Bug Fixes

• cli: agent activation: snippet rewrite, index-root walk-up, -n trap, stale-warning dedupe (#280)
• core: harden legacy ~/.vera detection (hidden files, unreadable dir) (#277)
• cli: report unproven vera launchers instead of ignoring them (#249)
• eval: match CPU vendor case-insensitively in host_cpu provenance test (#273)

Other Changes

• release: install cargo-zigbuild with its ziglang dependency in one venv
• repo: public-launch cleanup pass
• run CI on stock GitHub-hosted runners, reserve Blacksmith for releases (#278)
• release: build Linux targets with cargo-zigbuild (glibc 2.28, real musl C++ runtime) (#276)

Full Changelog: v1.4.0...v1.4.1

Features

• retrieval: enable vector filter-during-scan by default (#268)

Full Changelog: v1.3.1...v1.4.0

Features

• cli: port path-filter hint to vera grep via GlobMatcher (refs #250, PR #221 1be6ee4) (#258)
• ranking: add filename stem gating knobs with TDD (#257)
• vector: filter-during-scan eligibility map and flat scan (#255)

Bug Fixes

• tests: close remaining flakes from #262/#263 (LAST_HYDRATION_COUNT, cargo_bin, set_var) (#264)
• cli: recognize the shim by matching what the installer wrote (alternative to #260) (#263)
• setup: ask Direct3D 12 whether DirectML can run (#259)

Other Changes

• vector: harden filter-during-scan with typed fallback, scan guard and bulk matcher (#256)

Full Changelog: v1.3.0...v1.3.1

Features

• retrieval: toggleable hypotheses knobs for issue #196 (DEFAULT OFF) (#244)
• eval: add host CPU model and ranking env provenance with TDD tests (#242)
• retrieval: toggleable ranking signals for issue #196 with mechanism rationales (#239)
• cli: streamline Qwen first-run to single key and auto protocol (#237)
• cli: ship Qwen OpenRouter preset and setup hardening (#236)
• reranker: protocol/capability model and config plumbing (#230)
• consumers: update split-symbol consumers to bare+part_index (#226)
• core: split-symbol identity bare+part_index (#225)
• indexing: add completeness marker and flock guard for reuse race (#224)
• eval: ship reuse_index with identity gates (#223)
• references: record call receivers and filter callers by them

Bug Fixes

• ci: remove docs paths-ignore so docs-only PRs run validation (#247)
• cli: correct retrieval.max_output_chars help default to 0 = unlimited (#235)
• retrieval: harden m4 reranker review findings (#234)
• progress: harden m3 scrutiny findings (typed cancel, clamp, dedup, tracker reconcile) (#233)
• core: harden m2 review findings and document VERA_MAX_OUTPUT_CHARS (#232)
• reranker: pin resilience - permanent 4xx no-retry, capped Retry-After/OpenRouter waits, cancellation, degradation (#231)
• indexing: honest progress denominator (indeterminate until ParsingDone) (#229)
• retrieval: filtered-vector loss at retrieval core (#227)
• retrieval: wait out reranker 429 quota windows when a wait cap is set
• core: cap query-time reads at max_file_size_bytes, key RST cache on cycle state
• cli: let doctor diagnose broken configs, probe DirectML, remove cargo binaries
• glob: memoize the recursive matcher and hint wildcard dir prefixes
• discovery: stat and sniff through one open handle, loop short reads
• parsing: canonicalize the include-cycle stack seed

Performance

• retrieval: deduplicate vector-store open per query (#197) (#245)
• retrieval: implement #197 latency work with bounded caches and staleness proof (#240)

Other Changes

• cli,eval: harden Qwen preset identity and harness provenance (#241)

Full Changelog: v1.2.0...v1.3.0

Features

• embedding: API document prefixes with staleness guard, ROCm and config hardening

Bug Fixes

• docs: release notes and refreshed benchmark numbers for v1.2.0
• retrieval: narrow the definition-site gate to fixture paths
• retrieval: gate content-symbol definition boost by content role
• security: surface transitive unsound advisories in cargo-deny
• storage: read-only index opens, batched stats aggregates, honest vector counts
• cli: doctor blank-env scoring, uninstall exit code, non-TTY agent install
• mcp,serve: frame size cap, watcher update coalescing, per-repo watchers
• parsing: fence-aware markdown, Python inheritance, archive-token scoping
• retrieval: deep-search robustness, BM25 operator sanitizing, reranker shortfall

