NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
Packagist
Deterministic L2 Meta-Prompt compiler and briefing manager for coding agents.
Last release today
08 Oct 2026
Ships on a steady schedule
a new release about every 9 days
Nearly every release is documented
notes for 60 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
3 months old
96 releases · first in 2026
One column per month.
agent-recall-compiler 0.25.8
agent-recall-compiler 0.25.8
voku/agent-map ^0.20.0 || ^0.21.0 || ^0.22.0, so installs can resolve agent-map 0.22.0 (and agent-graph 0.3 through voku/agent-learning 0.18.32). No Recall behavior change.agent-recall-compiler 0.25.7
agent-recall-compiler 0.25.7
subject_patterns list of literal substrings naming what the rule is about. A constraint that declares it is selected by scope or tag only when the task description, an acceptance criterion, or a touched file also contains one of them; the selection reason is subject_match, and a task that is demonstrably not about the subject gets the exclusion reason no_subject_match. Directory-wide scopes such as src/ therefore stop pulling a rule into every task below them. Constraints without the field, tasks without readable files, files that do not exist yet, paths leaving the project root, files over 2 MiB and directories over 400 files or 8 MiB all keep the previous behaviour (selected, reason constraint_scope), so missing information can never hide a constraint. meta.json and the bundle record subject_patterns only when a selected constraint has them, and meta.json now records the real selection reason of each selected constraint instead of always constraint_scope. Motivation: after file-pattern scopes (0.25.5), a consuming project still had constraints whose only plausible scope is a whole tree; the outcome history shows most of their selections rated irrelevant or not_used.ConstraintSubjectMatchTest (16 tests): a task that never touches the subject, a touched file that does, the task text naming the subject without reading any file, fail-open without a reader and for an unreadable file, unchanged behaviour for constraints without patterns, scope and tag interplay, an excluded constraint that cannot block compilation, reader confinement (missing path, .., symlink out, oversized file, directory over the cap), the persisted reason and patterns in meta.json and the bundle, an unchanged persisted shape without patterns, and malformed manifests failing loudly. Disabling the gate fails 7 of the 16. Full suite 380 tests, 1981 assertions; PHPStan clean.agent-recall-compiler 0.25.6
agent-recall-compiler 0.25.6
voku/agent-map ^0.20.0 || ^0.21.0, so installs can resolve agent-map 0.21.0 (adds agent-map watch; no plan contract changes). The suite and PHPStan pass against agent-map 0.21.0.agent-recall-compiler 0.25.5
agent-recall-compiler 0.25.5
BlindSpotReviewer evaluates security_sensitive_context on what the task touches - the task_files and task_targets in meta.json - instead of on the whole Recall artifact text. system.md embeds the selected guidance prose, which names words such as role, sql and permission by construction, so the old scan raised the warning for 220 of 223 real reports (and 221 of 222 recorded reports were warn), which made the review gate carry almost no information. On the same reports the task-files signal fires for 14 (6%). The finding keeps its id and word-boundary matching, now reads "Task files or targets mention security-sensitive terms", and its evidence names the first five matching paths. A consumer that matched the old message text must match the finding id instead.* or ? is matched as a pattern instead of as a literal path: a pattern without a slash (*_UnitCest.php) matches the file name anywhere in the tree, a pattern with a slash (modules/*/lib/*.php) matches the whole path and everything below a matching directory, and * crosses directories. Directory-prefix scopes, the exact entries * and / (global) and every other existing behaviour are unchanged. Before this, an entry with * was compared as a literal path and never matched, so a pattern scope could only ever exclude guidance; the only such entry found in a mature consuming project was on a Finding, not on guidance, so no existing selection changes. Motivation: constraint manifests could only scope whole trees (lib/, modules/, infra/), which selected the same five constraints for nearly every task (1526 constraint selections, 82 percent of recorded outcomes