/release_gate.sh writes logs and receipts outside the checkout; the bv-graph-wasm Rust toolchain is pinned; installer, gate, WASM and dashboard smoke tests live under tests/scripts (0aca294f); smoke and installer artifacts are retained as evidence (8a6386e9).
• Tests: end-to-end coverage for robot scoping, search relevance, recipe execution, export flows and board/swimlane interaction, plus benchmark workloads isolated from tree state and a frozen real correlation history (087a847f, 1a400830, 257427df, f9d950a0).

Dependencies

• golang.org/x/text v0.38.0 -> v0.41.0 (fixes GO-2026-5970, reachable through norm.Form.Properties), golang.org/x/image v0.42.0 -> v0.45.0, golang.org/x/net -> v0.58.0; vendor/ regenerated with the three hand-patched files above preserved (1e8acace, #200).

Housekeeping

• Version metadata bumped to v0.24.0 (flake.nix, pkg/version fallback, README install examples, installer test defaults).

Published release: GitHub Release v0.21.2. This release publishes the profile-driven work staged under the tag-only v0.21.1, plus the final Nix packaging correction. The campaign completed all 50 requested iterations, but credits only improvements that survived command-level measurement, output-equivalence checks, focused causal tests, and the full Go verification suite.

Delivered capability: Faster Cold Robot Triage

• Profile the real bottleneck first. On the pinned 540-issue / 19-open Git fixture, history correlation accounted for 62.26% of command CPU and snapshot extraction for 58.49%. Within that path, profiles attributed roughly 100 ms to large allocations, 70 ms to clearing memory, and 70 ms to background garbage collection. This ruled out speculative graph and JSON tuning as the first lever.
• Recycle a blob only after its last reader is gone. Snapshot record sets contain slices that point directly into their Git blob buffer, so reusing a live buffer would silently corrupt history. The extractor already knew each blob's last use. Commit 22305d12 made that lifetime explicit: evict the record set first, then place its backing buffer in a one-slot, largest-capacity spare owned by the streaming git cat-file --batch reader. The next blob re-slices that storage when it fits instead of allocating and zeroing another large byte array. Only one spare is retained, so reuse stays bounded rather than becoming a memory cache.
• Preserve the algorithm, remove allocation churn. Git object order, the one-response-at-a-time protocol, line identity, event ordering, and the existing two-to-three-snapshot live window from #182 did not change. Four-snapshot lifetime tests prove that two live blobs never alias and that only an explicitly recycled buffer is reused. A real-history differential reproduced all 1,776 legacy events, and 20 independently normalized robot outputs were identical at SHA-256 5992ff99901b1c5abf8ccf3b1a9e3d2490a6f0eda1742cad938e7b3ff9809918.
• Measured result: ten independent low-load, interleaved cold-cache pairs reduced mean user CPU from 0.647 s to 0.573 s (11.44%), with 10/10 wins and exact sign p=0.00098. Five traced pairs reduced mean GC cycles from 25.8 to 13.0 (49.61%). Wall time improved 2.26% but missed significance (p=0.0547), while peak RSS increased 0.177%; therefore this release claims lower CPU and GC work, not lower wall latency or memory usage.

Delivered capability: Reused Graph Analysis

• What-if batches: 4b685960 added a stats-consuming path so --robot-insights can pass the completed GraphStats it already owns. Previously a planted 540-issue batch performed 540 redundant analysis-cache decodes through TopWhatIfDeltas -> computeWhatIfDelta -> Analyze. The focused same-worker benchmark moved from 4.4108 s / 1.648 GB / 8,530,586 allocations to 3.203–16.564 ms / 1.444–4.931 MB / 8,504–26,341 allocations. This is a scoped batch benchmark, not a whole-command latency claim.
• Cycle-break insights: 5b73a6b6 routes the handler's completed statistics into advanced-insight generation instead of launching another analysis merely to recover the same cycle list. Sentinel-cycle tests prove the supplied statistics control the result, while nil callers retain the ordinary analyze-on-demand fallback.
• Direct dependency decoding: a83bf01c reduced a direct-loader benchmark by 13.00% and allocations by 33.63% while replaying malformed and invalid-UTF-8 inputs through the standard library for exact behavior. It did not improve the dominant cold-correlation command path, so it is not included in the 11.44% release claim.

