The daemon/CLI rendezvous directory is created under %LOCALAPPDATA% (Windows) or
/tmp -- /private/tmp on macOS -- and every ancestor of it must pass the
private-directory walk. That ancestry is not always acceptable, and when it is
not, EVERY invocation fails, `config list` included, so the settings surface
cannot be reached either:
codebase-memory-mcp: secure daemon endpoint could not be created
#1623 narrowed the Windows side of this by admitting AppContainer package and
capability SIDs on ancestors, and named the remainder explicitly: a live local
group, Authenticated Users inherited from a secondary volume root, and orphaned
unresolvable SIDs still refuse, and "those need CBM_RUNTIME_DIR or a separate
change". #1621 is the POSIX shape of the same dead end -- /private/tmp/cbm-daemon-<uid>
refused with no way to move it.
There was no way to move it in a shipped build. The only relocation hook,
CBM_TEST_DAEMON_RUNTIME_PARENT, is compiled out unless CBM_ENABLE_TEST_SEAMS is
defined, so a test build started while the shipped build did not; CBM_CACHE_DIR
is no help either, because it moves the cache and never the rendezvous.
CBM_RUNTIME_DIR names the parent directory the rendezvous is created under. It
does NOT relax the check: the directory it names goes through exactly the same
validation as the default -- ancestors owned by you or root, not world-writable,
no allow-ACL; the rendezvous directory itself still forced to owner-only -- and a
value that fails is refused rather than silently replaced by the default. The
operator only chooses an ancestry that passes. cbm_safe_getenv never truncates,
so no half of an over-long value can become a runtime parent.
The override is resolved in cbm_daemon_bootstrap_endpoint_new(), the one function
every product endpoint goes through: the daemon, the MCP client, the local CLI,
the index worker, and the install/update/uninstall activation path in cli.c. No
call site can silently keep the default, and the detached daemon inherits the
value with the rest of its environment. An explicit parent still wins, so the
compile-time test seam and the lifecycle guards' isolated namespace behave
exactly as before.
Approach and variable name from #1576 by Leonardo trindade miranda, resolved one
layer lower so the activation path is covered too.
Refs #1574
Refs #1621
Co-Authored-By: Leonardo trindade miranda <tmonestudio@gmail.com>
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
Release evidence from run 31744302624 shows the tolerated Microsoft `!ml`
verdict is close to a coin flip per byte image rather than a property of the
code. Across the eight targets the stripped and unstripped candidates of the
SAME linker output disagreed on four, and in both directions:
linux-amd64 stripped microsoft-ml unstripped clean
darwin-arm64 stripped clean unstripped microsoft-ml
linux-arm64 stripped clean unstripped microsoft-ml
If the classifier were keying on something intrinsic to our code the siblings
would agree; they do not. So each variant is close to an independent draw, and
5 of 16 candidates drew the flag.
Two draws is not always enough. On that run linux-amd64-portable came back
microsoft-ml on BOTH candidates, leaving no clean binary to ship for that
target. A third independent draw at a ~31% observed per-candidate hit rate takes
the both/all-flagged case from roughly 1-in-10 per target to roughly 1-in-30.
The third candidate is `--strip-debug` (Apple: `-S`): debug information removed,
symbol table kept. Behaviourally identical to the other two — same linker
output, only metadata differs — but a distinct byte image, which is all
VirusTotal needs to scan it as its own file. Verified on the real v0.10.4
candidates: linux-amd64 gives three distinct hashes (294,634,656 /
294,623,208 / 293,746,096 bytes) and darwin-arm64 likewise, with the ad-hoc
signature verifying after strip.
Selection is unchanged in spirit and now ordered: smallest artifact first
(stripped, debug-stripped, unstripped), take the first CLEAN one, and only if
every candidate drew the tolerated verdict ship the smallest flagged one. A
hard verdict on any candidate still blocks the release before selection.
Cost is 24 objects per release instead of 16.
Derivation enforces that all three hashes differ — identical candidates would be
one draw wearing three hats, and the selector would believe it had alternatives
it does not have. Public claims in README, SECURITY.md and docs/index.html
updated from "both stripped and unstripped" to the three candidates.
All three release contract tests pass, including the native derivation test
which exercises the real strip and codesign path on this host.
