mirror of
https://github.com/abhigyanpatwari/GitNexus.git
synced 2026-09-19 00:03:33 +00:00
253 commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
dependabot[bot]
|
d80230a564
|
chore(deps): bump docker/setup-qemu-action from 4.2.0 to 4.3.0 (#3249)
Bumps [docker/setup-qemu-action](https://github.com/docker/setup-qemu-action) from 4.2.0 to 4.3.0.
- [Release notes](https://github.com/docker/setup-qemu-action/releases)
- [Commits](
|
||
|
Navid EMAD
|
506432017f
|
fix(zig): model callable-value references, and stop reporting their absence as exact (#3219)
Some checks are pending
CodeQL / Analyze (python) (push) Waiting to run
Gitleaks / gitleaks (push) Waiting to run
Publish / Classify release event (push) Waiting to run
Publish / RC guard (marker + release-PR skip) (push) Blocked by required conditions
Publish / ci (push) Blocked by required conditions
Publish / Publish to npm (push) Blocked by required conditions
Publish / Build & Push RC Docker images (push) Blocked by required conditions
Trivy Image Scan / Trivy (gitnexus-cli) (push) Waiting to run
Trivy Image Scan / Trivy (gitnexus-web) (push) Waiting to run
CodeQL / Analyze (javascript-typescript) (push) Waiting to run
Scorecard / Scorecard analysis (push) Waiting to run
* fix(impact): stop reporting 'exact' over unmodelled callable-value references
A function named in VALUE position — `bridge.accessor(Element.getNamespaceUri,
null, .{})`, `{ onClick: handler }`, a comparator handed to a sort — is
registered somewhere rather than called. The registration is modelled (a
`value-ref` site becomes a USES edge; Kythe `ref` vs `ref/call`, Joern
METHOD_REF), but the invocation THROUGH the stored value is not: it happens
later via a struct field, a registry lookup or comptime reflection.
impact()/context() nonetheless reported such a target as `epistemic: 'exact'`.
For lightpanda-io/browser that meant the DOM `Element.namespaceURI` accessor
came back with two internal callers, LOW risk and a claim of completeness —
worse than no answer, because 'exact' tells the reader not to look further.
tools.ts defines 'lower-bound' as "the walk provably missed callers", which is
precisely this case.
computeEpistemicBoundary now probes for inbound USES edges stamped with the
value-ref reason and hedges when it finds any, contributing a boundary note and
a new `causes.callableValueReferences` (unit: distinct referrer symbols).
Read from the graph, not from index metadata: unlike a dropped receiver — which
leaves no edge to find and therefore needs a persisted summary — a value
reference IS in the graph. So the signal needs no re-index and no analyzer
change, and it works on indexes written before this commit.
The cause gets its own slot rather than joining `dispatchBoundary`: a value
referrer is neither an implementation nor an interface-level consumer, and the
two differ in what the reader should do about them — a dispatch boundary is
irreducible, a callable value usually becomes traceable once the provider
models the store/load that carries it.
Language-neutral: every provider that emits a `value-ref` capture participates.
The writer and the reader now share VALUE_REF_EDGE_REASON, because drift
between them would fail silently — the probe would match nothing and every
answer would go back to claiming certainty.
* fix(zig): emit `value-ref` for a callable named in value position
`mapReferenceKindToEdgeType` has handled the registration-vs-invocation case
since #2437, and TypeScript, JavaScript and C++ all emit `value-ref`. Zig
emitted zero — so a Zig function handed somewhere as a VALUE was absent from
the graph entirely.
That is not a corner: Zig's JS bridge is built out of this one shape,
pub const namespaceURI = bridge.accessor(Element.getNamespaceUri, null, .{});
and `bridge.{accessor,function,indexed,…}` appears 2,047 times across 257 files
in lightpanda-io/browser — the project's whole JS<->Zig surface, none of it
reaching the graph. `impact` on `Element.getNamespaceUri`, which IS the DOM
`Element.namespaceURI` accessor, answered with its two in-file callers.
Three query rules, tagging `@reference.value-ref` on: a bare identifier in
argument position (`bridge.accessor(_tagName, …)`), a qualified one
(`bridge.accessor(Element.getNamespaceUri, …)`), and a const binding initialiser
(`pub const defaultHandler = onReset;`). Everything downstream already existed —
no new edge type, no schema change, no capture-machinery change, no baseline
edited.
One grammar detail is load-bearing: in tree-sitter-zig, call arguments are
DIRECT children of `call_expression` — there is no `arguments` node, only
builtins have one — so the callee has to be consumed explicitly by `function:`.
Without that binding the same rule also matches the callee of `foo(bar)` and
mints a USES edge shadowing the call's own CALLS edge. There is a test for it.
The rules are deliberately broad — `js.Bridge(Element)` and `register(count)`
match too — because the callable gate in the property-dispatch pass
(Function/Method/Constructor only) is the filter, the same design that keeps
TypeScript's `{ port: DEFAULT_PORT }` from registering anything. Measured on the
real corpus: all 3,169 emitted edges land on a Method (3,146) or Function (23).
Deliberately NOT modelled: the terminal invoke. The value reaches
`Accessor.init` -> a struct field -> `Factory.zig` reflection (`inline for`,
`@typeInfo`) -> `Caller.zig`'s `@call(.auto, func, args)` over `func: anytype`.
Resolving that needs comptime evaluation. The point is to stop dropping the
reference; the shortfall is now reported as `epistemic: "lower-bound"` by the
preceding commit instead of being papered over.
Measured on lightpanda-io/browser (698 .zig files), wiped index both times:
30,222 nodes / 71,070 edges -> 30,222 nodes / 74,229 edges. The delta is
entirely `USES` (3,169 after vs 10 before, and every one is a value-ref); CALLS,
ACCESSES, IMPORTS, HAS_METHOD, DEFINES and the rest are unchanged — purely
additive, no node invented.
The fixture joins the existing `zig-idioms` corpus rather than adding a new
one, because `bench/receiver-resolution` uses `test/fixtures/lang-resolution` as
its `--check` corpus. That gate, `scope-capture` (15 languages),
`zig-cross-file-resolution` and every other bench `--check` pass unchanged.
* fix(review): resolve qualified value references through their owner, and stop
hedging on registrations the analyzer already followed
Five findings from the review bot on #3219; four valid, all addressed.
1. QUALIFIED VALUE REFERENCES RESOLVED BY TAIL NAME (the serious one).
`bridge.accessor(Element.getNamespaceUri, …)` was resolved with
`findCallableBindingInScope(site.inScope, site.name, …)`, which never sees
the receiver and gives LOCAL bindings precedence. Reproduced:
const Element = @This();
pub fn getThing(...) // main.getThing
pub const JsApi = struct {
fn getThing(...) // JsApi.getThing
pub const thing = bridge.accessor(Element.getThing, null, .{});
};
emitted `USES JsApi → JsApi.getThing` — a WRONG edge, which is worse than
the missing edge this PR set out to fix.
`resolveValueRefTarget` now resolves a site carrying an explicit receiver
through `findClassBindingInScope` → `findOwnedMember` (the machinery
`receiver-bound-calls` already uses), gated on `CALL_TARGET_TYPES` because
`findOwnedMember` also answers with fields. A bare site keeps the lexical
walk, which is what an unqualified name means. When the owner cannot be
resolved the site is DECLINED rather than falling back: declining costs a
reference, falling back mints a confident edge to the wrong function, and
the missing reference is now reported as `lower-bound` anyway.
Cost on lightpanda-io/browser: value-ref edges 3,169 → 2,799 (−12%). Those
370 were tail-name coincidences, not registrations — all 94 `Element.zig`
`JsApi` entries survive, the cross-container case resolves and is correctly
attributed (`IntersectionObserverEntry.JsApi` → `IntersectionObserverEntry.
getTarget#0`, not the enclosing observer), and both acceptance probes are
unchanged: `getNamespaceUri` 7/3/LOW/lower-bound, `getTagNameLower`
31/10/HIGH/exact.
2. A FAILED PROBE READ AS "NO BOUNDARY". `.catch(() => [])` turned an
unanswerable query into count 0 and no note, so `exact` could be published
on the strength of a question that was never asked. It now returns `null`
and emits a boundary note. This file's own `loadMeta` comment states the
rule: a probe failing must never read as certainty.
3. NOT EVERY VALUE REFERENCE IS AN UNMODELLED INVOCATION. Where
`emitPropertyDispatchCalls` sweep 2 synthesized the dispatch (reason
`property-dispatch`), the walk did NOT provably miss the caller, so hedging
was noise over an answer that was computed — and a signal that fires on
every JS/TS hook table stops carrying information. A second probe excludes
those targets. Zig never sets a property key, so the motivating case is
untouched. The exclusion is symbol-level, not per-edge, because the graph
does not record which registration produced which synthesized call; that
residual is documented at the call site.
4. `LIMIT 50` SILENTLY UNDERSTATED THE COUNT. `rows.length` over a capped row
set published a ceiling as the documented symbol count. Replaced with
`COUNT(DISTINCT other.id)` — bounded work without a bounded answer, the
shape `countByType` in the same file already uses.
5. THE TEST DID NOT PIN THE CALLER COUNT its own comment promised. Added
`expect(result.impactedCount).toBe(2)`.
New regression tests: qualified references bind the written owner and not the
nearer lexical match; an unresolvable receiver emits nothing rather than a
wrong edge; a dispatch-modelled registration stays `exact`; a probe that cannot
run hedges instead of claiming certainty.
Known limitation, pinned by a test rather than left implicit: when a `@This()`
alias's NAME differs from its container's (`const Element = @This();` inside
`main.zig`), the receiver does not resolve and the reference is declined. The
provider has `rewriteZigThisAlias` for this, but it is applied to type nodes
and extending it to reference receivers would change existing CALL-site
behaviour. Lightpanda's `const Foo = @This();` in `Foo.zig` convention makes
the names coincide, which is why the corpus is unaffected.
* fix(review): bump the parse-cache schema and resolve module-owned value references
Review round 2 on #3219. Four inline items, all reproduced against the
worktree before deciding.
P1 — the new captures could stay INERT on a warm parse cache. Adding
`@reference.value-ref` rules to `ZIG_SCOPE_QUERY` changes
`ParsedFile.referenceSites`, which is a PARSE-TIME fact, but `SCHEMA_BUMP`
stayed at 93. A repo indexed before this branch and re-analyzed after it
replays the old, empty site list for every unchanged `.zig` file — `--force`
included, since shards are content-addressed — so no USES edge is emitted, the
boundary probe measures a real zero, and `impact` on a registered accessor goes
back to `epistemic: "exact"`. #3399 un-fixed on the incremental path most users
are on, with every cold-run test still green. DECISIONS D1-1's "zero changes to
… the schema" conflated the graph schema with the cache schema.
Bumped 93 -> 98, not 94: #3190 claims 94 and #3179 claims 94 through 97 in one
PR. `incremental-parse-cache.test.ts` re-pinned, with 93-97 added to the taken
list. Re-check against origin/main and open PRs immediately before merging.
P2 — a qualified value reference through a MODULE was declined with no hedge.
R1-2 resolves a written receiver with `findClassBindingInScope`, which requires
`isClassLike`; a namespace-only `@import` handle is not class-like, so
const dom_utils = @import("dom_utils.zig"); // no `@This()` in that file
pub const comparator = bridge.accessor(dom_utils.compare, null, .{});
resolved to nothing. That is not the conservative half of R1-2's trade-off: a
declined site emits NO edge, so there is nothing for the probe to read and
`impact` on `compare` reports `exact`. Silence, not a hedge — and the pass
comment claiming otherwise was wrong on this path.
`resolveValueRefTarget` now tries the second kind of owner a qualified name can
have. `findNamespaceValueRefTarget` reads the file's `namespace` import edges
for the handle and the target module's own `origin: 'local'` module-scope
bindings for the member — the same channel `receiver-bound-calls` Case 1
already trusts for `dom_utils.compare()`, with Case 1's three guards for Case
1's reasons: `isNamespaceNameShadowed`, local-origin bindings only, and two
distinct defs under one name resolve nothing. `CALL_TARGET_TYPES` gates module
owners exactly as it gates container owners. Language-neutral: it reads generic
namespace import edges, names no language.
Still declined, deliberately: a receiver this index knows under no name at all
(the `@This()`-alias case, owners outside the workspace). There the alternative
is a confident edge to a lexically-nearer function the source did not name.
P2 — `impact-callable-value-references.test.ts` was in neither vitest list.
It opens a real engine via `withTestLbugDB(poolAdapter: true)`, so TESTING.md
puts it in the `lbug-db` include list and the `default` exclude list; it was in
neither, so `default` also collected it into the parallel pool. Added next to
its `impact-epistemic-lower-bound` sibling in both arrays.
P3 — DECISIONS.md was stale against HEAD and embedded host paths. D2-3 still
described the `LIMIT 50` that R1-5 removed and D1-3 still described the
receiver-blind resolution that R1-2 replaced; both now carry explicit
"superseded by" pointers. The `~/code/...` and mise-node paths are replaced
with placeholders.
Two smaller corrections the new cause made necessary:
- `formatImpactResult`'s `lower-bound` header hard-coded "callers binding via
DI / dynamic dispatch", which now contradicts the value-reference bullet
printed directly under it. The bullets carry the cause; the header only
states that the count is a floor.
- `tools.ts` said a `causes.callableValueReferences` of 0 means "nothing was
missed". The dispatch exclusion is symbol-level, not edge-level, so a target
with both a followed registration and an unfollowed escape also reads 0. The
docs now say a 0 means "no unfollowed registration was proven".
Fixture: `src/webapi/dom_utils.zig` (namespace-only) plus three cases in
`Element.zig` — the module-qualified registration, a non-callable module member,
and a `u8` parameter shadowing the handle. The shadow case was verified to FAIL
with the guard disabled, so it is not passing for an unrelated reason.
Gates: `tsc --noEmit` clean, `npm run build` clean, prettier clean.
`resolvers/` 3,630 passed / 3 skipped; `unit/scope-resolution` 2,010 passed;
`impact-callable-value-references` 7 passed under `lbug-db`;
`incremental-parse-cache` 40 passed; eval formatters 104 passed. Bench --check
all PASS with no baseline edited: receiver-resolution, zig-cross-file-resolution,
scope-emission, callable-value-flow, scope-capture (15 languages), python-scope.
* fix(review): guard the class receiver against a shadowing binding, and pin the dispatchability partition
Review round 3 (`gitnexus-check` bot on `cf53bbaa`). Three findings, each
reproduced against the code before deciding.
R3-1 (Error, valid, REPRODUCED) — a CLASS receiver could resolve through a
shadowing value binding. `findClassBindingInScope` is a class-only walk: it
filters the scope chain by `isClassLike`, so it steps over a nearer binding that
is a value and keeps climbing — and past the chain entirely, into a
qualified-name fallback that answers with the unique workspace definition of the
name. A `u8` parameter named `Ticker`, in a file that neither declares nor
imports the `Ticker` container another file defines, therefore emitted
`register(Ticker.fire)` as a confident USES edge to that container's method.
That is exactly the wrong-edge failure R1-2 exists to prevent, arriving through
the class channel instead of the lexical one.
Fixed with `isOwnerNameShadowedBySomethingElse` — a sibling of
`isNamespaceNameShadowed` with one extra clause. The plain namespace guard could
NOT be reused: a container is often its own local declaration
(`fn make() { const Local = struct {…}; register(Local.go); }`), and reading that
binding as its own shadow suppresses precisely the resolutions this path exists
to make — the #2723 mistake, one channel over. So a scope that binds the name
answers immediately, and the answer is "not shadowed" only when one of that
scope's own bindings IS the def just resolved.
Both halves are pinned and both were verified to fail when the guard is
weakened: the parameter case fails with no guard at all, the local-container
case fails with the plain `isNamespaceNameShadowed`.
R3-2 (Warning; mechanism correct, unreachable today; fragility fixed instead) —
the dispatch exclusion could suppress an unfollowed registration. The bot is
right about the code: sweep 2 synthesizes CALLS only for a registration whose
site carried a `propertyKey`, while the exclusion zeroes the note on ANY inbound
`property-dispatch` CALLS edge. It is not reachable in the current rule set, and
the reason is measured rather than assumed: `@reference.value-ref` is emitted by
exactly three languages — JavaScript (2 rules), TypeScript (2), Zig (3) — every
JS/TS rule also captures `@reference.property-key` (both are object-literal
shapes) and no Zig rule does. A dispatchable registration is therefore always a
JS/TS one, an undispatchable one always a Zig one, and they cannot meet on one
symbol.
Rejected: splitting the edge `reason` into dispatchable / undispatchable. It is
the precise fix, but it is a graph-content change that churns whichever side
keeps the old literal — the Zig, TypeScript and probe suites all pin
'scope-resolution: value-ref' by hand as a drift canary — and it buys nothing
against a case no rule can produce.
What was actually wrong is that the exclusion's soundness rested on a
coincidence recorded nowhere, in files nobody reading `local-backend.ts` would
open. Fixed at both ends: the exclusion site now states the invariant, the three
facts it rests on and the two options for when it breaks; and
`value-ref-dispatchability.test.ts` fails the day it does — a JS/TS rule for a
bare callback argument, a Zig rule that grows a key, or a fourth language
emitting `value-ref` at all. Verified to fire (adding a property key to a Zig
value-ref rule fails the Zig case), and its rule splitter has its own guard test
so the suite cannot pass vacuously. `ZIG_SCOPE_QUERY` is exported for that test
only.
R3-3 (Nit, valid) — a test comment claimed the wrong epistemic result. The
`declines a qualified reference whose receiver cannot be resolved` case said the
shortfall shows up as `lower-bound`. It does not: with no edge there is no
evidence and the target stays `exact`. The pass docstring was corrected in round
2 and this comment was missed. It now says the decline costs the reference AND
the hedge, and why that is still the right trade.
Gates: tsc --noEmit clean, npm run build clean, prettier clean.
`test/integration/resolvers` 3,632 passed / 3 skipped (70 files);
`test/unit/scope-resolution` 2,015 passed (120 files);
`impact-callable-value-references` 7 passed under `lbug-db`. Bench --check:
receiver-resolution, zig-cross-file-resolution, scope-capture (15 languages),
scope-emission PASS with no baseline edited; callable-value-flow failed once on
its TIMING budget (2.006 > 1.9) with a byte-identical fingerprint, then passed
twice at 1.788 / 1.813 — machine load, not a regression.
* fix(review): let a file's own `@import` outrank the workspace-wide class fallback
Found by running the local preflight harness before pushing rather than after.
One of its five findings is a real hole in R2-2; the rest are documentation that
overclaimed.
R4-1 (valid, REPRODUCED) — the container channel preempted the file's own
`@import`. R2-2 added the module channel as a FALLBACK after
`findClassBindingInScope`, and that order is wrong: `findClassBindingInScope`
does not stop at the scope chain. When its `isClassLike` walk misses — and a
namespace handle binds a Module, so it always misses — it falls back to
`scopes.qualifiedNames`, a workspace-wide index, and answers with the unique def
of that name anywhere in the repo. A container named `dom_utils` in a file
`Element.zig` never imports therefore captured
`bridge.accessor(dom_utils.compare, …)`, binding
`Method:src/webapi/decoy.zig:dom_utils.compare#2` while the module channel that
would have answered correctly was never reached. R3-1's shadow guard cannot
catch it: the import binds at MODULE scope, which that guard treats as the floor.
Fixed by trying the module channel FIRST. An `@import` written in this file is
the strongest available statement about what the name means here and outranks a
global uniqueness guess; when the handle is not an import of this file the
channel answers nothing and the container path runs exactly as before. Pinned by
`decoy.zig` and a strengthened assertion on the existing module-owner test,
which fails on the old order.
R4-2 — the cause documentation named shapes nothing captures. `tools.ts`
illustrated `callableValueReferences` with "a callback argument", "a stored
function pointer" and `qsort(xs, n, sz, compareItems)`. Only Zig captures a call
argument or a const initialiser; JS/TS capture only object-literal property
values, and C has no value-ref rule, so the `qsort` example is counted in no
language. Both cause blocks now name the captured shapes and say that a bare
JS/TS callback argument is not among them, so a 0 does not rule it out.
R4-3 — the same block exempted itself from the re-index caveat this PR proves it
needs. "Read from the graph, so it needs no index-time metadata" is true of the
probe and false of the edges: an index built before a language emitted these
captures has none and reports 0 — the warm-cache failure R2-1 bumped
SCHEMA_BUMP for. It now says to re-analyze before reading a 0 as measured.
R4-4 — the dispatchability canary was narrower than its own promise. Its header
claimed it fails on "a fourth language emitting value-ref at all"; it reads
`languages/<dir>/query.ts`, so Vue — which owns no query and borrows
`emitTsScopeCaptures` / `emitJsScopeCaptures` — emits value-refs while the
assertion lists three languages, and a capture synthesized in code is invisible
to it. The case now asserts on query OWNERS, which is what it checks and is
sound because a delegating language inherits the rules it borrows; the header
states the synthesized-capture gap instead of letting a green tick imply it away.
