The §7a.2 per-ref profile overturned the assumption the whole arc was
built on: resolveOne owns only ~93s of the kernel-scale ~433s batch loop.
Loop-stage attribution (CODEGRAPH_RESOLVE_PROFILE, shipped here) named the
rest: backpressure folds 111.2s, count guard 93.9s, batch reads 54.6s,
deletes/inserts/marks ~84s, settle 85.7s.
- Non-progress guard O(remaining)→O(1): the per-batch COUNT(*) walked every
remaining pending row (O(N²/batch) per run, 93.9s). The cleanup queries
now return summed SQLite , and zero-removals-from-claimed-work
is the guard signal — the DIRECT evidence the count diff inferred (a
mismatched-name resolver makes keyed cleanup no-op ⇒ changes=0). A real
COUNT runs only on that suspicious path and arbitrates exactly as before.
- Batch reads OFFSET→keyset (54.6s→O(batch)): OFFSET re-walked the
accumulated failed-row prefix every read; seeking past the last-seen
rowid is prefix-independent and enumeration-order identical.
- WAL valve caps scale with DB size (env still wins): every fold re-writes
hot pages (#1231 in bounded form — 111.2s at the flat 256MB cap);
soft=clamp(dbSize/4, 256MB, 2GB) trades ~4× fewer folds for a transient
WAL ≈ project size.
- CODEGRAPH_RESOLVE_PROFILE: per-outcome resolveOne histogram + loop-stage
attribution, main + workers, off by default.
Gates: dubbo dump byte-identical; suite 2,491 passed / 4 skipped (kernel
required). Kernel-scale payoff run lands in the plan doc next.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
A truncate checkpoint started against an ACTIVE writer wins the lock and
then blocks that writer for its entire backfill; after the edge-index
recreate's multi-GB single-transaction burst that exceeds the writer's 5s
busy_timeout and fails the index with 'database is locked' (§7a.2 record
run, EXIT=1 at kernel scale — the small mid-resolution truncates folded in
ms and masked the hazard). Barrier truncates (backpressure/foldNow) are
collision-free by construction: the writer is awaiting the valve.
Dubbo gate: exit 0, peak 81MB (barrier folds carry containment), dump
byte-identical. Valve + sizing suites 27/27.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Four §7a.1 instrumented-run findings, each measured:
1. File-size trigger + truncate-at-barrier: a fully-backfilled WAL still
grows the FILE without bound — the writer only restarts at frame 0 when a
commit finds zero reader marks, which the instrumented run showed never
happens (file marched 361→721MB through two COMPLETE backfills; 22GB by
phase end). backpressure() now also trips at 4× the soft cap on raw file
size and TRUNCATEs at the parked barrier; the timer path truncates
opportunistically after complete backfills. Dubbo peak: 251MB → 69MB at
the same 16MB valve; dumps byte-identical under aggressive folding.
2. cgroup memory credit: memory.current counts reclaimable page cache — a
post-parse container read 57MB of headroom on a 6GB box and silently
disabled the pool. inactive_file is credited back (the docker-stats
working-set convention); the same run now reads a sane 4.4GB budget.
3. Pool at 2 cores reversed: sequential resolution measured FASTER than
pooled-6-on-2 at kernel scale (853s vs 1,150s), and synthesis is
Amdahl-bound by cFnPtrEdges (306s of 358s) so pooling it bought nothing.
cpuCap = min(ap−1, 6), no floor: ap=2 → sequential is the fast path.
4. Parse floor of 2: one parse worker at a 2-cpuset measured 34% slower
(493s vs 369s) — main + store-worker don't fill the second core. Floor
restores the baseline (373.5s measured).
Plus the observability §7a.1 burned three 25-minute cycles for: valve
armed/fire/timer-pass/heartbeat lines, checkpoint-worker error capture,
pool sizing decisions (incl. the disabled path), backpressure-hook
presence — all behind CODEGRAPH_SYNTH_TIMINGS / CODEGRAPH_WAL_VALVE_DEBUG.
Suite: 2,490 passed / 4 skipped (kernel required). Kernel-scale record
runs with this build follow in the migration plan §7a.1.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Three §7a.1 run-1 lessons (kernel-scale 2c/6GB: EXIT=137, WAL 22.2GB with
the backpressure hook DEPLOYED):
1. TRUNCATE at parked barriers: a completed passive backfill bounds the
un-checkpointed backlog but the FILE only stops growing when a commit
finds zero readers holding WAL marks — rare while pool workers cycle
(dubbo debug baseline: file climbed monotonically through six completed
pass-1 backfills). At a parked barrier the no-reader window is
guaranteed, so chop the file there with wal_checkpoint(TRUNCATE)
(off-thread, 2s busy_timeout — a racing reader degrades it to a no-op).
2. Futility latch: when backfill gives up (pinned reader), parking again at
every over-cap boundary burns a 20-pass checkpoint attempt — each a
worker thread + fresh connection against a multi-GB DB — per batch. Two
consecutive give-ups now disable parking for 60s; a pinned phase degrades
to pre-valve behavior instead of OOM-amplifying.
3. CODEGRAPH_WAL_VALVE_DEBUG=1 surfaces valve decisions without the
caller's verbose plumbing, and give-up lines print under
CODEGRAPH_SYNTH_TIMINGS — run 1 failed silently because give-ups were
verbose-gated.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Pool sizing used os.cpus().length, which enumerates the HOST's CPUs: inside
a 2-CPU cpuset it sized 6 resolver workers (the §7a.1 false-'sequential'
premise) and 8 parse workers, and at true 8-core concurrency six ~1GB
workers OOM-killed a 7GB container (oom_kill=5) mid-synthesis — sizing had
no memory term and no override knob.
resolvePoolSize (pure, matrix-tested): explicit CODEGRAPH_RESOLVE_WORKERS
override (0 disables, cap 16); CPU term max(2, min(availableParallelism-1,
6)) — cpuset-honest, floored at 2 so true 2-core boxes keep pooled
synthesis's ~2×; memory term floor(budget*0.7 / clamp(0.2*dbSize, 256MB,
1.5GB)) with budget = min(freemem, cgroup v2/v1 headroom). Parse pool's
core input switches to availableParallelism. Dev machines are unchanged
(still 6 workers); the 8c/7GB kernel-scale container now sizes 4.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
At kernel scale the pooled resolution/synthesis superphase grew a 22GB WAL
on a 4.6GB DB (cg1212, §7a.1): autocheckpointing is deferred for the run,
and the valve's timer-driven passive checkpoints stay perpetually partial
against the pool's continuous reads — no mechanism ever completed a
backfill, so the WAL accreted the whole phase's write volume, blowing disk
and feeding page-cache pressure into the 8-core/7GB container OOM.
The valve's writer-side backpressure() hard-cap backstop existed but was
wired only into the PARSE orchestrator. Thread it into the resolution batch
loop at the double-buffer's one pool-idle boundary (batch settled, next not
yet fanned out), after the edge-index recreate, and through the synthesis
insert loops. Parked there, the backfill completes; readers re-enter at
SQLite's backfilled mark and the next persist commit wraps the WAL.
Dubbo validation, same build: valve@16MB peak WAL 251MB (floor = the
single-transaction edge-index recreate) vs defaults 914MB; dumps
byte-identical (441,270 rows); wall unchanged (11s). Suite 2,479 green.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
On CRLF checkouts (every Windows autocrlf clone) the JS reference's
block-continuation strip /^\s*\*\s?/gm finds a line start after the \r and
its greedy \s* consumes the \n, leaving a bare \r in the docstring; the
kernel's (?m)^ pass matched after \n only and kept \r\n. Caught by the O2
Windows VM leg (6 kernel-tsjs-parity failures), reproduced on macOS by
CRLF-converting the fixtures.
js_multiline_strip now replicates the JS anchor set (\n, \r, U+2028, U+2029)
for all five line-marker passes; CRLF variants of every torture fixture are
pinned in kernel-tsjs-parity, derived in-memory so nothing can normalize
them away.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
R1-R6 are done; §0 is now the open-work list in recommended order
(merge → Windows VM leg → P1 resolution → C/C++ port → long tail), and
§0a carries everything a fresh session needs: where the work lives, the
build/gate commands, the proven add-a-language recipe, and the paid-for
traps (encoding-dependent error recovery → defer policy, node-ID-string
dedupe, UTF-16 positions/slices, the exact-seam contract, crate+wasm
grammar lockstep). §1/§6 keep the original expectations with SUPERSEDED
annotations pointing at the measurements that corrected them; §7a now
carries the R6 numbers that make it the top open perf item.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
cg1212 (Linux kernel, 2 CPU/6GB): completes in 26.4min vs the ~27min
baseline with all new machinery active — no regression; graph scale
identical (2,048,664 nodes / 6,405,964 edges). The §6 parse expectation
(6m → ~2m) was mis-premised: the tree is ~99% C, an unported T2
language, so it transfers to the C/C++ port (R7). Resolution remains
the kernel-scale wall (19.2m, 73% — P1). The tree's own Python tooling
files: 99/99 byte-parity. kernel-parity.mjs now skips dangling
symlinks (Linux dtc fixtures).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Python (codegraph-kernel/src/python.rs) and Go (src/go.rs) join the
native kernel, mirroring the wasm extractors bug-for-bug. Python:
decorated_definition docstrings/decorators (decorates refs only for
bare-identifier decorators — the call-kind quirk), function-in-class →
method, module-level assignments always extract as variable, from-import
per-name binding refs, self.x fn-ref candidates as bare names. Go:
receiver methods with Recv::name qualified names + contains edges to the
first earlier struct of that name, type_spec struct/interface
classification with embedding→extends and interface method nodes,
composite-literal instantiates keeping the package qualifier, top-level
var/const initializer walks attributed to the declared symbol (#693),
2-hop field chains (#1276), New().Method() re-encode (#645/#608), and
the GO_SPEC fn-ref layers.
Grammars: tree-sitter-python 0.23.6 + tree-sitter-go 0.23.4 crates, with
wasm vendored from the same tags (parser.c sha-matched) — both were
2023-era in tree-sitter-wasms.
Gates: extraction sweeps 100% (flask 83/83, django 3,035/3,038 +3
error-file deferrals, gin 99/99, prometheus 978/979 +1); full-init
dump-diffs byte-identical on flask (10,833 rows), gin (17,540), django
(360,794), and prometheus (213,758); torture fixtures enforced in npm
test. DEFAULT_ROUTED now covers typescript/tsx/javascript/jsx/java/
python/go. Full suite: 2,471 tests pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Kernel-routed files ship their flat tables from the parse worker to the
store worker as buffers (tryKernelExtractRaw → kernelBuffers on the
result → KernelStoreBundle); the store worker decodes and finalizes
(finalizeStoreBundle shared with the object path so filter semantics
can never drift). Files with applicable framework extract() hooks keep
the decoded path; the no-writer fallback materializes via
materializeKernelResult. Byte-identical dumps re-verified on dubbo,
excalidraw, express, gson; full suite green (2,467).
Measurement (plan §4d): dubbo's parse-loop wall is 94% store-writer
busy time — the many-core fresh-index wall is single-writer SQLite
ingest, not extraction or main-thread work. d2s improves the writer
lane ~11% (structured-clone deserialization avoided on the writer) and
frees the main thread; the remaining many-core gap is a
store-architecture arc (deferred index builds, multi-file
transactions, buffer→bind), out of the kernel project's scope.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Java joins the native kernel (codegraph-kernel/src/java.rs), mirroring
the wasm extractor's Java paths bug-for-bug: package namespaces,
imports, javadoc, annotations→decorates, type_list inheritance,
static-final constants, enum constants, anonymous classes (including
the TS side's 0-based-line quirk on the extends ref), method_invocation
calls with the this.field unwrap and the Foo.getInstance().bar() chain
encoding (#645/#608), static-member value reads, method-reference
fn-refs (#756), value-reference edges, and the full Lombok member
synthesizer (#912: Getter/Setter/Data/Value/Builder/ToString/
EqualsAndHashCode/Slf4j-family with taken-member dedup). The shared
docstring/textutil modules moved to crate level. Grammar:
tree-sitter-java 0.23.5, with the wasm grammar vendored from the same
tag (parser.c sha-matched) replacing tree-sitter-wasms' 2023-era build.
Gate (plan §4c): extraction sweeps 100% — gson 262/262, retrofit
341/341, dubbo 4,048/4,048 — plus a Java torture fixture in npm test;
full-init dump-diffs byte-identical on gson (49,766 rows), retrofit
(62,735), and dubbo (441,266 rows); all R2/R3 repos re-verified; Linux
container runs all 23 kernel tests green under CODEGRAPH_KERNEL_EXPECT=1.
The gate caught a real cross-language bug: fn-ref dedupe and value-ref
self-target checks must compare node ID STRINGS, not node-table rows —
ids collide for same-(kind, name, line) nodes, which minified one-line
bundles hit routinely (retrofit's website JS exposed it; latent in the
TS/JS walker since R2, never released). Fixed in both walkers.
