v1.38.0

Full Changelog: v1.37.1...v1.38.0

v1.37.1 — rerank reaches the MCP

1.37.1 — rerank reaches the MCP

search (MCP) gains an optional rerank parameter — same local-LLM
reordering the CLI's --rerank had (config rerank = true /
CODEGRAPH_RERANK=1 for the CLI default). The implementation moved to
codegraph-llm so both surfaces share one function; the LLM call runs on
spawn_blocking (the embedder-probe wedge class stays impossible), and it
degrades to the original order when no local model answers.

v1.37.0 — lean MCP payloads, semantic that can't dead-end, MCP↔CLI parity

1.37.0 — lean MCP payloads, semantic that can't dead-end, MCP↔CLI parity

Driven by the third field report (v1.36 head-to-head vs ripgrep on a 5-project
monorepo — structural queries won; three defects found, all fixed).

• MCP routes was unusable by its own client (232 KB single-line JSON,
  rejected by Claude Code): each route serialized the FULL graph node.
  Now: lean rows ({method, path, handler, file, line}) + limit/offset
  pagination + path_prefix/method filters; 560 routes ≈ 6 KB. Same lean
  sweep applied to every list-shaped MCP response that still leaked full
  nodes: implementers, dead_code, pinned callers. One shared lean
  serializer; full node detail stays behind get_node(id).
• semantic_search can no longer dead-end: with no embedder it degrades
  to lexical search and says so ("degraded": "lexical fallback — no embedder…") instead of hard-failing a tool it advertised. stats reports
  embedder_available up front so agents route around it. install.sh now
  builds with local-embed (bundled bge-small) by default — and
  indexstore on macOS — so the stock install answers semantic queries out
  of the box. CODEGRAPH_NO_EMBEDDER=1 forces lexical-only (also the
  deterministic test hook).
• MCP↔CLI naming parity: semantic-search/semantic_search →
  semantic, blast-radius/blast_radius → impact,
  trace-path/trace_path → trace, graph-query/graph_query →
  cypher, dead_code → dead-code (joins the earlier stats → status).
  Agents translate MCP names to CLI constantly; every advertised name now
  resolves.
• Regression tests: routes payload cap (<25 KB on 300 routes) + pagination +
  filters, alias resolution for every MCP name, forced no-embedder
  degradation with embedder_available:false in stats.

v1.36.0 — speed where it's visible + every fix is a release gate

1.36.0 — speed where it's visible + every fix is now a release gate

Driven by the second field report ("we should be 10–20× faster than grep,
we're tied"): the tie was the freshness probe, not the answer.

• Probe parallelized: the two root git ls-files spawn concurrently,
  nested-repo enumerations fan out on rayon, and the 8k-file stat sweep is
  parallel. Default-path query on an 8.7k-file monorepo: 0.21 s → 0.10 s —
  vs rg's 0.21 s without any freshness guarantee. Warm path (--no-autoheal
  / MCP between heals): 0.00–0.02 s — 10–100× rg. The honest framing: rg
  pays full scan per query; codegraph pays a freshness proof once and ~0
  after.
• Constant-path routes recovered: @Delete(SOME_CONST) endpoints are
  real — they now emit with the constant as a «symbolic» segment +
  path_unresolved: true instead of being dropped (66 real endpoints on the
  field monorepo; route accounting now closes exactly: 524 decorators =
  458 literal/bare + 66 symbolic).
• --indexstore early-return bug: with zero file changes the flag never
  reached the merge path — a re-merge no-op'ed. Fixed; field monorepo now
  carries a 46,110-edge Swift compiler oracle and codegraph audit measures
  96.1% precision (SelfThisMember 100%, SameFileUnique 96.6%,
  FieldTypeMember 83.3% on a small sample) — served via MCP stats.
• codegraph stats aliases status (agents guess the MCP tool name).
• Doc noise capped: identifier search keeps docs to ≤5 rows (one per
  file) when code answered the query.
• Release gate: scripts/release-qa.sh runs clippy-clean, the full
  suite, crash-recovery + determinism e2e, and the eval receipts — every
  field-reported bug fixed in 1.34–1.36 has a pinned regression test
  (enumeration classes, flock exclusion, MCP empty-graph/dead-root/
  generation-bump, --files machine contract, NestJS route shapes, DI
  narrowing, search ranking).

v1.35.0 — field-test fixes: the MCP can't lie, NestJS answers

1.35.0 — field-test fixes: the MCP can't lie, NestJS answers

Driven by a promom field-test report (11.6k files, TS+Swift+Python) that
found two red bugs and three weaknesses. All fixed, all regression-tested.