R4-5 — the BARE docstring described a lexical walk that is not one.
`findCallableBindingInScope` applies the callable predicate WHILE walking, so a
nearer parameter or local is stepped over: the defect R3-1 fixed on the container
channel, unguarded here, pre-existing since #2437 and reachable in JS/TS. Out of
scope for #3399, so behaviour is unchanged and the sentence now says what the
walk does rather than implying a guarantee it does not give.
Gates: tsc --noEmit clean, npm run build clean, prettier clean.
`test/integration/resolvers` 3,632 passed / 3 skipped (70 files);
`test/unit/scope-resolution` 2,015 passed (120 files);
`impact-callable-value-references` 7 passed under `lbug-db`. Bench --check:
receiver-resolution, zig-cross-file-resolution, scope-capture (15 languages),
scope-emission, callable-value-flow all PASS with no baseline edited.
* fix(review): honour the hub opt-in for value references, so `hub.fn` means one thing
Review round 5 (`gitnexus-check` bot on `1c7c05ff`). One finding, valid and
reproduced before fixing.
`findNamespaceValueRefTarget` accepted only `ref.origin === 'local'`. R2-2
recorded that as deliberate — "the `namespaceExportsIncludeImportedNames` hub
opt-in is a provider decision this language-neutral pass does not make" — and
that reasoning was wrong twice over. Zig sets the flag
(`languages/zig/scope-resolver.ts:41`, measured on ghostty and tigerbeetle before
it landed), and the pass does have the provider in scope: `runScopeResolution`
takes one and already forwards several of its hooks to other passes.
The result was the exact asymmetry R2-2 argued against in its own first
paragraph. A Zig hub declares nothing — every name it publishes it imported — so
requiring a local declaration declines every member reached through one:
// hub.zig
pub const scale = @import("dom_utils.zig").scale;
// Element.zig
const hub = @import("hub.zig");
pub fn callsThroughTheHub(v: u8) u8 { return hub.scale(v); } // resolved
pub const scaled = bridge.accessor(hub.scale, null, .{}); // declined
One name meaning two different things depending on whether a `(` follows it.
Fixed by forwarding `provider.namespaceExportsIncludeImportedNames` into the pass
and consulting the published channel when it is set — the same question
`receiver-bound-calls` Case 1 asks, through the same `lookupBindingsAt` read
`findExportedDefIncludingImportedNames` performs for the CALL form. The pass
still names no language; it asks the provider, which is the sanctioned hook.
Precedence is unchanged where it mattered: a locally declared member still wins
over a republished one, two distinct defs under one name still resolve nothing,
and `CALL_TARGET_TYPES` still gates the answer — pinned by `hub.DEFAULT_NS`, a
re-exported CONSTANT, which stays unregistered. Languages that do not opt in are
unaffected: the parameter defaults to `false`, and passing `false` was verified
to fail the new hub test, so it is load-bearing. The fixture republishes a member
(`scale`) that nothing else uses, so the hub assertion cannot be satisfied by an
edge another case emitted.
Gates: tsc --noEmit clean, npm run build clean, prettier clean.
`test/integration/resolvers` 3,634 passed / 3 skipped (70 files);
`test/unit/scope-resolution` 2,015 passed (120 files);
`impact-callable-value-references` 7 passed under `lbug-db`. Bench --check:
receiver-resolution, zig-cross-file-resolution, scope-capture (15 languages),
scope-emission, callable-value-flow all PASS with no baseline edited.
Also recorded in DECISIONS.md: `/autofix` returned "No successful autofix run"
because the `PR Autofix` run on `1c7c05ff` failed at `actions/upload-artifact`
with a 403 from GitHub's artifact storage, not because of anything in this
branch. This push triggers a fresh run.
* fix(review): inspect the module scope in the owner-shadow guard, and check the TSX suffix
The `gitnexus-check` pass on `bd6e577e` carried three findings of its own, and I
answered the later `1c7c05ff` pass without noticing them. Recording that as a
process failure too: bot reviews are per-head, and a later pass does not
necessarily repeat an earlier one's findings.
R6-1 (Error, valid, REPRODUCED) — the shadow guard stopped one rung short.
`isOwnerNameShadowedBySomethingElse` returned `false` on reaching the module
scope, justified as "a container declared there IS the binding, and the caller
already resolved it". That holds when the owner came from the scope chain and
fails when it came from the workspace-wide qualified-name fallback:
// Gauge.zig — never imported by Element.zig
const Gauge = @This();
pub fn read(self: *Gauge) u8 { … }
// Element.zig
const Gauge = @import("dom_utils.zig").DEFAULT_NS; // NOT a container
pub const level = bridge.accessor(Gauge.read, null, .{}); // → Gauge.zig's read
`findClassBindingInScope` steps over the module-scope binding because it is not
class-like, answers from `scopes.qualifiedNames`, and the guard waved it through.
Worth recording: the first fixture attempt did NOT reproduce. A local
`const Gauge: u8 = 3;` also claims the workspace qualified name `Gauge`, leaving
two candidates, and the fallback refuses to guess between two — so the shape
defeated itself. Binding the name by IMPORT claims no qualified name, the
fallback stays unique, and it fires. A negative result on the first shape was not
evidence the finding was wrong.
Fixed by inspecting the module scope as the last rung instead of skipping it.
The identity exemption is what makes that safe where `isNamespaceNameShadowed`
cannot do it (#2723: a namespace import writes its own name into the module scope
and would read as its own shadow) — the binding that IS the owner exempts itself,
and only a binding to something else answers `true`. `lookupBindingsAt` is
consulted at that scope and only there, because an imported alias lives in the
finalized channel rather than in `scope.bindings`.
R6-2 — the hub re-export finding, already fixed in `5d8fe9d8`; the same defect
restated on the later head.
R6-3 (valid, fixed) — the dispatchability canary omitted the TSX suffix.
`getTsScopeQuery` analyzes a `.tsx` file with `TYPESCRIPT_SCOPE_QUERY +
TSX_JSX_QUERY_SUFFIX`, and the test read only the base, so a `value-ref` rule
added to the suffix would be emitted in TSX analysis with the canary green. The
suffix is now exported and concatenated into the check; verified load-bearing by
adding an unkeyed `jsx_expression` value-ref rule to it, which fails the
TypeScript case.
Gates: tsc --noEmit clean, npm run build clean, prettier clean.
`test/integration/resolvers` 3,635 passed / 3 skipped (70 files);
`test/unit/scope-resolution` 2,015 passed (120 files);
`impact-callable-value-references` 7 passed under `lbug-db`. Bench --check:
receiver-resolution, zig-cross-file-resolution, scope-capture (15 languages),
scope-emission, callable-value-flow all PASS with no baseline edited.
* perf(bench): gate callable-value reference resolution on linear scaling
`resolveValueRefTarget` (#3399) replaced one lexical walk with four
channels, and the last of them — a qualified receiver resolved through
`findClassBindingInScope` — falls back to `scopes.qualifiedNames`, a
WORKSPACE-WIDE index consulted once per site. Keyed that is O(1); scanned
it is O(files) per site, and a registration table that costs O(sites)
today costs O(sites x files) tomorrow. Nothing in the suite can see that:
the fixtures are single-file, and the corpus it actually matters on is
lightpanda-io/browser, where `bridge.{accessor,function,…}` appears 2,047
times across 257 files.
`bench/value-ref-resolution/measure.mjs` builds two synthetic Zig corpora
of identical shape 4x apart in file count, and times ONLY the per-site
resolution loop — extraction, `reconcileOwnership` and finalize are setup.
`linear_factor` is `(t_large/t_small)/(N_large/N_small)`: measured
1.00-1.09 across runs, with `us_per_site` flat at 1.4-1.5 between the arms.
Correctness comes first, because a timing gate alone is satisfied by a fast
wrong answer: exact site/resolved/declined counts per arm plus an
order-independent sha256 over every (site -> resolved target) pair. The
corpus exercises all four channels — CONTAINER, NAMESPACE, HUB and BARE —
and carries two DECLINE controls per module (a non-callable namespace
member, a non-callable bare argument), so a widened callable gate moves
`declined` instead of hiding inside the timing.
Verified load-bearing rather than assumed: making the qualified lookup scan
a workspace-sized collection leaves the fingerprint IDENTICAL and takes
`linear_factor` to 3.752 against a slack of 1.375 — the regression class
this exists for is exactly the one no correctness gate can see.
Zig is the corpus because it is the only language whose provider sets
`namespaceExportsIncludeImportedNames`, so it is the only one that can
exercise the hub channel at all; the pass itself names no language.
`resolveValueRefTarget` is exported for the bench. Timing
`emitPropertyDispatchCalls` instead would fold the signal into edge
emission, and re-implementing the channel order in the bench would pin the
bench's idea of the function rather than the function.
* feat(zig): index every build package in the repo, not only the root one
Zig already had workspace setup — `loadZigBuildConfig` has parsed
`build.zig.zon` `.path` deps, the root `build.zig`'s named modules and each
build module's own `addImport` table since #1432, threaded through
`ScopeResolver.loadResolutionConfig` exactly as tsconfig is for TypeScript.
What it did not have is the part `tsconfigFor` supplies: per-package scope.
`loadZigBuildConfig` reads `<repoRoot>/build.zig{,.zon}` and nothing else, so
a repo laying its packages out as `packages/<name>/build.zig` — no root build
files at all — got `null`, and EVERY bare `@import("<module>")` in it went
unresolved. Cross-file resolution silently degraded to relative imports.
Measured on a two-package probe before this change: `config = null`,
`@import("core")` from `packages/app/src/main.zig` → `null`.
`loadZigWorkspaceIndex` discovers packages the way `findTsconfigFiles`
discovers configs — one bounded breadth-first walk that skips the hardcoded
ignore set — and `zigPackageFor` is the `tsconfigFor` analogue: the nearest
enclosing package governs a file, deepest-first, with no fall-through to an
outer package. Fall-through is what makes a vendored dependency's
`@import("config")` resolve to the outer repo's `config` module, the same
failure `loadTsconfigIndex` documents for a package declaring no `baseUrl`.
`resolveZigImportInternal` is UNCHANGED — impact analysis puts it at HIGH risk
with 6 dependents, and it does not need to move: it is handed one package's
config, and which config it receives is the only difference. Its 48 existing
tests pass untouched. A nested package's paths are rebased to repo-relative at
load time (`.path = "../core"` → `packages/core`, which the root-relative
reading rejects outright as an escape); the ROOT package keeps its raw
spelling, which is what `parseZigBuildZon` promises and its tests pin, so a
single-package repo is byte-identical.
`loadImportConfigs` — which runs unconditionally for every repo, Zig or not —
keeps calling the root-only loader, so no non-Zig repo pays for the walk. That
is the split TypeScript already has between the cheap `loadTsconfigPaths` and
the repo-walking `loadTsconfigIndex`, which `loadResolutionConfig` reaches only
during a language pass.
The `zig-monorepo` fixture carries the discriminating case rather than only the
happy path: `tool` binds the alias `core` to its OWN `src/core.zig`, so a
repo-wide flattened module map — the shape a workspace index invites — would
point `measure` at `packages/core`, a confident edge into a package `tool` does
not depend on. That is worse than the unresolved import this fixes, and only
per-package scoping keeps them apart.
Tests: 7 unit cases pinning the index (including `loadZigBuildConfig` answering
`null` on the same fixture, side by side) and 3 integration cases pinning the
EDGES — cross-package CALLS from the module root AND from a non-root file, and
`tool` not crossing over. Verified load-bearing: restricting the index to the
root package fails all three.
Second consumer caught by the integration test and fixed with it:
`populateZigWorkspaceStaticGating` reads the same `resolutionConfig`, per file
because two files of that pass can belong to different packages.
Gates: build clean, `tsc --noEmit` clean, prettier clean, eslint 0 errors.
`test/integration/resolvers` 3,768 passed / 4 skipped, `test/unit/scope-resolution`
2,015 passed. Bench `--check` with NO baseline edited: `receiver-resolution`
(whose corpus is `test/fixtures/lang-resolution`, where the fixture lands),
`zig-cross-file-resolution`, `value-ref-resolution`, `scope-capture` (15
languages), `import-target`, `callable-value-flow`, `scope-emission`.
* perf(zig): dequeue the package walk by head index, not `shift()`
`findZigPackageDirs` pushes children while it drains the queue, which keeps
the array in a mode where `Array.prototype.shift()` memmoves the whole
remainder rather than taking V8's left-trimming fast path — so the walk is
quadratic in the frontier, bounded only by `ZIG_SCAN_MAX_DIRS` (20,000).
Measured at that bound rather than estimated from the move count: 53 ms at
fan-out 4 and 81 ms at fan-out 20, against 0.8 ms with a head index — 66-106x,
and paid before a single config is read.
FIFO order is unchanged, and the package ordering never depended on the walk
anyway: `loadZigWorkspaceIndex` sorts by (dir length desc, localeCompare).
Memory is unchanged too — the entries were already retained by the pushes;
`shift()` only dropped the head.
`findTsconfigFiles` and `loadNodeWorkspacePackages` carry the same walk with
the same bound and are equally affected. Deliberately NOT fixed here: they are
pre-existing, outside this PR's diff, and reach TypeScript/Node import
resolution, which nothing in this change set covers. Filed separately.
Reported by gitnexus-check on
|
||
|
Gergő Magyar
|
e2da8d90ce
|
test(eval): run the benchmark offline against a scripted provider (#3235)
* feat(eval): a scriptable stand-in for Anthropic and OpenAI Every defect this harness shipped last round was invisible to its own tests for one reason: the tests exercised a layer BELOW where the code runs. The usage log was never written because the proxy is a subprocess with a constructed environment. The callback could not be imported because LiteLLM loads it by path, not as a package. Failures went unrecorded because only the async hook was overridden. CI or review caught all three; no unit test could, because each called the function directly instead of driving the path that calls it. This closes that gap without spending money. It speaks the two wire protocols the harness actually depends on - Anthropic Messages, streaming and not, and OpenAI Responses - so a run can go through the real sandbox, the real CLI, the real gateway and the real usage callback with only the model faked. The runner already supports pointing at it: --base-url is the same path the free-model proxy documentation uses. Scripted rather than simulated. A test decides what the model says, which tools it asks for, and exactly what usage it reports. That last part is what makes provider-native accounting testable at all: real cache hits are not reproducible on demand, but a declared cache_read of 44,000 is. One Reply served down both protocols is also the cleanest demonstration that the same billed work is stated as a sum on one side and as a whole on the other. Tool blocks are the mechanism for artifact-producing cells. The CLI runs what it is asked to run, so a scripted Write block makes it write that file inside the sandbox for real - no model deciding anything. The end-to-end test drives the real proxy against the mock and asserts the usage log records the provider's own arithmetic through the Anthropic-shaped translation. It SKIPS here, because litellm's console script is absent in this environment, so it is unverified until CI runs it - the same footing the bubblewrap canary started on, and that one found a real bug on its first CI run. Not yet built: driving a whole sweep against this. That needs a scripted reply sequence that carries a cell to a scored artifact, which is the next step and the point of the exercise. 668 eval tests pass, 17 skipped; the two test_model_gateway.py failures are the pre-existing environmental ones. * test(eval): run a real session against the scripted provider The mock only proves something once the harness runs against it. This adds the stand-in CLI and the first integration tests that use it, so a session goes through the real code with only the model faked. tests/fixtures/fake_claude.py does what the CLI does at the two boundaries the harness depends on: it calls ANTHROPIC_BASE_URL for a turn, EXECUTES the tool blocks that come back, and prints the stream-json sequence the parent parses. Everything between - the session runner, the event-stream parse, the usage extraction, the artifact capture, the scorer - stays real. Four tests, chosen for the layers that have actually broken here: the usage a provider reported survives to the row, a scripted Write produces an artifact parse_review_output accepts, the prompt the harness meant to send is what arrived, and an upstream 529 lands as a failed session rather than a usable measurement. Writing the stand-in found two things worth keeping. The prompt arrives on STDIN under "-p --input-format text"; scanning argv for a non-flag token picks up a flag's value instead, and the prompt-fidelity test is what caught it. And three of these tests had been holding a sandbox they never applied, since no command_prefix is passed - that implied coverage which was not there, so the sandbox is gone from them and stays only in the artifact test, which needs its review directory. What these do NOT cover, checked rather than assumed: making the stand-in write in place instead of atomically still passes. On the host-unsafe backend there is no read-only mount to refuse it, so the atomic-write requirement remains a bubblewrap mount property that only the real-sandbox canary can prove. Dropping cache_read from the recorded usage does fail, so that half is genuinely pinned. 672 eval tests pass, 17 skipped; the two test_model_gateway.py failures are the environmental ones. * fix(eval): the usage adapter read a shape the callback never receives Running the gateway against the scripted provider proved the accounting merged in #3220 does not work, and the same run showed why nothing had caught it. LiteLLM does not hand a logger the upstream body. It normalises usage into its own Chat-Completions-shaped object first, so an OpenAI Responses reply reaches the callback as prompt_tokens / prompt_tokens_details.cached_tokens - never the input_tokens / input_tokens_details the shipped adapter reads. Every field came back unknown. The observed call_type is "anthropic_messages" as well, because Claude Code calls the Anthropic-shaped endpoint, so canonical_provider returned None and normalize_usage would have refused outright. Both were assumptions about a boundary I had only read about. The unit tests agreed with them because their fixture was written in the same wrong shape, so producer and consumer were consistent and both wrong - the exact failure the producer/consumer round trip exists to catch, one layer further out. Adds a LITELLM_NORMALIZED adapter for the object that actually arrives. The arithmetic is still OpenAI's - prompt_tokens is the whole, the details are subsets - so ordinary input is recovered by subtraction. The Responses adapter stays for a raw upstream body, which the mock still serves and tests directly. An unrecognised provider is still refused rather than guessed. The fixtures now carry the measured shape, and the end-to-end test asserts it through a real proxy: 48k prompt tokens with 44k cached is read back as 3k ordinary rather than as silence. 676 eval tests pass, 16 skipped, none failing. * test(eval): run a whole sweep offline, with negative controls The layers between a model turn and a promotion decision had never been exercised together. Unit tests covered each alone, and the paid runs that would have covered the composition kept dying, so the contracts BETWEEN them went unverified - which is where this harness has repeatedly shipped bugs. Drives runner.main() the way the workflow does. Real task selection, hidden oracle capture, sandbox, CLI subprocess, artifact capture, scoring against the oracle, aggregation, health guard and promotion gate. Only the model is scripted. Getting to green meant satisfying nine real contracts nothing had exercised end to end, and each failure was the harness correctly refusing bad evidence: --unsafe-no-bwrap is restricted to the paired review arms; ce_* needs a plugin carrying ce-plan, ce-work and ce-code-review; candidate_* needs an overlay; the clone needs .gitnexus/meta.json with indexedAt and lastCommit; the evidence gate needs a Skill request with a non-error result; review findings need exactly ten fields with severity in critical/high/medium/low; and the hidden labels use a DIFFERENT schema from the review output - line_start/line_end, six fields. That last one only a real run surfaces. Three negative controls, because a scorer that cannot be wrong measures nothing. A finding in the wrong place is tp=0 fp=1 fn=1 and oracle-failed, while its evidence stays VALID - being wrong is a quality result, not a broken measurement. Approving defective code is a miss with no false positive, and precision is None rather than 0, because it is undefined with no predictions. One run cannot promote: the gate says it needs three valid paired runs. A fourth control exists because a mutation demanded it. Forcing skill_was_invoked_events to return True left every other test here passing, so nothing pinned the gate that separates measuring a SKILL from measuring a model. Writing it turned up behaviour worth recording rather than assuming: a skill-not-invoked row still carries its score AND still counts toward the arm median, because aggregate() drops EXCLUDED_ERROR_KINDS and evidence_valid=False and skill-not-invoked is neither. The health guard stops the sweep, so a single-run sweep cannot promote on it, but a mixed run's median would include a cell whose skill never ran. Pinned as-is so it cannot change silently in either direction; changing it is a promotion-semantics decision, not a test fix. Two provisioning steps are stubbed and neither is harness logic: the pinned runtime mounts (no node_modules in a worktree) and the sanitized graph build (needs the gitnexus CLI at a mounted path). Containment is host-unsafe here; bubblewrap stays with the real-sandbox canary. 681 eval tests pass, 16 skipped, none failing. Runs in ~18s. * fix(eval): an uninvoked skill must not move the arm's quality median Found by the offline sweep: a skill-not-invoked row still carried its score into the arm's quality median. aggregate()'s filter dropped EXCLUDED_ERROR_KINDS and evidence_valid=False, and skill-not-invoked is neither, so an arm could be credited for a review it never performed with the skill under test - which is the one thing an arm exists to measure. Excluded from the QUALITY metrics only. Cost and duration still count that row, because the session really ran and really was billed, and the promotion gate still sees it, because it has its own vocabulary for a candidate that never loaded its skill. Two wider fixes were tried and abandoned, both because the tests said so rather than because I reasoned it out first. Reusing the health guard's evidence_failed predicate also excluded transcript-missing rows, but test_aggregate_excludes_session_error_rows_from_medians pins those as counting: that session ran, only its transcript is unverifiable. Excluding the row from `valid` outright turned a candidate whose skill never loaded from keep_incumbent into insufficient_evidence - the safety property held either way, but the decision vocabulary is promotion semantics and not mine to change on a measurement fix. Mutation-checked: putting the rows back into the quality median fails the new test. Both directions asserted, since a filter that excludes everything would also pass - a wrong-but-valid review still moves quality, because being wrong is exactly what a quality median should reflect. 