irrelevant or not_used); a file-type scope such as *_UnitCest.php was inexpressible.BlindSpotSecurityMarkerTest (5 tests): guidance prose in system.md and the draft no longer triggers the warning, task files and targets do, evidence lists the matching paths capped at five, and word-boundary cases are unchanged. Restoring the whole-text scan fails the guidance-prose test. RecallScopePatternTest (5 tests) covers file-name patterns, path patterns, mixed scopes, plain guidance and the unchanged prefix/global controls; disabling the pattern branch fails its four pattern tests. Full suite 364 tests, 1928 assertions; PHPStan clean. composer ci stops earlier in this checkout at composer validate --strict because the gitignored local composer.lock is stale against composer.json.agent-recall-compiler 0.25.4
agent-recall-compiler 0.25.4
voku/agent-map ^0.20.0 (was ^0.18.0) so a consumer can also use voku/agent-edit ^0.3.0, which needs agent-map 0.20. Map facts and plans consumed by Recall are unchanged: agent-map 0.19/0.20 only add class_removal_plan@1.0 and the optional non_php_references residue block on plans.agent-recall-compiler 0.25.3
agent-recall-compiler 0.25.3
review blindspots <task-id>, while an already-authoritative plan that should continue executing after the check uses execute-plan-with-blind-spot-check. No new recipe, lifecycle state, or approval authority was added.c79e2afbdc44922c48567ff8c08c284fffbb1fd3 passed PHPUnit and PHPStan on PHP 8.3, 8.4, and 8.5 plus governed agent-loop context-explain dogfood and the released Learning identity bridge in Actions run #757.576cf64768ea502f36cbb496a4ada22a2da0ba22 passed post-merge main CI in Actions run #758. The review follow-up also pins that the self-audit text survives final BlindSpotPromptBuilder composition.agent-recall-compiler 0.25.2
agent-recall-compiler 0.25.2
CompiledContextExplainItem::$subjectId: the id of the owner record an explain item is about, carried as a typed field. For a learning_precedent item it is the LearningNote id, written from the note_id the fact payload already holds - so a selected precedent, one held back by covered_by_active_guidance and one omitted by the context budget each name their note. Until now a consumer relating the item to its note had to take the id apart (learning-precedent:<noteId>), which is Recall's own format and not a contract (#220, #221).null means "not recorded", never "no such record": an item that is not about an owner record (for example a map omission) and a selection-report.json persisted before this field existed both read back as null. The reader treats context_explain.subject_id like every other field, so a non-string value is rejected.No selection, rendering, authority or existing id/source_ref value changes; the new constructor parameter is trailing and optional. The persisted selection report remains bound to the compiled bundle by bundle_sha256 but has no digest of its own, so this field has the same integrity as the existing explain items and no more.
49114c661ed269db6df7377ebb6b1edfa7c69cf5 passed the PHP 8.3/8.4/8.5 matrix and the governed context-explain job (4 check runs); its merge commit efc70437a00a1af2121b4df618d326bd5f690949 passed the same on main (3 passed, 1 skipped). The changelog section itself was missed in #221 and is added by this release's prepare commit.agent-recall-compiler 0.25.1
agent-recall-compiler 0.25.1
measured/direct/unflinching tone, and an optional bounded focus through ReviewPromptOptions and the review CLI. Code-review prompts also accept the language option.composer ci passed: 352 tests and PHPStan.agent-recall-compiler 0.25.0
agent-recall-compiler 0.25.0
recall-log.draft.json starts with an empty guidance_outcomes list; a session adds rows only for guidance it has something to say about. Every selection is still recorded as a selection event.log-outcome accepts selected guidance without an outcome row and no longer requires guidance_outcomes_withheld_reason (still optional and still recorded when given). In real consumer history 82% of forced per-item judgements were invented not_used/irrelevant filler, which pushes guidance toward retirement.unknown row without a comment is still refused, now with a message that says to remove the row instead.agent-recall-compiler 0.24.0
agent-recall-compiler 0.24.0
