Skip to content

Local-first 🧠 memory for Hermes Agent that lives in your Obsidian vault and remembers project context. Nightly 😴 dream passes turn repeat corrections into confirmed preferences with measurable confidence. Adapters ship for Claude Code, Codex, and OpenClaw, with an MCP server for anything else.

TypeScript
25.9 KB
Raw Blame

Recall quality wave - implementation plan

Executed by the implement node with PARALLEL LANES. The substrate lands first (serial); lanes A-E then run concurrently over disjoint files; the integrator finishes. Every task is atomic and commit-sized.

Cross-lane contract

(a) Substrate signatures - land FIRST, every lane codes against the landed code

Lane A, B (and the integrator) code against exactly this surface, landed by S1/S2 as behavior-identical code. Nothing else may redefine it.

// src/core/brain/entities/canonical.ts (S1)
export type EntityLabelInvalidReason =
  | "empty" | "too-long" | "no-letter-or-digit" | "denylisted"
  | "shouted-heading" | "function-word-run";

export interface EntityLexiconHandle {
  readonly caseful: boolean;            // false for caseless-script corpora
  isCommonWord(token: string): boolean; // store-learned; consulted for caseful scripts only
}

export interface ValidateEntityLabelOptions {
  denylist?: ReadonlySet<string>;
  lexicon?: EntityLexiconHandle | null; // null/absent => byte-identical to v1.79.0 behavior
}
// EntityLabelValidation stays { valid: boolean; reason?: EntityLabelInvalidReason }.

Pinned predicate rules (inside canonical.ts, pure): labels with a structural code shape (contains _, ., /, any digit, or internal uppercase) bypass shouted-heading and function-word-run but keep empty/too-long/no-letter-or-digit/denylisted; case-based verdicts run only for caseful labels (some letter where l !== l.toLowerCase()); caseless labels admit on the structural gate alone. The hard-at-create vs advisory-heal policy split lives in the callers, not the predicate (see A2/A3).

// src/core/brain/derived-artifact.ts (S2) - the co-occurrence/tuning convention, extracted once
export interface DerivedArtifact {
  schema: string;       // "o2b.<name>.vN"
  generated_at: string; // ISO instant
  dataset_hash: string; // sha256 over the sorted input-set signature
}
export function deriveDatasetHash(parts: readonly string[]): string;
export function writeDerivedArtifact<T extends DerivedArtifact>(
  vaultDir: string, relPath: string, artifact: T,
): void; // atomic write, vault-identity guarded
export function readDerivedArtifact<T extends DerivedArtifact>(
  vaultDir: string, relPath: string, expectedSchema: string, datasetHash: string,
): T | null; // null on missing file, torn JSON, schema mismatch, or dataset-hash expiry (fail-open)

Option-type growth owned exclusively by the lane that lands it - no other lane touches these:

  • SearchOptions (src/core/search/types.ts) kind/attributes growth - owned by LANE C (kind?: "page" | "section", promoted-key property pushdown semantics, query-cache key inputs).
  • brain_recall_telemetry operation union (list | summary | gate_list | gate_summary | observed_reuse | cost) - owned by LANE D (grows dead_weight in src/mcp/brain/recall-tools.ts op union + tool schema enum + src/cli/brain/verbs/recall-telemetry.ts dispatch).

(b) Per-lane disjoint file ownership - overlap forbidden

Every file is owned by exactly one lane; the only shared files are the APPEND-ONLY registration files below. If two lanes must touch the same non-shared file, stop and escalate - do not coordinate silently.