Benchmark honesty: dubbo fresh-init on an 11-core Mac is ~flat
(parse-loop wall 5,020→4,394ms; total ~11.3s both arms) because that
wall is main-thread-bound (reads + store), not worker-CPU-bound — the
§6 expectation assumed otherwise. Where worker CPU binds the kernel
delivers: dubbo on a 2-CPU/6GB container drops 27.8-28.6s → 22.3-22.8s
(~1.25×). The identified lever for the many-core headline is decoding
kernel buffers directly into store rows (skipping per-node JS object
materialization); the buffer contract already carries everything.
DEFAULT_ROUTED now includes java. Full suite: 2,467 tests pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Gate evidence (docs/design/rust-kernel-migration-plan.md §4b):
- Graph parity, byte-identical (stronger than the §5 ≤0.5% bar): full
codegraph-init dump-diffs kernel-vs-wasm on express (13,712 rows),
excalidraw (89,898), and vscode (2,378,238 rows) — identical bytes.
Python control repo (flask) identical + timing unchanged. The parity
harness is now ORDER-sensitive (emission order drives rowids, which
drive resolution order) and dumps come from the new
scripts/dump-graph.mjs (natural keys, no rowids/timestamps).
- The one real find, caught by the vscode tier: tree-sitter error
RECOVERY is encoding-dependent — byte-identical grammar sources and
the same core (0.25.10) recover erroring files differently under
UTF-8 (native) vs UTF-16 (web-tree-sitter) parsing; proven by
reproducing the wasm tree with a native UTF-16 parse. Policy: the
kernel defers any file whose tree has_error() to the wasm extractor
(silent 'defer:' signal, per file) — parity by construction on
erroring files (incidence 0-0.42% across the gate repos), and the
harness fails if deferrals exceed 10% so a broken kernel can't hide
behind the fallback.
- Retrieval invariants: canonical excalidraw flow (mutateElement →
renderStaticScene) connects end-to-end on the kernel-indexed graph;
synthesized-edge families present. Agent A/B is vacuous under
byte-identical DBs (same justification as #1320-#1322).
- Perf: vscode init 105.4s → 82.1s (1.28×) on an 11-core Mac;
excalidraw on a 2-CPU/6GB Linux container (the CI-runner envelope)
6.2-7.1s → 4.3-4.8s (~1.5×). Linux arm64 in-container build: all 22
kernel tests green under CODEGRAPH_KERNEL_EXPECT=1. Windows VM leg
deferred (VM stopped; prlctl start needs Parallels Pro) — benign: a
missing .node falls back to wasm, and the release matrix builds and
gates the win32 prebuilds.
- Full suite: 2,465 tests pass WITH default-on routing, so the entire
extraction corpus now exercises the kernel for TS/JS wherever a
.node is staged.
DEFAULT_ROUTED = {typescript, tsx, javascript, jsx}. Override:
CODEGRAPH_KERNEL_LANGS (replaces the set) / CODEGRAPH_KERNEL=0 (kill).
Changelog entry added under [Unreleased].
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Replaces the R1 seed .scm emitter with a bespoke Rust walker
(codegraph-kernel/src/tsjs/) that mirrors TreeSitterExtractor's TS/JS
paths function-for-function: declarations (incl. #808 field/property
classification), qualified names, docstrings (#780 wrapper climbs),
signatures, imports/re-exports + per-binding refs, calls with
receiver-qualified callees (#1230 literal-receiver skip), instantiations,
decorators, inheritance, type annotations (#381), type-alias members +
tuple contracts (#359/#634), React component recognition (#841
forwardRef/memo/styled), object-of-functions / zustand-through-middleware
/ RTK Query endpoints + generated hooks / vuex + pinia store shapes,
function-as-value capture with the flush gate (#756), and value-reference
edges with the shadow prune (#895/#897). The generic query emitter is
deleted — extraction parity needs logic .scm can't express; future
languages get walkers too (migration plan §4a).
Positions and JS string-slice semantics are emitted in UTF-16 code units
natively, so kernel output is byte-identical to web-tree-sitter's — no
column diff class exists.
Parity evidence (macOS): scripts/kernel-parity.mjs (full-object multiset
diff per file) — this repo 353/353 files, excalidraw 643/643 (10,650
nodes / 10,726 edges / 68,307 refs), plus torture fixtures checked into
__tests__/fixtures/kernel-parity/ and enforced in npm test by
kernel-tsjs-parity.test.ts. The strict compare caught one real decoder
bug the loose harness missed: refs must NOT carry denormalized
filePath/language at the extractFromSource seam (the store fills them).
Perf: extraction 2.6× single-thread on excalidraw (487ms vs 1,255ms,
identical outputs). Routing stays opt-in (CODEGRAPH_KERNEL_LANGS) until
the R3 equivalence gate (large repo, DB dump-diff, retrieval invariants,
agent A/B, Linux/Windows) passes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Phase 0 of the Rust extraction-kernel migration (docs/design/
rust-kernel-migration-plan.md, now checked in with §3a recording the
shipped state):
- codegraph-kernel/ napi-rs crate: extractFile(path, content, language)
→ five flat buffers (meta/nodes/edges/refs/arena), one JS boundary
crossing per file. Node ids computed Rust-side, byte-identical to
generateNodeId (pinned by test vector). Reserved per-node metrics slot
for the Arc 3.2 code-metrics work.
- Generic .scm-driven emitter (@def.<kind>/@name/@ref.<kind> captures,
byte-range scope stack → ::-joined qualified names, contains edges,
refs attributed to the innermost enclosing symbol). Seed TS/JS queries
are smoke-level; R2 replaces them with the full port.
- Routing seam in extractFromSource with per-file wasm fallback.
DEFAULT_ROUTED is empty — no behavior change until a language passes
its equivalence gate (R3). Dev opt-in: CODEGRAPH_KERNEL_LANGS. Kill
switch: CODEGRAPH_KERNEL=0. Loader verifies ABI + kind tables before
routing; EDGE_KINDS became a runtime array because kind order is now
wire contract.
- Grammar-source parity: vendored TS/TSX/JS wasm grammars built from the
exact crate revisions (tree-sitter-typescript v0.23.2,
tree-sitter-javascript v0.25.0, checked-in parser.c, ts-cli 0.25.10) —
the tree-sitter-wasms builds were 2023-era, which the new
kernel-grammar-parity test caught on day one. Production TS/JS parsing
gets 2.5 years of grammar fixes; full suite green (2456 tests).
- Build/release wiring: scripts/build-kernel.sh + npm run build:kernel;
release.yml kernel prebuild matrix (continue-on-error — the kernel is
optional everywhere, bundles fall back to the wasm path); bundles stage
lib/kernel/codegraph-kernel.node; release job runs the kernel suites
with CODEGRAPH_KERNEL_EXPECT=1.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two hazards found by running today's full stack against the Linux kernel
(70,129 files) in the cg1212 repro container:
1. The parallel-synthesis fallback retried a worker-failed pass on the MAIN
thread. At multi-million-node scale a worker failure is usually a memory
ceiling, so the retry would OOM the process and take the whole index with
it. Above 1.5M nodes a failed pass is now skipped with a clear stderr
message (its synthesized edges are absent; the index completes). Below
that, the main-thread retry stays — small-scale worker crashes are
transient and the retry keeps coverage.
2. endBulkEdgeLoad rebuilt all four edge indexes in one synchronous span —
measured 79s at kernel scale, past the #850 liveness watchdog's 60s
stall window. A daemon-triggered re-index would have been SIGKILLed right
after doing the work. Now async with an event-loop yield between builds,
keeping each stall to a single index (~20s at kernel scale).
Validation: full Linux kernel index to completion in the repro container —
2,048,674 nodes / 6,405,964 edges, EXIT 0, zero passes skipped, on a 2-CPU
VM (worst case: pool disabled, sequential resolution + synthesis) in ~27min.
Phase walls: parse 6.0m, resolution 19.5m (incl. synthesis 6.3m, recreate
79s), maintenance 74s. Suite green (2444).
Also adds docs/design/native-extraction-kernel.md — the spike-validated
design for the native extraction kernel (Rust parse+walk over dubbo's Java:
202ms rayon / 1.07s single-thread vs 4.7s for the current wasm pipeline).
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The resolution persist's measured cost is B-tree maintenance on the edges
table's five indexes (offline replay of a 224k-edge resolution set: 2.8s with
all indexes, 1.1s with only the unique identity index, +0.3s to recreate the
rest). On big runs (same >=150k-ref gate as the resolver pool) the four
non-unique edge indexes are now dropped for the batch loop and recreated in
one pass each before synthesis.
Why this is safe:
- idx_edges_identity stays: INSERT OR IGNORE's dedup conflicts on it (#1034),
and its leftmost column is `source`, so the only mid-window edge reads —
resolution's supertype walks (implements/extends by source) — keep an index
via its prefix (verified with EXPLAIN QUERY PLAN).
- The window closes BEFORE synthesis, whose passes read kind-keyed, and on
every error path (finally).
- A crash inside the window heals on the next DatabaseConnection open —
schema.sql re-applies CREATE INDEX IF NOT EXISTS, same recovery as the FTS
bulk-load pattern this mirrors.
- Concurrent readers (a daemon serving the project mid-index) stay correct;
target/kind-keyed reads degrade to scans only for the window's duration.
dubbo (4,402 files): persists 4.0s -> 3.0s, fresh init 11.9s -> ~11.1s,
graph byte-identical. excalidraw (below the gate): untouched, byte-identical.
Recreation cost ~250ms, logged under CODEGRAPH_SYNTH_TIMINGS as
edge-index-recreate. Suite green (2444).
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The ~36 independent synthesis passes (callback/event/framework wiring) ran
sequentially on the indexer's main thread — 2.0s of a 4,402-file Java repo's
index, and the stage where kernel-class repos die (#1212). They now live in
an explicit registry (SYNTH_PASSES) and, when the resolver pool is alive
(>=150k-ref repos), fan out across its read-only workers: dubbo synthesis
2,024ms -> ~900ms (-55%), total fresh init 13.5s -> 11.9s. Graphs verified
byte-for-byte identical on both the pool path (dubbo) and the sequential
path (excalidraw).
Why this is safe: no pass's edges persist until the ordered merge, so every
pass sees the same committed post-resolution DB state in either mode, and
results merge in registry order regardless of completion order — the
first-seen dedup is unchanged. The pool now survives through synthesis
(destroy moved after it) instead of being torn down moments before the one
stage that could reuse it.
Robustness: a pass that fails on a worker (crash, OOM) is retried on the
main thread — a synthesizer blow-up now costs one worker instead of the
whole index, which is half the #1212 story on very large repos.
Also: ref-row cleanup deletes now run as one transaction with a cached
statement instead of one implicit commit per 500-row chunk (mechanically
fewer WAL commits; matters most on HDD-class storage). A set-based rewrite
of failed-ref parking was tried, measured ~zero on NVMe, and dropped — the
remaining persist cost is edge-index B-tree maintenance, not statement
dispatch.
SYNTH_PROGRESS_STEPS now derives from the registry (passes + fixed marks);
the pin test counts registry entries plus literal __mark sites.
Suite green (2444). Sequential-path timing unchanged on excalidraw.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fresh init on a 4,402-file Java repo (dubbo): 18.2s -> 13.5s (-26%), with
the resolution phase going 12.6s -> 7.9s (-38%). Graphs verified
byte-identical on both the pool path (dubbo) and the sequential path
(excalidraw).
Two changes:
1. The fastInit+pool path restored WAL for the resolver workers but left
wal_autocheckpoint at its default, so the persist loop inline-checkpointed
hot pages all phase long (#1231's pathology inside resolution — measured
at 58% of resolution wall). Checkpointing is now deferred behind the
bounded valve and folded once at maintenance, mirroring the deferWal path.
2. The resolution loop is double-buffered: batch k+1 is prefetched (OFFSET
past batch k's still-pending rows, under an explicit ORDER BY rowid) and
fanned out across the pool while batch k's ref cleanup runs on the main
thread. Batch settle-waits dropped 2572ms -> 117ms.
Correctness invariant found by the byte-identical gate and now documented in
the loop: batch k+1's resolution READS batch k's edges (resolveMethodOnType
walks supertype chains over extends/implements edges that resolution itself
inserts), so edges must persist BEFORE the next batch fans out; only the ref
cleanup overlaps.
Also extends the CODEGRAPH_SYNTH_TIMINGS instrumentation with phase labels
(grammar-init, parse-loop, fts-rebuild, resolver-reinit, resolution,
callback-synthesis) and pool timings (worker open/resolve, per-batch mode,
persist), so the next profile is one env var away.
Suite green (2444). Sync path timings unchanged. Pool floor re-validated:
forced-on at 40k refs is still net-slower, so the 150k threshold stands.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Three compounding defects (#1196) made a query bag of object-literal
keys (`profileInfo isTrialEligible quotaInfo billingMethod`) return
unrelated results while the defining files never surfaced:
1. Step 5b title-cased interior humps (profileInfo -> Profileinfo) and
then compared case-SENSITIVELY, dropping every row SQLite's
case-insensitive LIKE had just recovered. The hump lookup is now
case-insensitive with an explicit uppercase-at-match requirement.