682 eval tests pass, 16 skipped. * test(eval): run the offline sweep unstubbed in the job that can, and probe CLI identity Items 5 and 6 turned out to be one change. The containment (ubuntu) job already installs bubblewrap, the pinned Claude CLI, node_modules and a built GitNexus - everything the sweep's two provisioning stubs stand in for. So the stubs are not a property of the test, only of a machine that lacks those things. GITNEXUS_REQUIRE_FULL_SWEEP=1 makes the sweep run with nothing stubbed: real containment instead of --unsafe-no-bwrap, the real runtime mounts, the real sanitized graph. Set in that job, following the GITNEXUS_REQUIRE_BWRAP_CANARY pattern already there. The gate FAILS on a missing piece rather than degrading to the stubbed path, which is the point - a green tick that silently tested less is what the bubblewrap canary was written to prevent. Verified both states here: default green, and gate-on fails on this machine rather than skipping, since it cannot create user namespaces. Item 7 is an experiment, not an answer. Per-cell attribution needs an identifier that travels WITH the request, because one proxy serves the whole sweep and anything read from its environment is identical for every call. What the real CLI sends is not documented anywhere I can check, and guessing a wire format is exactly how the last three accounting bugs happened. So the probe drives the REAL pinned CLI against the mock and records the identity-bearing headers and body keys that arrive. It asserts only that a request was made; the recorded evidence is the deliverable, and the job log preserves it. Skips without CLAUDE_CANARY_BIN. Two guards caught this rather than review: the repo pins the containment job's env and its exact test list, so both had to be updated deliberately - which is the guard working, not friction. 682 eval tests pass, 17 skipped. * test(eval): make the offline sweep cross-task, so a scheduler change is checkable The sweep fixture had one task, and a single task cannot show the thing a cross-task scheduler changes: waves are per-task, so ordering, packing and a breaker spanning a task boundary are all invisible with one. A second task with its defect in a DIFFERENT file, and its own hidden labels, makes per-task routing observable. The scripted reply is now task-aware, which matters for the same reason: replying with the first task's finding scores the second task wrong. The load-bearing assertion is that each task scored against ITS OWN oracle. That is the dangerous failure mode of interleaving cells from different tasks - a mis-routed context or artifact scores one task against another's labels, and every row still looks green. Mutation-checked: pointing every cell at the first task's oracle snapshot fails it. This is the safety net the packed-scheduler wiring needs. Measured earlier against the real sweep_packed_cells, that change is worth -27% on a cold sweep and -37% weekly, with breaker fidelity holding at three injected failure positions - but it restructures a 125-line loop across ~92 names that also holds graph prefetch, reuse selection, oracle staging and the canary drop. Landing that on top of a one-task fixture would have been unverifiable, which is why this comes first and separately. 682 eval tests pass, 17 skipped. * fix(eval): commit the stand-in CLI's executable bit The file was created and chmod +x'd locally, but committed 100644 - so the mode existed only in my working tree. Any fresh checkout, CI included, gets a non-executable file and every cell dies with "required executable is not an executable regular file". Found by accident: checking out origin/main and back to compare a flaky test restored the file from the index and stripped the bit, which turned 5 green tests into 9 failures. Without that detour this would have failed on the first CI run instead. Same shape as the bugs this branch exists to catch - something that works only because of local state, breaking where the code actually runs. * fix(eval): apply code review findings Seven local reviewers and an independent cross-model pass. The headline is that a fix I added in this branch was worse than the gap it closed. Reverted the aggregate() quality-median filter. Excluding skill-not-invoked rows from the quality metrics left valid_runs and excluded_runs still counting them, so the promotion gate saw N clean runs while the median came from fewer. The dropped rows are systematically an arm's worst, so it biased toward PROMOTING - reproduced: one real run at 0.9 plus two uninvoked rows at 0.0 gave the gate 3 valid runs, zero exclusions and a 0.9 median, flipping keep_incumbent to promote. Three verdict fields compounded it: they are all() reducers still reading the wider set, so one uninvoked cell flipped a whole arm. Five reviewers found the two halves independently. Closing it honestly needs a scored-run count plus a paired-equality check in the gate, which is promotion semantics rather than an aggregation fix. The gap is now pinned by a test that states why the half-fix was reverted. Stopped forging the absence of CI. The runner refuses --unsafe-no-bwrap when CI is set because that mode runs sessions with bypassPermissions behind a boundary its own docstring calls "not a security boundary"; the sweep test deleted CI to get past it, so eval / locked pytest ran an uncontained agent sweep on the runner holding the checkout and credentials. It skips under CI instead - the containment job still runs it for real with GITNEXUS_REQUIRE_FULL_SWEEP=1. The stand-in CLI was lying in three ways. It never set is_error, so a refused write read as a completed one. It had no Skill branch at all, so honoring is_error revealed the evidence gate had been satisfied by a tool the fixture never ran - the gate was measuring the fixture, not a skill. And a reply with no usage became four zero-valued fields plus a fabricated cost, which is exactly the unknown-is-not-zero confusion the accounting it feeds exists to prevent. A provider failure also crashed the subprocess with no terminal result event. The identity probe never ran anywhere. test_mock_provider.py was in no job's file list, and the only job setting CLAUDE_CANARY_BIN runs a fixed list. My commit message claimed the next containment run would produce the answer; it would not have. Now wired in, with the CI-shape test updated to pin it. Also: the regex-miss fallback wrote a predictable name in shared /tmp through a symlink-following stage, now scoped to the test's own directory; and the canonical_provider docstring plus the callback comment still asserted a call_type branch the code no longer has. Deferred as design decisions rather than review fixes: the containment sweep uses the stand-in CLI rather than the pinned real one, the full-sweep path bypasses the gateway so native usage accounting is unexercised there, _normalize_litellm duplicates the Responses algorithm, and OPENAI_RESPONSES is now unreachable from canonical_provider. 682 eval tests pass, 17 skipped, ruff clean. * fix(eval): carry scripted tools over the Responses protocol Review round on #3235. Three real items; five more were already fixed in |
||
|
mengkaka
|
4154b63131
|
feat(indexing): add Objective-C semantic indexing support (#3179)
* docs: add Objective-C fork provider notes * feat(objective-c): add deterministic provider and grammar * feat(objective-c): finalize provider MVP * fix(objective-c): harden provider integration * fix(objective-c): normalize bare macro markers * docs(objective-c): integrate provider documentation * fix(objective-c): harden resolution and header classification * fix(objective-c): complete provider follow-ups * fix: address Objective-C review follow-ups * chore: format Objective-C grammar sources * fix(objective-c): harden review follow-ups * Address PR review feedback (#3179) Keep Objective-C chunking and macro recovery aligned with the grammar, and stop Community MEMBER_OF edges from leaking into symbol context. Co-authored-by: Cursor <cursoragent@cursor.com> * Address follow-up review on ObjC chunking and language fallback. Keep preprocessor directive text from changing file-scope brace depth, group real ivar nodes, skip header modifiers, and restore Rakefile/Gemfile detection through getLanguageFromFilename. Co-authored-by: Cursor <cursoragent@cursor.com> * Parse Objective-C headers with the objc grammar in embeddings. ensureAndParse and structural extraction now use the same content classifier as ingest, including method snippets from .h files, so Protocol/Category/Class chunks are not re-parsed as C++. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#3179) Keep file-scope macro elision off C line splices and @interface/@protocol/@implementation bodies, and attach ivar attributes to the following instance variable when chunking. Co-authored-by: Cursor <cursoragent@cursor.com> * chore(bench): rebaseline Objective-C CSV emit * feat(objective-c): add workspace resolution and linear emit benches Plain .h files are classified as C++, so the ObjC pass could not resolve #import of those headers. Load a C/C#-style workspace once per pass, and keep protocol-candidate USES linear. Refs #3179 Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#3179) - Compare LadybugDB labels() as a scalar when excluding Community MEMBER_OF edges. - Walk superclass members, skip file-static C sibling defs, and ignore comments in ObjC header/macro scans. Note: pre-existing failure in objective-c-provider integration (worker-pool ready timeout) not addressed by this PR. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#3179) Emit Objective-C declaration captures so compilation-unit siblings can share header/implementation bindings, and keep class vs protocol visibility groups distinct. Note: pre-existing failure in worker-pool startup (GITNEXUS_WORKER_READY_TIMEOUT_MS) not addressed by this PR. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#3179) Emit every comma-separated property/ivar declarator, and count @interface after a multiline block comment closes so in-declaration macros stay intact. Note: pre-existing failure in worker-pool startup (GITNEXUS_WORKER_READY_TIMEOUT_MS) not addressed by this PR. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: ximengkai <ximengkai@soyoung.com> Co-authored-by: Gergő Magyar <gergomagyar@icloud.com> Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergő Magyar
|
1e8bdd890a
|
fix(ci): look up fork prebuild PRs by head owner and branch (#3236)
* fix(ci): look up fork prebuild PRs by head owner and branch
commits/{sha}/pulls is empty for fork SHAs, so deliver-fork-prebuilds failed closed on every real fork PR. Resolve the open PR from workflow_run head owner+branch instead.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ci): use the same fork-PR lookup in autofix publish
pr-autofix-publish had the same commits/{sha}/pulls fallback, which is empty for fork SHAs. Share the pulls?head=owner:branch verifier and always run it before sticky comments or check runs.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ci): fail closed on ambiguous fork-head PRs
Untrusted artifact pr_number must not pick among sibling open PRs from the same fork branch. Parse paginated gh --slurp pages and require verify success before prebuild checkout/push.
Co-authored-by: Cursor <cursoragent@cursor.com>
* style(ci): prettier the fork-PR identity verifier
Root prettier --check fails on .cjs; lint-staged only formats .js.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
|
||
|
Gergő Magyar
|
b1d87c1f33
|
fix(eval): sweep evidence handling and measurement health, with guarded comparator reuse (#3207)
* fix(eval): cut skill-evolution wall clock without shrinking the gate
Reuse matching incumbent/CE cells, sanitize each SHA once, and default
dispatch workers to 3 so weekly review generations finish inside the
EventBridge window. Cap the sweep from leftover instance uptime so a
Friday dispatch still uploads evidence.
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* perf(eval): pipeline graph setup and correct the wall-clock cost model
The evolution sweep paid `sanitize` + `analyze --pdg --index-only` for every
unique task SHA on the critical path, one at a time, with nothing overlapping.
`_run_sweep` now starts the next unpaid SHA's clone template and graph snapshot
on a prefetch thread as soon as the current task's cells are dispatched, so
every SHA but the first hides behind a paid session wave. The thread is joined
before that SHA is used and before the trees tempdir is torn down, and a
prefetch failure is recorded against the SHA exactly as an inline failure is.
Tasks whose cells are all reusable comparator rows are not prefetched: they
never build a graph, so priming one would be pure cost.
Adds `measure_evolution_cost.py`, the cost model behind these numbers. It reads
the review corpus, the evolve defaults, and the workflow's workers default —
it does not start a session. Its first version charged `copy_isolated_tree`
once per paid cell, serially. `run_cell` clones inside its own pool worker, so
the clones in a wave overlap and only one is on the critical path per wave;
the model now charges `ceil(cells / workers)` waves.
Estimated review generation at workers=3: cold 21570s, weekly 7710s.
Wall clock is quantised by `ceil(cells_per_task / workers)`. A cold review task
is 9 cells, so workers=4 buys the wall clock of workers=3 and pays host
contention for it. Documented in the workflow's rollout checklist.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* perf(eval): price the benchmark against measured cell durations
The cost model assumed every cell runs the 1140s mean. Cells are not uniform:
the 41 rows in Actions run 33912693948's artifact are 826s at the median,
1262s at the mean, 2976s at p90, with two pinned at the 5400s session ceiling.
A wave waits for its slowest cell, so a mean understates every concurrent
schedule — the previous model called workers=3 cold 5.99h when the same
schedule against real durations is 10.33h.
session_durations.json carries the sample in submission order with its
provenance and its caveat: every cell in that run returned unusable evidence,
so the durations are real but a clean run may sit lower. It is the only live
artifact; the 2026-07-22 green run's has expired.
The model now simulates the schedule cell by cell rather than multiplying a
mean by a wave count, averaged over all 41 rotations of the sample so no
single alignment between sample order and cell index decides the answer. It
prices today's barrier (wave_makespan) against a continuously fed pool
(fed_makespan) and reports both, and it charges the proposer session — one
per generation, measured at 344.7s — which it had been omitting entirely.
Measurement only; no runtime behaviour changes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* perf(eval): price arms separately and stop inventing setup constants
Two errors in the model, both found by auditing it against the artifact it
claims to describe.
The arms are not interchangeable. `candidate_review` runs 1416s at the mean
against `review`'s 1204s and `ce_review`'s 1176s, and the weekly lane pays the
candidate arm and nothing else — reuse skips both incumbents. Pricing weekly
from a pooled sample charged it for arms it never runs: weekly is 4.59h, not
the 3.65h a pooled sample reported. Cells are also submitted run-major and
arm-minor, so at workers=3 every wave holds one cell of each arm and the
slowest arm sets the wave; the model now builds cells in that order.
The setup constants were invented. GRAPH_ANALYZE_SECONDS=600 and
TEMPLATE_SANITIZE_SECONDS=180 charged 3900s of per-SHA setup for a cold run —
more than the entire non-session time of the source run, which was 2541s for
41 cells and 5 SHAs. `duration_s` is the sum of a cell's Claude sessions
(runner_sessions.py), so that 2541s residual is every clone, graph build,
sandbox and teardown the sweep paid. The model now charges the measured
residual per cell, 62.0s, and no longer credits clone templates or graph
prefetch: both landed after that run and there is no measurement of them yet.
The residual bounds what they can be worth.
Cold 37452s (10.40h), weekly 16541s (4.59h), against a fed pool at 31683s and
16541s. Measurement only; no runtime behaviour changes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* perf(eval): charge sweep overhead where more workers cannot dissolve it
Three defects, found by auditing the model against the artifact again.
The overhead was charged inside the schedule. session_durations.json claimed
the residual was charged "per cell and serially - the pessimistic reading",
but task_cells folded it into each cell's duration, where the pool then
divided it by the worker count. The residual mixes per-cell work the pool
really does divide with per-SHA graph setup it cannot, and the artifact cannot
separate them, so it now sits outside the schedule: cold 11.09h, not 10.40h.
Alignment averaging weighted the shortest sample twice. The arm samples are 13,
14 and 14 long and the average ran over max()=14 offsets, so candidate_review's
first cell was counted twice and its last never. Averaging over lcm()=182
offsets weights every arm's sample evenly.
The wall assumed all 54 cells run. Replaying the sample's own error_kind
sequence through today's systemic_outage_streak trips the outage breaker at
cell 5 of 41. The source run executed all 41, so its runner did not break on
that sequence, but the current one would: these numbers price a HEALTHY sweep,
and a sweep with the sample's failure profile never reaches them. Stated on
generation_seconds and recorded next to the sample it qualifies.
Measurement only; no runtime behaviour changes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(eval): give the review agent somewhere it can actually write
Every review cell in the last recorded generation returned unusable evidence.
Not some — all 41, across all three arms and all six tasks, at $3653 for the
run. The transcripts say why, 127 times across 35 of 35 sessions:
EROFS: read-only file system,
open '/workspace/review-output.json.tmp.2.90a76e583b0c'
The review arm mounted the artifact as a writable FILE at
/workspace/review-output.json while binding /workspace read-only. The Write
tool writes atomically: it creates `<target>.tmp.<n>.<hex>` beside the target
and renames it. The parent was read-only, so the temp create failed and the
artifact was never written. A writable file inside a read-only directory is
not writable to anything that writes atomically. Agents tried
/proc/self/root/workspace/... and /proc/1/root/workspace/... to get around it;
all 41 artifacts came back 0 bytes.
The artifact now lives in its own writable directory bound at /review-output,
outside the workspace. That is what a rename needs, and it lets the workspace
get stricter rather than looser: the review phase may now change nothing there
at all (enforce_phase_workspace gained allowed_artifact=None), where before it
was entitled to one path inside it. The file is no longer pre-created — the
agent writes it, and absence is now meaningful evidence.
parse_review_output reported every one of these as "review output is not valid
UTF-8 JSON". The file was empty, and its except folded OSError, UnicodeError
and JSONDecodeError into that one string, so a sandbox that made writing
impossible was indistinguishable from an encoding fault. That is why this read
as an agent-quality problem for fifteen consecutive non-green runs. Each cause
now names itself: never written, empty, not valid UTF-8, not valid JSON with
the decoder's position. run_arm also keeps the FIRST error_detail, as it
already did for error_kind, so a phase-boundary violation is no longer buried
under the parse failure it causes.
The test double conflated sandbox.private_root with the clone, which put the
artifact directory inside the workspace and would have hidden the stricter
check. Regression tests pin the mount shape in the generated bwrap argv, the
contract path in the prompt, the four parse diagnostics, and the
untouched-workspace contract.
Verified by unit tests only: this container has unprivileged user namespaces
disabled, so bwrap cannot run here and the mount was not exercised end to end.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(review): close the artifact path in every layer that gates it
Code review of this branch found the relocated review artifact was fixed in the
bwrap mount and nowhere else. Four independent layers decide whether the agent
can write it, and three still named the old location.
Claude Code applies its own filesystem policy to its own tools, and
build_claude_settings listed only /workspace, /tmp and /home/agent under
allowWrite with denyRead ["/"]. The artifact used to live under /workspace, so
this list was correct until it moved. SANDBOX_REVIEW_OUTPUT is now in allowWrite
and allowRead; without it the bwrap bind grants a write the CLI then refuses.
The task corpus still ran `test -s review-output.json` from the workspace, in a
separate sandbox invocation that never sees the artifact mount. Every review
cell would have been stamped verify-failed with resolved=False no matter how
good the review was, which also made those rows permanently unreusable and so
silently disabled this branch's own comparator reuse for review arms. The verify
and hidden-oracle commands now read the location from
GITNEXUS_BENCH_REVIEW_OUTPUT and get the directory bound read-only, mirroring
the mount-plus-env-var shape _run_hidden_oracle already used.
host_text and host_path did not translate the new path, so the host-unsafe
backend told the agent to write somewhere that exists on neither backend.
Adding the mapping exposed a second defect: host_text substituted every
occurrence of a target, and "/review-output" appears twice in
"/review-output/review-output.json" - once as the directory and once inside the
filename. Matching is now anchored to a path boundary.
Comparator reuse had three ways to accept evidence it should have rejected. A
row with no runtime_digest passed the drift lock because the guard only compared
when both sides were bound, and the branch's own test asserted that as correct;
absence is now a mismatch and the test states the rule. materialize_reused_row
overwrote recorded_at with the copy time while the age check read that field, so
a row copied forward each generation refreshed its own clock and never aged out;
the first measurement time is now preserved and aged against. A future-dated
stamp passed a one-sided bound and is now rejected as corrupt.
RUNTIME_DIGEST never reached the runner at all: runner_environment builds a
fixed dict and process_control replaces the child environment wholesale, so the
digest the workflow exports was dropped and the lock it feeds was inert. The
instance-window deadline was also checked only after run_proposer returned,
buying a proposal the generation had no room to benchmark.
The graph prefetch thread was started without copy_context, so it never saw the
cancellation ContextVar the rest of the sweep shares, and the outage breaker
returned without setting cancel_event - together, a tripped breaker would block
on joining a prefetch that was never told to stop. Both fixed, with outage
checked before cancellation at the two exits so an outage keeps exit 1 instead
of becoming a Ctrl-C's 130.
Both bwrap canaries that actually execute a write still bound the pre-fix shape
against a file this branch no longer creates, so they would have errored rather
than caught anything. They now bind the directory and write atomically - temp
file beside the target, then rename - which is the exact operation that failed
with EROFS. A source-text assertion over inspect.getsource(run_arm) was replaced
with one that inspects the real mount, and a wall-clock assertion was pinned to
a fixed monotonic clock.
Not applied, and why: binding task-asset and dependency digests into comparator
reuse needs asset snapshots prepared before the reuse decision rather than
inside the per-task loop, and shipping the comparison without that would add a
guard that silently never fires. Forcing a paid canary cell per incumbent arm
and folding reused rows into the outage streak are behaviour decisions, not
fixes. Clone-template reuse still has no test. The cost model's per-cell
residual still shrinks with arm count, overstating weekly savings by at most the
2541s residual; the docstring now says so rather than inventing a split.
585 eval tests pass, ruff clean, 29 workflow contract tests pass. The two
test_model_gateway.py failures are pre-existing and fail on main.