attribution (seen_before_decision, also_prescribed_by) on guidance outcome events; the outcome draft carries an explicit attribution slot per selected guidance row.CompiledRecallOutput::suppressedLearningPrecedents(): relevant LearningNotes withheld from the briefing because their repository evidence needs review, so hosts can surface the otherwise silent suppression as Learning maintenance.log-outcome refuses a helpful guidance outcome without attribution. A bare helpful cannot distinguish guidance that changed a decision from guidance that merely matched it; real consumer history showed labels written after the credited fix and decisions already prescribed by loaded skills and Constraints.voku/agent-learning ^0.18.23, which owns the attribution semantics; a round-trip test proves Recall-written outcomes parse into Learning's causal-audit set.agent-recall-compiler 0.23.1
agent-recall-compiler 0.23.1
agent-recall-compiler 0.23.0
agent-recall-compiler 0.23.0
missingness-audit recipe surface acceptance-to-scope gaps as explicit implementation-authority and verification-scope blockers instead of leaving the agent to infer them.agent-recall-compiler 0.22.1
agent-recall-compiler 0.22.1
MapRecallProvider to Map's typed SearchIndexStore::semanticProvider() owner API.agent-recall-compiler 0.22.0
agent-recall-compiler 0.22.0
SearchReadinessInspector for Search currentness, capability, snapshot identity, and recovery projection instead of reconstructing Search metadata in Recall.agent-recall-compiler 0.21.0
agent-recall-compiler 0.21.0
voku/agent-map ^0.17.0, the released optional Search maintenance owner baseline used by first-party consumers.agent-recall-compiler 0.20.0
agent-recall-compiler 0.20.0
voku/agent-map ^0.16.0, aligning Recall with the released bounded Map preparation owner baseline used across first-party consumers.agent-recall-compiler 0.19.0
agent-recall-compiler 0.19.0
CompileRequest carry explicit OperatingPromptRequest selections so embedding hosts can use Recall-owned operating recipes without reconstructing --operating-prompt CLI JSON.InlineCompileTask to represent context-only inline compilation without fake code targets, matching the existing standalone compile semantics used by durable handoff construction.agent-recall-compiler 0.18.0
agent-recall-compiler 0.18.0
voku/agent-map ^0.15.0, the released Map preparation-owner baseline used by first-party consumers.chore(release): 0.17.11
chore(release): 0.17.11
voku/agent-map ^0.14.0.learning_root from .agent-loop/init.json in PathResolver.RecallDecisionEngine rejected every selected guidance whose target matched a retired proposal, including the proposal named in that retirement's superseded_by, so replacing a memory or skill row blocked its own successor. RecallRetirement now carries supersededBy (read from superseded_by), and only retirements that do not name the selected guidance still count as contradictions.agent-recall-compiler 0.17.10
agent-recall-compiler 0.17.10
RecallDecisionEngine rejected every selected guidance whose target matched a retired proposal, including the proposal named in that retirement's superseded_by, so replacing a memory or skill row blocked its own successor. RecallRetirement now carries supersededBy (read from superseded_by), and only retirements that do not name the selected guidance still count as contradictions.agent-recall-compiler 0.17.9
agent-recall-compiler 0.17.9
RecallPromptBuilder::buildSystemMd(). When a task declares known files, exact symbols, classes, methods, and line ranges are now rendered into system.md under ## Navigation Facts instead of collapsing into a solitary map index path, removing unnecessary coding-agent source rediscovery. Closes #174.composer ci passed with 328 tests, 1793 assertions, and PHPStan clean level max.agent-recall-compiler 0.17.8
agent-recall-compiler 0.17.8
decisions.md and assumptions.md artifacts in review prompts using a bounded head+tail projection, so recent decisions no longer disappear behind the 5 KB head truncation boundary.1efd48487431dd9c122c7a6257af8dd3f920647c passed PHPUnit and PHPStan on PHP 8.3, 8.4, and 8.5 in Actions run 34779676024.Merge branch 'feature/scoped-package-resources'
Merge branch 'feature/scoped-package-resources'
PackageResources: consumerSkills() exposes agent-recall-consumer, and maintainerSkills() exposes agent-recall-compiler-maintainer.continue-until-done prompt primitive and use case, ensuring structured task resumption across session boundaries.voku/agent-map ^0.13.0 beside existing 0.10, 0.11, and 0.12 compatibility lines.composer ci passed with PHPUnit (319 tests, 1768 assertions) and PHPStan Level Max static analysis with zero errors.agent-recall-compiler 0.17.6