2. Step 5b/5c's kind whitelist held only type-like kinds — dead code on
method-centric codebases. Callable kinds (function/method/component)
are fetched as a SEPARATE LIKE batch so hot single-word terms can't
crowd classes out of the length-ordered 200-row batch.
3. explore's named-symbol seeding was exact-name only; a field token
seeded nothing. A camelCase token with ZERO exact defs now seeds its
camel-infix definers (callables, hump-boundary or prefix, shortest
first, capped at 3) — bare lowercase words keep the #1252 stopword
guard untouched.
The reporter's acceptance query is a pinned e2e test (definer files
present, exact-name seeding unaffected). excalidraw probe: the
canonical flow query (mutateElement renderStaticScene) is byte-
identical; NL queries shift toward more-central callables
(useUIAppState/getDefaultAppState over observer periphery).
Fixes#1196
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
`.gitignore: /repos/` lists `repos/` as ONE ignored entry, while the
CLI hint (#1156) suggests `includeIgnored: ["repos/a/", "repos/b/"]` —
the child spelling. findIgnoredEmbeddedRepos tested the opt-in matcher
against the PARENT path only, which a child pattern never matches, so
the documented opt-in silently indexed nothing and init looped the
byte-identical suggestion back at the user (#1295).
Ignored dirs that don't match as a whole are now descended (the walk
was already bounded: depth 4 / 2000 entries, and only runs when
includeIgnored is configured) and each nested repo root is matched
individually — parent spelling opts in everything under the dir, child
spelling exactly the named repos. findUnindexedIgnoredRepos gets the
same per-repo check so the hint stops nagging about repos that are
already configured while still naming unopted siblings.
Fixes#1295
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
", ".join(sorted(x)) resolved by bare name to a project function named
join — one nested inside a DIFFERENT function, so scope alone rules the
edge out. Both defects from #1230, fixed independently:
1. Extraction: a member call on a LITERAL receiver (string, number,
collection, regex — across grammars) emits no call ref at all. A
literal's methods are the language's builtins, never project
symbols; the bare-name fallback let them exact-match any same-named
project function. Silent miss, never a wrong edge.
2. Resolution: matchByExactName filters out candidates nested inside a
same-file FUNCTION container unless the ref originates within that
container's line range. Class members (parent is a class-like node),
top-level symbols, and C++ namespace prefixes (no parent node) are
untouched.
requests re-index: byte-identical (813 calls edges). excalidraw: -27
edges, all literal-receiver refs by construction. The issue's repro is
pinned: join has exactly one caller (format_fields), report_missing
has zero project callees.
Fixes#1230
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
target.conn.Exec("insert") with `conn *sql.DB` emitted a BARE `Exec`
ref (the receiver chain was dropped for non-identifier receivers), and
exact-match then bound it to the only local `Exec` — an unrelated
interface's method — fabricating an internal dependency (#1276).
Extraction now keeps Go 2-hop selector chains (`base.field.Method`),
and a dedicated matcher resolves them EXCLUSIVELY via two inference
hops: base's type from the enclosing scope (#1108 machinery), field's
declared type from the struct's own declaration lines (comment-
stripped, per-line — chi's "the tree router" doc comment otherwise
donates a phantom type). resolveMethodOnType validates the target.
Package-qualified field types are followed only when the package is
in-module — `handler http.Handler` must not bind a same-named local
decoy. Failure at any hop leaves the ref unresolved: chained Go
receivers never fall through to the bare-name strategies (they were
never emitted before, so no prior recall depends on that path).
chi before/after: node count stable (1,181); 8 correct field-chain
edges gained (mx.tree.FindRoute/InsertRoute/routes, validated,
including the unexported `node` type); the removed edges are the
prior bare-name guesses on external receivers.
Fixes#1276
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
reproStore.notifyJoinGuildStatus() after `import { reproStore }`
resolved its calls edge to the exported CONSTANT (resolvedBy:'import'),
while the identical same-file call resolved to the method via
local-variable receiver inference (#1108) — so `callers <method>`
missed every cross-file use and a widely-used method could look
unused (#1292).
resolveViaImport's member-descend now handles imported VALUES alongside
the #825 static-member case: when the base resolves to a
constant/variable, the value's type is inferred from ITS OWN
declaration lines in the exporting file (the shared #1108 pattern
table: `= new T(...)` initializers and type annotations) and the member
is resolved AND VALIDATED on that type via resolveMethodOnType. A
failed inference or validation keeps the existing constant edge —
never a fabricated one. Calls only; plain member reads still reference
the value.
excalidraw control: byte-identical graph (10,653 nodes / 19,483 calls
edges before and after).
Fixes#1292
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The CLI's bare-symbol branch passes includeCode=true to the
codegraph_node handler, but the symbol-pinned-to-file branch didn't —
so exactly when a user disambiguated an overloaded name to one file
(the point of -f), they got Location + trail with no code (#1284).
Fixes#1284
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
A directory deletion arrives as ONE event on the directory's own path.
That path has no source extension, so handleChange dropped it at the
isSourceFile gate before ever scheduling a sync — and the files inside
may never get events of their own (Windows's recursive watcher reports
only the top-most removed entry; FSEvents can coalesce a tree deletion
the same way). Every child record then sat stale in the index until an
unrelated edit happened to trigger a sync (#1285).
A non-source path that no longer EXISTS on disk now schedules the
debounced sync; the sync's scan-diff removes whatever vanished (already
correct — verified: manual `codegraph sync` cascades fine). Events for
live non-source files stay fully ignored, so build churn schedules
nothing.
Fixes#1285
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The #1242 fix (WAL deferral + checkpoint valve, the 26x win on
HDD-class storage) was wired only into indexAll. CodeGraph.sync never
touched wal_autocheckpoint, so every incremental run kept the default
1000-page cadence and re-triggered the #1231 per-page checkpoint
thrash — a 7-file sync took 2m 2s at 0-2% CPU on the reporter's
hardware, because the cost scales with the EXISTING database's hot
pages, not the change size.
sync now mirrors indexAll exactly: defer autocheckpoint + start the
valve for the run, fold the store phase's WAL before the post-store
reads, restore the interval in the finally. Same kill switch
(CODEGRAPH_NO_WAL_DEFER=1). Idle valve cost is one timer, so
watcher-frequency syncs stay cheap.
Fixes#1248
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
SEC_ATTR UINT32 LostName(VOID) — an unknown attribute macro before a
typedef'd return type — misparses in tree-sitter's C grammar: the macro
becomes the type, the return type the declarator, and the PARAMETER
LIST is stored as the function name ("(VOID)"). The C++ grammar
recovers this shape via recoverMangledCppName, but in C the real name
never reaches the mangled string, so only a pre-parse blank can help.
Attribute macros are project-specific, so the blank keys on structure:
line-leading ALL-CAPS token followed by TWO identifiers then `(` — the
`MACRO Ret name(` definition shape. Plain typedef'd returns, ALL-CAPS
calls, #define lines, multi-word builtin returns, and mid-line uses are
all rejected by construction. Offset-preserving like the C++ blanks.
curl re-index: 5,531 C functions before and after, zero name changes;
7 nodes in memdebug.c improve start-line accuracy by 1 (the macro line
no longer counts as part of the definition).
Fixes#1211
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
An out-of-line member definition inside a namespace block takes its
qualifiedName from the declarator's receiver, which is spelled RELATIVE
to the enclosing namespace — so `namespace simulator {
ManifestStartup::Output ManifestStartup::Apply(...) {} }` indexed as
ManifestStartup::Apply while the class node carried
simulator::ManifestStartup. Fully-qualified call sites
(simulator::ManifestStartup::Apply(...)) never resolved; callers and
file impact came up empty (#1291).
The receiver-based qualifiedName now composes the active namespace
prefix, anchored at the first prefix segment the receiver re-spells
(so `namespace sim { void sim::M::f() {} }` doesn't double-prefix).
namespacePrefix is only ever non-empty for C++ — Go/Rust/Kotlin/Lua
receivers pass through unchanged.
leveldb re-index: node count byte-stable (3,044), calls edges +6,
namespace-qualified method names 947 -> 1,252.
Fixes#1291
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
template<typename T> T Box<T>::get() stored qualified_name Box<T>::get —
the <T> qualifier never matched the class node indexed as Box, so the
method didn't link to its class, while the inline form of the same
method produced Box::get. ICU-shaped multi-line template parameter
lists leaked whole <…> blocks (newlines included) into qualified_name,
exceeding NAME_MAX for downstream consumers.
extractCppReceiverType now applies stripCppTemplateArgs (the #1043
normalization for base-class refs) to the receiver qualifier.
fmt re-index: template-arg-in-qualifier names 25 -> 4 (remaining are a
FMT_BEGIN_EXPORT misparse artifact and gmock conversion-operator names,
both distinct pre-existing shapes), node count byte-stable at 7,536.
Fixes#1286
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
cache.Put("a", 1), store.Get("config", out), bus.Handle("user.created",
h) — any verb-named method with a string first arg — were indexed as
HTTP routes (38 of 82 route nodes were false positives on the
reporter's 200 KLOC Go codebase). A registration's first argument must
now start with "/" (every router style), or be a Go 1.22
"METHOD /path" mux pattern on Handle/HandleFunc — which now also
extracts the real method instead of ANY.
Validated on go-chi/chi (212 real routes retained, all path-shaped)
and golang/groupcache (0 route nodes).
Fixes#1259
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
codegraph's glyphs were ASCII on every Windows console while
@clack/prompts drew its Unicode frame around them, so one index block
mixed `|` and `│` rails (#398). supportsUnicode() now mirrors the
is-unicode-supported detection clack bundles (Windows Terminal, VS
Code, ConEmu/Cmder, Alacritty, xterm-256color, JetBrains, CI), so both
systems always pick the same glyph family.
The shimmer worker's raw fs.writeSync(1) bytes still decode through the
console codepage (OEM codepages mojibake UTF-8 even under Windows
Terminal — the #168 regression to avoid), so:
- the raw path gets its own supportsUnicodeRawWrites() that stays ASCII
on win32 unless CODEGRAPH_UNICODE=1, and
- the persistent "phase done" lines move from the worker to the parent,
written via process.stdout (wide-char console API, codepage-immune) at
phase transitions — the main thread is alive there, it's delivering
the progress callback. Only transient, self-erasing animation frames
remain on the raw path, so ASCII never lands in scrollback.
Fixes#398
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
List commands (status, query, callers, callees, impact, files) embedded
ANSI color codes even when stdout was a pipe, and NO_COLOR had no effect.
One switch now decides color for all codegraph-authored output:
--no-color > --color > NO_COLOR > FORCE_COLOR > stdout TTY > CI.
Piped init/index/sync also stop emitting shimmer animation frames
(\r + erase-line rewrites) and print one plain line per phase instead;
a TTY with NO_COLOR keeps the animation but drops the color codes.
The detection mirrors picocolors' so @clack frames and our own lines
agree within a run.
Fixes#1281
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* perf(index): ~34% faster fresh indexing, byte-identical graphs
Profiling a fresh init on a medium TS repo (excalidraw, 657 files) showed
the main thread as the critical path: per-row SQLite statement calls,
repeated import-resolution walks, and per-row FTS trigger firings, with
the parse workers ~75% idle behind it. This lands the semantics-preserving
tranche of fixes:
- Multi-row batched INSERTs (nodes/edges/unresolved refs/name segments)
behind cached per-batch-size prepared statements; row order preserved,
so rowid-based resolution determinism (#1015) is unchanged.
- storeFileBundle: one transaction per file instead of four; nested
transaction() calls now flatten (BEGIN-in-BEGIN previously threw, so no
caller depended on nested rollback).
- Dedicated store-writer thread for the fresh-DB bulk path (bundles
applied in file order on a single writer connection; main thread does
no DB work during the parse loop). Kill switch: CODEGRAPH_NO_STORE_WORKER=1.
- Bulk FTS mode: drop the nodes_fts sync triggers during the bulk load,
rebuild once at the end; crash inside the window self-heals on the
next open.
- Per-context memos for resolveImportPath/findExportedSymbol + a per-file
exported-symbol index, invalidated exactly where clearCaches() already
resets the resolver's own caches.
- Fast-init on completely fresh DBs (journal in memory, no fsync until
the index completes; interrupted init re-runs from scratch). Kill
switch: CODEGRAPH_NO_FAST_INIT=1.
- MaybeYield returns undefined on the not-due path so per-ref yield
checks stop paying a promise + microtask hop each.
- Parse pool prewarm for bulk indexing; compile-cache enabled at CLI and
worker entry points.
Excalidraw fresh init: 5.11s -> 3.36s median (n=5, warm cache, M-series).