  • Substrate: src/core/brain/entities/canonical.ts, tests/core/brain/entities.canonical.test.ts, src/core/brain/derived-artifact.ts, tests/core/brain/derived-artifact.test.ts
  • Lane A: src/core/brain/entities/lexicon.ts, tests/core/brain/entities.lexicon.test.ts, src/core/brain/entities/registry.ts, tests/core/brain/entities.registry.test.ts, src/core/brain/entities/label-hygiene.ts, tests/core/brain/entities.label-hygiene.test.ts, src/core/brain/doctor/entity-checks.ts, tests/core/brain/entities/doctor.entities.test.ts, src/core/brain/fact-extract.ts, src/core/brain/atomic-facts.ts, tests/core/brain/atomic-facts.test.ts, src/cli/brain/verbs/entity.ts, tests/cli/brain-entity.test.ts
  • Lane B: src/core/brain/link-graph/session-mining.ts, src/core/brain/link-graph/hallways.ts, tests/core/brain/link-graph/session-mining.test.ts, tests/core/brain/link-graph/hallways.test.ts, src/cli/brain/verbs/hallways.ts, tests/cli/brain-hallways.test.ts, src/core/brain/continuity/types.ts, tests/core/brain/continuity/record-envelope.test.ts, src/core/brain/sessions/import.ts, sessions import tests under tests/core/brain/sessions/
  • Lane C: src/core/search/schema.ts, src/core/search/store/keyword.ts, src/core/search/store/vectors.ts, src/core/search/store/documents.ts, src/core/search/store/chunks.ts, src/core/search/search.ts, src/core/search/indexer.ts, src/core/search/types.ts, src/core/search/query-cache.ts, src/core/search/pipeline/pool-filters.ts, src/core/search/result-filters.ts, src/mcp/search-tools.ts, src/cli/search/verbs/query.ts, tests/core/search/prefix-pushdown.test.ts, tests/core/search/schema-v14.test.ts, tests/core/search/kind-attributes-pushdown.test.ts, tests/core/search/hydration-two-tier.test.ts
  • Lane D: src/core/search/benchmark.ts, tests/core/search/benchmark-lang.test.ts, tests/core/search/recall-benchmark-zh.test.ts, tests/fixtures/recall-benchmark-zh/, src/cli/search/verbs/rerank-eval.ts, src/core/brain/dead-weight.ts, tests/core/brain/dead-weight.test.ts, src/mcp/brain/recall-tools.ts, src/cli/brain/verbs/recall-telemetry.ts, tests/cli/brain-recall-telemetry.test.ts, src/core/vault-scope/index.ts (additive accessor only if measured necessary). tests/core/search/recall-benchmark.test.ts and tests/fixtures/recall-benchmark/ are FROZEN for this wave.
  • Lane E: src/core/brain/permissions/resolve.ts, src/core/brain/permissions/document.ts, src/core/brain/intake/extract-intake.ts, src/core/brain/source-cleanup.ts, src/mcp/brain/ingest-tools.ts, src/core/brain/doctor/store-integrity.ts, src/core/install/ownership-measure.ts, src/cli/brain/verbs/secret.ts, tests/core/brain/permissions-target-canonicalization.test.ts, tests/core/brain/intake-stage-disposition.test.ts, tests/core/brain/delete-by-source-deny.test.ts, tests/core/install/ownership-research-pool.test.ts, tests/cli/secret-unlock-nudge.test.ts, tests/core/brain/bundle-wildcard-warning.test.ts
  • Integrator: README.md, docs pages, CHANGELOG.md, package.json + manifests synced by scripts/sync-version.ts, census registration fixes.
  • APPEND-ONLY shared (each lane appends only its own entries; ratchet tests are the arbiter; the integrator reconciles): src/cli/command-manifest.ts, src/cli/brain.ts (dispatcher cases), src/cli/brain/help-text.ts, census registries (VISIBILITY_SURFACE_REGISTRY, write-site/egress/vault-path/origin-channel/append-shard/progress registries).

(c) Dependency-wait rule

Lanes A and B wait on the substrate; C, D, E wait on nothing. A waiting lane starts by writing its failing tests against the pinned signatures (they fail on the missing surface, which is the expected failure), then polls git log -- <upstream file> every 30 s for up to 12 min; if the upstream commit has not landed after 30 min of waiting total, escalate to the orchestrator instead of inventing the surface.

(d) House rules

  • Strict TDD: write the failing test first, run it, and watch it fail with the expected error before implementing. Refactor only with all tests green.
  • Commit each atomic unit as its own conventional commit (feat(search): ..., fix(permissions): ...); never bundle two tasks.
  • Stage files BY NAME - never git add -A or git add .. Commit with git commit --only <paths>.
  • Format and lint your own files before EVERY commit; a green formatter + linter is a commit precondition, not end-of-PR cleanup. If the formatter touches files outside your lane, surface it, do not fold it in.
  • Run tests with a throwaway HOME: env HOME=$(mktemp -d) bun test <paths>.
  • Credential-shaped literals in fixtures come from tests/helpers/fake-credentials.ts assembled parts. Never write an identifier containing secret directly before a quoted value; never commit a credential-shaped literal.
  • chmod-denial tests guard with test.skipIf(CHMOD_CANNOT_DENY) from tests/helpers/platform.ts.
  • Windows path assertions are normalized on both sides (no raw separator comparisons).
  • Subprocess-spawning tests get explicit 20000 ms budgets.
  • Order-dependent assertions use sorted listings.
  • Census counts reconcile BY MEASUREMENT - never weaken an assertion to make a count fit.
  • No new MCP tool: the tool-count pin 115 (tests/core/install/tool-ceiling.test.ts) must not move; operations on existing tools are fine.
  • Committed artifacts are English; detection and verdict logic contains no hardcoded natural-language word lists or phrases (the standing house invariant - learned or operator-supplied vocabularies only).
  • All house rules apply on the branch feat/recall-quality only; never touch the project root checkout or another branch. Never commit .reviews/ (untracked operator scratch).