agent-recall-compiler 0.17.6
agent-recall-compiler 0.17.5
agent-recall-compiler 0.17.5
agent-loop consumer proof that pins released Learning 0.18.5 and exercises numeric task id 403.0.17.4 package source for numeric task-id lineage. Existing release tags remain immutable; the marker targets this release-ready commit only.chore(deps): allow voku/agent-map ^0.12.0
chore(deps): allow voku/agent-map ^0.12.0
agent-recall-compiler 0.17.3
agent-recall-compiler 0.17.3
ReviewReportReader::jsonPath() as the supported read-only owner projection for the canonical deterministic review JSON artifact path, so lifecycle hosts can describe missing or invalid review evidence without reconstructing Recall's private reviews/<task>.blindspots.json layout.ReviewReportReader::read() now resolves the report through the same owner path projection, keeping expected-path diagnostics and persisted evidence reads on one Recall-owned boundary.90086b1caac30f702a8caf912e19aef55b7a2ecb passed PHPUnit and PHPStan on PHP 8.3, 8.4, and 8.5 plus the governed installed agent-loop consumer dogfood before merge.5571a027a37037d9f797a17ab071b11030e617d8; this release candidate reruns the repository release gates on top of that exact owner change.agent-recall-compiler 0.17.2
agent-recall-compiler 0.17.2
voku/agent-map ^0.10.0 || ^0.11.0 and voku/agent-learning ^0.18.1.agent-recall-compiler 0.17.1
agent-recall-compiler 0.17.1
voku/agent-learning class alone as provider availability, then called the bounded owner lineage API with a missing root and broke clean consumers with Learning root must be an existing directory. Existing Learning roots still flow through the owner API and retain fail-closed stale/corrupt-state behavior.a45ee40b207807c6ad54fe304a5a4515a2a1d3ae passed PHPUnit and PHPStan on PHP 8.3, 8.4, and 8.5 plus the governed installed agent-loop consumer dogfood before merge.a69112e1cab536e30a9fff0ff985ea6b90a7ca13; this release candidate reruns the repository release gates on top of that exact fix.agent-recall-compiler 0.17.0
agent-recall-compiler 0.17.0
LearningNote candidate acquisition path with the bounded task-precedent query consuming voku/agent-learning 0.18.1 (LearningLineageService::precedentsForTask()). Precedent discovery is now scoped to the canonical task identity (TaskBrief::$id) rather than scanning all active notes repository-wide.LearningTaskPrecedentProjection modeling the bounded lineage envelope (taskId, precedents, identityIds, depthByIdentityId, relations, maximumDepth, maximumResults, truncated).learning_precedent_observation facts with query and selection telemetry (candidates_returned, candidates_considered, precedents_selected, selected_precedent_ids).voku/agent-learning dev dependency to 0.18.1.voku/agent-learning explicitly through Recall instead of falling back to global absence claims.agent-recall-compiler 0.16.0
agent-recall-compiler 0.16.0
CompiledRecallBriefing through CompiledRecallOutputReader::briefingForTask() so lifecycle hosts can consume the canonical task briefing with exact path/content/SHA-256 identity without knowing Recall's private system.md layout.InlineCompileTask -> CompileRequest -> RecallCompiler -> CompileResult, allowing embedding hosts such as agent-loop edit to compile task id, description, and concrete targets without constructing Recall CLI tokens or stabilizing Command\CompileCommand.todo-card-handoff and execution-dispatch operating prompt recipes in DiscoveryProductionHandoffPromptTest (Recall #110 work-package / dispatch lifecycle proof).CompileRequest now accepts exactly one task source: a persisted task-brief path or a typed inline task. Both routes continue through the same existing Recall compilation semantics.LearningNoteRecallProviderTest so Level Max PHPStan passes cleanly without iterable warnings.f84eff255c9f72c4edcca7ae3c003db6412f9424 passed repository CI before merge, including the typed target-aware inline compilation regression.agent-loop dogfood on its exact head before merge.agent-recall-compiler 0.15.0
agent-recall-compiler 0.15.0
voku/agent-map ^0.10.0, which moves its Make include to
resources/make/agent-map.mk and introduces PackageResources.skills/ to resources/skills/ following the shared repository layout convention. Introduce PackageResources as the single owner of package-shipped asset paths.LearningNoteProjectionSource in the test suite explicitly.