Graph dumps byte-identical across init, re-index, and sync paths; full
suite green (2403 passed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* perf(resolution): parallel reference resolution with canonical admission
Fan resolution batches across a pool of read-only worker threads, each
hosting a full ReferenceResolver over its own SQLite connection; results
are admitted on the main thread in chunk order, so edge insertion order,
row cleanup, failure parking, and deferred post-pass queues are exactly
the sequence the single-threaded loop produces. Per-ref inputs match the
baseline because the sequential path already resolves each batch against
the state committed BEFORE that batch.
Validated byte-identical on excalidraw (pool forced on) and apache/dubbo
(4,048 Java files): dubbo full index 39s -> 19s (2.05x) with identical
graph dumps (91,495 nodes / 223,953 edges).
The pool only engages when total pending refs clear a threshold (default
150k, CODEGRAPH_PARALLEL_RESOLVE_MIN to tune, CODEGRAPH_NO_PARALLEL_RESOLVE=1
to disable): measured on a ~58k-ref repo the workers' boot CPU contends
with resolution on the same cores and makes indexing slower, so small
repos keep the sequential path. When fast-init left the DB in
memory-journal mode, WAL is restored before resolution only when the pool
will run (readers + rollback-journal writers don't mix).
Also: sqlite adapter readOnly open support.
TreeCursor spine rewrite of the body walker was built, measured neutral
on real repos and equal in a 20k-child microbench (web-tree-sitter's
namedChild(i) is not quadratic in this binding), and rejected — per-node
JS<->WASM marshaling is the floor, which a traversal swap cannot remove.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Follow-up to #1299: the per-pass bar still parked on one number while a
single long pass ran — on C-heavy repos that's the fn-pointer dispatch
pass, which sweeps every C/C++ file four times (typedefs, registrations,
field propagation, dispatch sites) and dominates the linking phase.
The pass now reports a real fraction of its dominant work
(scannedFiles / files×4, at the same per-16-files cadence as its
cooperative yield), and the orchestrator surfaces instrumented passes'
fractions as fractional steps, throttled to whole-percent movement so
the UI message volume stays bounded. The mechanism is opt-in per pass —
any synthesizer that a real repo shows parking the bar can adopt the
same callback.
Verified on the 1,342-file C repo from the report: the linking bar now
moves through 88→89→90 where it previously sat at 88 for the whole
pass; graph byte-identical (50,520 nodes / 148,232 edges).
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Two first-run UX bugs surfaced by indexing a real 1,342-file C repo:
1. After 'Resolving refs' hit 100%, the ~40 dynamic-dispatch synthesis
passes ran with no progress surface, so the bar sat frozen at 100%
long enough to read as a hang (the C fn-pointer pass alone can hold
for a while on C-heavy repos). Synthesis now reports per-pass
progress through a new 'linking' IndexProgress phase, rendered as
'Linking dynamic dispatch'. The step total is pinned by a test to
the synthesizer's actual __mark() count so adding a pass without
bumping it fails loudly.
2. node:sqlite's ExperimentalWarning is emitted once per THREAD, so the
main process plus every parse worker printed it mid-index,
interleaved with the progress UI. All launch paths now pass
--disable-warning=ExperimentalWarning: both bundle launchers, the
Windows npm-shim invocation, and the CLI self-relaunch
(NODE_RUNTIME_FLAGS, deliberately excluded from the re-exec gate so
an older installed launcher never triggers a pointless re-exec, and
version-gated off nodes older than the flag).
Verified end-to-end on the same repo: zero warnings, live linking bar,
byte-identical graph (50,520 nodes / 148,232 edges). Full suite green.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
All seven published packages (@colbymchenry/codegraph + six platform
bundles) now have this repo's release.yml configured as their trusted
publisher on npmjs.com, so publishes authenticate via the workflow's
OIDC identity instead of a long-lived NPM_TOKEN. The runner upgrades to
npm 11 (trusted publishing needs >= 11.5; Node 22 bundles npm 10) and
setup-node no longer writes a token-referencing .npmrc.
README gains a 'Verified releases' section + badges: how npm provenance
and the GitHub Release attestations work and the commands to verify them
(npm audit signatures / gh attestation verify).
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
One-time, strictly opt-in prompt at the end of codegraph install and
codegraph upgrade to join the CodeGraph Pro beta waitlist (same list as
the getcodegraph.com homepage form). Nothing is sent unless the user
answers yes AND enters an email; either answer is recorded machine-wide
so no later install or upgrade re-asks, and --yes / non-interactive / CI
runs never see the prompt.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Every published artifact is now cryptographically verifiable as built by
this repo's Release workflow: npm publishes carry npm provenance (OIDC,
shows the Provenance badge on npmjs.com), and the GitHub Release bundles
+ SHA256SUMS get signed build attestations via
actions/attest-build-provenance, verifiable with
`gh attestation verify <file> -R colbymchenry/codegraph`.
pack-npm.sh now writes a repository field into the generated shim and
per-platform package.jsons — npm --provenance refuses to publish without
one matching the repo — and the root package.json gains the same field.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Post-batch cleanup deleted resolved refs (and parked failed ones) by
(from_node_id, reference_name, reference_kind) — no line/col. When one
caller had several call sites to the same callee and a batch boundary
split them, the first batch's cleanup removed every row with that key,
including later-batch siblings that were never attempted — their edges
were silently never created. On nlohmann/json this ate 422 real call
edges (write_cbor's 38 to_char_type calls indexed as 11).
Refs loaded from unresolved_refs now carry their row id through
resolution, and all three persist paths (sync resolveAndPersist, the
yielding retry pass, the batched drain loop) delete / mark-failed by
exactly that id. The key-tuple methods remain only as the fallback for
hand-built refs from the public API. Failed-parking gains the same
precision: outcome can differ per call site (receiver inference reads
the ref's line), so a sibling must not inherit another row's failure.
Also untracks the zz-scratch local test files that slipped into #1268
and gitignores the pattern.
Validation: red-green regression test (5 sites, batch size 2 — old code
kept 2 edges, fix keeps 5); nlohmann/json re-index is a strict superset
of the previous edge set (0 lost, 422 recovered, spot-checked against
source).
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(cpp): resolve explicit operator calls (a.operator+(b)) to the operator method (#1247)
tree-sitter-cpp can't parse an operator_name in field position: the
call_expression carries `function: <receiver>` plus an ERROR child
wrapping the operator_name instead of a field_expression callee, so the
extractor emitted a calls ref named just the receiver (`a`) and the edge
never resolved — while the operator method itself indexed fine.
Two-part fix, scoped to the explicit call form (infix `a + b` / `a[i]`
need receiver type inference and are tracked in #1258):
- extraction: recover the operator_name from the ERROR child and emit
`<receiver>.operator+` (`->` receivers normalized, `this->` emits the
bare name), like any other member call
- resolution: matchMethodCall's dot pattern now admits an operator
method part (cpp-gated; symbol chars failed the \w match), so
receiver-type inference + resolveMethodOnType validate the target —
a same-named operator on an unrelated class can't capture the edge
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(cpp): harden explicit operator-call recovery against real-world shapes (#1247)
Validated on nlohmann/json (dozens of explicit operator[] / operator* /
operator< call sites). Two refinements the synthetic fixtures missed:
- normalize spaced call-site operator names (`it.operator * ()`,
`other.operator < (*this)`) to the compact form definitions index as
- drop the ref for a complex receiver (`obj()->operator+`, member chains
ending in a call) instead of emitting a bare operator name: exact-name
fallback GUESSED among unrelated same-named operators (linked a
std::map operator[] call to an in-repo operator[]) — silent miss,
never a wrong edge
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(uninstall): remove the CLI binaries too, not just agent configs (#1071)
`codegraph uninstall` swept agent configurations and stopped — every
installed binary stayed behind, so `codegraph` still ran afterward. Three
disconnected paths each removed a fraction of an installation (uninstall:
configs; install.sh --uninstall: the bundle; npm preuninstall: configs +
npm's own package), and none cleared a shadowed second install — the
uninstall edition of the #1071 PATH shadow.
The uninstall now PLANS every install present on the machine — the bundle
layout(s) (running binary's own, the platform default, a custom
CODEGRAPH_INSTALL_DIR), the npm global package (found by asking
`npm root -g`, so nvm/fnm/volta prefixes resolve correctly), and the
bin-dir launcher link (only when it verifiably points into a detected
install) — confirms with the user, then removes them all. `--yes` skips
the prompt; the new `--keep-cli` flag keeps the old configs-only behavior.
Safety rules: a source checkout is reported, never deleted; a
project-local npm install is left to the project; on unix the default
install dir doubles as the machine state dir, so only the install
artifacts (versions/, current) are removed there — telemetry choice and
daemon records survive. Windows can't delete a running exe but can rename
it (the in-place upgrade's trick): a locked node.exe is renamed aside and
surfaced as a one-file leftover instead of failing the removal, and npm
is routed through cmd.exe (a direct .cmd spawn EINVALs on modern Node).
Planner/executor are split with injected side effects (the upgrade
orchestrator's convention) and unit-tested across the shadow case,
state-dir preservation, custom dirs, foreign-shim protection, and the
locked-exe dance; validated end-to-end on macOS against a fake HOME.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(uninstall): key path math on the target platform, not the host
Real-Windows validation caught it: the planner/executor used the host
path module, so win32 fixtures were meaningless on a POSIX host and
POSIX fixtures failed on the Windows VM. Same convention as
detectInstallMethod now — path.win32/path.posix chosen by the injected
platform.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(upgrade): route npm through cmd.exe on Windows — a direct npm.cmd spawn EINVALs on modern Node
Found while validating the uninstall change on the Windows VM: upgradeNpm
spawned npm.cmd without a shell, which every current Node rejects with
EINVAL (the CVE-2024-27980 hardening) — so `codegraph upgrade` on a
Windows npm install failed before doing anything. Verified live on the VM:
spawnSync('npm.cmd') → EINVAL; cmd.exe /d /s /c npm → works.
npmInvocation moves into the upgrade orchestrator (remove-binary imports
it from there — same direction as its existing imports, no cycle), and the
win32 test now pins the WORKING invocation instead of the broken one.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The recommended MCP config launches the local binary, so a server left
running drifts behind releases silently — users discover the version gap
only when something breaks. Per the reporter's preferred option 1, the
server now checks the latest GitHub release in the background on startup
(never blocking; reuses the upgrade command's release-redirect resolution
so the two can't drift) and surfaces a one-line notice on three surfaces:
one stderr line (the MCP host's server log), the initialize instructions
(with do-not-run-it-yourself guidance for the agent), and codegraph_status.
Discipline: results cache in ~/.codegraph/update-check.json shared across
every proxy/daemon on the machine — 24h TTL on success, 1h backoff after
failure, an outage never hides an already-known update, and a stale cache
re-kicks a background refresh so long-lived daemons keep noticing. The
initialize path is a memoized synchronous cache read (the respond-fast
contract holds), and both handshake answerers (session + proxy) share one
helper so they can't diverge. Never stdout.
Hardening: the latest tag arrives from a network redirect via an on-disk
cache and ends up inside agent-visible instructions, so only a canonical
vX.Y.Z rebuilt from PARSED semver fields is ever interpolated — a tag
carrying trailing text (parseSemver is not end-anchored) renders without
it, and a non-version tag renders nothing and counts as a failed attempt.
Off is off: CODEGRAPH_NO_UPDATE_CHECK=1 (dedicated) or DO_NOT_TRACK=1
(broad convention — already set by data-plane deployments) suppresses the
network call and the notice entirely. Documented in TELEMETRY.md.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
handleExplore's named-symbol seeding treats every identifier-shaped query
token as "a symbol the agent named" and grants its definition the
named-FIRST sort tier. Natural-language questions broke this assumption:
ordinary words exact-matched unrelated callables ("check" ->
WalCheckpointValve.check, "serve" -> query-worker serve, "initialize" ->
DatabaseConnection.initialize), and those collisions outranked — and within
the per-repo file budget fully displaced — the corroborated answer files,
forcing the agent back to Read/Grep. The >3-def single-pick fallback had
the same hole: on grpc, the #1064 flagship query "add a parameter to
NewClient" itself tiered balancerStateAggregator.add's file to slot #1.
Guard: a shape-precise token (camelCase, PascalCase, snake_case,
qualified) seeds unconditionally — it is an unambiguous symbol reference.
A bare lowercase word seeds only defs whose file another query token
co-names (that token is itself an exact symbol name defined in the same
file — the "check drain fire" sibling-bag shape), which an incidental
English-word collision never is. Applied by filtering cands ahead of both
branches so the overloaded-name fallback is covered too.
Validated per the retrieval playbook: deterministic probes on this repo
(collision queries fixed; sibling-bag and single-camelCase retained), and
baseline-vs-fixed probes on the #1064 repos — Alamofire and excalidraw
byte-identical, grpc improved (the add-collision file drops out and
clientconn.go + dialoptions.go lead). Full suite green.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Carries #1221 by @w0lan plus a hardening pass: property-receiver typing consults property-shaped declarations only (typed property / promoted ctor param / pseudoconstructor assignment / assignment-followed classic ctor and typed setter), so same-named locals and parameters can never mistype a property.