Lanes and tasks

Substrate (serial, lands first; behavior-identical)

Task S1: Admission-predicate extension in the leaf kernel

  • Files: src/core/brain/entities/canonical.ts, tests/core/brain/entities.canonical.test.ts
  • Acceptance: entities.canonical.test.ts adds the verdict-matrix rows for shouted-heading and function-word-run, the code-shape bypass, the caseful-only case rule, and caseless-script admit - plus a byte-identity test proving every v1.79.0 verdict is unchanged when lexicon is null/absent. Import-cycle ratchet stays green (canonical.ts still imports nothing).
  • Depends on: none

Task S2: Shared derived-artifact IO helper

  • Files: src/core/brain/derived-artifact.ts, tests/core/brain/derived-artifact.test.ts
  • Acceptance: derived-artifact.test.ts proves round-trip write/read, torn JSON -> null, schema mismatch -> null, dataset-hash expiry -> null (fail-open), atomic write, and vault-identity guard - mirroring the tuning-store test family. No existing module changes behavior (extraction only; co-occurrence/tuning keep their own code this wave).
  • Depends on: none

Lane A - admission + lexicon (t_11d51443)

Task A1: Store-learned lexicon module

  • Files: src/core/brain/entities/lexicon.ts, tests/core/brain/entities.lexicon.test.ts
  • Acceptance: entities.lexicon.test.ts proves the statistics fold over a corpus walk (listVaultPages at MAINTENANCE_LANE_REACH, Brain-root excluded, session transcripts excluded - scope asserted in the artifact header), writeEntityLexicon/loadEntityLexicon on the S2 helper with o2b.entity-lexicon.v1 + dataset hash, fail-open on a torn/stale artifact (mirroring the tuning-store test), top-K cap, and resolveEntityLexicon returning the pinned EntityLexiconHandle.
  • Depends on: S1, S2

Task A2: Creation gate + anchoring wiring

  • Files: src/core/brain/entities/registry.ts, src/core/brain/fact-extract.ts, src/core/brain/atomic-facts.ts, tests/core/brain/entities.registry.test.ts, tests/core/brain/atomic-facts.test.ts
  • Acceptance: entities.registry.test.ts proves upsertEntity resolves the lexicon beside the denylist and applies the tightened verdicts on the CREATE arm only (single-token function-word-run and multi-token prior-occurrence-free shouted-heading refuse with the new reasons; a single-token shout is a brand/acronym and stays advisory per design.md; the update arm is unaffected; a code-shaped label such as foo_bar/FTS5 still admits). atomic-facts.test.ts proves buildAnchorables/anchorEntityForms skip junk forms while valid aliases still anchor.
  • Depends on: S1, A1

Task A3: Store-wide heal wiring

  • Files: src/core/brain/entities/label-hygiene.ts, src/core/brain/doctor/entity-checks.ts, tests/core/brain/entities.label-hygiene.test.ts, tests/core/brain/entities/doctor.entities.test.ts
  • Acceptance: entities.label-hygiene.test.ts proves findMalformedEntityLabels resolves the handle and reports the new reasons (multi-token runs advisory), and the prune path (dry-run default, snapshot-gated) flows them through unchanged; doctor.entities.test.ts proves the doctor surfaces them under entity-label-malformed with the existing o2b brain entity prune remedy. No new repair command.
  • Depends on: S1, A1

Task A4: Lexicon rebuild verb

  • Files: src/cli/brain/verbs/entity.ts, tests/cli/brain-entity.test.ts, append-only src/cli/command-manifest.ts, src/cli/brain/help-text.ts
  • Acceptance: o2b brain entity lexicon rebuild writes only the derived artifact (dry-run report by default, no --confirm needed for a derived write); brain-entity.test.ts covers the verb; manifest-completeness ratchet stays green.
  • Depends on: A1