Performance

• indexing: overlap parse, embed, and store stages of full builds
• retrieval: cache the indexed file list across queries
• indexing: windowed parse-embed-store with staged publish
• retrieval: paged hydration for the BM25 rejection tail
• retrieval: hydrate BM25 candidates in paged batches

Other Changes

• lint: satisfy clippy 1.98 collapsible-if and chunks-exact lints
• add pull request validation workflow and correct the MSRV
• core: share index staging suffixes and deduplicate BM25 test

Full Changelog: v1.1.0...v1.2.0

Features

• local-models: pin revisions, verify digests, multi-shard weights
• retrieval: make reranking work under the potion backend, default off
• parsing: attach doc comments to symbol chunks, keep bash dispatchers whole
• parsing: extract bash case arms as named sub-symbols
• parsing: extract Lua function-valued assignments as symbols
• local-models: allow overriding the local reranker via env
• embedding: make potion-code-16M-v2 the default local backend
• core: honor pinned revisions for local embedding models
• cli: offer indexing on first search
• eval: pin Semble provenance and metric contract
• eval: configurable model lanes, task slicing, and full provenance (#141)

Bug Fixes

• security: descriptor-relative source reads via cap-std
• serve: keep in-flight model loads when the initiator disconnects
• parsing: bound RST include fan-out, cache expansions
• cli: align tests with reranking default, show sub-ms stage timings
• eval: match ground-truth paths by path-boundary suffix
• cli: repair no longer rewrites backend or reranker state, and trailing-period queries classify correctly
• core: install rustls provider before blocking prefix fetch
• cli: reject the structural query routes and sql cannot honour (#137)
• parsing: stop body-spanning containers swallowing their members (#131)
• embedding: apply jina-embeddings-v5's Query/Document retrieval prefixes (#95)
• embedding: pool jina-embeddings-v5 on the last token, not the mean (#94)
• serve: cache each model per slot instead of rebuilding it per request (#134)
• mcp: fuse multi-query search results instead of concatenating them (#124)
• retrieval: break RRF ties on a stable key so results stop varying per process (#127)
• retrieval: time query embedding and vector search separately (#113)
• core: re-root git scope paths onto the indexed directory (#138)
• cli: emit one JSON document from vera uninstall (#128)
• cli: exit non-zero when vera doctor reports a failing check (#136)
• indexing: do not follow symlinks out of the repository (#111)
• mcp: surface a stale-index warning in tool results (#130)
• core: single-source the Language wire name and honour retrieval.max_rerank_batch (#125)
• embedding: single-source the CodeRankEmbed query prefix (#129)
• cli: honour --client and --scope on vera agent sync (#110)
• parsing: Rust inline mod swallows every symbol inside it (#126)
• parsing: advance chunk coverage monotonically past nested symbols (#115)
• indexing: write hashes and the freshness stamp after the stores are rebuilt (#123)
• retrieval: contain indexed file paths within the project root (#112)
• retrieval: resolve the reranker provider before acquiring its runtime (#90)
• docs(parsing): pin the naming convention for dotted JSX components (#88)
• parsing: select the signature grammar by path, like indexing does (#87)
• indexing: make update publication cancellation-safe (#83)
• upgrade: verify the version changed, not that it equals the release tag (#63)
• indexing: treat unreadable tracked files as stale (#77)
• retrieval: bound the KNN candidate pool instead of clamping silently (#65)
• parsing: parse .tsx with the tsx grammar and index JSX call sites (#58)
• parsing: name const-assigned functions in TS/JS (#59)

Performance

• storage: exact flat-SIMD vector scan sidecar, sqlite-vec fallback
• retrieval: reuse search stores across queries
• retrieval: rerank pooled candidates with coherence, coverage, and definition signals
• reranker: length-sort local rerank batches to cut padding waste
• mcp: build the watcher's runtime and provider once, not per debounce cycle (#133)
• retrieval: build the reranker lazily, only for queries that rerank (#132)
• cli: let the update cache answer the install method before detecting it (#114)
• storage: resolve KNN rowids in one query instead of one per hit (#86)
• indexing: parallelize update path and batch metadata writes/reads (#69) (#73)
• indexing: delete BM25 documents for all changed files in one pass (#84)
• storage: seek the index when deleting a file's vectors (#81)
• storage: index the lower() lookups instead of scanning (#80)