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
The build+VT dry-run (run 31744302624) blocked all eight targets on:
BLOCKED: objects/3eb486a4... completed with only 48/75 decisive engines (< 50)
That object is the windows-arm64 STRIPPED candidate with ZERO detections. It
was classified "hard" — the same bucket as a real malware verdict — purely
because 48 engines answered instead of 50. Its own sibling scanned clean with
66.
How many engines return a decisive result is VirusTotal fleet availability on
the day, not a property of our binary. These artifacts are ~300 MB and many
engines skip or time out at that size, so the count sits near the floor and
moves run to run: on this run one object landed at 48 and the other fifteen
spread 59-68. A floor there makes shipping a lottery decided by someone else's
infrastructure, which is exactly what our CI-determinism rule forbids, and it is
not something we can influence by fixing anything on our side.
Classification now depends only on what engines FOUND:
hard - two or more engines, any non-Microsoft engine, any label that is
not `!ml`, or anything suspicious. Still blocks the release.
microsoft-ml- exactly one Microsoft `!ml`. Tolerated and disclosed.
clean - nothing flagged.
The decisive-engine count is still recorded per candidate in the results
manifest and printed as a NOTE when it is below the reference value, so the
evidence is unchanged; only its power to block is gone. The selector keeps
rejecting an incoherent response (total < completed).
Public claims updated to match in README.md, SECURITY.md and docs/index.html —
all three promised "at least 50 decisive engines" as a publication requirement,
which would no longer have been true.
Contract test updated: a clean verdict now stands regardless of engine count,
and the below-reference NOTE must still be reported. All three release contract
tests pass (vt-gate-policy, release-gate-chain, candidate-derivation).
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
Routes the MCP index_repository handler and the graph UI's POST /api/index
through a single decision function. The UI route previously checked only that
root_path was a directory, so an operator's configured boundary held on one
entry point and not the other; it now canonicalizes first and applies the same
policy, answering 403 with the reason.
The decision is two-tier, because a bare default-deny would refuse every
first run and a bare opt-in leaves the default open:
- Breadth is always enforced, with nothing configured. Filesystem, drive and
share roots, top-level system trees, the home directory itself and
credential directories are refused as indexing roots out of the box.
- Containment in a declared root applies once CBM_ALLOWED_ROOT is set or a
grant exists, and is evaluated first so a path outside a configured root is
reported as exactly that.
Three things the tests caught, each a real defect rather than a test fix:
- On macOS /etc, /tmp and /var are firmlinked under /private, so
canonicalizing "/etc" yields "/private/etc" and counted two deep — sailing
past a minimum of two, missing the very path being guarded. Depth now
discounts a leading "private" component.
- An earlier draft refused any root containing the cache directory. That was
over-claimed: the indexer only parses recognised source files and a graph
database is binary SQLite it would never extract. Refusing a whole root is
also the wrong remedy where the concern does hold — not walking the cache
is. Removed, with the reasoning recorded at the site.
- Rewording the refusal to "outside every allowed root" broke an assertion
matching "outside the allowed root", and that test's early return skipped
its CBM_ALLOWED_ROOT cleanup, leaking the variable into every later test in
the suite. The original wording is kept and guidance appended instead.
Worth remembering: these contracts match strings, not properties.
Docs updated in the same change, since both env-var tables said "unset
imposes no restriction" and that is no longer true: CONFIGURATION.md and
README.md describe the two tiers, and CONFIGURATION.md lists the always-
refused roots along with the two limits that matter — this constrains scope
rather than sensitivity, and the credential list is a denylist that raises the
cost of a mistake rather than closing the class. SECURITY.md's supported-
versions table was still on 0.8.x and now reads 0.9.x.
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
Daemon lifecycle and Windows correctness, verified on a real Windows 11
ARM64 VM through the maintained test-infrastructure/vm drivers, plus the
macOS and Linux arm64 suites and the container lint gate.
Daemon lifecycle:
- daemon start/stop/status subcommands. `daemon start` launches a
PERMANENT daemon (spawn shape is byte-exact argv; survives idle
periods and session ends) and reports an already-active daemon
instead of failing. Permanence is honored at every stop latch:
last-committed-client disconnect, host initial-client window,
coordinator release, and application final-session close — a
permanent daemon also keeps admitting new sessions after its last
one closes.