* fix(review): bind reuse to its environment and keep the health canary real
Applies the five findings the previous review round left open.
Comparator reuse ignored the environment a row was measured in. TaskReuseBinding
carried the task and oracle identity but not the task-asset or sandbox-dependency
digests, and this branch itself changes sandbox_dependencies in the review
corpus - so a reused comparator could be measured against one dependency set and
compared against a candidate built on another, handing the gate a false
comparison. Closing it needed the digests to exist before the reuse decision, so
asset snapshots are now prepared for every task up front instead of lazily
inside the per-task loop. That also removes the concurrent TaskAssetCache.prepare
the prefetch thread could otherwise race, which the file's own "plain dict,
read-then-write race" comment warned about. Both digests fail closed on either
side, matching the runtime digest.
The broken-incumbent canary could not fire when reuse was working. It read
`resolved`, which counts reused rows, so an arm whose cells were all reused
always looked healthy - in precisely the run where a broken environment would go
unnoticed. aggregate now also reports `resolved_fresh` and the canary reads it.
That count would be vacuous if an arm were reused end to end, so the sweep keeps
one paid cell per incumbent arm and says which one it kept.
Reused rows did not participate in the outage streak, so a run of failures could
carry across them and trip on stale history. A reused success now resets the
streak the way a paid success does.
The cost model charged sweep overhead per cell, which credited a weekly
generation for shrinking work it still performs: it pays one arm instead of
three but builds exactly the same graphs. Overhead is charged per SHA now.
Weekly is 5.20h rather than the 4.80h the per-cell rate reported; cold is
10.86h. The residual still cannot be split between per-SHA and per-cell work
from one artifact, so session_durations.json records that assumption and the
direction it errs in, rather than leaving a number nobody can trace.
Clone-template reuse - the branch's core speedup, taken on essentially every
multi-cell sweep - now has a test that builds a real sanitized template, asserts
the cell runs against the copy with the template's HEAD, and fails if run_cell
re-clones. A second test asserting only on a namespace built inside the test was
written and deleted: it exercised nothing, which is the failure this review
round penalised elsewhere.
589 eval tests pass, ruff clean, 29 workflow contract tests pass. The two
test_model_gateway.py failures are pre-existing and fail on main.
* refactor(eval): consolidate duplicated harness logic after the review round
Simplification pass over the branch. Behavior-preserving throughout; three
reviewers, nine findings applied, two skipped.
The review-artifact block in _run_hidden_oracle was unreachable. That function
runs only in run_arm's non-review branch, while the directory it probes for is
created only in the review branch, and each sandbox serves exactly one arm - so
`review_artifact.parent.is_dir()` could never be true. It was added an hour
earlier to make the hidden oracle resolve the moved artifact; the oracle never
runs for review tasks, so the guard was dead on arrival. Deleting it also
removes the duplication it had with the verify-command wiring.
EXCLUDED_ERROR_KINDS is now one definition. runner.py and comparator_reuse.py
each carried the same six-member frozenset, kept in sync by a comment. Only one
direction is possible: runner already imports from comparator_reuse, so the
reverse import fails at module-init with a circular-import error. That is now
stated where the alias lives, so nobody tries it the other way.
ensure_task_graph and prefetch_next_graph shared ten keyword parameters, passed
through two call sites and forwarded whole between them. They now take a
GraphBuildEnv, mirroring TaskCellContext, which already bundles per-cell state
in this file. Its ready_keys() replaces an inline four-set union at the call
site.
Smaller consolidations: _sha256_file's hand-rolled chunk loop becomes
hashlib.file_digest (3.11+, already used in runner_artifacts); _copy_owner_only
reuses task_assets._write_all and COPY_CHUNK_BYTES instead of repeating the
short-write retry; its stat-then-open existence check becomes the O_EXCL failure
it was already relying on, which is atomic rather than merely narrow; and
runner_environment reads the digest through comparator_reuse.current_runtime_digest
instead of re-parsing the environment variable.
Three test docstrings summarised the branch's own history ("the branch's core
speedup", "the regression that produced fifteen runs") rather than the invariant
under test. Rewritten to state the constraint, which is what survives the merge.
Repaired the indentation left behind by the outage-streak edit and flattened the
prefetch dispatch from three nested conditionals to one.
Skipped: consolidating comparator_reuse._real_directory onto proposer_sandbox's
same-named helper - they differ, the sandbox one rejects any symlink in the
resolved path while this one checks only the leaf, so sharing it would tighten
behavior rather than preserve it. That needs a decision about which policy the
reuse path wants, not a simplification.
589 eval tests pass, ruff clean, 29 workflow contract tests pass. Unrelated and
pre-existing: two test_model_gateway.py failures, and
test_process_control.py::test_timeout_kills_term_ignoring_descendants_before_they_write,
which is a TERM-to-KILL timing flake (passes 2 of 3 in isolation) in a file this
branch does not touch.
* refactor(eval): name the reuse directory check for the promise it makes
The simplification pass left one finding open: comparator_reuse and
proposer_sandbox both defined `_real_directory`, same name and same shape, with
different guarantees. The sandbox one rejects every symlink hop in the path; the
reuse one checks only the leaf and resolves through parents. Sharing the name
invites a consolidation that would silently tighten one of them.
They should not be merged, so the name stops claiming they could be.
proposer_sandbox guards a mount root, where a symlink hop changes what an
untrusted session is handed. comparator_reuse guards a data directory whose
contents are already validated one file at a time - reads go through
_regular_file, which lstats and rejects symlinks, and writes through O_NOFOLLOW.
A symlinked parent therefore grants nothing those guards do not already cover,
while refusing one would reject a symlinked artifacts directory or macOS's /var
for no gain.
Renamed to _resolved_directory, with the reasoning recorded at the definition,
and a test that pins both halves: a symlinked parent is accepted and resolved, a
symlinked leaf is still refused. Behavior is unchanged.
591 eval tests pass, ruff clean. The two test_model_gateway.py failures are
pre-existing and fail on main.
* test(eval): measure the sweep scheduler instead of modelling it
measure_evolution_cost predicts wall clock from a model of what
sweep_task_cells does. This runs the real thing - real threads, the real wave
barrier, the real outage breaker - with only the paid agent session replaced by
a sleep, and times it.
Durations are the measured per-arm samples divided by 5000, so a 1416s cell
takes ~0.28s. The shape is kept on purpose: the median cell is 826s against a
5400s ceiling, and that spread is the entire reason a barrier costs anything.
Uniform random sleeps would erase the effect under test. All schedulers consume
one identical seeded plan, so a comparison cannot be an artifact of one of them
drawing luckier cells.
The model survives contact: it tracks real execution within about 10%, and
workers=1 - which runs without a pool at all - sits at 0.95, so the residual
above 1.0 at higher worker counts is per-wave thread overhead rather than a
modelling error. Two structural claims that were arithmetic are now observed.
Weekly is flat from workers=3: 3.59, 3.59, 3.59, 3.60, 3.59, 3.59 across w=3..8.
workers=4 buys nothing over workers=3 on cold, 7.68 against 7.78.
Two prototype schedulers are measured beside it, deliberately before any
production code exists. A continuously fed pool per task is worth more than the
model claimed on cold, -27.3% against a predicted -17.9%, and exactly nothing on
weekly, +0.0%, because a weekly task is one wave with nothing to feed. One pool
across all tasks beats both: -40.7% weekly and -42.9% cold at workers=3, rising
to -65.7% and -63.9% at workers=8. It also subsumes the fed pool, since packing
across tasks is a fed pool.
That reorders the backlog. Cross-task packing moves from second to first: it
dominates on both profiles, and it is the only thing that moves weekly at all.
Raising the worker count is worth nothing until it lands - under the barrier
weekly does not improve from w=3 to w=8, and speedup against serial is 1.58x for
three workers and only 2.40x for eight.
The bound on all of it: sleeping threads do not contend. Real sandboxed sessions
compete for CPU, page cache and disk, and the duration sample was itself
measured at workers=1, so it carries no contention either. These speedups are
upper bounds. The ordering is trustworthy because the schedulers were compared
under identical conditions; the magnitudes are not. The packed prototype is also
a bare ThreadPoolExecutor with no breaker folding, no per-task graph lifecycle
and no reuse binding - which is the actual cost of building it, and is not
measured here.
* test(eval): carry the sweep invariants into the packed prototype
The first packed prototype was a bare ThreadPoolExecutor. It reported -43% and
none of the invariants the shipped scheduler holds, so it priced an idea nobody
could ship. This one carries them: a global submission order continued across
task boundaries, in-order folding, the real outage breaker, and per-task graph
readiness gating behind a serial builder.
The fidelity check first reported the two schedulers tripping on different
cells, 17 against 16. That was my instrumentation, not a divergence -
sweep_task_cells folds an entire wave before it evaluates the breaker, so the
last cell folded is not the cell that tripped. With the harness mirroring the
breaker's own evaluation the two agree exactly, across failures starting at
cell 0, 4 and 12, with overrun inside the workers-1 bound the wave docstring
promises.
Two results worth the exercise.
Head-of-line blocking, not the barrier, is what a naive in-order design pays.
Holding submission to `workers` cells beyond the fold pointer leaves the
faithful scheduler at -8.1% cold and -2.7% weekly: one slow cell stalls the
pointer, the window cannot slide, and it reproduces the wave almost exactly.
That is the number to quote if anyone proposes the obvious implementation.
But the overrun bound turns out to be set by the worker count, not the window.
Only `workers` cells can be running when the breaker trips; everything queued
behind them short-circuits on the halt flag. Overrun is 3 at an unbounded
window exactly as at 6, and the trip cell never moves off 16. So H2 does not
have to trade breaker fidelity for speed - a wide window takes -42% with the
semantics intact. The tension I assumed was there is not, and window=12 already
captures 97% of it.
Still an upper bound: sleeping threads do not contend, and the sample was
measured at workers=1. What this establishes is that the invariants are
affordable, which was the thing blocking H2. Not built here: the trees tempdir
lifecycle, reuse-row binding, and the cancel_event path.
591 eval tests pass, ruff clean.
* test(eval): put the scheduler comparison under real CPU contention
Every Phase 2 number so far came from sleeping threads, which contend for
nothing, against a duration sample measured at workers=1, which contains no
contention either. That was the standing caveat on the whole result, so this
measures it.
A cell now waits for its API share and then burns a fixed number of sha256
rounds in a subprocess. Work-bounded rather than wall-clock bounded, so it takes
longer when cores are busy - that is the effect under test. A subprocess because
Python threads burning Python would measure the GIL rather than the machine.
Calibrated at 519k rounds/s, stable within 2% across three probes.
The first run of this was worthless and is recorded as such: on a 24-core host
with 3 to 6 workers nothing ever contends, since cpu_fraction 0.5 at 6 workers
is about 3 cores of demand out of 24. It measured an absence. Re-run pinned with
taskset to 4 and 2 cores.
The packing advantage survives. It holds between -40% and -47% across every host
size and CPU fraction tested, including a genuinely oversubscribed 2-core box at
cpu_fraction 0.5 with 6 workers.
But contention erodes packing more than it erodes waves, for a structural
reason: packing is what creates the concurrency. Moving from 24 cores to 2 at
cpu 0.5 and 6 workers, the faithful scheduler slows 13% while the wave slows
3.7%, and the gain narrows from 45.0% to 39.8%. Packing and a higher worker
count are therefore not independent wins - packing spends the contention
headroom first, so raising workers has to be re-argued after it lands rather
than added to it.
Three things this still does not measure, and they bound the result. The real
CPU fraction of a benchmark cell is a guess informed by roughly 180 tool calls
per session; nobody has profiled one. The evolution runner's core count decides
which column applies and is unknown here. And the burn is sha256, pure CPU,
while real cells run vitest and analyze, which are memory and IO heavy - so this
is a floor on contention, not a ceiling.
591 eval tests pass, ruff clean.
* perf(eval): add a packed sweep scheduler, and correct the bound I claimed for it
sweep_task_cells finishes one task before starting the next and drains a wave
before refilling it, so a task with fewer cells than workers leaves workers
idle and one slow cell stalls its whole wave. sweep_packed_cells feeds every
task's cells through a single pool instead. Measured against the review corpus
it is worth about 40% of a cold sweep, and it is the only change that moves a
seeded weekly run at all - there a task is three cells and a wave is never full.
The breaker keeps its exact meaning. Cells carry a total submission order
continued across task boundaries, a folder walks results in that order, and
consecutive systemic failures are counted there, so a doomed run aborts on the
same cell it would have under waves. Verified at three failure positions.
This commit also corrects a finding from the Phase 2 prototype. I claimed the
overrun bound was set by the worker count rather than the submission window,
and that packing therefore cost nothing in breaker fidelity. That was derived
from a window sweep that only ever injected failures at one position. Driving
the real function at other positions shows the halt flag does not bound overrun
at all: the folder walks in order, so a slow early cell lets workers race ahead
and the trip is detected after those cells have already paid. An unbounded
queue overran by 11 cells where waves overrun by 2.
So the window is load-bearing and the trade is real, measured at workers=3 with
failures injected at four positions:
window 3 -> -8% wall, overrun 2 (the wave scheduler's own bound)
window 6 -> -27% wall, overrun 4
window 12 -> -42% wall, overrun 9
window 54 -> -44% wall, overrun 11
Overrun is wasted paid sessions at roughly $70 each. The default multiplier is
2, keeping the worst case within twice the wave bound while taking most of the
gain; the curve is in the constant's comment so raising it is an informed
decision rather than a guess.
Not wired in yet: _run_sweep still calls sweep_task_cells per task. Moving the
per-task graph, trees tempdir and reuse binding out of that loop behind
await_ready is the larger and riskier half, and it belongs in its own change.
595 eval tests pass, ruff clean.
* fix(eval): judge harness health on execution, not on how many tasks resolved
broken_incumbent_arms infers "the environment is broken" from an arm resolving
zero tasks. That inference does not hold: a reviewer can be wrong about every
task in a hard corpus while every process, mount and capture worked perfectly.
Actions run 33962002890 is exactly that shape - 51 cells, all resolved=False
with error_kind=oracle-failed, median score 0.212, and a healthy harness.
Someone already knew this, and patched it by excluding review arms at the call
site. That leaves the unsound inference in place for workflow and
workflow_direct, and leaves review arms with no health check at all - so the
run that genuinely was broken, 33912693948, where the mount made an atomic
write impossible and all 41 artifacts came back empty, could not have been
caught here either.
So this replaces the inference rather than adding another exemption. aggregate
now classifies fresh rows into execution failures (the process or its tooling
did not complete), evidence failures (it completed but produced nothing
trustworthy or scoreable), and admissible measurements. An arm is unhealthy
only when it has fresh attempts, zero admissible measurements, and at least one
execution or evidence failure. Resolution count is no longer consulted. Arms
with only reused rows report current health as UNKNOWN rather than good.
With the inference corrected, review arms are checked again, which is what lets
the empty-artifact case be caught at all.
Deliberately unchanged: comparator reuse eligibility, quality denominators,
promotion thresholds, model settings, skill prompts and scheduler behaviour.
Failures that stop being called infrastructure failures still surface in the
counts and reasons - an agent-originated failure must not vanish from reporting
because it was reclassified. broken_incumbent_arms and its tests are left in
place; deleting behaviour belongs in its own change.
Seven regression tests, built from both runs' shapes and labelled as
reconstructed from logged observations, since 33962002890's results.jsonl did
not survive the instance shutdown. They pin: a badly-scoring reviewer is
healthy; an all-zero score is still a valid negative; empty artifacts are
unhealthy; one admissible cell keeps an arm healthy while its failures stay
visible; reused rows alone leave health unknown; reused successes do not mask
fresh failures; and a parseable artifact does not excuse a failed session.
602 eval tests pass, ruff clean.
* fix(eval): pin the health guard below the breaker, and stop calling mixed runs healthy
Two corrections to the health-classification patch.
The regression I wrote could not have proved what it claimed. A fixture of 41
empty artifacts aborts through the outage breaker long before finalization:
review-evidence-invalid is in SYSTEMIC_ERROR_KINDS and the limit is 5, so it
trips at cell 5 through the pre-existing path. It demonstrated failure
detection, not the new guard. The decisive test now uses ONE fresh unusable
cell, asserts the streak stays under the breaker threshold, and only then
requires finalization to abort - leaving the new check as the only thing that
can catch it. Removing the call makes that test fail; restoring it passes.
The accurate defect statement is narrower than the last message claimed. Review
arms were excluded from the final incumbent-health check while the consecutive-
failure breaker gave them separate, partial coverage. They were not unguarded.
Second: "one admissible cell plus two execution failures" was asserted as
healthy. That converts "not wholly unusable" into "ran reliably", which is how
a partly-broken sweep passes review. Arms now report UNKNOWN, OBSERVED_OK,
DEGRADED or UNUSABLE. Only UNUSABLE is fatal, so eligibility and promotion are
untouched - this changes what is reported, not what is allowed.
The guard is extracted as enforce_measurement_health so it can be driven
directly, and it now reports a status line per arm. It names no cause: an empty
artifact establishes that evidence is unusable, not that a mount rejected the
write, so it prints cause=undetermined rather than guessing EROFS. It still
runs after report.md and promotion.json are written, so a failing sweep leaves
its evidence behind.
ce_review is named explicitly at the call site. It is a comparator rather than
a candidate, so it is absent from CANDIDATE_ARMS.values(), and dropping the
review exclusion alone would have left it unclassified.
The wiring test reads _run_sweep's compiled code object for the referenced
global rather than matching source text. It is honest about its limit: it
proves the call exists and would catch its removal, but no test here drives
_run_sweep end to end, which needs bwrap and a sandbox.
broken_incumbent_arms is marked LEGACY and NON-AUTHORITATIVE with removal
tracked. It has no production caller.
608 eval tests pass, ruff clean. The two test_model_gateway.py failures are
test_locked_litellm_translates_messages_to_offline_responses and
test_openai_gateway_never_leaves_proxy_output_on_an_undrained_pipe; both fail
identically on origin/main in this environment, checked directly rather than
carried forward as an inherited label.
* fix(eval): review artifact path, evidence classification, comparator reuse
Extracted from the combined skill-evolution branch. This is the runtime change
set: everything that alters how a sweep executes and what it records. The
packed scheduler and its measurement harness were separated onto
perf/skill-evolution-packed-scheduler, which is purely additive.
Correctness. The review artifact was mounted as a writable FILE inside a
read-only workspace while the agent's Write tool writes atomically - temp file
beside the target, then rename - so the temp create failed EROFS and the
artifact was never written. Four layers gate that path and three named the old
location: the CLI's own allowWrite/allowRead policy, the task corpus verify
command run in its own sandbox invocation, and host_text/host_path for the
host-unsafe backend. Fixing the translator exposed a second defect, since
"/review-output" appears twice in "/review-output/review-output.json"; matching
is now anchored to a path boundary. parse_review_output folded OSError,
UnicodeError and JSONDecodeError into one message, so an artifact that was
never written looked like an encoding fault; each cause now names itself.
Health classification. broken_incumbent_arms inferred a broken environment from
an arm resolving zero tasks, which a reviewer facing a hard corpus falsifies -
Actions run 33962002890 is exactly that shape. Arms are now classified from
fresh execution and evidence outcomes as UNKNOWN, OBSERVED_OK, DEGRADED or
UNUSABLE, and only UNUSABLE aborts. Resolution count is not consulted. The
guard names no cause: an empty artifact establishes unusable evidence, not that
a mount rejected the write.
Comparator reuse. Reuse accepted evidence it should have rejected: a row
without a runtime_digest passed the drift lock, recorded_at was overwritten with
the copy time so a row could outlive its own max_age, and the binding ignored
task-asset and dependency digests although this change alters
sandbox_dependencies in the review corpus. Closing the last one required
preparing asset snapshots before the reuse decision, which also removes the
concurrent TaskAssetCache.prepare the prefetch thread could race.
These three concerns share aggregate() and _run_sweep, which is why they ship
together: separating them further would mean hunk-level surgery on a function
all three modify, and the risk of a silent omission outweighs the reviewability
gain.
592 eval tests pass at this base. The two test_model_gateway.py failures,
test_locked_litellm_translates_messages_to_offline_responses and
test_openai_gateway_never_leaves_proxy_output_on_an_undrained_pipe, fail
identically on origin/main in this environment.
Known gap, and the reason this is not ready to merge: no test drives _run_sweep
end to end. enforce_measurement_health is unit-tested including the
below-breaker unusable case, and the caller wiring is pinned structurally by
reading _run_sweep's compiled code object, but interruption semantics, exit
precedence and persisted artifacts are not exercised through the real path.
* fix(eval): address PR review feedback (#3207)
- aggregate: count admissible rows directly instead of subtracting the
execution and evidence counters, which double-charged a row that is both
a session error and invalid review evidence and could report UNUSABLE for
an arm holding real measurements.
- run_proposer: bound the session timeout by what is left of
--max-runtime-seconds, so clearing the sweep minimum cannot start a
full-length session past the instance window.