Other Changes

• deps: tantivy 0.25 -> 0.26.1, MSRV 1.86
• remove dead code and deduplicate indexing helpers
• cli: remove runtime config forwarding helper
• docker: use ephemeral buildx builders and cap Blacksmith job runtime
• publish the openvino docker image (#61)
• build ROCm image from digest-verified GHCR base mirror (#55)
• make buildx prune non-blocking so the daemon retry always runs
• retry failed docker publishes via the runner docker daemon
• checkout the requested tag in docker recovery workflow
• scope docker concurrency per variant, least-privilege permissions
• reuse release binary for docker images, drop redundant cache
• size release runners to the macOS baseline, add docker cache repair
• move musl build to 8vcpu to tighten the wall-clock band
• balance release runner sizes to equalize build wall-clock
• serialize docker publishes, pass tag via env instead of template expansion
• speed up release builds and fix rocm docker publish

Full Changelog: v1.0.1...v1.1.0

Features

• Merge pull request #39 from citron07r/fix/index-no-progress
• cli: add --no-progress to vera index and gate progress on an interactive stderr

Bug Fixes

• Merge pull request #47 from VeraTools/fix/agent-sync-non-destructive
• Merge pull request #40 from citron07r/fix/setup-non-interactive
• Merge pull request #43 from citron07r/fix/upgrade-verify-applied
• Merge pull request #42 from citron07r/fix/vector-error-context
• Merge pull request #33 from citron07r/fix/coreml-reranker-cpu-provider
• cli: quiet auto-sync, CRLF-safe snippet refresh, doc accuracy
• setup: reject empty API credential env values
• cli: stop agent sync from overwriting user-edited markdown
• storage: log KNN clamp and test error chain formatting
• setup: honor complete API credentials in non-interactive mode
• upgrade: report unconfirmed upgrades as not applied
• upgrade: verify the applied version instead of trusting the installer exit code
• storage: clamp KNN k to the sqlite-vec limit instead of losing vector search
• setup: make flag-driven setup work without a terminal
• retrieval: run the reranker on CPU under CoreML

Other Changes

• add CodeRabbit configuration tuned to project guidelines
• upgrade: single-source the version-match decision
• migrate release and docker workflows to Blacksmith runners
• point repository references at the VeraTools org

Full Changelog: v1.0.0...v1.0.1

Features

• fix(build): fail loudly when vendored grammars are missing
• Merge experimental graph augmentation (off by default)
• retrieval: experimental structural graph augmentation (VERA_GRAPH_AUGMENT)
• retrieval: rerank path-glob scoped searches
• parsing: index TS/JS/Java class methods and Java enum methods
• search: repeatable --path filter, symbol-type aliases
• reranker: support Voyage rerank endpoints
• config: support embedding model aliases
• reranker: prefer API reranking over local when credentials configured
• Merge pull request #31 from freaksdotcom/bhigh/vera-mcp-cancel
• Merge pull request #30 from freaksdotcom/bhigh/vera-mcp-progress-reporting
• update: limit files per run
• update: report progress
• serve: add idle-timeout model cache with per-request default
• serve: cancel in-progress inference when client disconnects
• serve: add vera serve command with pure inference API
• retrieval: parallelize hybrid search and strengthen ranking signals
• retrieval: add BM25 stemming, stronger definition ranking, concept augmentation
• retrieval: add potion backend and cached search
• eval: semble benchmark integration, full-pipeline adapters, benchmark_root scoping
• retrieval: adaptive RRF weighting, identifier BM25, file saturation decay
• structural: index explicit type relations
• structural: add agent-oriented search intents
• search: add git scopes, explain-path, and AST queries