- daemon stop refuses while sessions are active and lists the blocking
peers (pid/role) that must finish first; an idle daemon drains
through the activation-shutdown machinery with the ACK ordered after
connection interrupts. A second stop is idempotent. The wire ops are
no-cohort first-frame requests with peer fingerprint authentication,
so stop/status never conflict with an exact-build admission gate.
- One-shot CLI commands now execute through the daemon (index workers
keep their local supervised path). A cold CLI run that had to spawn a
temporary daemon prints a hint that `daemon start` removes the
per-command startup tax; a warm daemon is recycled silently.
- Hooks are connect-only fail-open: with no daemon present the hook
emits a visible, rate-limited notice (Claude-dialect systemMessage
plus stderr for other dialects) and always exits 0 — augmentation is
never allowed to block the caller's tool use.
- Version skew: a newer-build client automatically drains an
older-build permanent daemon (strict semantic-version triples only;
dev builds never auto-drain) and the build-conflict message names
`cbm daemon stop` as the manual escape hatch.
Windows IPC/runtime (real-VM verified):
- ipc(win): persistent pending overlapped ConnectNamedPipe. The accept
path used to destroy its listening pipe instance on every 20 ms poll
timeout; a client attaching in the teardown window was severed or left
on an orphaned pipe object whose HELLO no server handle could ever
read, absorbing the connect until the client's own timeout expired.
The pending connect now survives poll timeouts and nothing is
destroyed while a client could be attaching.
- ipc(win): drain-before-close for final responses. Closing a named-pipe
server handle can discard a just-sent response before the peer reads
it (POSIX stream sockets never lose buffered data on close). A bounded
cbm_daemon_ipc_connection_drain (read-until-peer-EOF; no-op on POSIX,
immediate on interrupted connections) now precedes close in
runtime_worker_finish and runtime_reject_inline, so hello-conflict,
capacity and disconnect acknowledgements reliably reach the peer.
- runtime: CLOSE_INTENT wire frame. A Windows named-pipe client has no
transport half-close, so close_begin now announces departure with an
explicit frame (ordered after APPLICATION_CANCEL, before the local
interrupt); the server releases the client's admission on receipt
instead of waiting for the handle to close. Admission-drop timing is
now identical to POSIX shutdown() semantics on every platform.
- runtime(win): client close cancellation. close_begin serializes with
request publication under the send lock, best-effort sends the active
token's APPLICATION_CANCEL frame, then interrupts local I/O; the
server cancels MCP/subprocess work promptly. Contract tests accept
both correct outcomes (interrupted transport or decoded CANCELLED).
- runtime: activation acknowledgement ordering. The activation ACK is
the requester's license to act on "snapshotted and draining", so every
connection interrupt is now initiated before the ACK is sent; a
session could previously get one more request serviced after the
requester observed the ACK.
- service(win): deadline-bounded private-file prepare. The conflict-log
prepare retry loop (100 x Sleep(2), which rounds up to the ~16 ms
timer granularity) burned ~1.6 s against permanently obstructed paths,
stalling hello rejections past the client's timeout. The retry budget
is now a 250 ms deadline; transient share collisions still retry.
- subprocess(win): cmd.exe /C payload encoder quotes metacharacters
correctly (root cause of the git-on-Windows failure cluster).
- watcher: SHA-256 buffer sizing (CBM_SZ_64 -> CBM_SZ_128) and a native
Windows stop/unwatch cancellation test with exact-image verification.
- httpd: send_all writes in bounded 64 KiB slices. A single giant
nonblocking send() on Windows is absorbed wholesale into AFD kernel
buffering regardless of SO_SNDBUF, so send deadlines and interrupts
could never engage against a slow peer (and the full payload was
pinned in nonpaged pool). Slicing restores a deterministic
backpressure point; a test hook pins SO_SNDBUF for the deadline and
interrupt tests.
- ui/http: shutdown lifecycle — interrupt checks, response-wide send
deadline, explicit connection states, refusal to free a server while
a listener-owned connection is active.
Windows long-path support:
- Central path-aware wide conversion (canonicalize via GetFullPathNameW
and prepend the extended-length prefix for absolute paths >=240) at
the compat chokepoints (cbm_fopen/compat_fs/mkstemp/mkdtemp), sqlite
store opens, and the daemon build-fingerprint/log paths. Deep managed
installs (a 64-hex generation directory routinely exceeds MAX_PATH)
now index, stage and activate correctly.