- comparator reuse: hold one O_NOFOLLOW descriptor for the size check,
digest and copy, and prove it is the inode that was checked, closing the
swap window a concurrent writer of the reuse directory had.
- Drive the review-artifact mount assertion through run_arm and the
clone-template assertion through run_cell, instead of rebuilding the
expected values in the tests (also removes the CodeQL unnecessary lambda).
- Assert the workflow invokes run-evolution.sh rather than that its YAML
mentions --max-runtime-seconds, which only appears in a comment.
- Correct the parse_review_output failure-mode claim: the fold was empty
artifacts reported as "not valid UTF-8 JSON"; a never-created file raised
FileNotFoundError.
- prettier: wrap the over-long readFileSync call flagged by PR autofix.
Note: pre-existing failure in tests/test_model_gateway.py::test_locked_litellm_translates_messages_to_offline_responses (local LiteLLM proxy never becomes ready in this environment) not addressed by this PR.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(eval): address the second round of PR review feedback (#3207)
- Refuse a symlinked `transcripts` component on both sides of comparator
reuse. O_NOFOLLOW guards the leaf only, so a link there redirected the
read or the copy out of the results directory; checked per component as
evolution._require_directory_chain does.
- Base the paid incumbent canary on the cells this sweep PLANS. Reuse
selection accepts any prior run index, so a results directory produced
with more runs left extra keys, the equality never held, and the canary
stopped firing. Extracted as drop_canary_reuse_key and unit-tested.
- Start the runtime clock in main(). --max-runtime-seconds is measured from
/proc/uptime before exec, so parsing, task I/O, preflight and gateway
setup were being handed back to the sweep out of the upload reserve.
- Do not fall back to shutil.copytree when the managed clone copy was
cancelled or timed out; that fallback is for a filesystem that cannot
reflink, and copytree cannot be cancelled.
- Assert the review session's writable mount, not only the verifier's
read-only one: the EROFS bug is about the agent's write.
- Exercise ref isolation in the copy_isolated_tree test rather than
comparing an initial HEAD a shared namespace would also match.
- Point the stale-symlink fixture at the sentinel via os.path.relpath, and
skip the reuse symlink tests where symlink creation needs privilege.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(eval): close the runtime-cap gap and pin the reuse directory
Both were left open on #3207 as approach decisions rather than nits.
Runtime cap: run-evolution.sh computed the budget in its own
`uv run python -c` and passed a number, so the script's remaining
provenance work and the CLI's own startup were spent by nobody and charged
to the sweep — out of the upload reserve the cap exists to protect. The
script now passes --max-runtime-from-instance-window and evolve reads
/proc/uptime itself, on the line after it starts the clock the budget is
measured against, so no interval exists to lose. Also removes an
interpreter start from the script and lets --dry-run print the real argv.
Reuse directory: _real_child_directory lstat-checked `transcripts` and
returned its pathname, so a concurrent writer could rename the directory
and leave a symlink before the name was used again — O_NOFOLLOW guards
only the leaf. Every artifact is now resolved against a held descriptor:
_open_real_directory opens with O_DIRECTORY|O_NOFOLLOW (check and open in
one syscall), and _open_regular / _copy_owner_only take dir_fd. The reuse
path is therefore POSIX-only; _require_openat says so and fails closed,
which the runner already treats as "run a paid cell". _resolved_directory
still tolerates a symlinked reuse root, unchanged and still tested.
evolution._require_directory_chain is still lstat-per-component. It guards
a different surface (candidate overlay reads) that neither review raised,
so it is left alone rather than widened into here.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(eval): sample the proposer budget where it is spent, digest what is copied
Four findings against
|
||
|
Gergő Magyar
|
f48bf81256
|
perf(parse): tighten the dispatch-round memory bound and unclamp the worker-pool override (#3200)
* docs(parse): record why dispatchGroups is a required interface member Review finding #10 argued dispatchGroups should be optional to match `getQuarantinedPaths?` / `getStats?`. Those are compatibility accommodation for WorkerPool shapes that predate them, not a convention for new members; optional here would force a `?.` plus an unreachable fallback at the single production call site. Documenting the decision so the next reader does not re-litigate it from the neighbouring optional markers. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit addaab647377f3c4553f752fa3ca1388bcb9ca81) * refactor(parse): simplify round accounting and dispatch setup Simplification pass over the dispatch-rounds change. Behavior preserved: identical graph on a full analyze (51,286 nodes / 163,092 edges). - Drop `roundMissBytes`. `roundBufferedBytes` counts the same bytes plus the cache hits, so it is always the greater of the two and the first disjunct of the close condition could never fire on its own. One counter, one reset, one check. - Measure round bytes with `Buffer.byteLength(content, 'utf8')` instead of `String.length`. UTF-16 code units undercount non-ASCII source by up to 3x, so the cap meant to bound main-thread retention was letting a CJK-heavy repo hold well past its nominal budget. Matches `estimateItemBytes` in the pool. - Reset the durable ParsedFile directories for a round's chunks concurrently. Each targets its own chunk-hash directory, and running them serially put N round trips of fs work on the critical path the round exists to shorten. The try/catch stays inside the mapped callback, so one failure still degrades that chunk alone. - Skip the quarantine filter entirely when nothing is quarantined, which is every run without a worker death. It was an identity copy of every group. - `dispatchChunkParseRound` takes `DispatchGroup<...>` rather than re-declaring that shape inline; the type was already imported and used in its body. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit 527d5b6e0ca8ae7bbc6a414c5ac7e27fd85e9995) * refactor(parse): count round misses with the same idiom startRound uses `drainRound` hand-rolled a reduce to count 'miss' entries while `startRound`, one function above, filters the same predicate over the same union. Same integer, one idiom. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit 7eaa193b0cb5fa515844f36ae1401d6fb2fed7b8) * fix(parse): honor GITNEXUS_WORKER_POOL_SIZE above the auto sizing cap The auto pool size is bounded by source bytes so a tiny repo does not spawn a full idle pool. That bound was also clamping the operator's env override, because the env value is read inside `resolveAutoPoolSize()` and the result went through `Math.min(..., workProportionalCap)`. `DEFAULT_POOL_SIZE_CAP`'s own comment offers `GITNEXUS_WORKER_POOL_SIZE` and `--workers <N>` as equivalent escape hatches for operators on bigger machines. They were not. Measured on a 30MB corpus, where the byte-derived cap is 16: --workers 24 -> pool: 24/24 active GITNEXUS_WORKER_POOL_SIZE=24 -> pool: 16/16 active (silently ignored) Both are deliberate operator input, so both now bypass the work-proportional cap, which goes back to bounding only the auto default. After the fix, on the same corpus, with identical graph output (51,286 nodes / 163,092 edges): GITNEXUS_WORKER_POOL_SIZE=24 -> pool: 24/24 active GITNEXUS_WORKER_POOL_SIZE=4 -> pool: 4/4 active unset -> pool: 16/16 active Verified by hand against the pool's own throughput log; not covered by an automated regression test, since the pool size is only observable through that log line and not through the progress stream a test can read. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit 17ed08608c878079b2927da25cfd39c1608a02a2) * fix(parse): bound the durable-reset fan-out and pin the pool-size override Review follow-ups on #3200. The round's durable ParsedFile directory resets went out as one unbounded `Promise.all` — one recursive rm + mkdir per miss chunk, all at once. A round can hold hundreds of small packs, and those resets compete for descriptors with the chunk prefetch this loop already has in flight. `readFileContents` degrades a losing read SILENTLY by documented contract, so a dropped file would vanish from the chunk, from the graph, and from the chunk hash — shipping a narrowed index with exit 0. Now routed through `mapConcurrent` at the same width the file reads use, which keeps the pipelining win and caps in-flight descriptors. An operator's pool size is now also bounded by the number of files there are to parse, so `GITNEXUS_WORKER_POOL_SIZE=100000` on a five-file repo cannot become the literal thread count. This applies to `--workers` and the env var alike, so the parity the previous commit established is intact. It does NOT shrink an incremental re-analyze: `totalParseable` counts every parseable file in the scan, not the changed ones. Adds the regression test a reviewer asked for. The existing coverage (`worker-pool-resilience` calling `resolveAutoPoolSize` directly, `analyze-worker-pool-size` mocking `runFullAnalysis`) never reaches `runChunkedParseAndResolve`'s `effectivePoolSize`, so both stayed green through a revert of the fix. The new test drives the real parse phase with a worker double that writes a per-`threadId` marker, and counts them: verified it fails on the reverted line with `expected [ 'worker-1' ] to have a length of 3 but got 1`, and passes on HEAD. Also corrects the `GITNEXUS_PARSE_ROUND_BYTES` docstring, which still described the cache-miss counter deleted two commits ago. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(parse): skip caching a chunk with a stale durable generation; warn on over-subscription Closes the two findings left open by the review of #3200. When `prepareDurableParsedFileChunk` fails, the previous generation's shards are still on disk, so a later warm hit would union them with the new ones. The chunk is now recorded and its parse-cache write skipped -- the same posture `finalizeWorkerChunk` already takes for a quarantined chunk, and for the same reason: do not cache what we cannot vouch for. The next run re-dispatches into a directory it can actually clear. Bounding the reset fan-out removed the correlated trigger; this closes the individual case. Pool size over-subscription now warns rather than caps. Silently capping is precisely what the override exists to prevent, so an operator's number is still honored -- but an exported GITNEXUS_WORKER_POOL_SIZE applies to every analyze in a long-lived caller (watch auto-sync, the MCP server), including small incremental ones, and that is easy to set once and forget. The warning names the host's usable core count, so it is a hardware fact rather than an invented threshold. `resolveHostParallelism` is extracted from `resolveAutoPoolSize` rather than re-deriving the cgroup-aware fallback at the new call site. Tests: the stale-generation skip is pinned by a new case asserting nothing is written under any key; verified it fails without the guard with `expected 1 to be +0`. 60 unit and 49 integration tests pass across the affected suites. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * test(parse): guard dispatch-round cadence with a bench, not a wall-clock budget Round boundaries are deliberately invisible to graph output — batching that changed output would be a bug — so nothing in the repo could see the #3196 win regress. It would have come back as a silent ~1.5x on every cold analyze. Two earlier attempts to pin it as a unit test failed for that exact reason: one scraped a logger line the progress stream does not carry, the other asserted graph content that is identical either way. Extracts the round-close fold into `createRoundBudget`, so the decision is a shared unit the bench measures rather than a copy that drifts. The parse loop is streaming and cannot know chunk sizes up front, so an accumulator is the honest shape — not a planner. Four deterministic arms, one ratio, no millisecond gate: - layout_fingerprint — pack membership. Every cache key derives from it, so drift needs a SCHEMA_BUMP, never a lone re-baseline. - packs / single_file_packs — the FLOOR. `rounds` only asserts something while the corpus over-splits (774 packs where the byte budget needs 5). This is bench/import-target's lesson, where four heap arms read 0 B and passed every ceiling: a ceiling says "not too big", nothing said "still measuring". - rounds — the regression signal, both directions. - cjk_rounds vs ascii_rounds — pins UTF-8 byte accounting. The two corpora share a UTF-16 length and differ only in encoded size, so String.length collapses them to equal. This is the arm no unit test could be. - pack_scaling_ratio — (t_4n/t_n)/4, min-of-15. A ratio because wall-clock is runner-speed-dependent and this repo has the scar: callable-value-flow's ms gate failed twice at 2.07 and 1.975 against 1.9 with correct code, on a sub-11ms measurement. Every arm verified to fail before being recorded: close-every-chunk reads 774 rounds, disabling the close reads 1, reverting roundFileBytes to String.length takes cjk_rounds 8 -> 3, and shrinking the corpus trips the shape floor. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs(bench): record the analyze phase breakdown and the rejected optimizations Where analyze time actually goes, measured while landing #3194/#3196/#3200, plus the two optimizations that looked compelling and were measured away. The headline is that the parse work is done: a one-file-edit re-analyze is 36.5s, of which parse is 2.8s (8%). scopeResolution is 40% and the unlogged graph emit + FTS rebuild is 49% — neither is incremental, and the ~18s sits outside the phase runner so every phase log is blind to it. Also records the trap that invalidated an earlier measurement: a non-git corpus never records a schema fingerprint, so every run is a forced rebuild and any "warm" number taken that way is fiction. Rejected, with numbers: more workers (16/20/24 land inside run-to-run spread) and bundling the worker entry (~250ms on a normal filesystem; the 8.6s that motivated it was a 9p-mount artifact). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Parafee41
|
1c1cbf111e
|
fix(zig): resolve cross-file static gates (#3185)
Some checks failed
CodeQL / Analyze (javascript-typescript) (push) Has been cancelled
CodeQL / Analyze (python) (push) Has been cancelled
Gitleaks / gitleaks (push) Has been cancelled
Publish / Classify release event (push) Has been cancelled
Scorecard / Scorecard analysis (push) Has been cancelled
Trivy Image Scan / Trivy (gitnexus-cli) (push) Has been cancelled
Trivy Image Scan / Trivy (gitnexus-web) (push) Has been cancelled
Publish / Build & Push RC Docker images (push) Has been cancelled
Publish / RC guard (marker + release-PR skip) (push) Has been cancelled
Publish / ci (push) Has been cancelled
Publish / Publish to npm (push) Has been cancelled
* fix(zig): resolve cross-file static gates * Fix Zig workspace import alias enrichment * Handle extensionless Zig workspace imports * Reuse Zig import resolution for static gates * Document Zig workspace static gating * Harden Zig workspace reference enrichment * Benchmark Zig cross-file static gating * Enforce linear Zig benchmark scaling --------- Co-authored-by: Gergő Magyar <gergomagyar@icloud.com> |
||
|
Gergő Magyar
|
a957c5d757
|
Merge branch 'main' into fix/skill-evolution-gate | ||
|
Gergo Magyar
|
1054e3e038 | fix(eval): make evolution evidence valid and bounded | ||
|
Gergő Magyar
|
c5c4fbe43c
|
fix(zig): vendor tree-sitter-zig so npm i -g no longer warns on peers (#3180)
* fix(zig): vendor tree-sitter-zig so npm i -g no longer warns on peers Published overrides do not apply to dependents, so the Zig optionalDependency kept warning that tree-sitter@0.21.1 does not satisfy peerOptional ^0.22.1. Load it from vendor/ like Dart/Kotlin/Swift instead. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(ci): add zig parse snippet to prebuild validate The six zig prebuild jobs failed at "Validate the .node loads and parses" because snippets[GRAMMAR] was undefined and tree-sitter threw "Input must be a function". Co-authored-by: Cursor <cursoragent@cursor.com> * chore: drop Unreleased changelog note from the Zig vendor PR CHANGELOG.md is owned by the release process, not individual PRs. Co-authored-by: Cursor <cursoragent@cursor.com> * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33949409205 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33949616521 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33949829377 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33950025077 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33950220655 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33950400275 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33950607933 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33950882912 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33951170483 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33951386477 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33951624305 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33951813452 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33951998309 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33952225172 * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33952524393 * fix(ci): stop native prebuild rebuild loops PR path filters and source checks see the cumulative diff, so generated binaries kept rebuilding the original source change. Skip output-only synchronize events using their exact before/head range, failing closed when Git cannot compare it. Exercise the workflow against real commit histories, including multi-commit source pushes and merge-ref drift. * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33953226606 * fix: address Zig PR review feedback (#3180) Check for the vendored package without loading its native binding so a broken installed Zig grammar fails the parsing test instead of skipping. Match the optional child descriptor in the Zig metadata declaration and include Zig in the two optional/vendored grammar comments. Validation: 159 targeted tests, TypeScript, metadata type fixture, and formatting passed. Injected native-load failure now fails instead of skipping; explicit Zig opt-out still skips. * chore(vendor): rebuild native prebuilds (tree-sitter-zig) Built by https://github.com/abhigyanpatwari/GitNexus/actions/runs/33956342308 --------- Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: gitnexus-release-bot[bot] <gitnexus-release-bot[bot]@users.noreply.github.com> |
||
|
Gergo Magyar
|
7c68905aac | Merge branch 'main' into pr-2785-feedback | ||
|
Gergő Magyar
|
a348bc3957
|
fix(build): build the web UI from prepack, not from every npm ci (#3166)
* fix(build): build the web UI from prepack, not from every npm ci gitnexus-web is a separate ~650-package tree (React, Vite, LangChain, Mermaid). Because `prepare` built it, every `npm ci` in gitnexus/ also installed and Vite-built a second product. On CI that install ran uncached inside an execSync timeout, so a healthy-but-slow install was SIGTERM'd mid-flight and surfaced as `spawnSync /bin/sh ETIMEDOUT` -- repeatedly killing node floor compat, a job that only import-links the CLI dist and never needs the UI. The UI is only needed inside the published tarball, so build it from prepack instead. `npm run build` and `prepare` are now CLI-only; pass --web (or npm run build:web) to include it. Jobs that pack or publish install gitnexus-web in their own visible step, and the in-script fallback install is untimed so a slow install can no longer be killed halfway and reported as a build failure. The tsc/vite timeout default goes 300s -> 600s so the remaining bounded steps have headroom. Default build on this machine: 30s, no gitnexus-web work. * fix(build): enforce web package artifact integrity Co-authored-by: Cursor <cursoragent@cursor.com> * refactor(build): clarify web packaging helpers without changing behavior Keep the same opt-in, fail-closed, and pack/publish preserve rules while trimming comments, sharing the test harness, and reading index.html directly instead of probing it first. Co-authored-by: Cursor <cursoragent@cursor.com> * ci: skip prepare on typecheck so a cold shared install cannot cancel the job quality/typecheck's 10-minute budget was spent on an uncached gitnexus-shared npm install plus a full prepare tsc that tsc --noEmit does not need. Co-authored-by: Cursor <cursoragent@cursor.com> * ci: stop typecheck-web from canceling before the npm cache can save Hashing gitnexus-shared into the web cache key forced a cold 650-package install; the 10-minute job then canceled and never wrote a warm cache. Co-authored-by: Cursor <cursoragent@cursor.com> * ci: give format the same 10-minute budget as lint A cold root npm ci already took 4m19s and canceled prettier at the 5-minute cap. Lint does the same install and needed 7m41s on that run. Co-authored-by: Cursor <cursoragent@cursor.com> * ci: stop installing TypeScript 7 just to compile gitnexus-shared A dedicated npm ci in gitnexus-shared took 7 minutes to add two packages (TypeScript 7's optional per-platform binaries) and cancelled typecheck, Windows pack, and coverage shard 1. Compile shared with gitnexus's tsc. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#3166) - Run tsc via execFileSync so the compiler path is never interpolated into a shell. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#3166) - Run tsc as node typescript/bin/tsc so Windows never has to execFile a .cmd shim. Co-authored-by: Cursor <cursoragent@cursor.com> * Launch tsc via node and lib/tsc.js on every OS. The npm .bin/tsc shim is tsc.cmd on Windows, which execFileSync cannot spawn. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(ci): lock eval containment against a dedicated shared npm ci Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergo Magyar
|
6925fb344d |
feat(eval): evolve review skills against historical PRs
Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergo Magyar
|
167642ec9d |
feat(ci): let a dispatched evolution run start from a blank slate
Seeding is unconditional today, so the next run would inherit the rejected proposal from a generation whose proposer could still read the hidden oracles. That taint propagates: each generation stages the previous proposal, so one contaminated proposal survives until the artifact expires. Scheduled runs still always seed — memoryless weekly runs would re-propose the same rejected candidate forever. Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergo Magyar
|
bc7f4a907e | Merge branch 'main' into pr-2785-feedback | ||
|
Gergo Magyar
|
d541105340 |
fix(workflow-bench): repair the harness defects the verbose proposer logs exposed
The first fully-logged skill-evolution run failed for five deterministic reasons that had nothing to do with the candidate under test. Each is fixed at the layer that actually owns the contract: - Strict provider adapters materialize omitted optional string arguments as "". The MCP alias normalizer now treats a blank optional alias as absent (a blank REQUIRED target is still rejected), and a trusted PreToolUse hook strips blank strings before Read/GitNexus tool calls. - MCP semantic errors rode home in a successful envelope and logged as result=ok. SessionProgress now inspects the payload and reports them as semantic-error. - Claude Code's nested sandbox overlays absent root dotfiles with device nodes, which the provenance snapshot read as unauthorized workspace changes. Those names are excluded at the workspace root and hidden from git via an immutable excludes file. - The proposer could not read /evidence from Bash (missing allowRead entry) and had no offline gitnexus runner, so it fell back to npx and hit the network. Both are now mounted; ripgrep is installed in CI. - selected-rows.json advertised host artifact names that do not exist in the mount. Rows now name their staged patch_file/transcript_files, the prompt describes the real layout, and oversized bundles compact artifacts before dropping evidence rows so no row is silently lost. Also replaces two benchmark scenarios that main already satisfies (trivial-version-alias, inv-bug-pdg-note) with non-vacuous ones, verified to fail against a pristine checkout. Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Navid EMAD
|
932d937085
|
feat: add Zig language support (#1432)
* feat: add Zig language support
Adds a Zig LanguageProvider grounded in @tree-sitter-grammars/tree-sitter-zig 1.1.2.