PHPStan does not apply a constructor @param docblock to a promoted property,
so active() was reported as returning array rather than
list<LearningNotePrecedentProjection>; 0.14.0 shipped with composer ci red.VerificationPlan analysis fingerprint parameter type to array<string, mixed>|null
for agent-map 0.10.0 AnalysisFingerprint compatibility.agent-recall-compiler 0.14.0
agent-recall-compiler 0.14.0
voku/agent-learning LearningNotes as deterministic, low-authority precedents through the Learning owner's public projection API. Recall selects only explicit scope/tag matches, keeps precedent prose bounded, preserves replay identity, and exposes selection, omission, suppression, and review-needed reasons without treating solved cases as current task authority.ReviewAuditPreparer as the typed Recall-owned boundary for deterministic review evidence preparation, so embedding hosts no longer need to reproduce CLI/private review preparation behavior.BundledOperatingPromptManifest so hosts can locate the shipped operating-prompt manifest through a Recall-owned API instead of reconstructing package paths.agent-learning does not change provider snapshots when no active owner projections exist for the current repository.main head 39d44d519b7c7301f769c2a37935675513070edb passed repository CI in Actions run 33581972446.c9f0ea5a4fdf9083d46b14e1bbc0a0478fefa9fc passed PHP 8.3/8.4/8.5 PHPUnit + PHPStan, governed context-explain dogfood, released voku/agent-learning 0.14.0 compatibility, and completed review with all threads resolved.agent-loop context-explain dogfood on its exact head before merge.agent-recall-compiler 0.13.16
agent-recall-compiler 0.13.16
voku/agent-map ^0.9.0 and run the governed agent-loop context-explain dogfood against Map 0.9.0 instead of the retired pre-0.9 dependency line.2b02a3013ef2c2d12ab89556c217b725a5f5aec2 passed the PHP 8.3/8.4/8.5 PHPUnit + PHPStan matrix and the governed installed agent-loop context-explain dogfood using released agent-map 0.9.0 in Actions run 33387614396.agent-recall-compiler 0.13.15
agent-recall-compiler 0.13.15
OperatingPromptCatalog embedding API with immutable recipe, argument, validation, and deterministic preview projections so hosts can build prompt tooling without parsing Recall manifests or rendered CLI output.OperatingPromptRecallProvider through the same public catalog used by embedding hosts, keeping recipe parsing/rendering and validation under one semantic owner while preserving legacy schema-1.0 custom manifests.agent-recall-compiler 0.13.14
agent-recall-compiler 0.13.14
plan-as-draft operating recipe so an apparently complete plan can be challenged as a minimum baseline and strengthened without inventing authority, scope, or missing evidence.KanbanContextProjection embedding input so a host can pass the bounded board facts Recall already consumes without persisting a second orchestration-owned context file; the standalone --kanban-context <path> contract remains supported.agent-loop context-explain dogfood on exact PR #121 head ff1cbbbc333088df84cf7eaf8c30b5d8743913c2 in Actions run 32905912065.agent-recall-compiler 0.13.13
agent-recall-compiler 0.13.13
CompiledRecallOutputSuperseder beside the compiled Recall output model so lifecycle hosts no longer implement Recall-owned output archival themselves. Existing output is archived by the persisted identity digest when available, fallback names remain collision-safe, and dangling symbolic links are treated as occupied archive paths instead of being overwritten or causing the rename to fail.aede433affb0d5481c290ba65a2b5fc73f69cafc passed PHPUnit and PHPStan on PHP 8.3, 8.4, and 8.5 plus governed agent-loop context-explain dogfood in Actions run 32704270415. CodeRabbit's dangling-symlink collision finding was reproduced, fixed, regression-covered, and automatically resolved on the final head.agent-recall-compiler 0.13.12
agent-recall-compiler 0.13.12
PRE_EXISTING, INTRODUCED, and UNKNOWN_ORIGIN validation failures when practical, and reconciles final completion claims against actual artifacts and evidence.NO_FURTHER_INVESTMENT and authority boundaries.fe3159fd54bc55ac30480698fd469aa679a80672 passed PHPUnit and PHPStan on PHP 8.3, 8.4, and 8.5 plus governed agent-loop context-explain dogfood in Actions run 32664987840.9dc9528ccde4c7abf534defd04ba2ed0ab5b2deb passed PHPUnit and PHPStan on PHP 8.3, 8.4, and 8.5 plus governed agent-loop context-explain dogfood in Actions run 32665686309.7572702f8e5fd239ea3afdf6ddb93b81d65e10eb passed PHPUnit and PHPStan on PHP 8.3, 8.4, and 8.5 plus governed agent-loop context-explain dogfood in Actions run 32692426818; AccessLint and CodeRabbit were green and CodeRabbit reported no actionable comments.agent-recall-compiler 0.13.11
agent-recall-compiler 0.13.11