- activation transaction: its own file APIs and the component-walking
ancestry validators now operate in the extended-length namespace;
the launcher path is canonicalized (and prefixed when deep) once at
entry so every downstream exact-string comparison stays
form-consistent.
- Executable self-resolution uses the wide APIs (GetModuleFileNameW,
GetFileAttributesW) so non-ASCII install paths survive argv[0]
resolution.
Windows launcher install/uninstall transaction:
- FileRenameInfoEx names are NUL-terminated in an over-allocated
buffer. FileNameLength governs per the contract, but filter drivers
read FileName as NUL-terminated and appended adjacent heap bytes to
created names — a flaky, garbage-suffixed rename target. Both the CLI
and the launcher rename helpers are fixed.
- Uninstall retires state via rename-aside (.cbm ->
.cbm-retired-v1-<tag>-<pid>) with the retired tag shortened to 16 hex
chars so the bare rename target stays under the FileRenameInfoEx
NT-conversion ceiling at guard depths; 64 bits still uniquely
identify the generation.
- When the running launcher's mapped generation backings pin .cbm
against rename, the backings are relocated to activation-<pid>-N
.retired tombstones beside the install (a mapped image may be renamed,
never deleted; the launcher's liveness-guarded sweep reclaims stale
tombstones). Every relocation is recorded, and a FAILED uninstall
reverses the moves after restoring .cbm — via MoveFileExW with
extended-length paths on both arguments, since the deep generation
target is beyond the handle-based rename's bare-path reach — so a
restored install keeps its generation backings and stays runnable.
- After a committed uninstall the retired tree's backings are relocated
out so the tree is shallow enough for the detached cleanup's rd, and
the cleanup's working directory strips the extended-length prefix
(CreateProcessW lpCurrentDirectory silently ignores prefixed paths).
- Files created under Administrators-default-owner directories
(CopyFileW destinations, CREATE_NEW tombstones, probe directories)
are explicitly owner-stamped so the exact-owner validators hold on
runner images; guard fixtures stamp hand-built trees the same way.
Diagnostics, tests and infra:
- diagnostics: discovery is now an always-delivered JSON control record
(new cbm_log_control) that survives CBM_LOG_LEVEL suppression and
paths containing spaces; placement honors $TMPDIR with /tmp fallback
via a diagnostics-local helper; the soak parser reads the JSON record;
documented in docs/CONFIGURATION.md. Red-first coverage for suppressed
log levels, TMPDIR-with-spaces, and native Windows output-contract
assertions.
- tests(win): daemon_ipc/daemon_frontend fixtures now build endpoint
parents with production-shaped ancestry (LocalAppData on Windows, via
th_secure_runtime_parent_new) — the runtime ancestry validation
correctly refuses temp roots whose ancestors grant mutation rights to
Authenticated Users (C:/msys64/tmp, GitHub-runner work dirs) — and
drive the documented startup-owner publication flow before reading
generation-bound endpoint addresses. This turns the 26 Windows
failures previously visible in CI's full-test job green without
weakening any validation.
- tests(win): the launcher guard covers the full permanent-launcher
contract including failed-uninstall restore and immediate reinstall
after uninstall; new daemon lifecycle and reworked hook-augment
guards run the start/recycle/stop flow end to end.
- tests: CBM_SKIP_PERF is now actually consumed by the test runner
(it was set by CI but never read, so perf suites ran everywhere);
four throughput/bench suites are classified as perf, the heavy
store_arch suite moved to the slow-timeout tier, and two
wall-clock-sensitive assertions were rewritten as invariant checks
with coarse hang-detector backstops.
- build/test infra: build-dir safety contract, UI dev-proxy security
contract, soak daemon-recovery contract, path-safety helper, the
Windows VM worktree-sync contract wired into scripts/test.sh, and
vm/win.sh guards building its clean embedded-UI product in an
isolated BUILD_DIR so it cannot clobber the incremental test build.
provision-windows.sh now installs Node.js for the guards UI build.
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
The per-language chapters were LLM-drafted to establish format and
dimension mapping — their symbol targets are not yet ground-truth-derived
and must be regenerated at execution time per §3.1/§12. Make that status
explicit at the top of Appendix A so the drafted symbol names are not
read as validated.
Supersedes #420 (same diff; recreated for DCO author/sign-off).
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
Add CBM_ALLOWED_ROOT to the environment-variable tables in README.md and
docs/CONFIGURATION.md: when set, index_repository refuses a repo_path that
resolves outside it; unset imposes no restriction.