The grammar's published peerOptional `tree-sitter@^0.22.1` is suppressed via an
npm `overrides` entry that aliases the peer to the bundled `tree-sitter@0.21.x`;
load-time smoke testing confirmed ABI compatibility.
v1 capabilities:
- .zig file detection + Prism syntax mapping
- Top-level + nested function_declaration as Function/Method
- struct/enum/union (anonymous in the grammar — owner name resolved from the
enclosing variable_declaration in class/field/method extractors)
- container_field as struct/union fields and enum variants
- top-level const/var as Variable nodes
- free + member call_expression as @call edges
- @import("./foo.zig") local-file resolution; std and external packages
return empty (no ghost edges)
- pub keyword detection for export checking
- no heritage hooks (Zig has no inheritance; queries never emit @heritage.*)
Generic extractor changes (backward-compatible):
- field-extractors/generic.ts and method-extractors/generic.ts: empty
`bodyNodeTypes` falls back to the type declaration node itself as its own
body container — needed because Zig's struct_declaration directly contains
its container_field children. (The `extractOwnerName` hook this commit
originally introduced now exists upstream; Zig just configures it.)
Out of scope (deferred):
- usingnamespace, build.zig.zon package graph, comptime/anytype
- scope-resolution hooks (emitScopeCaptures, interpretImport, …): Zig is
classified `experimental` and uses the generic fallback resolution path
- cross-package imports (std, deps)
Tests:
- new fixture test/fixtures/sample-code/simple.zig
- Zig describe block in tree-sitter-languages integration test
- simple.zig added to parsing.test.ts fixture-existence list
- Zig added to ingestion-utils detection unit test
- Zig smoke case in parser-loader-abi.test.ts
- Zig grammar registered in the grammar-literal validation gate
* feat(zig): integrate build.zig.zon resolution + Union label from PR 1096
Ports the additive pieces of grgisme's standalone Zig provider PR
(https://github.com/abhigyanpatwari/GitNexus/pull/1096) onto the rebased
provider:
- build.zig.zon `.path` dependency resolution for bare-name
@import("pkg") (parseZigBuildZon / loadZigBuildZon in
language-config.ts, resolveZigImportInternal in import-resolvers/zig.ts,
wired through ImportConfigs.zigBuildZon). `.url` deps and
repo-escaping paths return null cleanly. 13 unit tests.
- `union(enum)` containers now produce `Union` nodes (not Struct):
'Union' added to ClassLikeNodeLabel + CLASS_LIKE_LABELS, the Zig query
tags @definition.union, CONTAINER_TYPE_TO_LABEL maps union_declaration
to 'Union'. The label was already plumbed graph-wide on main.
- /^build$/ entry-point pattern (build.zig).
- zig-basic lang-resolution fixture + resolvers integration test.
Adapted to current main while porting:
- labelOverride relabels container-nested fns Function → Method
(mirrors isKotlinClassMethod); the structure phase no longer derives
Method from the legacy method-extraction path for plain
@definition.function captures.
- IMPORTS/CALLS edges require scope-resolution hooks
(emitScopeCaptures / interpretImport) since the legacy DAG removal;
Zig does not implement them yet, so the integration test documents
that with a skipped import-edge case. The resolver itself is wired
into the resolver factory and becomes live when the hooks land.
Not ported: named-bindings extractor (the legacy namedBindingExtractor
API no longer exists) and the bespoke field extractor (the generic
factory's extractOwnerName / empty-bodyNodeTypes hooks cover Zig).
Co-authored-by: Garrett Griffin-Morales <grgisme@gmail.com>
* feat(zig): scope-resolution hooks — IMPORTS and CALLS edges (Ring 3)
Implements the registry-primary scope-resolution path for Zig, the
prerequisite for cross-file edges since the legacy DAG removal. Adds the
standard per-language stack under languages/zig/:
- query.ts: scope query (containers as Class scopes, blocks, functions),
declarations (container anchors placed on the container node itself so
the def lands in its own Class scope and the name binding auto-hoists
to the parent — populateClassOwnedMembers needs the class-like def
among the class scope's ownedDefs), @import statements (#eq?-gated
builtin), parameter/constructor type bindings, and call/constructor
reference sites. The grammar is required lazily (optionalDependency).
- captures.ts: emitZigScopeCaptures — groups query matches, drops the
plain-variable group for container/import bindings (their dedicated
rules bind the name), and relabels container-nested fns
@declaration.function → @declaration.method (labelOverride parity).
- interpret.ts: namespace-kind imports (const x = @import("…")) and
type bindings — self-parameter convention marks the receiver, Zig
sigils (*, ?, [], error unions, const) stripped from type names while
dotted qualifiers (mod.T) are preserved for Case-3 namespace-prefix
receiver dispatch.
- simple-hooks.ts: parameter bindings stay function-local (Go
rationale), local-over-import merge precedence, bounds-check arity
(always 'unknown' today — no synthesized arity metadata).
- scope-resolver.ts: emit-side wiring; build.zig.zon threads through
loadResolutionConfig into the same resolveZigImportInternal the legacy
resolver config wraps. fieldFallbackOnMethodLookup off (statically
typed). Registered in SCOPE_RESOLVERS.
The resolvers integration test un-skips the import-edge case and gains
CALLS assertions: free call (main → helper) and receiver-bound method
dispatch through a namespace-qualified constructor
(var p = pioneer.Pioneer{…}; p.tick() → main → tick).
* fix(zig): Union is class-like + missing-grammar warning (review pass)
Self-review findings on the Zig branch:
- scope/walkers.ts `isClassLike` and finalize-algorithm's
CALLABLE_OR_TYPE_LIKE did not include 'Union': a `union(enum)`
container's methods got no ownerId from populateClassOwnedMembers, so
method dispatch on union receivers silently dropped. Widened both
sets; the zig-basic fixture gains a Tag method + a CALLS assertion
(main → isEnergy) that fails without the widening (verified by
reverting).
- optional-grammars.ts now lists tree-sitter-zig with an npm `probe`
(it is an optionalDependency, not vendored): users with .zig files
and no prebuild get the standard one-line stderr warning instead of
a silently degraded index.
- Deduplicated the container-method predicate: `isZigContainerMethod`
+ ZIG_CONTAINER_TYPES now live once in languages/zig/captures.ts and
feed both the provider labelOverride and the scope-capture relabel.
- README language matrices: Zig row now claims Type Annotations,
Constructor Inference, and Config (build.zig.zon) — all true since
the scope-resolution hooks landed.
* fix(zig): anchor @declaration.variable to the binding identifier
`(variable_declaration (identifier) @declaration.name)` matched EVERY
identifier child of the node, so `const first = target;` also declared a
phantom local named `target` in the enclosing block. That phantom shadowed
the real function for later references (and starved callable-value-flow
seeds of a target). The `.` anchor pins the pattern to the first named
child — the bound name.
Regression test in resolvers/zig.test.ts pins the capture set.
* feat(zig): callable-value-flow captures + main's per-language conformance gates
Post-rebase catch-up: since this branch forked, main added three "every
registered language must appear here" tests. Each needs a Zig entry:
- callable-value-flow (#2522): Zig now emits `@callable-flow.*` facts via
`synthesizeCallableFlowCaptures` (ZIG_CALLABLE_CAPTURE_OPTIONS in
zig/captures.ts). tree-sitter-zig's `call_expression` carries arguments
as direct children with no wrapper node, which the shared helper could
not decompose, so this adds a language-neutral `extractCallArguments`
hook (mirror of `extractFunctionParameters`; `undefined` = shared path).
Zig joins the provider matrix as 'matrix' with a real assign→copy→
argument→invoke case.
- external-import-conformance (#2953): `@import("std")` beside a decoy
`src/std.zig` resolves to nothing; the decoy stays reachable via the
relative spelling. Zig holds the property (no suffix fallback), so it is
a case, not a KNOWN_GAPS entry.
- import-target-index-reuse contract (#2909): membership-probe-only
fixture (minimumScans: 0, same shape as Rust).
* fix(zig): address gitnexus-check review findings
One commit per the bot's list so each item is easy to check off:
- parser-loader: Zig row gains `userSkippable: true`, so
`GITNEXUS_SKIP_OPTIONAL_GRAMMARS` (=1 or a list naming `zig`) disables it
at analyze time like swift/dart/kotlin — as `optional-grammars.ts` already
documented. Covered in parser-loader-skip-optional.test.ts.
- export-detection: `zigExportChecker` stops at the first declaration it
reaches. A non-`pub` fn inside `pub const T = struct {…}` was reported
exported because the walk continued up to the wrapper.
- import-resolvers/zig: a `.path = "."` dep normalizes to '' and no longer
grows a leading slash (`/src/main.zig` could never match).
- language-config: build.zig.zon parsing strips `//` comments string-aware
(a `//` inside `.url = "https://…"` survives) and matches braces while
skipping string literals, so a commented-out `.path` cannot declare a dep
and a `}` in a comment/string cannot truncate the block.
- method-extractors/configs/zig: the leading `self` receiver is excluded
from `parameters` (Rust parity). Fixing that exposed a worse bug: the
`parameters` node is a plain child of `function_declaration`, not a
`parameters:` field, so `childForFieldName('parameters')` was always null
and every Zig method had no parameters, no receiver and `isStatic: true`.
One `zigParameterList` helper now feeds all three readers.
- variable-extractors/configs/zig: container (`struct`/`enum`/`union`) and
`@import` bindings are skipped via the same predicate the scope captures
use (`isZigContainerOrImportBinding`), instead of the comment merely
claiming they were.
- tree-sitter-languages.test.ts: the "missing grammar" case now forces the
absent-binding path through the loader's runtime opt-out on a fresh module
instead of passing vacuously when the package is installed.
New: test/unit/zig-extractors.test.ts (exports, receiver/parameters,
variable guard); zig-import-resolver.test.ts gains the `.` dep, comment and
brace cases.
The `createFieldExtractor` heads-up needs no change: the added branch is
unreachable for every existing config (none has empty `bodyNodeTypes`).
* fix(zig): address second gitnexus-check review pass
- import-resolvers/zig: `..` above the repository root now returns null
instead of aliasing a same-named root file (`../bar.zig` from `main.zig`
is not `bar.zig`); the stale "extension is stripped and re-added" comment
is corrected to what the code does.
- variable-extractors/configs/zig `extractType`: read the `type:` field only.
The positional fallback returned the INITIALIZER of `const f = target;`
as its type and gave up on compound annotations (`*Foo`, `?[]const u8`).
The comment claiming 1.1.2 has no `type` field on variable_declaration
was wrong (verified by AST dump) — and it is what led the review to
suspect the callable-flow `extractAssignment` callback, which was
already correct for `extern var f: T;`.
- receiver detection: only a FIRST parameter named `self` is the receiver.
`emitZigScopeCaptures` tags first-position parameters
(`@type-binding.first-parameter`), `interpretZigTypeBinding` requires the
tag as well as the name, so `zigReceiverBinding` no longer turns
`fn f(a: u32, self: T)` into an instance method.
- resolvers/zig.test.ts: both suites `describe.skipIf(!zigAvailable)`
(Swift/Dart pattern) — the grammar is an optionalDependency.
- tree-sitter-languages.test.ts: the Zig parsing case gates on
`isLanguageAvailable` instead of a catch-all `return`, so an installed
grammar that fails to load fails the test; comment no longer calls Dart
and Swift npm optionalDependencies (they are vendored).
- test/helpers/literal-collectors: `DIR_LANG` gains `zig`, so literals under
`languages/zig/**` are validated against the Zig grammar alone rather than
against every grammar.
- walkers.ts `isShapeLike` doc: Union IS included (via isClassLike, wired by
Zig's union member container); Typedef remains the only deferred one.
- language-config `ZigBuildZonConfig.pathDeps` doc: values are the raw
`.path` strings; the resolver normalizes.
Tests: zig-import-resolver (+1), zig-extractors (+3).
* fix(zig): address third gitnexus-check review pass
- language-config `parseZigBuildZon`: the `.dependencies = .{` header and
the per-entry `.<name> = .{` headers are matched only OUTSIDE string
literals (per-offset string mask + `matchZonHeader`). A `.name` or
`.description` value spelling `.dependencies = .{ .fake = .{ .path = … } }`
used to be taken as the block and returned the fake dep instead of the
real top-level one.
- import-resolvers/zig: an absolute import (`@import("/foo.zig")`) returns
null. The path walker skipped every empty component, so the leading `/`
vanished and `/foo.zig` resolved as importer-relative `src/foo.zig` — an
in-repo edge for an import Zig rejects as outside the module path.
- tree-sitter-languages.test.ts: the Zig parsing case gates on the PACKAGE
being installed (`createRequire().resolve`, minus a deliberate
`GITNEXUS_SKIP_OPTIONAL_GRAMMARS` opt-out), not on `isLanguageAvailable`,
which is false for absent AND for installed-but-broken bindings — so a
load failure (ABI mismatch, bad export) now fails the test instead of
skipping it, as the comment already claimed.
- walkers.ts `isShapeLike` doc: `Union` sits in `isClassLike` because that
is the label set the ownership walkers consult, not because unions
inherit — Zig has no inheritance and no heritage hooks. The previous
wording ("inheritance-capable owner") said otherwise.
- Not re-fixed (already addressed in the second pass, findings carried
over): "receiver = any parameter named self" — `interpretZigTypeBinding`
only sources a first-position parameter as `self`; `zigReceiverBinding`'s
doc now states that invariant. "`DIR_LANG` has no zig entry" — it does.
Extended one level out: `BASENAME_LANGS` (`zig.ts`) and `PREFIX_LANGS`
(`ZIG_`) map to the Zig grammar too, so extractor configs and the
export-detection set are validated against Zig alone. That immediately
caught a dead `childForFieldName('parameters')` in
method-extractors/configs/zig `zigParameterList` (there is no such field;
the named-child lookup was already the one doing the work) — removed.
Tests: zig-import-resolver (+2: absolute path, header inside a string);
both fail on the previous code.
* fix(zig): address fourth gitnexus-check review pass
- language-config: `parseZigBuildZon` only accepts a `.path` that is a
DIRECT field of a dependency entry. Nested blocks inside the entry body
are blanked (string-aware, offsets preserved) before the `.path` regex
runs, and a match starting inside a string literal is rejected, so
`.dep = .{ .url = "…", .meta = .{ .path = "x" } }` no longer becomes a
path dep. Regression test in zig-import-resolver.test.ts (fails on the
previous code).
- tree-sitter-languages test: the "grammar is absent" case now drives the
loader's real `source.load()` catch branch — `node:module` is stood in
with a `createRequire` whose require throws MODULE_NOT_FOUND for
`@tree-sitter-grammars/tree-sitter-zig` and delegates everything else —
instead of the `GITNEXUS_SKIP_OPTIONAL_GRAMMARS` opt-out, which has its
own test. It also asserts the opt-out flag is NOT set and that other
grammars still load (non-fatal optional failure).
Not re-fixed:
- "Return type is read from the wrong tree-sitter field": tree-sitter-zig
1.1.2 has NO `return_type` field on function_declaration — the type after
`)` is the `type` field (AST dump: `builtin_type "i32" field=type`; the
proposed `childForFieldName('return_type')` is null for every function).
A pin test in zig-extractors.test.ts asserts both the grammar fact and
that `returnType` is extracted (`void`, `!*Counter`).
- "`DIR_LANG` has no zig entry": it does (added in the first pass and
answered again in the third); the finding is carried over unchanged.
* fix(zig): address fifth gitnexus-check review pass
- tree-sitter-languages test: the Zig "functions, structs, enums, and
imports" case now asserts the `import.source` capture for
`const std = @import("std");` (the fixture's only import), so a query
change that drops Zig import matching fails it instead of passing
unchanged.
Not re-fixed:
- "Return type is read from the wrong tree-sitter field": carried over from
the fourth pass unchanged. tree-sitter-zig 1.1.2 has no `return_type`
field on function_declaration; the return type IS the `type` field, and
the pin test added in the fourth-pass commit
(zig-extractors.test.ts, `childForFieldName('return_type')` is null,
`returnType` = `void` / `!*Counter`) proves it.
- "`DIR_LANG` has no zig entry": carried over unchanged for the third time;
the entry exists since the second-pass commit.
* feat(zig): export fn visibility, opaque containers, named test blocks, member ownership
Ports the parts of upstream PR #305 (closed, unmerged) that our Zig
provider lacked, plus two gaps found while porting.
- `export fn` / `export var` (C-ABI linkage, never `pub`) are exported;
the pub/export predicate is now shared by the export checker and the
method/variable extractors' visibility (`hasZigVisibilityKeyword`).
- `const H = opaque { … }` is a Struct-labelled container (it may own
methods, never fields) in both the structure queries and the scope
query; ZIG_CONTAINER_TYPES is the single source for the extractor
configs.
- `test "name" { … }` blocks are Function nodes named by the string
node WITH quotes, so `test "add"` beside `fn add` cannot merge onto
Function:<file>:add; `test_declaration` joins FUNCTION_NODE_TYPES and
the Zig method config names it in the enclosing-function walk, so
calls inside a test attribute to the test. Anonymous `test {}` and
decl-tests `test add {}` are scopes without a node (an empty-name hook
result stops the walk instead of falling through to the identifier of
the function under test).
- Empty container bodies (`struct {}`, `opaque {}`) no longer mint a
nameless Property: tree-sitter-zig 1.1.2 recovers them as a
container_field with a MISSING identifier; #not-eq? guards in both
queries and the field extractor drop it.
- Owner walk (`findEnclosingClassInfo`): an anonymous container bound
by the enclosing `variable_declaration` takes the binding identifier,
same shape as the Go `type_spec` branch. Before this NO Zig member
had an owner — zero HAS_METHOD / HAS_PROPERTY edges for Zig.
Not ported from #305, deliberately: `builtInNames` (a bare-name call-site
drop filter; `alloc`/`free`/`append`/`print` are the most common user
method names in Zig and `std.*` receivers are already external via the
import binding), `usingnamespace` (removed in Zig 0.15), `@cImport`,
`build.zig` ignore, and the web-app changes.
* test(zig): absent-grammar case owns GITNEXUS_SKIP_OPTIONAL_GRAMMARS
The loader parses the opt-out variable lazily once per module copy, so
under `GITNEXUS_SKIP_OPTIONAL_GRAMMARS=zig` (or `all`) — a supported way
to run — the fresh loader took the opt-out branch and the "not the
opt-out path" assertion failed before the absent-binding path ran.
Clear the variable for the fresh module and restore it in `finally`,
instead of returning early: the branch stays exercised in every
environment. Verified: fails on the previous code under `=zig`, passes
with and without the variable now.
* fix(zig): declare the Union relation pairs — analyze aborted on any union
`gitnexus analyze` exited 1 on every Zig repository that declares a
`union` (including the zig-basic fixture itself): the member-ownership
commit made `union_declaration` a MEMBER_OWNER, so HAS_PROPERTY /
HAS_METHOD edges are emitted FROM a Union node, but `Union` was not in
LINKABLE_LABELS, so the schema's scope-bridge cross product never
generated a `FROM Union` pair and LadybugDB rejected the edge
(`labelPair: "Union|Property"`). Resolver tests stayed green because
they never write to the DB.
- `Union` joins LINKABLE_LABELS (also bridges `Tag{…}` constructor
references); the three hand-written `→ Union` target pairs move to the
generated half, per the STRUCTURAL_PAIR_DDL rule.
- structural-pair-coverage gains an optional-grammar corpus with
zig-basic (`Union|Property`, `Union|Method` sentinels), skipped when
the grammar is absent.
- Rust `union_item` note updated: the three gates it cited are widened.
Note for reviewers: the DDL fingerprint changes (#2808), so existing
indexes are rebuilt on next analyze.
* feat(zig): resolve path deps through the dep's build.zig and src/root.zig
The bare-name resolver only knew `src/<name>.zig` and `src/main.zig`.
`zig init` has written `src/root.zig` for libraries since 0.12, so the
default library layout never resolved. Now: the root the dep's own
build.zig declares (`b.addModule("<name>", .{ .root_source_file =
b.path("…") })`, name-matched module first), then src/root.zig,
src/<name>.zig, src/main.zig. `normalizeZigDepPath` is shared by the
loader and the resolver.
* feat(zig): Const/Variable defs, member imports, receiver typing, generic type constructors
Coverage gaps found by indexing idiomatic Zig against the branch:
- Const / Variable nodes: ZIG_QUERIES had no @definition.const /
@definition.variable, so `pub const VERSION`, error sets and type
aliases were absent and zigVariableConfig never ran. Rules are gated on
the literal `const` / `var` keyword — tree-sitter-zig 1.1.2 parses
statement assignments (`x = 5;`, `x += 1;`, `_ = expr;`) as keyword-
less `variable_declaration`s, and the scope query minted a phantom
local per assignment and one `_` per discard. Container and @import
bindings are skipped via `shouldSkipDefinitionCapture`.