Lane B - miner + hallways (t_13a82a31)

Task B1: Session-mining ledger in the import loop

  • Files: src/core/brain/link-graph/session-mining.ts, src/core/brain/continuity/types.ts, src/core/brain/sessions/import.ts, tests/core/brain/link-graph/session-mining.test.ts, tests/core/brain/continuity/record-envelope.test.ts, sessions tests under tests/core/brain/sessions/
  • Acceptance: session-mining.test.ts proves the precision detector families (symbol from code spans/fences and qualified-identifier shapes, url reusing fact-extract patterns, path in both slash directions, qualified-ident), family-scoped case-preserving side keys (NFC + trim only - never normalizeEntityName), secret-shaped-token refusal (credential URLs, token/key query forms), and dedupe at (family, key, sessionRef) granularity. record-envelope.test.ts covers the additive continuity record kind (no schema bump); sessions tests prove capture-boundary and private-turn withholding are respected and import counters stay never-silent.
  • Depends on: none

Task B2: Hallways pass + artifact

  • Files: src/core/brain/link-graph/hallways.ts, tests/core/brain/link-graph/hallways.test.ts
  • Acceptance: hallways.test.ts proves entity->session incidence and pairwise PMI (co-occurrence's math) over the ledger, Brain/link-graph/hallways.json written through the S2 helper (o2b.hallways.v1, dataset-hashed, fail-soft read), a reviewable proposals file with dismissals (bridges pattern), and separation from the vault-derived co-occurrence graph (no shared artifact, no merge).
  • Depends on: B1, S2

Task B3: Hallways CLI + thresholded promotion

  • Files: src/cli/brain/verbs/hallways.ts, tests/cli/brain-hallways.test.ts, append-only src/cli/brain.ts, src/cli/command-manifest.ts, src/cli/brain/help-text.ts
  • Acceptance: brain-hallways.test.ts covers mine|show|promote|dismiss; promotion is offered only for keys seen in >= N distinct sessions and flows through upsertEntity/relateEntities exclusively - a test proves a promoted label failing the tightened gate is refused by the admission path (the pinned coupling), and that ambient auto-write never happens without the promote step. Manifest and visibility-surface ratchets stay green.
  • Depends on: B2, S1

Lane C - store v14 migration (t_71cbf5bf + t_82851eda)

Task C1: Sargable path range predicates

  • Files: src/core/search/store/keyword.ts, src/core/search/store/vectors.ts, src/core/search/store/documents.ts, tests/core/search/prefix-pushdown.test.ts
  • Acceptance: prefix-pushdown.test.ts proves keywordTopK, semanticTopK, countDocumentsUnderPrefix (and the pending-vector census sharing the substr form) use d.path >= ?1 AND d.path < ?2 with ?2 = prefix || U+FFFF, that the caller prefix passes through canonicalNotePath (a win32 backslash prefix now matches), that assertSafePathPrefix refusals are unchanged, and that results are byte-identical to the substr form on fixture stores.
  • Depends on: none

Task C2: v14 migration - kind column, taxonomy indexes, attributes table

  • Files: src/core/search/schema.ts, tests/core/search/schema-v14.test.ts
  • Acceptance: schema-v14.test.ts proves the v14 entry appends to MIGRATIONS and runs inside the single open transaction; chunk_kind is a VIRTUAL generated column (page for chunk_index = 0, else section) with idx_chunks_kind; idx_documents_page_type exists; document_attributes(document_id, key, value) is created and backfilled from chunk 0 (v12 recipe) with promoted keys (type, tags, pinned, visibility, aliases) skipped and bounds enforced (64 keys/doc, 16 values/key, 256 bytes/value - overflow decrements, never fails); embeddings, chunk_vec, chunk_vec_map compare byte-identical before/after; the expected-schema gap check self-derives green; refusal above LATEST still names o2b search reindex.
  • Depends on: none (in-lane order after C1)

Task C3: Indexer attribute writes

  • Files: src/core/search/indexer.ts, src/core/search/store/documents.ts, tests/core/search/schema-v14.test.ts (extended)
  • Acceptance: upsert writes attributes computed from the already-parsed frontmatter (same parse that feeds fmNotices, so degradation notices are untouched); the content_hash fastpath skips unchanged documents exactly as today; a schema-v14 extension test proves freshen/update/delete keep the side table consistent and bounded.
  • Depends on: C2