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
Documents the current behavior as implemented in src/discover/discover.c
and src/discover/gitignore.c: where .cbmignore is read from, the glob
features the parser supports, the layered precedence against built-in
skip lists / .gitignore hierarchy / git global excludes, and what
negation can and cannot override today. Planned negation unification
(un-skipping built-in dirs, non-negatable safety core, shared predicate
for auxiliary walkers) is listed in an explicitly not-yet-implemented
subsection. Linked from the README "Ignoring Files" section.
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
Add docs/SECURITY-DISCLOSURE.md describing how vulnerability reports are
handled end to end (private fix, cross-platform validation, reporter
verification, patched release, then a GitHub Security Advisory with a CVE
and credit).
Update SECURITY.md: add GitHub private vulnerability reporting as the
preferred channel, replace the over-tight 48h/7-day commitment with honest
best-effort targets for a solo-maintained project, add a safe-harbor
statement, and refresh the stale supported-versions table (0.5.x -> 0.8.x).
Signed-off-by: Martin Vogel <martin.vogel@datadice.io>
Removes the heaviest vendored grammar (66 MB) along with its language
wiring and test fixtures; the supported-language count moves to 158
across README, site, and manifest. The grammar manifest additionally
records the canonical-source decisions for the five
registry-disagreement grammars and the per-directory license files
restored earlier this week.
Describe the Hybrid LSP layer as a lightweight C implementation
structurally inspired by and compatible with major language servers,
rather than a clean-room re-implementation of their algorithms, across
README, site, and llms.txt. Drop intelephense from the reference list.
Java, Kotlin, and Rust join the Hybrid LSP sections: badge and intro counts go 6 -> 9 language families, the capability tables gain the three new resolver rows (README + docs/index.html), the clean-room reference list adds Eclipse JDT and rust-analyzer, llms.txt and the landing-page JSON-LD/FAQ are aligned, the stale test-count badge moves to 5,577, and the sitemap lastmod is refreshed.
The legacy "Deploy from a branch" Pages build ran Jekyll on every push to
main and failed on the planning docs under /docs — Jekyll's Liquid parser
hit `Unknown tag 'data'` in EVALUATION_PLAN.md, so the website build went
red after nearly every commit.
/docs is a hand-written static site (index.html + robots/sitemap/llms.txt),
so it never needed Jekyll. Deploy it with a GitHub Actions workflow that
uploads /docs as-is and only triggers on docs/** changes, and add a
.nojekyll guard. Switch the Pages source to "GitHub Actions" for this to
take over from the legacy build.
Co-authored-by: DeusData <joerge@datadice.io>
Address peer-review feedback on question provenance/external validity.
- Anchor the five dimensions to Sillito, Murphy & De Volder's catalogue of
questions programmers ask during software-evolution tasks (FSE'06/TSE'08),
with LaToza & Myers (2010) and Ko et al. (ICSE'07) as support; add a
D1-D5 -> Sillito-group mapping table
- Adopt published repo-level QA benchmarks for the major languages: reuse and
compare against SWE-QA (primary), with CoReQA, CodeRepoQA and RepoQA as
references; note snippet-level sets (CodeQA/CS1QA/CoSQA) as the contrast
that motivates structural retrieval
- Policy: anchor to Sillito; reuse SWE-QA for major languages; generate the
multilingual remainder from independent LSP/git ground truth, never from
the model under test
Rename TEST_PLAN.md to EVALUATION_PLAN.md and rebuild it as a peer-review
evaluation specification covering all 159 supported languages.
- Graph (MCP) vs Explorer (grep/glob/read) per language, 5 bespoke questions
mapped to 5 universal dimensions (D1-D5) for cross-language aggregation
- Sequential, main-channel execution: one language at a time, Explorer the
only subagent, per-language index lifecycle (cold index -> eval -> delete)
to avoid concurrent-indexing memory blowups; index time is a key metric
- LLM-as-a-Judge grading (single disclosed model, blind A/B, median of passes)
- Per-language node-type and edge-type histograms with every one of the 32
edge types listed, zeros kept (a zero is a real extraction-gap signal)
- Cross-repo intelligence + semantic/similarity deep-dive for the 9 LSP
languages, with recall/precision ground truth
- Single result tree (eval-results/), retiring the versioned-dir scheme
- Master repo assignment for all 159 languages; per-language chapters in
Appendix A
Surface previously undocumented capabilities on the site, README, and
package metadata: semantic vector search (bundled nomic-embed-code
embeddings, fully local), SEMANTICALLY_RELATED / SIMILAR_TO clone
detection, cross-repo intelligence, data-flow tracing, change-impact
analysis, and ADR management. Broaden the cross-service entry to cover
gRPC/GraphQL/tRPC and pub/sub channels. Add a "star on GitHub" nav CTA.