Closed workstreams: Measurement and Negative Evidence

• Fixed registry-backed CPU-profile shutdown in 854a8070 and 0a0838ec, ensuring profiling completes before robot dispatch exits instead of leaving zero-byte or truncated profiles.
• Ran 50 bounded optimization passes. PGO, GOAMD64=v3, fixed GC pacing, smaller transport buffers, direct event fusion, prefetch windows, alternate record tables, and other promising candidates were rejected when end-to-end confidence intervals crossed zero, system CPU or tails regressed, memory worsened, or the active profile placed too little cost in the proposed seam. These results and retry conditions remain in the pinned hotspot table and negative-evidence ledger.
• Retained snapshot frontier/allocation refinements have differential coverage, but no additional release-wide percentage is assigned to them. The only accepted cold-triage result from this campaign is buffer reuse's 11.44% user-CPU and 49.61% GC-cycle reduction.

Representative commits

• 22305d12 — recycle evicted Git blob buffers on the measured cold-correlation path.
• 4b685960 — consume completed graph statistics across what-if batches.
• 5b73a6b6 — reuse completed cycle data in advanced insights.
• 854a8070 and 0a0838ec — make robot CPU profiles complete and trustworthy.

Fixed and Packaged

• Document the explicit Nix unfree-package opt-in required by the project's OpenAI/Anthropic license rider, keeping the published flake instructions consistent with its corrected nonfree metadata.
• Publish five platform archives with individual SHA-256 sidecars, aggregate checksums, an SPDX SBOM, and the DSR build manifest. Publicly downloaded macOS and Windows binaries reported bv v0.21.2; the Go module proxy resolved v0.21.2 to the tagged source commit.

Changed — BREAKING (robot JSON semantics)

• Strict count semantics in triage output (#165). quick_ref.open_count / project_health.counts.open now count ONLY issues whose status is exactly open (previously: every non-closed issue, i.e. open+in_progress+blocked+deferred), and quick_ref.blocked_count / counts.blocked now count ONLY issues whose status is exactly blocked (previously: every non-closed, non-actionable issue). Both now always equal the corresponding counts.by_status entries. The legacy aggregates are preserved under semantically accurate names: not_closed_count / counts.not_closed (old open_count) and not_actionable_count / counts.dependency_blocked (old blocked_count), with the partition invariant not_closed == actionable + not_actionable. Consumers that relied on counts.open meaning "non-closed" must switch to not_closed; consumers that relied on counts.blocked meaning "dependency-blocked" must switch to dependency_blocked / not_actionable_count. This also aligns the triage counts with the bundled viewer's strict status tiles.

Changed

• Better .gitignore handling (#179, revisiting #34/#151). bv no longer unconditionally appends .bv/ to the repo's committed .gitignore. It now prefers .git/info/exclude (no repo litter, invisible to collaborators, shared across linked worktrees — worktree .git pointer files and their commondir indirection are resolved), skips writing entirely when the project is not a git repository or when .bv is already covered by the repo .gitignore, .git/info/exclude, or the user's global gitignore (core.excludesFile from the global git config, or the $XDG_CONFIG_HOME/git/ignore default). A new BV_NO_GITIGNORE environment variable (any non-empty value) disables all ignore-file management. Everything remains pure file I/O — no git subprocess, no git-binary dependency — and bv still never deletes or rewrites existing ignore entries; appending to .gitignore survives only as a last-resort fallback when .git exists but the exclude file is unusable.

Added

• Reliable light/dark theme selection (bv-128, idea from PR #178). New --theme flag (light | dark | auto) plus a top-level theme: key in ~/.config/bv/config.yaml, resolved once at startup with the precedence --theme → BV_THEME → config → auto-detect. The resolved preference is now applied to the global lipgloss renderer (lipgloss.SetHasDarkBackground), fixing the pre-existing BV_THEME override, which only touched per-model renderers while the bulk of the UI (package-global styles, badges, glamour markdown) kept auto-detecting — and auto-detection falls back to dark whenever the terminal doesn't answer the background query (common over SSH and in tmux/screen), leaving light terminals with near-white, unreadable text and no working escape hatch. Invalid --theme values warn on stderr and resolve to auto-detect.

• Bounded robot liveness for the triage history prologue (#166). The git-history correlation step of --robot-triage / --robot-next now runs under a hard budget (default 10 s), overridable via --robot-history-timeout-ms or BV_ROBOT_HISTORY_TIMEOUT_MS (0 = unbounded). On timeout the in-flight git subprocess is killed (the correlation package now threads a context.Context into every git invocation via exec.CommandContext) and triage proceeds without history. The outcome is surfaced as meta.history_status (ok | error | timeout; omitted when history generation was not attempted).

Performance release: a profile-driven overhaul of the agent-facing robot path makes repeat --robot-* calls roughly 25× faster (warm triage ~2.3 s → ~0.09 s) and cuts git subprocesses from ~340 to 3 — all while keeping output byte-identical (verified by golden and differential tests). Also bundles the correctness fixes and a fresh-eyes bug-review pass found during the optimization work, plus a routine dependency refresh.