Task C4: Kind + promoted-key pushdown on the query path

  • Files: src/core/search/types.ts, src/core/search/query-cache.ts, src/core/search/store/keyword.ts, src/core/search/store/vectors.ts, src/core/search/pipeline/pool-filters.ts, src/core/search/result-filters.ts, src/mcp/search-tools.ts, src/cli/search/verbs/query.ts, tests/core/search/kind-attributes-pushdown.test.ts
  • Acceptance: kind-attributes-pushdown.test.ts proves: kind pushdown on the keyword lane (WHERE before ORDER BY/LIMIT) and the semantic lane (post-KNN filter inside the 4x overfetch); promoted keys (type, tags, pinned, visibility) pushed as SQL equality/EXISTS with normalized values (type via normalizeSchemaToken; attributes stored as String(value) scalar form preserving matchesAny semantics); mixed filter maps push the promoted half and keep the rest post-rank with canDropRows overfetch preserved; absent filters leave the SQL byte-identical (defaults test); the query-cache key grows for kind. MCP kind argument and CLI --kind surface covered; existing lane-output tests stay green.
  • Depends on: C2, C3

Task C5: Two-tier hydration cut

  • Files: src/core/search/store/chunks.ts, src/core/search/search.ts, tests/core/search/hydration-two-tier.test.ts
  • Acceptance: hydration-two-tier.test.ts proves the pool hydrates metadata-only (no c.content) and content hydrates for the rankCap ids after assembleRankedResults; final results (content, scores, order) are byte-identical on fixture stores; peak-hydration row count asserted against the rankCap, not the pool.
  • Depends on: C4 (single writer on search.ts at a time)

Lane D - eval + dead weight (t_693fa22a + t_1ee67e5b)

Task D1: Language field + per-language report split

  • Files: src/core/search/benchmark.ts, src/cli/search/verbs/rerank-eval.ts, tests/core/search/benchmark-lang.test.ts
  • Acceptance: benchmark-lang.test.ts proves lang?: string validates like answer (non-empty string when present, default "en"), byLang aggregates fold from perQuery with the same reduce as the overall report, and rerank-eval renders the split. Report change is additive (return value, not a record).
  • Depends on: none

Task D2: Bundled ZH fixture + offline CI gate

  • Files: tests/fixtures/recall-benchmark-zh/ (self-authored ZH vault + ~12 queries, 2-3 with answer), tests/core/search/recall-benchmark-zh.test.ts
  • Acceptance: the new gate runs runRerankEvalGate for the ZH set on the offline stack (local provider + bundled jieba/tiny-segmenter FTS) with ZH pins measured on the first run then frozen; tests/core/search/recall-benchmark.test.ts and tests/fixtures/recall-benchmark/ are byte-untouched (EN denominators preserved).
  • Depends on: D1

Task D3: Dead-weight fold

  • Files: src/core/brain/dead-weight.ts, src/core/vault-scope/index.ts (additive included-paths accessor only if measurement shows one is needed), tests/core/brain/dead-weight.test.ts
  • Acceptance: dead-weight.test.ts proves the complement math: full set = walkVaultScope included paths, retrieved set = union of recall_telemetry top_artifacts paths + recall_observed_use sourceRefs paths; vanished (deleted-but-indexed) pages are never reported; the window states the earliest matching record createdAt; the report is a read-only deterministic fold (no emit).
  • Depends on: none

Task D4: dead_weight surfaces with CLI parity

  • Files: src/mcp/brain/recall-tools.ts, src/cli/brain/verbs/recall-telemetry.ts, tests/cli/brain-recall-telemetry.test.ts
  • Acceptance: brain-recall-telemetry.test.ts covers the dead_weight op (union grows to seven) following the observed_reuse envelope ({vault_path, window, pages_total, pages_retrieved, dead_count, dead (capped), gate_caveat}), with since/until args honored; remote-reach filtering drops hidden rows before total (requestRefView/gatedOwnerScopeView); CLI subcommand parity; tool-ceiling pin stays 115.
  • Depends on: D3

Lane E - repairs (files disjoint from every feature lane)

Task E1: Canonicalize target-scoped permission rules (PR #247 M5)