Bug Fixes

• eval: harden benchmark harnesses against partial runs
• core: parse git scope paths as NUL-delimited records
• cli: cancellable indexing and local-setup reranker hygiene
• serve: load only the needed provider per request
• local-models: harden model asset and runtime setup
• indexing: close staleness and removal gaps in incremental updates
• retrieval: honor cancellation and harden hybrid search
• retrieval: correct exact-match and query-classification edge cases
• mcp: correct operation nouns in git-scope schema descriptions
• benchmarks: drop fake vector-only mode from run_vera_benchmarks
• eval: fail loudly on adapter errors, guard empty sets, harden agent-bench reruns
• build: assert vendored grammar scanner.c exists
• serve: honor return_documents and guard eviction of in-use models
• mcp: correct search_code and path filter schemas
• parsing: TS interface methods and type-relation correctness
• core: git revision validation, update atomicity, exact_paths separators
• cli: agent snippet truncation, sync refresh gating, backend wizard
• build: fail loudly when vendored grammars are missing
• local: validate ONNX model integrity and repair safely
• indexing: read source files with lossy UTF-8 decoding
• search: clamp grep byte index to char boundary
• filters: handle Unicode paths in glob matching
• Merge pull request #22 from flier268/master
• Merge pull request #22 from flier268/master
• Merge pull request #29 from freaksdotcom/bhigh/vera-bounded-inference
• Merge pull request #25 from citron07r/fix/coreml-reranker-cpu-fallback
• Merge pull request #27 from citron07r/fix/openai-embedding-token-limit
• Merge pull request #26 from citron07r/fix/intent-breaks-bm25
• Merge pull request #28 from citron07r/fix/path-glob-directory-prefix
• retrieval: prevent scoped BM25 candidate starvation
• preserve indexing APIs and review edge cases
• indexing: cancel synchronous pipeline stages
• indexing: handle cancellation edge cases
• update: bound and cancel remote work
• inference: bound abandoned remote work
• search: match bare directory path globs as directory prefixes
• embedding: auto-truncate on OpenAI token-limit errors
• search: keep reranking on for intent searches; borrow in no-intent path
• search: keep --intent prefix out of BM25 queries
• local: stop CoreML reranker fp16 CPU fallback
• serve: fix cancellation gaps in local inference providers
• serve: prevent empty-key auth bypass and redact internal errors from API responses
• lint: resolve clippy warnings and fmt violations from vera serve
• local: recover from onnx provider session failures
• eval: filter tasks to corpus instead of failing on missing repos
• retrieval: harden bm25 and local indexing
• setup: prompt for api credentials
• resolve type relation generics, structural false negatives, and minor correctness issues
• eval: keep file-backed runs out of stdout
• search: tighten structural heuristics and add eval gating
• search: detect stale indexes and improve exact-match hints
• agent: keep managed Vera snippets current
• Merge pull request #18 from maxwellsocial/fix/voyage-ai-batch-limit
• cli: surface global output flags in help
• search: preserve generated regex hits and global flags

Performance

• retrieval: skip rerank when pool has no surplus over result limit
• local: scale CoreML batch size from unified memory
• namespace-aware embedding cache key prevents cross-model poisoning (#7)
• serve: remove unnecessary clones in embeddings and rerank handlers
• retrieval: trim vector candidate overfetch

Other Changes

• eval: share search_with/index_with between the two Vera adapters
• eval: single run_full_adapter helper for the three full-pipeline lanes
• benchmarks: extract shared metric helpers into bench_common.py
• benchmarks: prune archival result JSONs, keep v1 full-suite record
• benchmarks: prune 23MB of archival result JSONs
• eval: drop unreferenced Stability subcommand
• indexing: single chunk_file_for_update helper in update.rs
• mcp: dedup repeated tool schema properties
• core: share CompactResult and truncate_to_budget via presentation module
• cli: single AgentClient::META table for variants and names
• cli: split clap definitions into cli.rs, fix two lints
• retrieval: extract exact-match augmentation from search_service
• retrieval: split ranking.rs into query/score modules
• core: split local_models.rs into focused modules
• parsing: split extractor.rs into focused modules
• parsing: merge lisp define extraction behind keyword tables
• parsing: RawSymbol::at constructor, shared name-kind tables
• parsing: table-drive node classification, drop debug prints
• core: env override helper, preset builder, misc small dedups
• embedding,storage: merge do_embed paths, api_err and collect_rows helpers
• core: drop dead helpers, use rusqlite OptionalExtension
• embedding: single concurrent embed path, shared store setup
• retrieval: share chunk mapping, result key, and content predicates
• retrieval: remove dead search wrappers, collapse fuse_rrf_multi
• repo: drop patch artifacts, tracked pycache, spikes, superseded benchmark runs
• deps: semver-compatible dependency refresh
• deps: bump crossbeam-epoch for RUSTSEC-2026-0204, drop unused serde_json
• serve: load embedding and rerank models per-request instead of at startup
• structural: narrow impl search to Rust traits
• surface: converge on agent-first search workflows

Full Changelog: v0.12.13...v1.0.0

Bug Fixes

• cli: surface global output flags in help
• search: preserve generated regex hits and global flags

Full Changelog: v0.12.12...v0.12.13

Bug Fixes

• local: align CoreML reranker and search filters

Full Changelog: v0.12.11...v0.12.12