discovery-first and production-ready-handoff operating recipes for repository re-grounding and fresh-agent execution handoff without inventing authority, commands, or scope.execute-plan-with-blind-spot-check for executing an existing authoritative plan only after a bounded evidence-backed self blind-spot pass; disproved assumptions may change only the smallest affected plan step, while missing Done When or executable verification evidence fails closed instead of being invented.agent-loop enter <task-id> and agent-loop finish <task-id> plus the canonical next action instead of teaching a stale internal phase choreography.agent-loop context-explain dogfood for the discovery/handoff recipes.agent-loop context-explain dogfood for execute-plan-with-blind-spot-check on exact head eaa0fab86cf58c4a2f622dca490ad134b50f052f.agent-loop PR #272 ran current Loop head 3c38e332c8f4ee87466fed67166ecff4af532f53 against exact Recall candidate 285ba3413a0a92c3dd34251c6bac3e6e574dc73a; prompt-primitives dogfood run 32662042560 installed the candidate catalog and proved workflow plan accepts the new L1 recipe. The probe PR was then closed unmerged because no Loop product change or minimum-version bump was required.agent-recall-compiler 0.13.10
agent-recall-compiler 0.13.10
PHPUnit\Framework\TestCase is attributed to every file inheriting it; matching incoming relations against those shared ids grew with the square of the inheriting files. On a 311-file repository one test file's fact carried 88,179 relations (33.7 MB) instead of 949 (0.39 MB), and the compiled bundle reached 124 MB. Symbols listed under more than one file are now excluded from that matching, since a symbol listed by several files is declared by none of them. Files that declare their own symbols are unaffected.agent-loop dogfood on exact head c6ce6edb896499db8c2bd28eea1b0bc1e53da977; no review threads or submitted review findings remained before merge.agent-recall-compiler 0.13.9
agent-recall-compiler 0.13.9
RecallFactsDocumentReader own the standalone facts-document path through path() and readFromOutputDirectory(), so hosts no longer need to know the private facts.json filename.agent-loop dogfood; CodeRabbit reported no actionable review thread on the final head.agent-recall-compiler 0.13.8
agent-recall-compiler 0.13.8
RecallFactsDocumentReader and RecallFactsDocument so hosts consuming a standalone compiled facts document can read its canonical bundle identity and typed facts without decoding Recall JSON themselves.CompiledRecallOutputReader now delegates facts parsing to the same owner reader, leaving one implementation of the persisted facts schema inside Recall.1.0 document shape, raw lowercase 64-hex bundle digest, JSON-object root, and list-shaped fact collection while preserving open fact payload meaning for the host.agent-loop context-explain dogfood.agent-recall-compiler 0.13.7
agent-recall-compiler 0.13.7
CompiledRecallOutputReader, CompiledRecallOutput, and RecallFact as the typed read side of compiled Recall output for lifecycle hosts.current task fallback and validation of its recorded output hashes, including rejecting unsafe recorded relative paths before hashing.selected_constraints are exposed as their semantic IDs.agent-loop exercised the API against the existing lifecycle call sites and forced the reader to preserve real constraint, fact-provenance, missing/corrupt, task-fallback, and output-integrity semantics instead of simplifying fixtures around them.agent-recall-compiler 0.13.6
agent-recall-compiler 0.13.6
ReviewReportReader, ReviewReportArtifact, and ReviewReportPaths so PHP lifecycle hosts can consume the current blind-spot report and its exact persisted SHA-256 identity without reconstructing Recall's file layout or JSON schema.ReviewReportWriter now uses the same canonical path owner as the reader without changing the existing review artifact layout.null.agent-recall-compiler 0.13.5
agent-recall-compiler 0.13.5
test-driven-development L2 operating recipe for evidence-backed RED/GREEN/REFACTOR work without turning test-first technique into workflow authority.CompileRequest, RecallCompiler, and CompileResult APIs so PHP lifecycle hosts can embed governed Recall compilation without reproducing CLI option spelling, provider composition, or artifact rules.agent-recall-compiler 0.13.4
agent-recall-compiler 0.13.4
continue-until-done now treats slice-by-slice self-confirmation as an internal continuation check only. Autonomous execution may continue while existing authority remains valid and evidence supports the next step, but it cannot satisfy human, owner, reviewer, accepted-risk, or approved-contract decisions by confirming itself.HUMAN_DECISION_REQUIRED / BLOCKED, while successful completion requires the configured Done condition to be satisfied by observed evidence.agent-loop context-explain dogfood.4c621e1e7ff511a7e1cc5c5aa70059b5aa532c5a passed main-branch CI run #374 after merge. CodeRabbit was rate-limited for #82, so no external-review claim is made.agent-recall-compiler 0.13.3