Added

• TUI: left-click now focuses a panel and selects the row under the cursor, with precise geometry inversion for split and single-column list views (drags and out-of-range clicks are ignored) (bv-162).

Performance — Robot Path (--robot-triage / --robot-next / --robot-plan / --robot-insights)

• Correlation (the dominant former hotspot): replaced ~2 git show per commit with 2 batched git log calls; extract bead lifecycle via deduped --raw blob snapshots instead of git log -p; added a persistent correlation result cache keyed on HEAD + beads-hash + opts so repeat calls skip git extraction entirely. Cold-path: split the cache so working-tree bead edits no longer bust the HEAD-only extraction, and added content-addressed per-commit event and co-commit caches keyed by immutable commit SHA, so advancing HEAD only processes the new commits (#160, #161).
• Analysis disk cache: no longer rewrites the whole cache file on a read-hit; switched to goccy/go-json streaming codec; adopted a columnar struct-of-arrays on-disk shape (v2, ~44 % smaller).
• Loader / datasource: parse issues.jsonl once on the robot path (fused load + validation via a typed probe), memoize the data hash per invocation, and added a size-gated (≥ 4 MiB) parallel JSONL parse for large stores.

Fixed

• triage: exclude unblocked parent-child rollup edges from blocked_by (#158).
• watch: coalesce --watch-export with adaptive backoff and skip exports via the canonical content + dependency hash (#159).
• Fresh-eyes review (output-preserving): restored commit BeadIDs dropped from the HEAD artifact cache; rejected a fresher empty/non-issue JSONL from shadowing the real issues.jsonl; fixed parallel-parser ParseStats merge and an off-by-one max-capacity line boundary; stopped a transient git failure from poisoning the co-commit cache; fixed a possible readBlobs deadlock on a cat-file parse error; fixed a negative-index panic and a per-commit cache self-wipe.

Dependencies

• Updated 7 direct dependencies (git.sr.ht/~sbinet/gg, mattn/go-runewidth, golang.org/x/{image,sync,sys,term}, modernc.org/sqlite) and re-vendored canonically. See UPGRADE_LOG.md.

Patch release focused on updater correctness, release artifact reliability, and hardening fixes found during fresh-eyes review passes.

Updater & Versioning

• Ignore update notices whose release tag is equal to or older than the running version, including tags with stray whitespace.
• Clear stale in-session update banners when a later update check resolves to the current version.
• Label the footer badge as Update vX.Y.Z so it is not confused with the running bv --version value.
• Avoid redundant update checks after a recent successful check.
• Accept stable release archive names in the updater.

Build & Release

• Align GoReleaser archive names with README latest/download aliases, publish Windows as a zip archive, and update the GoReleaser config to v2 syntax.
• Keep direct download aliases reproducible and version-aliased.
• Keep GoReleaser snapshot builds from double-prefixing versions as vvX.Y.Z-next.

Tests

• Harden background worker tests by waiting on worker state instead of fixed sleeps.

Robustness

• Preserve completed history searches and coalesce queued background refreshes in the TUI.
• Resolve historical --as-of dates from commit history and filter stale data sources by true newest time.
• Harden correlation batch-stat locking, git stream progress callbacks, EOF handling, temporal author filters, and tombstone lifecycle classification.
• Safely pair agent blurb removal markers.
• Recover stale unreadable instance locks.

Search, Hooks, Analysis & Export

• Score all bead statuses in hybrid search mode.
• Reject malformed hook timeout values.
• Normalize loaded feedback data for analysis.
• Escape interactive graph HTML, validate Cloudflare project names, keep Cloudflare suggestions nonempty, and fail stale or zero-issue deployment verification.
• Avoid shell browser launchers on Windows and strengthen GitHub deployment helper behavior before force-with-lease pushes.

Patch release focused on dependency freshness, vendored reproducibility, and release metadata cleanup.

Dependencies

• Update Go module and vendored dependencies, including modernc.org/sqlite 1.50.1, fsnotify 1.10.1, pgregory.net/rapid 1.3.0, golang.org/x/* packages, goldmark 1.8.2, chroma 2.24.1, and terminal-width/runtime helpers.
• Verify the local /dp/toon-go dependency is already at the latest local commit used by go.mod.
• Update Rust/WASM lockfiles and move bv-graph-wasm to getrandom 0.4 with the wasm_js feature.

Build & Release

• Refresh fallback, Nix, README latest-download aliases, and changelog release metadata for v0.16.1.
• Restrict ACFS notification workflow triggers to main.
• Keep transient .beads/.write.lock files ignored and out of release commits.