GEO/SEO: add a question-style semantic-search section, extend the
JSON-LD featureList and FAQPage, add semantic keywords, and refresh the
sitemap lastmod.
Correct the language count to 159 (was 155/158) everywhere while keeping
the 157 vendored-grammar figure, update the Linux-kernel index stats to
the measured 4.81M nodes / 7.72M edges, and make trace_path the canonical
tool name (trace_call_path noted as an alias).
Resolved 5 conflicts where the php-lsp branch and main both added LSP
infrastructure but with different APIs:
- Makefile.cbm: kept both TEST_PHP_LSP_SRCS and main's py_lsp/scope/
type_rep/stack_overflow test source variables in ALL_TEST_SRCS.
- internal/cbm/cbm.c: kept both #include lsp/php_lsp.h and lsp/py_lsp.h;
the per-language dispatch already supports stacking PHP after PYTHON.
- internal/cbm/lsp_all.c: kept both lsp/php_lsp.c + php_stdlib_data.c
AND lsp/py_lsp.c + python_stdlib_data.c in the compilation unit.
- src/pipeline/pass_calls.c: took main's version entirely.
find_lsp_resolution + resolve_single_call(ctx, call, lsp_calls, ...)
is language-agnostic — it matches any CBMResolvedCall whose caller_qn
hits the textual call's enclosing function and whose callee_qn
short-name hits the textual callee. PHP LSP populates the same
result->resolved_calls array via cbm_run_php_lsp, so the unified
resolver handles PHP for free. Dropped the php-branch's
qn_short_name + lsp_override_resolution helpers as superseded.
Locked confidence floor at main's CBM_LSP_CONFIDENCE_FLOOR = 0.6f
(was 0.5f on the php branch).
- tests/test_main.c: kept all suite externs and RUN_SUITE invocations
for php_lsp + py_lsp/py_lsp_bench/py_lsp_stress/py_lsp_scale +
scope + type_rep.
Auto-merged: src/pipeline/pass_parallel.c.
Brings in 52 commits from main since the branch was forked at c768b61
(#207 watcher follow-up): #208 through #281, dependabot bumps, Pine
Script, AUR docs, INHERITS edges, get_architecture wiring, codeql
revert, the entire python-lsp subsystem.
Suite: 3249 passed, 0 failed (was 2971 on main; +278 new tests from
php-lsp matches the branch's claimed test count).
BENCHMARK_PYTHON.md updated through Round 10. The stress-test suite
went from 22 patterns (19 hard / 3 KNOWN GAP) to 43 patterns (all
hard-asserted PASS). Bench held at 100% (52/52) across the surface
expansion.
Patterns added in Round 10 cover the framework / real-world cases
that distinguish parity-with-LSP from passing-toy-fixtures:
SQLAlchemy 2.0 Mapped[T], Pydantic BaseModel, async with / async
for, Callable[..., R], next(iter), **kwargs annotation, dict-literal
dispatch tables, lambda call-site inference, recursive types,
classmethod-to-method chains, dataclass default_factory, nested
match patterns.
The original Phase 11 plan called this category "compiler-grade".
The actual line where compiler-rebuild starts is now narrower:
constraint solving for ParamSpec, code-flow narrowing across
basic blocks, custom metaclass __call__, and full descriptor
protocol. None of those are needed for what's resolved on the
bench fixture or the stress suite.
BENCHMARK_PYTHON.md ends the parity push with the in-process bench
results table (81% -> 79% -> 95% across rounds), an explicit
stopping-point list of features that would require building a
constraint solver / code-flow engine / plugin system, and an
"achievable next steps" list for future iterations that stay below
the compiler-rebuild line.
The 81% -> 79% jump going from a 65-line fixture to 136 lines is
deliberate: the smaller fixture was biased toward the resolver's
strengths. The 79% -> 95% jump comes from instance-attribute typing
(Round 4) which closed the largest single gap on real-world code
shapes.