Co-authored-by: Roman Wolan <roman.wolan@morizon-gratka.pl>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* chore: ignore .kommandr/ directory
* fix(sync): resolve cross-file refs when an edit adds or removes the satisfying symbol (#1240)
Incremental sync scoped reference resolution to the changed files' own
refs, and a completed pass deleted every ref it failed to resolve — so
a symbol change in one file could never repair references in UNCHANGED
files, in either direction, until a full re-index:
- New-export case: a.ts imports/calls `greet` before b.ts defines it.
The failed refs were deleted at index time; when b.ts later gained
`greet`, nothing revisited a.ts — the calls/imports edges stayed
missing while status reported a clean index.
- Removal case: when a re-index (or file deletion) dropped a symbol,
the incoming edges cascade-deleted and the callers — whose resolved
refs had been consumed — never got a chance to rebind to an
alternative definition or reconnect when the symbol returned.
Fix, sharing one lifecycle:
- Schema v8: unresolved_refs gains status ('pending'/'failed') and
name_tail (last dotted segment, so `h.greet` is findable by `greet`).
Both resolver persist paths now park unresolvable refs as failed
instead of deleting them. All pending-work readers (batched drain,
non-progress guard, #1187 orphan sweep, status pendingRefs) filter to
pending, preserving their invariants and keeping status honest.
- Sync retry: after scoped resolution, failed refs whose name tail
matches a symbol name now present in the changed files are re-resolved
through a per-ref-yielding path (watchdog-safe, #1091 class). Names
matching >500 failed refs are skipped as external/builtin noise (#999
rationale).
- Removal side: createEdges stamps each resolution edge with its
originating reference (metadata.refName, + refKind when kind promotion
rewrote it). When the #899 restore misses a target or sync deletes a
file, the dropped edge is resurrected as exactly that ref — re-resolved
in the same sync (rebinding to an alternative definition) or parked
failed until the symbol reappears. Edges without the stamp (pre-upgrade,
synthesized) still drop silently: reconstructing from the target's plain
name would strip receiver context and risk a rebind a full re-index
would never make.
- Pure-removal syncs clear resolver caches so a long-lived daemon can't
resolve resurrected refs against the pre-removal graph.
Validated: issue repro now yields a graph byte-identical to a full
re-index; move/remove-readd/file-deletion scenarios all rebind or heal;
baseline-vs-new A/B on express and gin shows identical node/edge counts
and no timing regression (DB grows ~25% from the parked ref rows — pure
cache, reset by any full re-index). 8 regression tests added.
Fixes#1240
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Two fixes to make `codegraph upgrade` trustworthy in the terminal it ran in:
1. detectInstallMethod checked the bundle layout before the node_modules
path check, but the npm thin-installer's per-platform package IS a
complete bundle inside node_modules — so every npm install misdetected
as a standalone bundle, and upgrade curled install.sh into ~/.codegraph:
a second install that never wins the PATH race against npm's shim,
leaving `codegraph -v` permanently on the old version. Path-based
checks (_npx, node_modules) now win over layout sniffing, so npm
installs upgrade through npm again, in place.
2. After a successful swap, runUpgrade now probes the PATH-resolved
`codegraph --version` and reports the real outcome: a green
confirmation that this terminal already serves the new version, a
loud shadow warning naming the fix (`which -a codegraph`) on
mismatch, or the old soft new-terminal hint only when the probe is
inconclusive. Replaces the unconditional "open a new terminal if the
version looks unchanged" hedge. Skipped for npm-local installs, whose
binary PATH never serves.
Companion to #1239: the misdetection also broke its post-upgrade
`install --refresh` for npm users (the spawn resolved the stale shim).
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The liveness watchdog judged the main thread by heartbeat silence alone,
which cannot distinguish a true wedge (the #850 infinite loop it exists to
kill) from one long synchronous SQLite statement on severely degraded
storage — so it SIGKILLed valid, in-progress indexes (observed on a
150-IOPS throttled rig, and latent on real HDDs at scale).
The CLI index/init paths now hand the watchdog the project's DB + WAL
paths. On a silent timeout the watchdog child stats them first: if they
advanced during the silence, the block is a slow store making forward
progress — defer and keep watching; if not, kill at the base timeout
exactly as before. Deferral is bounded by a hard cap (10× the timeout) of
continuous silence so a wedge coinciding with unrelated file activity, or
I/O hung beyond any legitimate statement, still dies. The daemon path is
unchanged (no progress paths — pure heartbeat).
Validated with real spawned processes (defer-on-progress, kill-on-static,
hard-cap kill) and on the throttled rig: a 150-IOPS index under a 10s
watchdog window — 6× tighter than production, with store stalls measured
at 10-20s — completes cleanly where the old watchdog killed it, while
true-wedge kill latency is unchanged.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Parse timeouts are now judged by the worker's own clock: the base timer
only marks a job late (after a long synchronous store stall, Node runs the
timers phase before the poll phase, so the timer fired before an
already-delivered result was processed — killing workers over parses that
took milliseconds, even on 0-byte files); a result arriving before a 3×
hard-kill backstop is accepted, timed-out files are retried, and
CODEGRAPH_PARSE_TIMEOUT_MS overrides the budget. Grammar WASM bytes are
pre-read once on the main thread and handed to every worker, so
spawns/respawns load grammars from memory instead of re-reading a
saturated disk.
Bulk indexing defers WAL auto-checkpointing for the whole run: the default
1000-page interval re-writes hot B-tree/FTS pages into the main DB file
over and over — ~95% of all disk I/O under throttled measurement. A
WalCheckpointValve bounds WAL growth with off-thread PASSIVE backfill
passes (never blocking the writer or the #850 watchdog heartbeat), pauses
the writer for a full backfill if the disk truly can't keep up, and folds
the WAL at the parse→resolution boundary so post-parse reads never page a
bulk-write-sized WAL. Opt out with CODEGRAPH_NO_WAL_DEFER=1; tune with
CODEGRAPH_WAL_VALVE_MB.
Measured at 150 IOPS (HDD class): commons-lang 1526s → 59s with 0 dropped
files (was 8); guava-scale completes in 7.6 min with a full graph where
v1.3.1 needed 25 min for a repo 5× smaller. Unthrottled: no change.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
closureCollectionEdges scanned every method/function node in every
language, but its dispatcher patterns ({ $0( / { it( ) are Swift/Kotlin
trailing-closure syntax — on PHP/JS repos the pass can never emit an
edge, yet .push(/.add( fired its append gate on nearly every function.
Per match it computed the line via src.slice(0, idx).split('\n'), which
is O(source) per match and goes quadratic on match-dense generated
functions (two-byte content roughly doubles it). On a 12,860-file
PHP/JS app that was 20+ minutes of the "Resolving refs" tail — frozen
at 97% — and a #850 watchdog kill; profiled on CRMEB it was 127s of a
166s index for zero edges.
- Skip nodes whose language isn't swift/kotlin before any file I/O.
- makeLineAt(): lazy newline index + binary search, shared with the
emitter passes' per-file lineOf.
- Yield every 256 regex matches inside the scan's match loops so a
single pathological function can't starve the watchdog.
CRMEB (ThinkPHP, 2,913 files): 166.8s -> 72.7s, graph byte-identical.
Alamofire: closure-collection edges byte-identical (9 edges, 4 fields).
Drupal core control: graph byte-identical.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Large-codebase indexing died at the end of "Resolving refs" two ways:
watchdog kills of healthy work (24k-file Java on Windows, #1212 — third
iteration of the #1091/#1122 class) and hard OOMs (Linux kernel scale,
where v1.3.0 could not complete at any watchdog setting). Root causes:
~31 of 37 dynamic-edge synthesis passes ran start-to-finish with no
yield points, several materialized whole-graph snapshots (kotlin
expect/actual opened with getAllNodes() — 2M nodes in one array; the
C fn-pointer pass retained every C file's contents twice plus every
function node), and the post-index WAL checkpoint ran minutes of
synchronous IO on the main thread, killing even a successful index at
the finish line.
The pipeline tail now follows the same discipline as the rest: never
hold O(graph) in the heap, yield everywhere.
- All synthesis passes stream node-kind scans (cursors, not arrays) and
yield on time-budgeted checkpoints; language gates skip passes whose
filters a project's file languages provably can't satisfy.
- kotlin expect/actual filters SQL-side; c-fnptr caches are LRU-bounded,
units stream one file at a time, and the all-functions array +
write-only id map are gone; spring reads each .java once, not twice.
- runMaintenance moved to a worker thread (own SQLite connection);
per-file store commits chunk with yields behind a serialized flush
chain (preserving #1015 file-order determinism); resolver warm-up
streams the DISTINCT name set; resolution batch-tail and merged-edge
inserts run in bounded sub-transactions.
- Daemon: fixed a socket-handoff race that could leave a fresh MCP
session permanently silent (client-hello tail unshifted into a
flowing stream with zero listeners — the long-standing #662 test
flake was this real bug); first tool call no longer queues behind
the query pool's cold start (pool.ready gate).
Validation: Linux kernel (70,129 files, 2.05M nodes, 6.4M edges) fully
indexes in 27m8s on a 2-core/6GB container at default heap + default
watchdog; llvm-project (180k files) completes under 1GB RSS including
kill-and-sync recovery; synthesized-edge and full-graph parity are
byte-identical vs baseline on elasticsearch/redis/vim; the ex-flaky
daemon test passed 25/25 under load. Env-gated diagnostics kept:
CODEGRAPH_SYNTH_TIMINGS pass/phase timings, CODEGRAPH_MCP_DEBUG hop
tracing. Design record: docs/design/main-thread-stall-followup.md.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Ships Nix language support and a batch of fixes staged under [Unreleased],
including the Java/Kotlin (Spring) resolution performance fix (#1180).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
On a large Java/Kotlin Spring monorepo, reference resolution — not extraction —
dominated a full index (Spring Boot's ~9,650-file tree: extraction 62s,
resolution ~26min). The Spring framework resolver ran an uncached
getNodesByKind('constant') full scan + canonicalConfigKey() filter for EVERY
dotted `calls` ref (every list.add(), builder.build(), receiver.method()),
because the config-key branch gated only on "dotted java/kotlin", not on ref
kind. With ~1,100 constant nodes × ~200k dotted calls that is ~200M wasted
row-fetches/allocations.
Fixes, one theme — config-key constants bind config `references`, never `calls`:
- frameworks/java.ts: gate the Spring config-key branch on
referenceKind === 'references' (what @Value/@ConfigurationProperties emit) so
the `calls` flood skips the scan.
- name-matcher.ts: a `calls` ref no longer resolves to a yaml/properties config
node via matchByQualifiedName (service.process() vs the yaml key
service.process) — a wrong edge that also hid the real callee; it now falls
through to method resolution.
- resolution/index.ts: cache getNodesByKind in the resolver context (same
lifetime as nameCache). Fixes the same uncached-per-ref scan in the Drupal
hook_ resolver and is defense-in-depth for the Spring :prefix branch.
Measured (Spring Boot): resolution 269s→16.5s on a 4.3k-file module (16×) and
~26min→44.7s on the full 9.6k-file tree (~35×); graph byte-identical, full suite
passes. Adds a regression test (same key, two ref kinds, opposite outcomes).
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A Git super-repo whose `.gitignore` excludes its child repositories indexed
~nothing at the parent: CodeGraph respects `.gitignore` by default (#970,
#1065), so the excluded children were skipped and `codegraph init` printed
"Done" with 0 nodes — even though `init` inside each child worked fine. The
empty index was silent and unexplained.
`init`/`index` now detect the gitignored child repos they skipped when an
index comes up empty of symbols, name them, and — in an interactive terminal
— offer to index them (writing an `includeIgnored` entry to codegraph.json and
re-indexing on the spot); non-interactive runs print the exact codegraph.json
snippet to add. Gated on nodesCreated === 0, so a project that deliberately
keeps gitignored reference clones out of a working index is never nagged.
- extraction: findUnindexedIgnoredRepos — the inverse of discoverEmbeddedRepoRoots
(bounded, skips default-ignored dirs, respects existing includeIgnored)
- project-config: addIncludeIgnoredPatterns — create/merge codegraph.json,
idempotent, refuses to clobber malformed JSON
- cli: wire the detect-name-offer flow into both `init` and `index`
- tests: +13 covering detection, config writing, and the no-nag gate
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Carries #1133 forward onto current main (rebased for conflicts). A lean Unreal-Engine-style `.h` whose only C++ signal is `class ENGINE_API Foo : public Bar` (no public:/virtual/namespace/template) was misdetected as C and its class + inheritance edge silently dropped; looksLikeCpp now recognizes the export-macro-annotated class/struct shape, matching what blankCppExportMacros already recovers.
Fixes#1159.
Co-Authored-By: robertyluo <robertyluo@tencent.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Carries #1133 forward onto current main (rebased for conflicts). A lean Unreal-Engine-style `.h` whose only C++ signal is `class ENGINE_API Foo : public Bar` (no public:/virtual/namespace/template) was misdetected as C and its class + inheritance edge silently dropped; looksLikeCpp now recognizes the export-macro-annotated class/struct shape, matching what blankCppExportMacros already recovers.