• MCP empty-graph bug (red): codegraph init wired the USER-GLOBAL MCP
  registration with one repo's absolute --path — the last-initialized repo
  won globally, and when that repo moved every project got a confidently
  empty graph ("no callers", nodes: 0) while the CLI was healthy. Fixed
  three-deep: registration is now cwd-following (no --path); a dead root
  refuses to serve at startup; and an EMPTY graph refuses to answer at all —
  every tool returns a diagnosis instead of clean emptiness (stats stays
  reachable and reports EMPTY_GRAPH). Regression suite proves a generation
  bump under a running server is served fresh on the next call.
• NestJS routes (red): decorator routes (@Get(':id'), bare @Get())
  never matched the leading-/ requirement, and @Controller('prefix') was
  never joined (including the @Controller(…)\nexport class shape where the
  decorator hangs off the export_statement). e2e/spec files polluted the
  answer with test traffic. Object-form @Controller({ path: 'x', … }) is
  read too, and a constant path (@Delete(SOME_CONST)) is skipped instead of
  fabricating "/" — a made-up path is a wrong answer, not a route. promom:
  120 noisy routes → 494 fully-prefixed real ones, zero spec noise, zero
  fabricated paths.
• TS DI caller resolution (was ~0%): this.profileService.getUserProfile()
  with constructor(private readonly profileService: ProfileService) never
  resolved when the class NAME exists in two apps of a monorepo (global
  unique-or-drop). The caller file's import now disambiguates the type —
  same evidence class as ImportNarrowed, unique-or-drop preserved. promom:
  0/23 → fully resolved caller list (+465 edges). Swift resolution untouched
  (its tests pin the behavior).
• search noise: identifier search ranks ANY code symbol above Document
  fragments, and collapses doc hits to one row per file.
• stale audit: stats no longer serves an outdated audit under
  measured_precision — stale numbers move to stale_audit_not_current.
• index lock → OS flock (std File::try_lock, Rust 1.89 — zero deps):
  the kernel releases the lock the instant the owner dies, so the PID-stamp /
  dead-owner-steal machinery from 1.34 is deleted outright. Measured: kill -9
  mid-index → next query self-heals and answers in 69 ms.
• callers --files: unpinned name-level questions return the UNION across
  same-name definitions again (dominant-definition narrowing was a measured
  recall loss), definition files stay as labeled evidence, human notes moved
  to stderr. Eval receipts (SCIP ground truth, 68 questions): P 0.66 /
  R 0.98 / answer-rate 100% / 319 B vs grep 0.63 / 0.94 / 100% / 2,701 B —
  ahead of grep on every measured dimension. (The 1.33 "R 0.87" receipt was
  stale: the harness predated the v1.33 --files format.)

v1.34.0 — zero-false-negative enumeration + ops hardening

1.34.0 — zero-false-negative enumeration + ops hardening

Enumeration moved from directory-walking to git ls-files (tracked +
untracked-unignored), with the walker as fallback. This is not just faster —
it is MORE COMPLETE: git knows which files are tracked, and tracked beats
gitignore. Measured on a live monorepo: 26 real docs (SRS, bug
investigations) lived under a docs/* ignore pattern but were committed
anyway — the walker dropped them forever, the git tier indexes them.
Staleness probe on a 3.6k-file repo: no-op index 0.19s.

• Nested plain repos (monorepo of independent .git checkouts, no
  .gitmodules): ls-files silently skips their subtrees — untracked dirs
  carrying a .git are enumerated recursively; any level that can't take
  the git path falls the whole enumeration back to the walker.
• .codegraphignore applied on the git listing (root-level, cascades
  into nested repos) — no more walker fallback just because the file exists.
  Submodules (.gitmodules) still force the walker: dropping submodule
  symbols would be a false negative.
• Walker parity: hidden(false) — dot-dirs carry real content
  (.claude/ agent docs, .github/), and git enumerates them; junk
  dot-dirs stay excluded via EXCLUDE_DIRS.
• Engine-version gate: the rebuild stamp now includes the release
  version alongside PARSER_VERSION — every upgraded binary rebuilds
  automatically; a resolver change nobody remembered to stamp can no longer
  serve a stale-engine graph.
• Cross-process index lock: MCP server, its watcher thread, and parallel
  CLI runs serialize on a PID-stamped lock file; dead owners are stolen
  instantly (kill -0 liveness + stale-age window), so a kill -9 mid-index
  never bricks the next query — verified live: the orphaned lock was stolen
  by the very next search, which self-healed and answered.
• Binary sniff: a NUL byte in the first 8KB (generated blobs with code
  extensions — valid UTF-8) keeps the file manifested but contributes zero
  symbols.
• Ambiguous-candidate ranking (MCP callers): strongest resolved evidence
  first, cross-language ties broken toward the language family the textual
  evidence lives in. Ranking only — never changes which edges exist.
• CODEGRAPH_MCP_CONCISE=1: drops per-response coaching fields
  (_hints, explainer notes); coverage/_fallback/truncation notes always
  stay. Measured 265→166 B on a callers answer.

Receipts: branch-switch storm (150 files) heals partially in 0.12s with
exactly reversible node counts; git-vs-walker parity proven on an identical
tree (1200 == 1200 nodes); verify-determinism byte-identical on a 3.6k-file
repo; zero phantom edges on 71k- and 25k-edge live graphs.

v1.33.1

Full Changelog: v1.33.0...v1.33.1

v1.33.0

Full Changelog: v1.32.0...v1.33.0

v1.32.0

Full Changelog: v1.31.1...v1.32.0

v1.31.1

Full Changelog: v1.31.0...v1.31.1