agent-recall-compiler 0.13.3
todo-card-handoff L2 recipe for self-contained durable task handoff to coding agents without chat, Session-private context, hidden reasoning, or prior-agent memory. The recipe requires repository anchors, verified/inferred separation, bounded next steps, observable acceptance, and the repository's existing task owner instead of inventing another backlog format.5934f4ecce271d00b2229e04bea4108e5c782bc4 passed main-branch PHPUnit and PHPStan on PHP 8.3, 8.4, and 8.5. PR #80 also passed the governed installed agent-loop context-explain dogfood. CodeRabbit was rate-limited for #80, so no external-review claim is made.agent-recall-compiler 0.13.2
agent-recall-compiler 0.13.2
system.md while preserving first-seen order. The regression was independently reproduced by voku/httpful#34 and voku/phonetic-algorithms#6.agent-loop dogfood, so the candidate under test cannot be confused with consumer-local state.agent-loop context-explain dogfood. CodeRabbit reported no actionable review findings.regression-hunt now treats minimum_findings as a bounded probe budget instead of a required defect count. Completed evidence-backed probes may return
regression-hunt now treats minimum_findings as a bounded probe budget
instead of a required defect count. Completed evidence-backed probes may
return CLEAN, missing material evidence may return BLOCKED, and the
recipe explicitly forbids manufacturing defects to satisfy the numeric floor.RegressionHuntPromptContractTest.Breaking: blind-spot review reports may carry the exact Contract revision and implementation snapshot they reviewed, so governed consumers can reject
review blindspots accepts paired --contract-revision and
--implementation-snapshot sha256:<digest> options. Standalone unbound review
remains supported; repository snapshot semantics stay with the orchestrator.Allow voku/agent-map ^0.8.0 in addition to the existing 0.5, 0.6, and 0.7 lines. This is a compatibility-only release: Recall providers, facts, bundle
voku/agent-map ^0.8.0 in addition to the existing 0.5, 0.6, and
0.7 lines. This is a compatibility-only release: Recall providers, facts,
bundle digests, prompt rendering, and outcome semantics are unchanged.agent-map 0.8.0 line before this compatibility release
is tagged; no path repository or development-branch alias is used.The opt-in prompt guidance-gaps diagnostic now classifies each material authority failure as AUTHORITY_MISSING, AUTHORITY_NOT_SURFACED, AUTHORITY_STAL
prompt guidance-gaps diagnostic now classifies each material
authority failure as AUTHORITY_MISSING, AUTHORITY_NOT_SURFACED,
AUTHORITY_STALE, AUTHORITY_CONFLICTING, or AUTHORITY_INCOMPLETE.AUTHORITY_NOT_SURFACED is evidence-bound: agents may use it only when an
applicable authority is proven to exist but was absent from usable context.
Its remediation points at manifest, scope, retrieval, installation, or
routing of that existing authority instead of manufacturing replacement docs.agent-loop#113 self-shape dogfood: the
self-shaping lifecycle document already existed, but the consuming project's
Recall document manifest did not register it. Registering the existing
authority changed the real governed Recall bundle without rewriting the doc.agent-loop consumer, AccessLint and CodeRabbit. The exact merged commit
passed main-branch PHP/PHPStan CI before this release candidate was prepared.Added the explicit prompt guidance-gaps L2 diagnostic for spec-driven, human-in-the-loop implementation sessions. The generated project-specific promp
prompt guidance-gaps L2 diagnostic for spec-driven,
human-in-the-loop implementation sessions. The generated project-specific
prompt keeps a running implementation-notes.html journal that separates
design decisions, deviations, tradeoffs and open questions from actual
guidance gaps where the expected spec, docs, skill, workflow, tool contract,
code or tests did not provide usable authority.HUMAN_DECISION_REQUIRED / BLOCKED boundary
instead of inviting the coding agent to guess.The canonical agent-recall-consumer skill now teaches the released 0.12 outcome-honesty contract: unedited unknown rows are placeholders, explicit unk
agent-recall-consumer skill now teaches the released 0.12
outcome-honesty contract: unedited unknown rows are placeholders, explicit
unknown requires a reason, and a caller without evidence must declare
guidance_outcomes_withheld_reason rather than fabricating not_used or irrelevant. Selection events retain outcome_withheld_reason.Breaking: log-outcome no longer accepts an unedited recall-log.draft.json as guidance feedback. compile pre-fills every selected row with outcome: unk