Fixes#1159.
Co-Authored-By: robertyluo <robertyluo@tencent.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The integration tests that shipped with #1158 use inputs too small to trigger
the tree-sitter class_specifier collapse, so they pass with OR without the fix
(verified: 0/3 reproduce on the pre-fix build) — only the offset-preserving unit
tests actually guard the behavior. This adds a real guard: a ~240-in-body-macro
reflected class (the density of the real CharacterMovementComponent.h) plus a
UENUM whose values carry mid-line UMETA. Asserting the decorated members and the
enum are extracted flips false->true across the three blank passes — verified the
same assertions FAIL on the pre-fix source (2a06d9a) and PASS on the fix.
The full class collapse is emergent from real engine-header content that can't be
shipped (Unreal source is EULA-licensed); this reproduces the recoverable-member
signal the collapse leaves, which regresses if any of blankCppAnnotationMacroCalls
/ blankCppApiPrefixMacros / blankCppInlineAnnotationMacros is reverted.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Four gaps in the MyBatis mapper extractor, all reported and reproduced by
@ESPINS in #1182 and verified against main:
1. Single-quoted attribute values (namespace/id/refid/resultType/parameterType)
were dropped — the regexes hardcoded double quotes. Now accept either quote
via a backreference.
2. Tags inside <!-- ... --> produced phantom statement/include symbols. A
length-preserving, CDATA-aware pre-pass blanks comments before scanning,
keeping offsets/line numbers intact.
3. Legacy iBatis 2 <sqlMap> files had zero statement coverage (the root finder
gated on a <mapper namespace> root). It now also recognizes <sqlMap>
(namespaced and namespace-less DAO.method ids) and iBatis's extra
<statement>/<procedure> verbs — closing the gap with no new dependency
(option (c) from the issue; the batis-xml parser route is declined).
4. Two statements sharing a qualifiedName AND a start line (a vendor-split
databaseId pair on one line) collided on the node id, so INSERT OR REPLACE
silently dropped one. The id-hash now folds in the statement's byte offset;
the stored qualifiedName/startLine are unchanged so the Java<->XML bridge is
untouched.
Gaps 1 and 2 follow @ESPINS's fix-mybatis-quotes-comments branch. Tests add
extractor-level coverage for all four gaps plus a DB-level e2e that proves
iBatis statements land and both vendor-split nodes survive a real indexAll.
Co-authored-by: Jimin Lee <dlwlalsggg@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The shared daemon's inactivity backstop (#692) reaped the daemon after
maxIdleMs (default 30 min) of no inbound query bytes whenever a client was
still connected — without ever checking whether that client was actually
alive. lastActivityAt is fed only by inbound socket data and MCP has no
keepalive, so a genuinely-live session that just hadn't queried CodeGraph in
30 min tripped it. The daemon then exited, and the proxy's onDaemonLost
degrades that session (and every other session sharing the daemon) to an
in-process engine for the rest of its life. On one dev machine over a day the
backstop fired 20 times on live sessions (clients=1) and the liveness sweep
caught 0 real dead peers — net harm.
The backstop exists only to catch a phantom client (one counted but gone,
whose socket-close was never delivered). It now consults the peer pids the
daemon already tracks: after the inactivity window it sweeps provably-dead
peers, then reaps the daemon only if NO remaining client can be proven alive
(every one is an unknown-pid connection the sweep can't verify — the sole
phantom class it can't catch). One provably-alive client keeps the daemon up.
Extracted the decision into Daemon.backstopShouldExit(isAlive) so it's unit-
testable with an injected liveness probe, mirroring reapDeadClients. All #692
guarantees preserved; the only behavior change is that a provably-alive quiet
session is no longer reaped.
- daemon-client-liveness.test.ts: 7 new deterministic cases for
backstopShouldExit (live kept, phantom reaped, mixed protects the live one,
dead-peer swept-then-held, within-window, zero-client).
- mcp-daemon.test.ts: the integration test that asserted the backstop reaps a
live connected client (it encoded the bug) now asserts the opposite — a
live-but-quiet session survives several backstop windows with its lockfile
intact and no backstop shutdown logged.
Validated end-to-end on the built bundle: a quiet session's daemon stayed up
across 4 backstop windows (maxIdle=3s), same pid throughout, zero backstop
fires. Found while fixing #1185.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
An MCP host that kills the launcher chain within the server's first ~100ms
while keeping the stdio pipes open (config probe, cancelled request, startup
timeout; Rust hosts that kill a child without dropping its stdio handles) left
the server orphaned: it booted already reparented to init, so the PPID
watchdog's "ppid changed" baseline was captured as 1 and could never fire, and
stdin never EOF'd. The process lingered — idle, ~30MB — until the host itself
exited, accumulating one per abandoned launch (the pile-up reported in #1185).
Reproduced on released 1.2.0/macOS: SIGKILL the launcher at +50ms → permanent
orphan; at +150ms the old late baseline had already run and reaped it.
Three-part fix:
- Capture process.ppid at the earliest line of the CLI entry (early-ppid.ts)
and use it as every watchdog baseline, shrinking the blind window to the few
ms before our first JS runs.
- Thread the real host pid down the bundled path: the npm shim and the
standalone sh launcher set CODEGRAPH_HOST_PPID (an outer launcher's value
wins), so the watchdog polls the host directly. Previously only the
--liftoff-only relaunch set it, leaving the entire npm/standalone install
base with hostPpid=null.
- Never-initialized backstop (startup-handshake.ts): a serve --mcp that
receives no MCP traffic for CODEGRAPH_STARTUP_HANDSHAKE_TIMEOUT_MS (default
15min, 0 disables) shuts down — the catch-all for a kill landing in the
residual pre-JS window. Disarmed on the first byte, so a quiet-but-live
session is never touched.
Also scrub CODEGRAPH_HOST_PPID from the detached daemon's env — it has no host,
and a stale pid must not leak into anything it spawns.
Validated end-to-end on the built bundle: the +50ms early-kill orphan is now
reaped while the host still holds the pipes open, and all six normal
lifecycle paths (clean close, SIGTERM/SIGKILL child, host exit/SIGKILL,
fd-holding adversarial host) stay clean. New coverage in
startup-handshake.test.ts, mcp-startup-orphan.test.ts, and npm-shim.test.ts.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
An indexing run killed mid-"Resolving refs" (crash, Ctrl-C, the #1122
watchdog kill) left the refs it never reached parked in unresolved_refs.
The git-scoped sync fast path only re-resolves changed files' refs, so
those files' call edges were missing permanently — a too-small blast
radius clustering by package/module (the #1187 field report: 3 of 10
caller files for a Spring @Resource-injected method) — until a full
re-index.
- sync() now sweeps leftover unresolved refs with the batched resolver
after its scoped pass, including on no-change syncs, so a bare
`codegraph sync` recovers a wedged index (and heals pre-fix indexes
on the first post-upgrade sync)
- the scoped pass deletes unresolvable rows too (parity with the
batched path), making "rows at rest" a sound orphan signal
- drop the batched loop's early break that abandoned all later batches
when one batch was all-unresolvable (its rows WERE consumed — that
early stop could orphan the rest of the table at init)
- surface the state: `codegraph status` warns, `status --json` gains
index.pendingRefs, and MCP codegraph_status tells agents the blast
radius is incomplete until the next sync
Verified end-to-end on a 2,414-file synthetic Spring repo: SIGKILL
mid-resolution reproduces the reporter's exact 3-of-10-callers state;
a bare sync now heals it to 10/10 with the edge count converging to
the clean-init total; a healthy-index sync stays a no-op.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Adds ArkTS (.ets, HarmonyOS/OpenHarmony) as a first-class language:
full TypeScript-grade extraction via the harmony-contrib tree-sitter
grammar (MIT, vendored byte-identical from the tree-sitter-arkts 0.2.0
npm tarball), plus the ArkUI constructs that make HarmonyOS apps
traceable:
- @Component/@ComponentV2 structs with decorators from both grammar
positions; members extract as class members with qualified names.
- build() component trees: child instantiation edges via
arkui_component_expression, no synthesizer needed.
- Attribute chains emitted dot-prefixed and resolved ONLY against
@Extend/@Styles/@AnimatableExtend/@Builder helpers (unique-or-drop) —
bare-name fallthrough produced 36,840 wrong edges (17% of calls) on
the OpenHarmony samples monorepo. All four grammar chain shapes
handled, including the detached-chain forms.
- .onClick(this.handler) method-reference bindings.
- ohpm workspace modules: bare imports follow oh-package.json5 file:
deps (ambiguous names dropped), honoring each module's main entry —
which also lets .ts consumers resolve .ets modules.
- ArkUI dynamic-dispatch bridges, all provenance:'heuristic' with
wiring-site metadata: assignment-gated state->build() re-render
(V1 @State family + V2 @Local/@Provider/@Consumer),
@ohos.events.emitter emit->subscriber pairing on static event keys
(numeric ids same-file, named constants same-module, fan-out capped),
and router.pushUrl literal urls -> the target page's @Entry struct.
- $r/$rawfile resource intrinsics treated as built-ins; arkts joins the
web language family, value-reference edges, re-export chase, and the
other TS-applicable gates.
Also ships a language-agnostic index-completeness guard: indexAll
stamps index_state (indexing -> complete/partial/failed), reconciles
discovered vs accounted files (a loaded run silently dropped 37 files),
and codegraph status surfaces truncated/partial indexes in human and
--json output.
Validated on HarmoneyOpenEye (82 files), CoolMallArkTS (528, modular
ohpm + ArkUI V2), and openharmony/applications_app_samples (11,693
files, 202,890 nodes stable across re-index, attribute false-positive
audit 36,840 -> 588 residual all-plausible). Supersedes PRs #656 and
#988 with credit — both informed this implementation.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Follow-ups noted in #1173:
- cloudposse/atmos remote-state: module.M.outputs.X emits a scoped
module.M:remote-output.X candidate; the resolver bridges it to the
target COMPONENT's own output when every gate holds — the module
source is the stack-config remote-state module, the component name is
static (a literal, or component = var.X whose variable declares a
literal default in the same directory), and exactly one directory in
the repo matches the component name and declares that output. Dynamic
(each.value) or ambiguous wiring stays unlinked. On
cloudposse/terraform-aws-components: 254 remote-state bridge edges,
every one re-derived from a matching source declaration (789/789
cross-directory output edges explained: 528 local-module + 254
remote-state + 7 checker-artifact false alarms under deprecated/);
coverage 66.4% -> 69.1%.
- provider aliases: provider "aws" { alias = "east" } is addressed as
provider.aws.east so aliased and default configurations stop
colliding; provider = aws.east on a resource/data block (and the
values of a module's providers map) reference the selected
configuration, resolved same-directory first then up the module tree
— the one construct Terraform genuinely inherits from parents. The
selection is no longer misread as a resource reference (aws.east).
- moved/import/removed blocks reference the resource addresses they
name (anchored to the file node — no phantom symbols), so a
refactor's paper trail joins the graph; check-assert conditions
contribute their references while check-scoped data blocks keep
indexing as before. Scoped module candidates are suppressed there:
module.a.aws_x.b names a resource inside a module instance, not an
output. +91 edges on cloud-foundation-fabric's moved-heavy stages.
Also fixes a latent test bug from #1173: cg.getNodeById is not public
API (cg.getNode is) — it only passed because the asserted edge list was
empty.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* feat(extraction): add Terraform and OpenTofu language support
Index .tf, .tfvars, and .tofu files via the tree-sitter-terraform dialect
of HCL (vendored from @tree-sitter-grammars/tree-sitter-hcl, Apache-2.0).
Symbols extracted:
- resource / data → class (qualified "type.name" / "data.type.name")
- module → module (qualified "module.name")
- variable → variable (qualified "var.name")
- output → variable (qualified "output.name")
- provider → namespace
- locals → constant per attribute (qualified "local.key")
References resolved cross-file:
- var.X, local.X, module.M[.out], data.T.N[.attr], <type>.<name>[.attr]
- built-ins skipped: each.*, count.*, self.*, path.*, terraform.workspace
The Terraform framework resolver disambiguates same-named candidates
across modules by preferring the one in the same directory as the
reference site, then by closest common-ancestor path, falling back to
the generic name matcher only when neither applies.
Validated on two Terraform monorepos (277 and 470 .tf files): indexing
runs in 1.3s and 2.4s respectively, query latency stays under 200ms,
and cross-module references resolve to the correct module 100% of the
time on inspected samples.
18 new extraction tests; full suite 1146/1148 green (2 pre-existing
flaky skips, 0 regressions).
* feat(terraform): bridge the module boundary and enforce directory scoping
Builds on #706. The module declaration was a dead end: module.M.out
resolved to the declaration and stopped, module inputs never reached the
child module's variables, and impact could not cross the boundary — on
real multi-module repos that breaks the core blast-radius question
("what breaks upstream if I change this module's variable/output").