- Imports: `const X = @import("x.zig").X` (named / alias), `const X =
ns.X` where `ns` is an @import binding of the file (promoted to a
named import), and `pub usingnamespace @import(...)` (wildcard, with
`expandsWildcardTo`). All three lost the file-level IMPORTS edge.
- Receiver typing: `var x: T = undefined` / decl literals `const x: T =
.init()` (annotation), `var c = T.init()` / `mod.T.init()` (call
return), `List(u8){}` (instantiation literal); `normalizeZigTypeName`
drops the comptime argument list.
- Generic type constructors `fn List(comptime T: type) type { return
struct {…}; }`: the returned container is a Struct/Union/Enum named
after the fn, owns its members (HAS_METHOD / HAS_PROPERTY), binds in
the module scope beside the Function def, and is emitted ahead of it
so a named import binds the type.
- `export` vs `pub`: `visibility` is now `pub`-only (Zig-module fact);
`isExported` keeps `pub|export` (visible outside the unit, as C's
external linkage). `export fn` without `pub` is not reachable from
other Zig files.
- Extractor configs share `zigContainerName`; ast-helpers' owner walk
learns the type-constructor shape.
Tests: zig-idioms fixture (10 resolver cases), extractor/interpret unit
cases for each rule.
* fix(zig): address sixth gitnexus-check review pass
- Windows absolute `.path` deps (`C:\x`, `C:/x`) return null from
`normalizeZigDepPath` like POSIX ones; a `/`-only check let them
through as repo-relative.
- `parseZigBuildZon` accepts the `.dependencies = .{` header only at
brace depth 1 (a direct field of the file's `.{`), so a same-named
field nested in an earlier struct cannot hijack the block.
- `importsExecuteWhereWritten: false` on the provider: `@import` is
compile-time name lookup (as C `#include`, Rust `use`); a body-level
`@import` is no longer marked `runsOnlyWhenCalled`.
- Namespace imports record the MODULE as `importedName`
(`zigModuleNameOf`: last path segment without `.zig`), per the shared
contract; the local handle stays `localName`.
- Keyword-less `<ident> = @import(…)` (`_ = @import("x.zig")` in a test
block) is a `side-effect` import: file edge, no binding. Only
`const`/`var` declarations bind a name or feed alias promotion.
- `extractZigFunctionName` doc: an empty name is falsy, so the enclosing-
function walk skips the test node and continues to the File; it does
not "end" there.
Not re-fixed: "DIR_LANG has no zig entry" — carried over for the fourth
pass in a row; `test/helpers/literal-collectors.ts` has had `zig` in
`DIR_LANG` (line 91) and `BASENAME_LANGS` since the second-pass commit.
Regression tests: absolute-path spellings, nested `.dependencies`
decoy, namespace/side-effect interpretation, function-scoped `@import`
not deferred (all four fail on the previous source).
* fix(zig): address seventh gitnexus-check review pass
- The scope query's `@import` binding rules are keyword-gated (`"const"` /
`"var"`, first-child anchored) like every other binding rule, and the
keyword-less `<ident> = @import(…)` statement has its own
`@import.side-effect` rule. Tree-sitter queries cannot express "no
keyword child", so that rule also matches the keyword shapes and
`emitZigScopeCaptures` drops those (they are the binding rules'
matches). Behaviour is unchanged from the sixth-pass fix — the existing
side-effect test covers it — the query text now carries the guard the
finding asked for.
Not re-fixed: "DIR_LANG has no zig entry" — fifth pass in a row;
`test/helpers/literal-collectors.ts` has had `zig` in `DIR_LANG` since
the second-pass commit. Left for a human reviewer to close.
* fix(zig): resolve @import of the repo's own build.zig modules (F3)
Bare-name imports were resolved through build.zig.zon path deps only, so
the module a repo's ROOT build.zig declares for itself —
`b.addModule("lightpanda", .{ .root_source_file = b.path("src/lightpanda.zig") })`,
imported by name from 378/567 Lightpanda files — never produced an IMPORTS
edge, and nothing reached through `lp.X` resolved. A repo with a build.zig
but no build.zig.zon got no resolution config at all.
- language-config: `parseZigRootModules` (static scan of the root build.zig:
`addModule("<name>", …root_source_file = b.path("<p>.zig")…)`, and
`createModule`/`addModule` bindings named via `addImport("<name>", m)` or
`.imports = &.{ .{ .name, .module = m } }`; generated / `.url` / computed
modules are skipped) → `ZigBuildZonConfig.rootModules`.
- `loadZigBuildZon` → `loadZigBuildConfig`: reads the zon AND the root
build.zig; null only when neither contributes.
- resolver: root modules are consulted before path deps; std/builtin/root
still never resolve.
- fixtures: zig-idioms gains a Lightpanda-shaped root module (+ decoy
`addOptions().createModule()`); new zig-rootmodule (build.zig, no zon).
Corpus (Lightpanda): IMPORTS 3014→3389 (378 edges to src/lightpanda.zig,
was 0), CALLS 13885→13989, ns.f() 79.3%→83.3%,
param.m() type=ns-qualified 43→46/417.
* fix(zig): import every @import in expression position; resolve @import("x").f()
Both query sets only saw `@import` as the value of a const/var or under
`usingnamespace`, so an @import in any other position produced no file
edge: Lightpanda's `pub const Interfaces = .{ @import("a.zig"), … }`
registration table (288 modules), call arguments
(`CounterEnum("size", @import("ArenaPool.zig").BucketSize)`), comparison
operands (`JsApi == @import("x.zig").JsApi`) and member-call receivers
(`try @import("dump.zig").root(...)`) — 417 of 3,401 in-repo import pairs
had no IMPORTS edge, and the 80 inline-receiver calls resolved 0 times.
Scope query: a catch-all `@import.inline` rule matches every `@import`
builtin; `emitZigScopeCaptures` drops the ones a binding rule (or the
keyword-less side-effect rule) already claimed (by string-node id) so a
bound import is never doubled, emits the rest as side-effect imports once
per distinct source per file, and binds a member-call receiver as a
namespace import whose local name is the builtin's own text — the
`@reference.receiver` text on that call is identical, so the shared
namespace-receiver lookup (Case 1) resolves the member in the imported
module.
ZIG_QUERIES: the three variable_declaration/usingnamespace-anchored
`@import` rules collapse into the same single builtin rule (the structure
phase only skips import matches; one match per builtin keeps
tree-sitter-languages' exact-capture assertion intact).
Lightpanda corpus (zig-corpus-check, before → after): IMPORTS 3014 → 3426,
in-repo pairs missing 417 → 5 (4 under a default-ignored `cache/` dir, 1 a
commented-out import the census regex counts), `@import(..).f()` 0/80 →
72/80 (the 8 left are `@import("root")` and non-import builtins the census
mislabels), CALLS 13885 → 13959; every other line unchanged.
* feat(zig): model file-structs — a file with top-level fields is a Struct named after the file
In Zig every file is a struct; one that declares top-level fields is an
instantiable type whose name is the file stem (`Page.zig` declares `Page`,
`@typeName` agrees), and its top-level `fn`s taking `self` are its methods.
Lightpanda spells 413 of 567 files this way and, before this, `page.getArena()`
on a `page: *Page` parameter resolved 23 of 993 times (2.3 %) — `impact` on
`Page.getArena` reported 0 callers for 159 call sites, and 2,395 top-level
fields were ownerless Property nodes.
Definition phase: `((source_file (container_field …)) @definition.struct)` +
the class extractor names it from the file path (`zigContainerName(source_file,
filePath)`); top-level fns/fields are owned through the new
`LanguageProvider.resolveFileTypeOwner` hook (consulted by
`findEnclosingClassInfo` when the walk reaches the tree root, and by the
method/field extractors' owner lookup) — ids become `Method:<file>:Page.getArena#0`
/ `Property:<file>:Page.session` with HAS_METHOD / HAS_PROPERTY edges.
Scope phase: `emitZigScopeCaptures` emits a Class scope over the whole file
(same range as the Module scope, nested under it — the pair `canParentScope`
already admits) plus a Struct def anchored on it; member NAME bindings are
hoisted back to the Module scope by `zigBindingScopeFor` so `Page.init()`
(namespace member) keeps working, while ownedDefs stay in the Class scope so
`populateClassOwnedMembers` stamps the owner. The file-level `const Page =
@This();` alias no longer mints a Const (it would shadow the Struct); `@This()`
aliases in type position (`self: *SigHandler` in Sighandler.zig, nested
`Self`) are rewritten to the container name so receivers resolve. A namespace
import of a `.zig` file gets a NAMED twin of the file stem so `x: *Page` in the
importer binds the type as well as the module.
Shared, additive: `resolveFileTypeOwner` hook; `filePath` threaded to
`ClassExtractionConfig.extractName` / `extractOwnerName`; nameless
`definition.struct` passes `getLabelFromCaptures` like `definition.class`
already did (extractor synthesizes the name).
Lightpanda corpus (zig-corpus-check): param receivers typed by a file import
23/993 → 993/993; `self.m()` 99.2 → 100 %; annotated locals 11.6 → 23.2 %;
CALLS 13,885 → 15,689; HAS_METHOD 1,477 → 8,003; ownerless Property 2,395 →
276; Function/Method 8,378/1,518 → 2,562/7,330; no row regressed.
Fixture `zig-filestruct` (Page/Session/Sighandler/util) + unit tests pin the
shape, the stem naming, the alias rewrite, the namespace twin and the
unchanged namespace-file behaviour.
* test(zig): expression-position import case sees the file-struct type twin
* fix(zig): stop reading member calls `x.f(arg)` as direct calls named `f`
tree-sitter-zig spells `field_expression` as `object:`/`member:`; the shared
callable-flow reader only knew `property`/`field`/`method`, so every Zig
member call collapsed to a DIRECT call named after the member and the
solver fanned each argument out to every same-named callable
(4,761 cap warnings on Lightpanda, `Global.deinit -> Global.deinit`
self-loops through `pub const release = deinit;`).
Shared (grammar-neutral, receiver-gated):
- `memberParts` also reads `member` (only C/C++ `offsetof_expression` and
JS `class_body` expose that field, without a receiver field).
- A member call is a field-stored-callable invoke only when a MEMBER store
(`o->run = handler`, `self.f = target`) or a declared callable-typed field
is visible — a same-named plain binding no longer gates it.
- `direct-callee-name` requires a direct designator: `.init(x)`,
`' '.join(x)`, `string.Join(x)` name no callee to seed by simple name.
Zig: formals are numbered without the leading `self`, so `r.run(target)`
joins `cb` and yields `Runner.run -> target`.
Goldens for python/csharp regenerated: the only drift is the dropped
`direct-callee-name` on `' '.join(...)`, `text.strip().ljust(...)`,
`string.Join(...)`.
Lightpanda: cap-warnings 4761 -> 2, cvf self-loops 10 -> 0,
CALLS 13885 -> 13857 (28 removed, all callable-value-flow: 10 self-loops,
17 same-name fan-out, 1 lost `on -> TypeErased.start`; +1 correct
`Arena.alloc -> allocator`).
* fix(zig): bind container field types so `self.field.m()` resolves (F5)
A container's field types were never bound on its Class scope: the scope
query's `container_field` rule captured only the name, and
`emitZigScopeCaptures` synthesized no `@type-binding.field` group. The
compound resolver reads member types from that scope
(`typeOfMemberOnClass` → `classScope.typeBindings.get(field)`), so
`self.session.name()`, `self.counter.incr()` — Lightpanda's dominant
cross-object call shape — resolved 9 of 2803 times (0.3 %).
- query.ts: capture `type: (_)? @declaration.field-type` on
`container_field` (enum variants have none).
- captures.ts: per typed field, push a `@type-binding.field` group (name =
field, type = the type text, `@This()` aliases rewritten to the container
name like parameter types); anonymous inline containers are skipped. The
binding lands on the container's Class scope — the file's Class scope for
a file-struct — since `zigBindingScopeFor` hoists only declaration names.
- query.ts/captures.ts: `const page = self.page;` / `var s = self.session;`
one-level field aliases become `@type-binding.alias` bindings whose "type"
is the RHS path; the resolver's member-alias branch re-resolves it as a
receiver chain. Import aliases (`const Counter = counter.Counter;`) are
dropped — they are named imports.
- interpret.ts: `@type-binding.field` → 'annotation',
`@type-binding.alias` → 'assignment-inferred' (an annotation on the same
binding wins).
Corpus (Lightpanda, zig-corpus-check): self.field.m() 11/2803 (0.4 %) →
1433/2803 (51.1 %); ident.m() bound=local-field-access 29/1902 (1.5 %) →
307/1902 (16.1 %); chain.m() 146/5022 (2.9 %) → 642/5022 (12.8 %);
CALLS 15838 → 18066. self.m() / free f() / ns.f() unchanged.
Tests: unit (zig-extractors) — one @type-binding.field per typed field with
sigils stripped and aliases rewritten; the binding hosted on the container's
Class scope (file-struct: the file's Class scope, not Module) with the
written spelling as declaredSpelling; field aliases bound to the RHS path
and never for import aliases. Integration (zig-idioms `holder.zig`,
zig-filestruct `Page.zig`): `viaField → incr` ×2 into counter.zig,
`viaAlias → get/twice`, `sessionName → name` / `sessionLabel → name` into
Session.zig. All fail without the change. The optional-payload capture
`if (self.opt) |c| c.incr()` is not asserted (F6).
* fix(zig): `pub const X = @import(…)` at file scope republishes X (reexportsName)
Lightpanda's `lightpanda.zig` is one long list of `pub const Arena =
@import("Arena.zig");`, and most files name their types through it (`const
lp = @import("lightpanda"); const Arena = lp.Arena;`, `arena: *lp.Arena`).
The scope side treated those bindings as plain imports of the hub file, so
the hub never published the names it re-exports and a third file's `const
Arena = lp.Arena;` (promoted to a named import of `Arena` from the hub) found
nothing.
`emitZigScopeCaptures` now marks named/alias import groups whose declaration
is a file-level `pub const` — the `@import(...).X` form, the alias promotion
`pub const Bar = ns.Bar`, and the file-struct type twin of `pub const Arena =
@import("Arena.zig")` — and `interpretZigImport` sets the shared contract's
`reexportsName: true` on them (the Python `__init__.py` shape, consumed by
`buildReexportClosures`). Private and fn-local bindings stay unflagged.
Not covered here: a receiver ANNOTATED with the dotted hub path (`arena:
*lp.Arena`) — Case 3 of the receiver-bound pass looks the member up with
`findExportedDef`, which only sees locally declared names; following
re-exports there is a shared change left for a follow-up.
* fix(zig): type receivers through `const X = <type expr>;` aliases (F7)
`const LocalAlias = Local;`, `const T2 = Thing;` (alias of an alias/import)
and `const B = util.List(u8);` (an INSTANTIATED generic type constructor)
were plain `@declaration.variable` bindings, so `LocalAlias.mk()`,
`var l = LocalAlias.mk(); l.go()`, `T2.make()`, `B.init()`, `B{}` and
`var x: B` all typed nothing (review repro r3-flow b1..b9; Lightpanda:
`pub const Proto = HtmlElement;` x68, `const Allocator = std.mem.Allocator`
x104, `pub const KeyIterator = GenericIterator(...)`, fn-local
`const R = ...(...)`).
Model: a `@type-binding.alias` binding of the alias NAME to the value's type
text — Rust's `let x = y` / JS's `const B = Foo`, source
'assignment-inferred' — NOT a TypeAlias def. Reasons: (1) the shared
machinery already chains typeBindings (`followChainedRef` in the extractor,
`followChainPostFinalize` after propagation), so `var l = LocalAlias.mk()`
and `var x: B` reach the target through the alias with no new shared code;
(2) nothing shared follows a `TypeAlias` def to its target — `isShapeLike`
only makes the alias itself a member owner (TS object-type aliases) — so a
relabel would have needed language-named shared code; (3) graph node ids
are UNCHANGED: every alias stays `Const:<file>:X`. `normalizeZigTypeName`
already drops the comptime arguments, so `util.List(u8)` binds `util.List`
and resolves through the namespace import (Case 3). The identifier /
member shapes also take `var` (`var cur = orig; cur.go()` — the cursor
idiom, same binding as Rust's `let x = y`).
Heuristic, stated as such: a CALL value is kept only when the callee's last
identifier is TitleCase (Zig's naming convention for types), because the
grammar cannot tell `util.List(u8)` from `util.makeThing()` and the latter
belongs to the call-return rules; the call-return group is dropped for the
same TitleCase shape so the two never race on match order. Import bindings
(`const Stack = @import("x.zig").Stack`), promoted namespace-member aliases,
enum/decl literals (`.foo`) and the `type:` annotation of
`var b: T = undefined;` are excluded.
Not done: the two-hop `pub const bridge = js.Bridge(T); bridge.accessor()`
chain. `js.Bridge` is a Function that RETURNS `bridge.Builder(T)` (a call,
not a container), so the alias binds `js.Bridge`, Case 3 finds a Function
with no members in js.zig, and Case 3b is skipped for a namespace head.
Following that hop needs a namespace-member return-type route in shared
code (or a Zig `resolveQualifiedReceiverMember` hook that re-implements
member lookup without the model); left for a follow-up.
Corpus (Lightpanda, `harness/zig-corpus-check.mjs`): CALLS 15838 -> 16051
(+213, 0 removed), `ident.m() bound=local-alias` 6/63 -> 36/63,
`local-call` 159 -> 176, `module/unknown` 110 -> 116, `local-other`
271 -> 273; `self.m()`, `free f()`, `ns.f()` unchanged or up.
* fix(zig): one alias rule set — F7's alias rules subsume F5's field-access alias rules
* fix(zig): type locals through try/catch/orelse, return types and payload captures
F6 of the gitnexus-check review. Three gaps in the value flow that types a
local receiver, all measured on Lightpanda:
1. `@type-binding.call-return` needed the `call_expression` as the DIRECT
value child, so `const p = try Page.init(…)` (2,551 sites), `… catch
return` (410) and `… orelse return` typed nothing. The rule is now one
keyword-gated declaration match; `emitZigScopeCaptures` unwraps `try`,
`catch`, `orelse` and parentheses (`zigUnwrapValue`) and decides what the
value types (`zigCallReturnTypeOf`): a module-level receiver still names
the type (`Counter.init()` → Counter, Rust `Foo::new()`); a free call
binds the callee name (`makeThing`); a member call on a fn-LOCAL receiver
(parameter / local / payload — Zig forbids shadowing, so "declared in the
fn" is exact) binds the compound `node.asElement()` the shared resolver
walks to the method's return type — instead of typing `el` as `Node`. A
TitleCase callee (`List(u8)`) is a type constructor and binds nothing.
2. No `@type-binding.return` existed. `fn make() !*Thing` now binds
`make ↦ Thing` in the enclosing scope (Module for free fns, the container's
Class scope for methods, where the compound resolver reads it). Builtins,
`type`, `@TypeOf(…)` and comptime type parameters (`?*T`) bind nothing;
`@This()` / `Self` returns name the container. `normalizeZigTypeName` now
strips the error union BEFORE the payload's sigils, so
`Allocator.Error!*Page` → `Page` (it used to leave `*Page`).
3. Payload captures had no binding at all. `populateZigRangeBindings`
(registered as `populateRangeBindings`) types `for (items) |it| / |*it|`,
`for (items, 0..) |it, i|`, `if (opt) |v|`, `if (call()) |v|`,
`while (it.next()) |x|` from the SUBJECT's written type minus one layer
(`[]T` element, `?T` payload) — declining when the layer is not visible
(`ArrayList(T)`) — and the same projection for `const t = items[i]` /
`opt.?` / `ptr.*`. `catch |err|` and `switch` prongs are skipped.
Corpus (Lightpanda, gate before → after): CALLS 15838 → 17853;
`ident.m() bound=local-try/catch/orelse` 11/1298 → 584/1298;
`local-call` 159/1282 → 398/1282; `payload` 82/1085 → 233/1085;
`other-recv:call_expression` 5/991 → 457/991; `local-other` 271 → 364;
`self.m()` 100 %, `free f()` 97.5 %, `ns.f()` 83.4 % unchanged; nothing down.
Not covered: `const t = ns.f()` (a namespace fn's return type across files —
Case 3 has no path from a namespace head to a callable's return binding),
and expression receivers (`items[i].run()`, `o.?.run()`).
* fix(zig): resolve leftover fixture merge markers (Page.zig)
* fix(zig): reconcile F6 value inference with F7 aliases and F5 field bindings
- A fn-local TitleCase receiver (`const R = generic.List(u8); var l =
R.init();`) is a type alias (F7), not a value local: `R.init()` names the
type `R` like `Counter.init()` does at module level, so `l` chains
R → util.List → push. F6's local-receiver rule now excludes TitleCase heads.
- The F6 unit helper only collects the value-inferred / return kinds it
owns; F5 field and F7 alias bindings for the same names are asserted in
their own suites.