log-outcome no longer accepts an unedited recall-log.draft.json as guidance feedback. compile pre-fills every selected row with outcome: unknown and no comment for a session to complete afterwards; logging that verbatim meant the component that selected the guidance also graded it. An unknown outcome now requires a comment saying why the guidance could not be judged.helpful, irrelevant and harmful require a justifying comment, and helpful and harmful cannot be recorded with applied: false.guidance_outcomes_withheld_reason. Selection is a fact about the compiler and usefulness is a fact about the session; a caller with no evidence was previously forced to invent a bucket, and both not_used and irrelevant push a rule towards retirement in agent-learning's staleness policies. Silent omission is still refused, so an accidental drop does not look like a deliberate one.outcome_withheld_reason for the selections a declared withholding explains, so a downstream gate can distinguish a declared absence from a dropped one after the draft is pruned. The field is additive and absent on existing records.agent-loop context-explain dogfood, before merging PR #50.agent-loop run, which is where the consuming workflow close gate was found to demand a judgement per selection and had to learn to accept a declared withholding.Align the canonical agent-recall-consumer skill with the live Recall CLI: compact .agent-loop defaults, current compile/outcome/review/reflection comm
agent-recall-consumer skill with the live Recall CLI: compact .agent-loop defaults, current compile/outcome/review/reflection commands, and the agent-loop project-layout boundary are now documented from the owning package.Cli dispatch table and rejects retired roots.agent-loop context-explain dogfood before merge.Recall-specific coding instructions and reusable operating-prompt assets now live with agent-recall-compiler: the first-party catalog ships beside ski
agent-recall-compiler: the first-party catalog ships beside
skills/agent-recall-consumer/SKILL.md instead of requiring a separately
versioned generic skill repository. Callers still select every recipe and
provide its arguments explicitly; bundling the catalog adds no hidden prompt
selection or defaults.docs/operating-prompts.md now makes the ownership boundary explicit:
tool-neutral engineering principles may stay with their generic owner, while
Recall commands, review primitives, manifest schema, and Recall-consumed
first-party recipe assets are reviewed, tested, and released with this tool.BundledOperatingPromptCatalogTest, which compiles the shipped
adversarial-review recipe through the real CLI and proves the falsification
contract still permits CLEAN after evidence-backed probes find no defect.Added a context-light review first-draft primitive for manual or automated falsification-oriented review without requiring task/run state. The library
review first-draft primitive for manual or automated
falsification-oriented review without requiring task/run state. The library and
CLI surface deliberately allows CLEAN only after concrete attempts to break
the implementation find no evidence-backed defect, and rejects unexpected CLI
arguments instead of silently ignoring them.review code <task-id> prompts now inherit the same first-draft
falsification lens while keeping their existing project evidence and dominant
installed review-lens routing.UNKNOWN or BLOCKED. Model confidence,
prior rationale, reviewer consensus, prompt construction, and unexecuted commands
are not verification, and changing an approved contract requires a separate
approved re-plan rather than weakening requirements inside the generated L1.agent-loop context-explain dogfood passed with the candidate, and
external review found and closed the unapproved contract-change escape hatch
before merge.Compile-time TaskBrief augmentation now preserves the complete governed task semantics when CLI targets or inline operating prompts are added. In part
TaskBrief augmentation now preserves the complete governed task
semantics when CLI targets or inline operating prompts are added. In particular,
the governed Run binding and approved acceptance criteria no longer disappear
while the requested target/prompt collection is extended.compile CLI regressions for both --target augmentation and
inline operating-prompt augmentation from a governed Contract, asserting the
canonical task-context fact retains exact acceptance criteria and Run lineage.Added approved task acceptance_criteria as an optional governed task input. Criteria are carried through direct and governed Contract parsing, canonic
acceptance_criteria as an optional governed task input.
Criteria are carried through direct and governed Contract parsing, canonical
task-context facts, effective task scope, and rendered system.md briefings.agent-loop dogfood passed on the
exact release candidate before merge.Your coding agent can read these notes before it upgrades. Set up the MCP server →