- module blocks now wire across the boundary through :-scoped refs only
the Terraform resolver understands: module.M:var.<input> → the child's
variable node, module.M:output.<o> → the child's output node (emitted
alongside the module.M declaration ref), and module.M:file → the local
source directory's entry file (imports). Registry/git sources emit no
file ref and resolve nothing — an out-of-repo module stays a visible
boundary instead of a guess.
- .tfvars top-level assignments reference the variable they set, walking
up to the nearest ancestor directory (envs/prod.tfvars → root vars).
- Resolution now enforces Terraform's real scoping: same-directory only
(no cross-module fallback by common path prefix, no single-candidate
anywhere-in-tree binding), and terraform refs never fall through to
the generic name matcher — var.X can never legally bind outside its
module directory, so the fallback could only add wrong edges.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs(terraform): README language table + changelog entry + agent-eval corpus
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Javier Rodríguez Fernández <jfernandez@freepik.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
CUDA rides the C++ grammar via the Metal (#1121) dialect pattern:
blankCudaConstructs (offset-preserving) blanks execution-space specifiers
(__global__ family), __launch_bounds__(...), and <<<grid, block>>> launch
configs — which otherwise lex as shift operators and destroy the
host→kernel call edge entirely. Gated by .cu/.cuh extension OR by content
(looksLikeCudaSource), because much real CUDA lives in .h/.hpp headers:
cutlass launches most kernels from headers and flash-attention's launch
templates are .h. Safe by construction — no CUDA marker is valid C++
anywhere, and the launch blank is bounded + brace-balance-checked so a
stray <<< (committed merge-conflict markers) can never blank real code.
All real-world launch styles connect: plain, templated
(k<T, 256><<<...>>>), function-pointer (auto kernel = &fn<...>; with
branch reassignments each linked), dim3{...} brace-init configs, and
kernels defined through name-in-first-argument macros
(DEFINE_FLASH_FORWARD_KERNEL style — gtest TEST_F / PYBIND11_MODULE
shapes deliberately excluded by the two-lone-identifiers rule).
Two general C++ resolution wins the flow validation forced out:
- namespace blocks now prefix contained symbols' qualifiedNames
(prefix-only — no namespace nodes, avoiding #1093-style crowd-out), so
ns::fn(...) calls resolve; previously every namespace-qualified C++
call was a permanently dead edge. cutlass: +30,864 edges (~10%), node
count byte-identical.
- templated callees (fn<T, 256>(args)) strip template args at extraction
(mirroring #1043 for base classes), so they match their definitions.
Validated on llm.c (165 host→kernel launch edges, was 0),
flash-attention (run_flash_fwd → flash_fwd_kernel → compute_attn traces
in one codegraph_explore call), and NVIDIA CUTLASS; fmt as the plain-C++
control (unchanged). A/B n=2/arm: Read/Grep displacement decisive on all
three repos (flash-attention Reads 29,13 → 5,2).
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
escripts (.escript) index like any module — the ELP grammar has a
first-class shebang node, so no source transform is needed; main/1 and its
helpers get full function/call extraction.
OTP application resource files (<app>.app.src and compiled <app>.app) join
the graph as Erlang terms the grammar parses natively. They route by full
suffix (their last-dot extension, .src, is far too generic for the
extension map). The application tuple yields structure: {mod, {Mod, _}}
links the app to its callback module — the app's entry point — and
{applications, [...]} / {included_applications, [...]} connect umbrella
sibling apps, resolving through the OTP app-name == module-name convention;
kernel/stdlib and other out-of-repo apps stay unresolved.
App-file refs resolve only ever to MODULES: validation on emqx caught the
ssl OTP-app dependency resolving to a test helper FUNCTION named ssl (the
same defect class as the earlier -behaviour gate), so the matchReference
module-only gate now covers every ref an .app/.app.src file emits.
Validated on emqx: 2 app.src + 6 escripts indexed, entry-module and
umbrella-dependency edges all namespace-targeted post-gate, escript
functions extracted; a stray legacy/module.src stays unknown.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Calls hidden inside -define bodies were invisible: the extractor consumed
pp_define without walking the replacement, and macro use sites produced no
edges, so a call path routed through a macro (ejabberd's SQL upsert macros,
logging wrappers) was completely dark.
The macro's constant node now participates in the graph. The -define body's
calls are attributed to the MACRO — true exactly once, instead of a per-use
duplicate that would explode on logging macros — and each use site links
in: ?MACRO(...) with arguments emits a `calls` ref (inlined code joins the
call chain), a bare ?CONSTANT read emits `references` (answering "where is
this macro used" without polluting call paths). Compiler-predefined macros
(?MODULE, ?LINE, ?FUNCTION_NAME, ...) are excluded, macro-use arguments
keep walking so a call nested in ?assertEqual(ok, do_thing()) still
attributes to the enclosing function, and macro-to-macro chains connect.
Validated: node counts unchanged on cowboy/ejabberd/emqx; edges +26/+7.3K/
+42K with honest hub shapes (?T i18n, ?SLOG logging, ?QOS_1 protocol
constants); 40/40 sampled edges precise; +1.3s index cost on emqx's 2,273
files. The payoff chain on ejabberd: set_password_scram_t → ?SQL_UPSERT_T →
ejabberd_sql:sql_query_t — database writes through SQL macros now trace
end-to-end.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
gen_server:call/cast/send_request now connects to the TARGET module's
handle_call/handle_cast for every statically-named target, not just self:
a bare atom reaches the module of that name (OTP's {local, ?MODULE}
convention names a server after its module), and a macro defined as a bare
atom (-define(STORE, kv_store)) resolves the same way, alongside the
existing ?MODULE / -define(SERVER, ?MODULE) self paths. A registered name
that matches no module emits a qualified ref that never resolves — silent,
never guessed. Pid, var, and tuple targets ({global, Name}, {Name, Node})
stay unlinked.
Validated on emqx: 53 new edges, 53/53 precise (each source line is a real
registered-name gen_server request; each target module self-registers under
that name, macro-indirected registrations included). Nearly all are
test-suite → handler links — production code goes through API wrappers the
self path already covers — which is exactly the tests-exercising-this-
handler linkage blast-radius and test-gap reporting consume. ejabberd
yields zero (it always wraps): no false positives invented.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Bridges the OTP callback boundary: a framework call through a variable
module — cowboy's Handler:init / Middleware:execute folds, a plugin
manager's Mod:callback(...) — now links to the repo's implementers of the
behaviour declaring that callback, so codegraph_explore connects flows
end-to-end across behaviour dispatch instead of stopping at it.
Precision gates: the callback arity must match the site, exactly one
in-repo behaviour may declare that (name, arity) — a collision bails
(cowboy's init/2 is declared by five handler-flavored behaviours and
correctly stays silent) — the implementer must export the callback, and
above the fan-out cap the site is skipped entirely (ejabberd's gen_mod
with ~230 implementers stays a visibly dynamic boundary). Behaviour
discovery scans -callback declarations in every module so implementer-less
behaviours still gate ambiguity. Edges carry provenance:'heuristic' with
synthesizedBy:'erlang-behaviour' and the wiring site, rendered as dynamic
dispatch in explore.
Validated per the dispatch-family playbook: cowboy 38 edges (middleware
chain, stream-handler folds, sub-protocol upgrade), ejabberd 598, emqx 843;
36/36 sampled edges precise (target declares the via-behaviour and exports
the callback); node counts unchanged; ~1.4s added on emqx's 2,273 files;
zero-control clean. The cowboy request flow connects in one explore call.
Includes an Erlang comment stripper (%-comments, string/atom/$-char aware)
for the dispatch-site scans.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Vendored WhatsApp/tree-sitter-erlang 0.19 (the ELP grammar, ABI 14) with an
Erlang-shaped extractor: multi-clause/multi-arity functions merged into one
symbol, -spec signatures, records with fields, -type/-opaque aliases, -define
macros, -include/-include_lib file edges, and -export-driven visibility.
Modules wrap in a namespace so remote mod:fn(...) calls resolve through the
existing qualified-name matcher as mod::fn with zero resolver changes.
-behaviour declarations link to the behaviour module — gated to namespace
targets only (bare-name fallthrough linked -behaviour(supervisor) to an
unrelated macro constant on emqx). OTP indirection with static targets is
followed: spawn/apply/proc_lib/timer/rpc MFA-argument callees, and
gen_server:call/cast(?MODULE | ?SERVER) to the module's own
handle_call/handle_cast. Var-module dispatch and message sends stay
deliberately unlinked. codegraph_explore also normalizes Erlang-native query
spelling (mod:fn/3, init/2) so named symbols resolve as typed.
Benchmarked on cowboy (189 files), ejabberd (414), emqx (2,447): extraction
PASS on all three; with-codegraph arms reached 2/2/0 file Reads vs 10/5+/19
without, fastest on the largest repo.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Vendored patched govindbanura/tree-sitter-vbnet grammar (MIT, ~20-fix patch
+ new external scanner for XML literals and multi-line LINQ continuation;
provenance + rebuild instructions in docs/grammars/tree-sitter-vbnet.md),
vbnet extractor with VB-specific call/index disambiguation, Inherits/
Implements heritage, As New instantiation, events, Declare P/Invoke, and
MustOverride abstract members.
Parse health on five real repos: PolicyPlus 100%, CompactGUI 100%,
staxrip 95.2%, SCrawler 87.2%, PCL 87.5% (upstream grammar: 3-18%).
Retrieval A/B (sonnet): 26-43% faster with 0-5 file reads vs 7-20 without.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Programs, sections/paragraphs (reconstructed extents over the grammar's
flat header stream), PERFORM/THRU/GO TO/CALL call edges, COPY copybook
imports incl. standalone .cpy fragments, DATA DIVISION records/fields/
88-levels with write-site impact references, and CICS flows: EXEC
LINK/XCTL program targets (literal + same-file VALUE deref), EXEC SQL
INCLUDE, and pseudo-conversational RETURN/START TRANSID hops resolved
to the owning program via a CICS framework resolver. Fixed and free
source format (free format via a scanner wide-mode sentinel).
Grammar: vendored wasm built from a patched yutaro-sakamoto/
tree-sitter-cobol (EXEC blocks as an external-scanner token, copybook
fragment entry point, single-quote continuation, COPY REPLACING
pseudo-text, NOT=, CALL GIVING, ENTRY, FREE, bitwise ops, abbreviated
relations, COBOL-2002 usages, and more). Patch + provenance + upstream
PR draft in docs/grammars/. Parse health: AWS CardDemo 43/44 native
(upstream: 9/31), 44/44 through preParse; copybooks 28/29; CobolCraft
free-format 17/17 (upstream: 0); NIST COBOL85 unchanged at 373/382.
Copybook members resolve to files like C includes (basename index,
name-matcher short-circuit so compiler-supplied members stay honestly
unresolved): CardDemo imports 5 -> 285. Impact proof: ACCT-CURR-BAL
(CVACT01Y copybook) surfaces its 4 writer programs cross-file.
Also: run-all.sh now neutralizes the ambient prompt-hook in both A/B
arms (CODEGRAPH_NO_PROMPT_HOOK=1); COBOL corpus entries for agent-eval.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
CFML joins the #1108 receiver-inference family: new/createObject/typed-arg/property(inject) declarations type the receiver, variables./this. fields scan whole-file, method QNs re-scoped to Class::member in all three extraction paths. 1,649 typed edges on fw1/ColdBox/CFWheels, 1,649/1,649 audit-consistent, inherited methods resolve via #1152 extends edges.
Co-authored-by: ghedwards <125586+ghedwards@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
extends="coldbox.system.web.Controller" (dotted) and extends="../base" (relative) now resolve to the right component via directory-corroborated matching; >=1 corroborating segment required, ties yield no edge. fw1 14->47, ColdBox 21->242, CFWheels 60->201 inheritance edges; 394/394 audited path-consistent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
.metal was absent from EXTENSION_MAP, so Metal Shading Language files were
silently skipped. MSL ≈ C++14, and the C++ grammar extracts its functions,
structs, type aliases, and call edges at parity with plain C++ — except MSL's
post-declarator [[attribute]] annotations, which misparse struct fields into
spurious extends refs from the struct to the field's own type (a wrong
inheritance edge whenever the repo typedefs float3/float4x4 itself, common in
shared ShaderTypes.h). blankMetalAttributes blanks them pre-parse,
offset-preserving, following the blankCppExportMacros pattern (#1061), gated
to .metal files only — in regular C++ the attribute position is legal syntax
the grammar parses natively. The preParse hook gains an optional filePath
param to support the gate.
Validated on llama.cpp's ggml-metal.metal (10.7k lines: 130 kernels vs 113
`kernel void` ground-truth lines, rope_yarn resolves its 4 kernel callers)
and SDL's shaders (PQtoLinear ← GetOutputColor), 0 bogus extends edges.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Five hardening fixes to the #1136 MEDIUM (graph-derived) tier:
- #1141: updateNode() now writes the segment vocabulary like insertNode()
does — framework post-extract renames (NestJS route prefixing) left the
new name permanently unsearchable (the old rows orphaned, the backfill
gated on an EMPTY vocab, so even a full re-index re-created the drift).