  • Files: src/core/brain/permissions/resolve.ts, src/core/brain/permissions/document.ts, tests/core/brain/permissions-target-canonicalization.test.ts
  • Acceptance: the new test proves a rule target written with .\ or backslash separators (Notes\a.md) matches a stored POSIX path (Notes/a.md) via canonicalNotePath normalization on both sides, and an unnormalized target in the permissions document draws a warning; existing permission tests stay green.
  • Depends on: none

Task E2: Ingest stage verdict stages entity pages and manifest row (PR #247 M4 residual)

  • Files: src/core/brain/intake/extract-intake.ts, tests/core/brain/intake-stage-disposition.test.ts
  • Acceptance: the new test proves that on a stage disposition the entity pages and the manifest row are staged into review (Brain/pending/) rather than published - only the deny arm publishes nothing, and the docblock states the disposition contract; the deny-before-any-write behavior is pinned regression-free.
  • Depends on: none

Task E3: Deny-consulted evidence for delete_by_source and schema mutations (PR #247 M3 residual)

  • Files: src/core/brain/source-cleanup.ts, src/mcp/brain/ingest-tools.ts, src/core/brain/doctor/store-integrity.ts, tests/core/brain/delete-by-source-deny.test.ts
  • Acceptance: the new test proves delete_by_source and the ingest-tools schema-mutation lanes resolve the document against the credential subject and refuse on deny (same refuseDocumentDeny discipline as update/append), and the store-integrity doctor reports the covered lanes; the CHANGELOG claim wording ("update or append") is corrected via the integrator's 1.80.0 entry.
  • Depends on: none

Task E4: Wire the research-pool use site to the vault config (PR #244 M10/L1)

  • Files: src/core/install/ownership-measure.ts, tests/core/install/ownership-research-pool.test.ts
  • Acceptance: the new test proves resolveResearchPoolEnv receives the vault-resolved configuration (not raw process.env alone), so a raw $secret: string no longer counts as a configured research key in doctor, and the refusal names the unresolved reference - making the v1.78.0 claim true on the live path.
  • Depends on: none

Task E5: Custody-flow documentation (PR #244 M4 + L7)

  • Files: src/cli/brain/verbs/secret.ts, docs page for the ingest manifest (rollback note), tests/cli/secret-unlock-nudge.test.ts
  • Acceptance: the new test proves the unlock flow prints an export-before-wrap/unlock nudge; docs state that the content-manifest schema move is one-way for older binaries (schema 3 rollback note). CHANGELOG wording rides the integrator's 1.80.0 entry.
  • Depends on: none

Task E6: Bundle import wildcard-allowlist warning (PR #244 L2)

  • Files: src/mcp/brain/ingest-tools.ts (bundle import warning surface) or the bundle import CLI path it shares - measured and kept within lane E's disjoint set, tests/core/brain/bundle-wildcard-warning.test.ts
  • Acceptance: the new test proves importing a bundle whose allowlist contains * emits a named warning (AAD still refuses tampering); no behavior change for non-wildcard allowlists.
  • Depends on: none

Integrator

Task I1: Census reconciliation

  • Files: append-only registration files (src/cli/command-manifest.ts, census registries) wherever lane appends left drift; measured-count fixes in tests
  • Acceptance: the full census family is green by measurement: manifest-completeness, visibility-surface, progress-census, write-site-census, egress-census, vault-path-census, origin-channel-census, append-shard-census, entity-read-census, tool-ceiling (115 unchanged), import-cycle ratchet. No assertion weakened; counts reconciled by measurement.
  • Depends on: S1-S2, A1-A4, B1-B3, C1-C5, D1-D4, E1-E6

Task I2: Docs and README

  • Files: README.md, affected docs pages, repair-lane CHANGELOG claim wording (folded for I3)
  • Acceptance: docs describe the new user-visible surfaces (hallways, entity lexicon rebuild, kind/property pushdown semantics including the as-indexed vs as-on-disk divergence, rerank-eval language split, dead-weight view); links resolve; English throughout.
  • Depends on: I1

Task I3: Version bump + CHANGELOG + full local gate

  • Files: package.json, manifests via bun run scripts/sync-version.ts, CHANGELOG.md
  • Acceptance: CHANGELOG.md gains the ## [1.80.0] - <date> entry and its compare link reference; version is 1.80.0 with sync-version.ts --check green - all inside this PR before any push; the CI validate job's gates pass locally in CI order (Bun 1.4.0, Python 3.11). Windows job is left to CI.
  • Depends on: I2