Total: 11 LSP commits on this branch (Phases 0-11 plus Rounds 1-4),
2892 / 0 passing, 95% resolution on the parity fixture.
Edge counts (laravel/framework):
- Total edges: 196,979 (baseline) -> 152,614 (-22%)
- CALLS edges: ~83,000 (baseline) -> 50,150 (-40%)
The drop is from suppressing name-fallback misroutes when the receiver
is statically typed to a vendor class that isn't indexed. The Phase-4
capabilities (generic templates, narrowing, property tracking, full
parent-walk method dispatch, expanded stdlib) cover most of phpactor's
type-resolution surface for in-project receivers.
100 unit tests, all passing. Total project tests: 2913 / 0 failed.
Final phase of Python LSP integration. Adds:
- docs/BENCHMARK_PYTHON.md — methodology document covering target
metrics (per-file LSP overhead < 10 ms, resolution ratio ≥ 40% on
application code, ≥ 70% with stdlib, < 5% spot-check false positive
rate), the declared accepted-failure-class ceiling (metaclass
__call__, __getattr__/__getattribute__, dynamic setattr, runtime
importlib, monkey-patching, eval/exec, descriptors beyond the
standard property/classmethod/staticmethod, dynamic type() classes),
and instructions for running scripts/benchmark-index.sh against the
largest Python repos in datadice (falkemedia: 232 files, api-hub:
188 files).
- tests/test_py_lsp_bench.c — in-process benchmark wired into
scripts/test.sh. Runs cbm_extract_file on a 65-line Python fixture
covering imports, dataclasses, classmethods, single-inheritance with
super(), typed parameters, attribute chains, and stdlib calls.
Asserts soft floor (resolved >= calls/4) and hard time budget
(< 50 ms under sanitizers). Initial run reports:
bench: 65 lines, 16 calls, 13 resolved (81%), 2.45 ms
81% beats the 40% application-code target and the 70%
application-plus-stdlib target. 2.45 ms under ASan + UBSan is well
inside the 10 ms (no-sanitizer) production budget.
This wraps the 11-phase plan documented in PYTHON_LSP_PLAN.md. Total
test count: 2874 / 0 (100% pass). Repository-level benchmarks via
scripts/benchmark-index.sh remain a manual operation against a fresh
production binary; methodology is now in place for when that runs.
PHP score moves from 9/12 (75%) to 10/12 (83%); the two remaining
PARTIALs (Q6 search recall, Q8 trace tool disambiguation) are non-LSP
issues tracked separately per PHP_LSP_PRE_FLIGHT.md §3 and §4.1.
Headline change is a graph-correctness win on laravel/framework: the
collide-set $prompt->value() misroute is gone, and unindexed-vendor
receiver calls no longer produce wrong name-fallback edges (total
edges drop ~14% from 196,979 to 169,268, trading recall for precision).
Pre-flight against laravel/framework establishes that the PHP graph has
systemic call-attribution drift on short names shared between a global
function and a method (e.g. value(), make(), get()). Spot-check shows
~20-30% of typed-receiver method calls misroute to the global helper
because the existing extractor cannot resolve receiver type.
Plan adds internal/cbm/lsp/php_lsp.{c,h} as a third sibling to go_lsp.c
and c_lsp.c. In-process C type resolver, no PHP runtime, no new
dependencies. Uses already-vendored tree-sitter-php grammar and the
existing CBMArena / CBMScope / CBMType / CBMTypeRegistry primitives.
Phase 1 success metric is collide-set attribution correctness, not
benchmark tier alone. Q6 (text search) and Q8 (trace tool
disambiguation) tracked as separate tickets.
- Delete cmd/codebase-memory-mcp/assets/ — skills and codex instructions
are embedded as C string literals in src/cli/cli.c since v0.5.0
- Move BENCHMARK.md to docs/ to reduce root clutter
- Fix setup scripts: replace stale go build with scripts/build.sh
- THIRD_PARTY.md: replace Go module dependencies with vendored C
libraries (sqlite3, mimalloc, mongoose, yyjson, xxhash, tre),
fix CGo reference to static compilation
The snippet handler now uses exact QN + suffix matching only (no fuzzy/auto-resolve).
Updated 3 tests to match: fuzzy_suggestions, fuzzy_last_segment, auto_resolve_enabled.
Martin Vogel