- #1142: new CodeGraph.healSegmentVocabIfEmpty() — the hook opens the
graph without sync, so a database migrated from pre-vocab schema kept
the MEDIUM tier dormant until some unrelated sync ran. The hook heals
on first use (one SELECT when populated; lock-aware, defers to a
running sync) and records noop-vocab-empty when it can't.
- #1144: a name whose only nodes are file/import kind is skipped instead
of falling back to surfacing an import statement as a matched symbol;
import specifiers no longer enter the vocab at all (shared
isSegmentableKind gate across insertNode/updateNode/rebuild page query)
since they can never be surfaced and only inflate rarity statistics.
- #1145: plural variant folding is keyed on English plural spelling —
bare-s plurals no longer mint a bogus -es sibling (services→servic),
unambiguous sibilant-es plurals no longer mint a bogus -s sibling
(classes→classe), trailing -ss singulars no longer strip (class→clas);
genuinely ambiguous endings (caches/databases) still emit both keys.
- #1146: getSegmentCoOccurrence folds variants to their original word
inside the SQL (CASE mapping + COUNT(DISTINCT word)) so a plural pair
of ONE word can't tie with a genuine two-word match and crowd it past
the pre-fold ORDER BY/LIMIT; the JS re-check stays as the honesty layer.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
gate('high-keyword'/'high-token') sat outside the injection guard, so an
errored or empty codegraph_explore still counted as a HIGH-tier success.
The gate telemetry is the measured recall/precision funnel that decides
whether the tiered gate design survives — a delivery failure must degrade
it toward noop-*, not inflate the high tiers. Failures now record
noop-explore-keyword / noop-explore-token. Doc enum updated (including
the noop-vocab-empty outcome the #1142 fix adds next).
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
extraction/index.ts bounds every git call it makes; worktree.ts,
git-hooks.ts, and the installer's npm install -g did not, so a stuck
subprocess blocked the caller indefinitely. Worst case was the daemon:
gitWorktreeRoot/gitCommonDir run (memoized) on the main event loop while
serving MCP clients, where an unbounded git hang would trip the 60s
liveness watchdog and SIGKILL a healthy daemon. git calls get 5s, the
interactive npm install 120s. Regression tests assert the option through
a mocked child_process plus a per-file call-site sweep.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The multilingual structural-question gate (#1134) matches stems as open
prefixes (left boundary only) so derived forms fire without enumeration.
Four English stems have common non-structural completions — callus,
calligraphy, Connecticut, connective, affectionate, Tracey — that
false-fired the HIGH (full-explore) tier. Those four now enumerate their
structural suffixes and re-assert the right boundary; callbacks/callable/
call sites are included so no structural form regresses. Also documents
the verified-unfixable Korean homograph class on the unsegmented table
(#1140): segmentation can't split 구조대 from 구조가, and a denylist would
break 구조대로.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The #850 liveness watchdog was killing valid `codegraph init`/`index` runs
at "Resolving refs 0-2%" on large collision-heavy repos (18-25K-file Java
monorepos on slower hardware). #1105's cooperative yielding assumed a
500-ref sub-chunk is always cheap, but per-ref cost is unbounded: a
colliding method name (`execute`, `process`, ...) whose candidate set
misses the 5,000-entry name LRU re-fetches every same-named row
(unbounded SELECT + materialization, measured 8.8ms at just 4K collisions
on an M4 — linear in collision count), and receiver-type inference
re-split the whole source file per ref (~20% of total index CPU). A dense
pocket multiplied that past the 60s window and the heartbeat starved.
Three guards, no behavior change:
- resolveBatchYielding checkpoints after EVERY ref (maybeYield is a ~ns
time check when under budget), so a slow pocket can never run more than
one ref past the yield budget.
- resolveMethodOnType's ref-independent candidate filter is memoized per
(language, Type::method) on the resolver context; per-ref
disambiguation (import FQN #314, call-site file #1079) stays outside
the memo.
- Receiver inference reads lines through a per-file LRU (shared and C++
inferrers), and skips generated/minified lines >10K chars instead of
regex-scanning them per ref.
Measured on a 4,028-file synthetic Java bank repo (392K refs): mid-loop
max event-loop stall 1528ms -> 546ms under cache thrash, total init
250.9s -> 96.8s at default config.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The keyword gate (#1126) can never know a repo's domain nouns. This adds
the graph-derived tier the design discussion converged on: symbol names
are split into prose segments at index time (name_segment_vocab, riding
the insertNode write path), and the hook verifies a prompt's plain words
against them — "the state machine des commandes" → OrderStateMachine, in
any language whose technical nouns are Latin script.
Confidence now decides HOW MUCH to inject, not just whether:
- HIGH (keyword, or index-verified code token): full explore injection,
unchanged — the validated adoption lever.
- MEDIUM (segment matches only): a ~500-byte pointer naming the matching
symbols; the AGENT writes the explore query. Never runs explore, so a
fuzzy match can't inject 16KB of wrong-feature context.
- Silent otherwise, as before.
Precision is derived from the repo's own naming statistics plus measured
FP fixes: co-occurrence (≥2 words on one name) always qualifies; a single
word must be ≥5 chars, cluster across 2–25 names (singletons are prose
coincidence: "deploy to production" → matchesNonProductionDir), match a
multi-segment name, and not be an English function/filler word (the one
place a word list is honest: identifiers are English, so only English
prose collides). Every candidate is re-verified against nodes before
being surfaced — vocab rows are proposals, deletions leave orphans by
design, a full index rebuilds from scratch, and sync heals pre-upgrade
databases (batched + yielding; emptiness captured at sync ENTRY so the
sync's own writes can't mask the backfill).
Schema v7 migration is DDL-only (instant; none of the #1067 row-churn
hazards). Gate outcomes roll up as anonymous usage counters
(prompt-hook-gate-<outcome>, names only, never content) through the
existing telemetry pipeline — recall becomes measurable, and the counters
are the agreed kill-criterion data for ever revisiting a local classifier.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(prompt-hook): fire the structural gate for Latin-script, Cyrillic, and JA/KO prompts (#1126)
The prompt-hook's keyword gate only knew English and simplified-Chinese
keywords, so a structural question in French (or Spanish, German, Italian,
Portuguese, Russian, Japanese, Korean, traditional Chinese) silently
no-op'd unless it happened to contain an identifier-shaped code token —
the #994 symptom, resurfaced for every other language.
Root causes fixed:
- JS \b is ASCII-only: a keyword whose first/last char is accented or
non-Latin (où, qué, Cyrillic, kana) can never match \bkeyword\b —
the same mechanism behind #994. Keyword matching now uses Unicode
lookaround boundaries ((?<![\p{L}\p{N}_]) … (?![\p{L}\p{N}_])).
- Bare-stem English entries never matched their own derived forms
(\barchitect\b can't match "architecture", \bdepend\b can't match
"dependencies"). Stems are now matched as word prefixes (leading
boundary only), which also lets one shared stem cover the Romance/
Germanic spellings that coincide.
- The "CJK" set was simplified-Chinese-only: Japanese (呼び出し, 仕組み,
実装 — and 追跡 ≠ 追踪), Korean, and traditional-Chinese terms are now
in the unsegmented substring set.
Code-token extraction and the graph-verification path are unchanged;
non-structural prose stays a zero-cost no-op in every language.
Fixes#1126
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(prompt-hook): extend the gate to tier-2 languages (VI/TR/ID/PL/UA/NL/CS/RO/HU/EL/Nordics/FI/HI/AR/FA/HE/TH)
The first pass covered the 10 largest languages; this closes the rest of
the major-developer-population set (~29 total). Notable per-language
mechanics the curation had to respect:
- Agglutinative languages (Turkish, Finnish, Hungarian) need stems, not
exact words — suffixes attach to everything (akışı, riippuu, működik).
- Indonesian me-/di-/ber- prefixes block leading-boundary stems, so
affixed forms are listed explicitly (memanggil, dipanggil, berfungsi).
- Arabic/Farsi/Hebrew are spaced but proclitics attach to the word
(وكيف = and-how), so they join the substring class with Thai.
- Ukrainian і/и spellings diverge from Russian (архітектур ≠ архитектур).
- Excluded terms that collide with English or code words: NL "pad",
SV "var", CS "tok", Catalan "com" (matches every .com domain) — with
regression tests pinning the exclusions.
Vietnamese was the sharpest gap: spaced Latin with heavy diacritics —
exactly the ASCII-\b failure class #1126 reports.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The managed-reasoning removal (e5897d03) stripped the CLI/MCP wiring but,
despite its stated intent, left the offload modules and their test suite
behind. The dead code still shipped compiled inside the platform bundles,
and its Windows browser-opener was flagged by a security report (#1114)
for routing the login URL through `cmd /c start`, where cmd re-parses
shell metacharacters. Unreachable since 2026-06-20 and never wired in any
tagged release — but delete it for real: src/reasoning/ (config,
credentials, login, reasoner), __tests__/offload.test.ts, the now-inert
CODEGRAPH_OFFLOAD_DISABLE guard in dynamic-boundaries.test.ts, and the
stale reasoner reference in the FILE_SECTION_PREFIX comment.
Closes#1114
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Lua method-call syntax (lg:Log()) is byte-identical to the Luau type-annotation
shape (lg: Logger), and the receiver-type scan starts on the call's own line —
so any PascalCase method call self-matched as "type = Log" before the scan
reached the real declaration, silently dropping the calls edge whenever two or
more classes shared a method name.
The annotation pattern now rejects a capture followed by any of Lua's three
call forms; its leading [\w.] lookahead alternative prevents backtracking from
shrinking the capture to dodge the gate. Gated rather than dropped: the pattern
is the only type source for Luau typed params and annotated locals whose
initializer isn't T.new().
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Completes the #1125 fix. The same typed-parameter gap fixed for TS/JS existed
in every other language whose localReceiverTypePatterns only matched
keyword-anchored locals (let/var/:=/= new) and never the bare parameter form:
- Rust: the `:`-annotation pattern required `let`, so `fn use(lg: &Logger)`
didn't match. Dropped the `let` anchor (still covers `let lg: Logger`),
keeping the `&?mut?` handling — now covers params and closures `|lg: T|`.
- Go: only `lg := T{}` / `var lg T` matched; a parameter/method-receiver
`func use(lg Logger)` / `func (l Logger) M()` (name-before-type, no keyword)
didn't. Added a PascalCase-guarded `ident Type` pattern — the guard plus the
existing enclosing-scope bound (excludes package-level struct fields) keep
the keyword-free shape from matching unrelated pairs.
- Dart: the type-before-name pattern's trailing `[=;]` missed a parameter's
`)`/`,`. Widened to `[=;,)]`, mirroring Java/C#.
- PHP: only `$lg = new T` matched; a typed param `function use(Logger $lg)`
(also `?Logger`, `\App\Logger`, `&$lg`, `catch (E $e)`) didn't. Added a
type-before-$var pattern. Reserved words can't be class names, so the
looser lowercase-allowing capture yields no wrong edges.
Every pattern still relies on resolveMethodOnType validating the inferred type
actually declares the method (no edge on a mis-inference) — the same safety
net the already-covered languages use. Verified with a deterministic probe:
all four now disambiguate two same-named methods via the typed param (Java +
Kotlin as passing controls), full suite green (1930), no regressions.
Adds a parameterized regression test (Rust/Go/Dart/PHP), associating method to
type by qualifiedName so it holds where the method sits outside the type's
line range (Rust impl, Go decl).
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The local-variable receiver-type inference from #1108/#1110 covered typed
parameters for every language except TypeScript/JavaScript (+ TSX/JSX). The
TS/JS `:`-annotation pattern required a leading `const|let|var`, so it only
matched a local's own annotation (`const lg: Logger`) and never a bare
parameter (`function use(lg: Logger)` / `(lg: Logger) =>`). With a second
class sharing the method name — the case where a same-name fallback can't
paper over it — `lg.log()` resolved to no edge, dropping it from callers and
impact/blast-radius. TS/JS is the most common language pair in the userbase,
so this was a real precision gap.
Replace the keyword-anchored pattern with the keyword-free
`\b${r}\b\s*:\s*([A-Z][\w.$]*)`, mirroring Kotlin/Swift/Scala. It's a strict
superset (still matches `const lg: Logger`) plus the typed-parameter case,
and the capture stops at `<` so a generic-typed param
(`repo: Repository<User>`) still yields `Repository`. resolveMethodOnType
already validates the inferred type declares the method, so the looser match
produces no edge on a mis-inference — the same safety net the other
languages rely on; Swift already ships this identical bare-colon pattern with
the same theoretical ternary/dict-literal exposure.
Adds a regression test using two ambiguous classes + typed params, asserting
each call routes to its OWN class's method (verified to fail without the fix
and pass with it — a single-class version would pass either way via the
same-name fallback, which is why the collision is load-bearing).
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