* fix(zig): give function-local and anonymous containers an identity (F8)
`const R = struct {…}` declared inside a fn (Lightpanda's reflection.zig
has ~20, one per builder) all collapsed onto one `Struct:<file>:R` with one
`R.get`; anonymous containers (`std.sort.pdq(…, struct { fn lessThan … }
.lessThan)`, `const byte_size = struct { fn it … }.it;`, `?struct { min,
max }` field types) had no identity at all, so their fns were OWNERLESS
Methods (`Method:<file>:lessThan#3`) that collided across a file.
`zigContainerName` now yields the graph IDENTITY on both phases:
- function-local named: `<enclosing callable>$<name>` — `Reflect.string$R`
(Java local-class `$` chain; `populateClassOwnedMembers` leaves it whole);
- anonymous: `<host>$<ordinal>` — `build$1`, `Outer$1`, `Page$1`
(javac's `Outer$1` numbering per host, in source order);
- a `test` host is keyed `test@L<line>` (its string does not survive the
class extractor's qualified-name normalization).
`zigContainerBindingName` keeps the spelling code writes (`R`) for scope
bindings and `@This()` alias rewrites (`@declaration.binding-name`).
Structure phase: bare `(struct|enum|union|opaque_declaration)` rules mint the
local/anonymous nodes via the class extractor; `shouldSkipDefinitionCapture`
keeps exactly one rule per container (`zigContainerAnchor`); a new
grammar-neutral `resolveContainerTypeOwner` provider hook lets the shared
owner walk name a container from context, so `Method:<file>:Reflect.string$R
.get#0` and its HAS_METHOD source agree by construction. Scope phase: the
wrapper group splits name/binding-name for locals and anonymous containers
get synthesized `@declaration.<kind>` defs (`is-synthetic`).
Lightpanda: ownerless Methods 14 → 0, fns without a node 55 → 0, ownerless
Properties 276 → 5, HAS_METHOD 8003 → 8074, HAS_PROPERTY 7275 → 7403,
CALLS 15838 → 15857, Struct 1905 → 2199; no resolution bucket dropped.
* fix(zig): re-add the implicit receiver on the call side so both method-call spellings reach the callback formal
`extractFunctionParameters` sliced the leading `self` off the formals, which
lined up `r.run(target)` (target@0 ↔ cb@0) but lost the explicit spelling
`Runner.run(&r, target)` (&r@0, target@1 ↔ cb@0): the callback never joined
its formal and `run → target` was missing (PR #1432 review by koriyoshi2041).
Formals are numbered once per function while the receiver differs per call
shape, so the fix lives in `extractCallArguments`: keep `self` as formal 0
and prepend the receiver as actual 0 when the callee is a member call on a
VALUE receiver — chain head is a fn-local name that is not TitleCase, the
same value-vs-type rule F6 uses. Namespace / type / decl-literal receivers
(`Runner.init(cb)`, `helpers.apply(cb)`, `List(u8).init`, `.init(cb)`) get
no prepend. Known residual gap, documented: a module-level value receiver
(`global_runner.run(cb)`) is not fn-local and still misses.
Tests: the F2 integration case now asserts both spellings plus a namespace
call; the unit contract pins `self@0, cb@1` and the per-call actual index.
* fix(zig): address eighth gitnexus-check review pass
- Named dependency modules: `parseZigBuildModuleRoots` scanned
`addModule("<name>", .{ … })` with a `[^}]*` regex, so a nested field
before `.root_source_file` (`.imports = &.{ .{ … } }`) ended the match
at the inner `}` and demoted the module to an unnamed fallback — the
first exe/lib root in the file then answered `@import("<name>")`. The
named lookup now walks the balanced `addModule(…)` argument list
(same scanner as `parseZigRootModules`, comment-stripped, string-aware);
the unnamed fallbacks are unchanged. Regression test in
`zig-import-resolver.test.ts`.
- Type-position `@import`: `var x: @import("m.zig").T = undefined;` was
read as an import binding of `x` on both sides — the query rules match
the `type:` child like a value, and `isZigContainerOrImportBinding`
scanned every named child — so `x` was never declared and became a
named import of `T`. The helper now skips the `type:` field, and
`emitZigScopeCaptures` drops binding-rule matches whose `@import` sits
in the annotation (`isZigTypePositionImport`) without claiming the
source, so `x` binds as a variable and the file edge survives as a
side-effect import. Regression tests in `zig-extractors.test.ts`
(variable extractor + scope captures).
- Union in the class-capture skip guard: the parse-worker's inline
class-like predicate lacked `Union`, so a `Union` definition bypassed
`shouldSkipClassCapture` unlike every other `ClassLikeNodeLabel`.
Added the label; no Zig behavior changes (Zig defines no skip hook), so
no test.
- File-owned method ids in `findEnclosingFunctionId`: the arity lookup
used `findEnclosingClassNode` while the owner came from the file-owner
aware `cachedFindEnclosingClassInfo`, so a Zig file-struct's top-level
fn produced `Method::Page.get` without the `#<arity>` suffix. It now
uses `findEnclosingClassNodeOrFileOwner`, the definition-phase lookup.
Consistency fix only: `ParseWorkerResult.calls` / `.assignments` (the
sole consumers of this id) are merged but not read since #942 — CALLS
edges come from the scope pipeline, whose ids were already right — so
no observable graph change and no test.
Not re-fixed:
- `test/helpers/literal-collectors.ts` `DIR_LANG` has no `zig` entry
(raised for the seventh time): the entry exists (`zig:
SupportedLanguages.Zig`, added by the second-pass commit), so
`languages/zig/**` literals are already validated against the Zig
grammar alone; documented in the PR body since the fifth pass.
* fix(zig): keyword-gate the constructor type-binding rules
The three `@type-binding.constructor` rules (`const p = T{…}`, `mod.T{…}`,
`List(u8){…}`) matched any `variable_declaration` with an identifier and a
`struct_initializer`, keyword or not — and tree-sitter-zig 1.1.2 parses a
re-assignment `p = T{…};` (and `_ = T{…};`) as the same node type. So an
assignment minted a constructor binding for `p` in its own block, and one
for `_`. Zig's static typing makes the extra binding redundant (`p`
already carries its type from its declaration: annotation, constructor or
inferred value), so it cost little, but it declared nothing and stood out
against every other binding rule (`@declaration.variable`, the import
rules, the call-return rules), which are keyword-gated for exactly this
shape. Split each rule into `"const" .` / `"var" .` variants, like the
call-return rules.
Regression test in `zig-extractors.test.ts`: `p = T{…}`, `q = mod.T{…}`,
`l = List(u8){}` and `_ = T{…}` after their declarations yield only the
three declaration bindings (fails on the previous query). The zig,
callable-value-flow, grammar-literal and tree-sitter-languages suites are
unchanged.
Raised twice by gitnexus-check (passes on 2026-08-18 12:42 and 12:56).
* fix(zig): re-baseline the callable-flow capture fingerprints, keep `await f<T>(x)` a direct callee
The `benchmarks (GITNEXUS_BENCH)` CI job gates two capture fingerprints
that this branch's shared callable-flow change (3e62b99a) moved without
re-baselining: `bench/python-scope/baseline-fingerprint.txt` and four
languages in `bench/scope-capture/baselines.json` (csharp, cpp,
typescript, kotlin). Both `--check` runs pass on origin/main and failed
on this branch; every other language matched its baseline on the same
run.
Drift, verified by dumping the canonical matches on both trees:
- csharp / kotlin / python: exactly the intended change — a MEMBER call
(`string.Join(x)`, `.forEach { }`, `' '.join(x)`, `.ljust(w)`) no
longer carries `direct-callee-name`; the argument fact is unchanged.
- cpp: `choice.select(1)` (cpp-deleted-overload) drops an INDIRECT
invoke + its synthetic `@reference.call.free` that were gated only by
the same-named free `select` binding; the site is a genuine method
call already captured as `@reference.call.member`. capture_groups_fp
4605 -> 4601.
- typescript: `await svc.verify<T>(x)` (member) and `initializer()(cb)`
(call-of-call) drop the name as intended. But `await verifyToken<T>(x)`
— a DIRECT call — lost it too, because tree-sitter-typescript parses
`await f<T>(x)` as `call_expression(function: await_expression(f),
type_arguments, …)` and the new direct-designator gate saw an
await_expression, not `f`. `wrappedExpression` now unwraps
`await_expression` (no named field, so the field-based unwrap missed
it), restoring parity with main for the direct spelling while the
member spelling stays nameless. Regression test added; it fails
without the unwrap on both assertions.
Gates run locally: scope-capture --check (15 languages), python-scope
--check, import-target --check, tsc, eslint, prettier, the callable-flow
/ golden / tripwire / resolver test files (33 files), full suite with
coverage (82.9/71.5/89.1/86.4 vs 26/23/28/27 thresholds).
* test(zig): fail CI when the optional Zig grammar is absent
`@tree-sitter-grammars/tree-sitter-zig` is an optionalDependency, so every
Zig suite gates on `isLanguageAvailable(Zig)` and the ABI load-smoke accepts
a clean load failure for an optional grammar. Both are the right contract for
a platform with no prebuild, and together they leave a hole: if the grammar
never installed on any CI runner, this PR would merge with all eight Zig
resolver suites plus the structure-phase suite reported green-by-skip, having
never executed the native Zig parser once.
Close it with the `GITNEXUS_REQUIRE_FTS` idiom already used for the FTS
suites. `GITNEXUS_REQUIRE_ZIG=1` declares "this runner has a prebuild, the
grammar MUST be here", and a missing grammar becomes a failure instead of a
skip. tree-sitter-zig@1.1.2 publishes prebuilds for {darwin,linux,win32}-
{x64,arm64}, so the flag is set on two required jobs that all run on covered
platforms: the sharded ubuntu `tests` job and the three-OS `abi-assert` job.
- test/helpers/optional-grammar.ts: the registry mapping a language to its
require-variable, plus `describeGrammarPresence`, a presence assertion that
FAILS when required-but-absent. Deliberately a separate test rather than
flipping the suites from skip to fail: a skipped suite reports success, so
only a failing test can turn "Zig never ran" into a red job.
- parser-loader-abi.test.ts: the optional exemption is revoked for a language
the environment declares required, so an ABI-broken Zig binding fails the
smoke instead of passing as a clean absence.
- optional-grammar-gate.test.ts: pins the two ways the gate could silently
never fire — reading a variable name CI does not set, or accepting a value
CI does not write.
Nothing changes for a run that leaves the variable unset: local runs and any
future prebuild-less platform still skip. Verified both directions --
`GITNEXUS_REQUIRE_ZIG=1` alone: 149 passed, 0 skipped; with
`GITNEXUS_SKIP_OPTIONAL_GRAMMARS=1` forcing the grammar away it fails 2 tests
with an actionable message; with the skip flag but no require flag it is
green-by-skip exactly as before.
Addresses the test-only blocker in the gitnexus-check review of #1432.
* fix(zig): type the optional-grammar gate by grammar key, not language
The gitnexus-check finding on parser-loader-abi.test.ts:155 is right, and
none of the gates caught it: `tsconfig.json` includes only `src/**/*`, so
neither `tsc --noEmit` nor CI typechecks the test tree, and vitest strips
types without checking them. Confirmed with a scoped tsc run over the file:
`error TS2345: Argument of type 'string' is not assignable to parameter of
type 'SupportedLanguages'`.
Not fixed with the proposed `key as SupportedLanguages` cast, which would
assert something false: `listGrammarSources()` yields one row per SOURCES
entry, including variants like `typescript:tsx` that are not enum members.
`isOptionalGrammarRequired` now takes the grammar KEY it is really given,
and the registry keeps a `satisfies Partial<Record<SupportedLanguages,
string>>` so every key we write is still pinned to a real language.
Two new cases cover the failure mode the type error was pointing at — a
registry key that can never match what the ABI smoke passes, leaving the
gate configured-looking and permanently inert: every OPTIONAL_GRAMMAR_ENV
key must be a key `listGrammarSources()` yields and must be marked optional
there, and an unregistered variant (`typescript:tsx`) must not be required
even with the variable set.
Same blind spot, two more latent errors in files this PR adds, both fixed:
`Parser.Language` is not an exported member (use the `setLanguage` parameter
type, as parser-loader-abi.test.ts already does), and the `ParsedImport`
filter did not narrow the union, so `localName` was read through a `!` on an
arm that has no such property — now a type predicate. The one remaining
error under the same probe, in `resolvers/callable-value-flow.test.ts:319`,
predates this branch (authored 2026-07-17, on main) and is left alone.
structural-pair-coverage's optional-grammar case switches from
`it.concurrent.each` to `it.concurrent.for`: only `for` passes the test
context as a second argument (`each`'s callback is `(...args: T[])`), and
that context carries the dynamic `skip()` the per-language gate calls.
Behaviour is unchanged — grammar present: 10 passed; grammar forced away:
9 passed, 1 skipped.
Whole test tree typechecking is a separate, much larger job: the same probe
over `test/**` minus fixtures reports 734 pre-existing errors across the
repo. Out of scope here.
* fix(zig): address tenth gitnexus-check review pass
- `normalizeZigDepPath`: normalize backslashes BEFORE the absolute-path
check. A UNC dep (`\\server\share\dep`) used to slip past the check and
normalize to the repo-relative `server/share/dep`; root-relative `\dep`
had the same hole. Both now return null. Regression case added to the
absolute-spellings test with the files those misreadings would resolve.
- `bindPayloads`: a pointer capture `for (pages) |*p|` now records `*Page`
(declaredSpelling) instead of `Page` — the `*` is an anonymous payload
child before the identifier. Method dispatch is unchanged (`rawName`
strips the sigil), but a deref projection `const q = p.*;` now sees the
pointer layer. New fixture fn `viaPtrCaptureDeref` + assertion; fails on
the previous code (verified by stashing the src fix).
- `optional-grammar-gate.test.ts`: renamed the `typescript:tsx` case — the
key IS a registry row; what makes it inert is the missing gate entry. Now
also asserts a key no registry yields.
- `structural-pair-coverage.test.ts`: header updated — ten tables (not
eleven) are absent from every rule's target side; `Union` left the set
when Zig made it linkable.
- `language-classification.ts`: doc comment now names zig in the
experimental set (added after Ring 1).
Not re-fixed (invalid findings):
- "owner-hook contract wired to an undeclared variable": stale-diff read —
`findEnclosingClassInfo` declares `resolveFileTypeOwner` /
`resolveContainerTypeOwner` as optional parameters (ast-helpers.ts:905,
917) and parse-worker threads them at every call site; tsc compiles clean.
- "optional Zig grammar added unconditionally to the parsing fixture
suite": the cited block only `fs.readFile`s the committed fixture file to
assert it is non-empty — no parser or grammar load is involved.
* fix(scope-resolution): mark construction-site CALLS edges in reason (opt-in), enable for Zig
PR #1432 human review, item 2: a Zig struct literal `T{ .f = x }` (no
parens) is modelled as a CALLS edge to the type — the Rust `T { .. }` /
Go `T{}` shape — and nothing on the edge told it apart from an invocation
(`get_next_spawn → SpawnRequest` from seven `return SpawnRequest{ … }`).
`ScopeResolver.markConstructionSites` (default off): when set, the edge
emitted for a `callForm === 'constructor'` site gets ` (constructor)`
appended to its reason, in both emit paths — `local-call (constructor)` /
`import-resolved (constructor)` in the free-call fallback and
`scope-resolution: call (constructor)` in the reference bridge. The Zig
resolver opts in. `Reference` gains an optional `callForm`, copied from
the site by `buildReference`, so the bridge can see the form.
Why `reason` and not a property or edge type: relationships carry no
arbitrary properties, a new column changes the relation DDL and moves
SCHEMA_FINGERPRINT, and `reason` is the channel the IMPLEMENTS `-pointer`
receiver form already uses. Why opt-in: the unsuffixed strings are a
pinned contract asserted verbatim by the other language suites
(php/cpp constructor calls expect exactly `import-resolved`); every
non-Zig edge stays byte-identical.
Tests: `references-to-edges-call-form.test.ts` pins both vocabularies
and the default-off behaviour; `zig.test.ts` asserts
`Reflect.string → Accessor` / `Reflect.url → Accessor` carry
`local-call (constructor)` next to a plain invocation, and that every
marked edge targets a Struct.
* feat(zig): track qualified struct literals (`mod.T{…}`) as construction sites
PR #1432 re-test (issue comment on
|
||
|
Gergo Magyar
|
3c2648b3c9 |
fix(eval): trim oversized proposer evidence instead of aborting
Selected rows can exceed the 2MiB sandbox bundle even when each file is capped; shrink the seed and stage path so a fat prior artifact no longer kills the evolution job. Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergo Magyar
|
b580c96501 |
fix(ci): wait for host OOM guard before evolution preflight
The runner stamps job processes at 500 faster than the host rewrite; a single read failed a live OpenAI dispatch before any model work. Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergo Magyar
|
37415cb1ca |
feat(eval): route skill evolution through OpenAI
Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergo Magyar
|
45f97045e0 |
refactor(eval): simplify sweep internals and needrestart check (#2785)
Reuse the incumbent skill digest instead of walking the tree twice, and keep the needrestart grep a literal match that actionlint accepts. Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergo Magyar
|
b7e621df4c |
fix(ci): gate evolution runs on runner readiness (#2785)
Prevent paid scheduled work until host survival protections and the proven three-worker rollout are explicitly in place. Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergo Magyar
|
999b7bbede |
fix(eval): close skill evolution review gaps (#2785)
Keep promotion decisions monotonic and evidence-bound while preserving paid sweep results, redacting live failures, and hardening prior-run seeding. Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergő Magyar
|
9e905a3606
|
Merge branch 'main' into fix/skill-evolution-gate | ||
|
dependabot[bot]
|
cc066656df
|
chore(deps): bump the codeql-action group with 3 updates (#3151)
Bumps the codeql-action group with 3 updates: [github/codeql-action/init](https://github.com/github/codeql-action), [github/codeql-action/analyze](https://github.com/github/codeql-action) and [github/codeql-action/upload-sarif](https://github.com/github/codeql-action). Updates `github/codeql-action/init` from 4.37.7 to 4.37.9 - [Release notes](https://github.com/github/codeql-action/releases) - [Changelog](https://github.com/github/codeql-action/blob/main/CHANGELOG.md) - [Commits]( |
||
|
Abhinav Pandey
|
c3eb5991c1
|
fix(eval): preserve complete evolution evidence | ||
|
Abhinav Pandey
|
aeb853b9cb
|
fix(ci): harden evolution evidence reuse | ||
|
Abhinav Pandey
|
d232671278
|
Merge origin/main into fix/skill-evolution-gate | ||
|
Gergő Magyar
|
e1e2464960
|
feat(kotlin): bind Spring config consumers on Kotlin sources (#3126)
* feat(kotlin): bind Spring @Value and ConfigurationProperties consumers Kotlin sources were skipping the Java-only config-binding attach path, so mixed JVM apps under-reported blast radius for Kotlin placeholders. Capture from the live AST, serialize on the existing side channel, and reuse the shared binder. (U1-U4) Co-authored-by: Cursor <cursoragent@cursor.com> * fix(review): require exact Spring annotation FQNs and skip raw-string escapes Reject similarly named third-party imports and leave Kotlin triple-quoted bodies undecoded so capture stays fail-closed. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(kotlin): use only grammar-valid Kotlin string and class-name nodes The coverage shard failed the #1920 literal gate on invented string node types and a Java-style name field. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(kotlin): fail closed on non-literal prefixes and persist unresolved config markers Reject constant and boolean @ConfigurationProperties arguments, scope nested Value shadows to their owner, and rewrite drifted consumer files when a config key is deleted. Co-authored-by: Cursor <cursoragent@cursor.com> * test(kotlin): add config-consumer capture benchmark and fix JVM feature stamps The Kotlin capture path had no performance or behavior guard: a file-wide lexical shadow regression silently dropped two of every three facts. The new bench arm fingerprints @Value / @ConfigurationProperties facts from an explicit-import control against a wildcard-import corpus whose files each declare a sibling nested `Value` type, so parity between the arms is the regression gate, and CI runs it with scaling + widening budgets. Broadening spring.config-bindings to .kt also means Kotlin-only JVM repos now stamp the feature, which the two exact-map orchestration expectations still denied. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(kotlin): let an explicit import shadow the Spring wildcard import importedAs accepted a star import of the Spring package even when the same simple name was explicitly bound to another type, so a file importing com.example.Value alongside the Spring annotation package emitted a false @Value consumer fact. Kotlin resolves the explicit import first, so the wildcard branch now only applies when the name is otherwise unbound. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
ChunxueLi
|
72ff2b9453
|
feat(jvm): fold Java static wildcards and Kotlin star imports for route constants (#3110)
* feat(java): expand wildcard static imports + POSIX-space import resolution - `import static a.b.C.*` records the class FQN at extract time (ModuleConstants.wildcardImports) and expandJavaWildcardStaticImports materializes the bare-name bindings once the repo constants map exists, mirroring explicit single-member static imports. Wired in both the group-side prepareRepo fold and the ingestion-side parse-impl pass so the two query surfaces cannot diverge. Without this, wildcard-imported route constants (~693 routes on our monorepo) silently failed to fold and their provider contracts were dropped. - resolveJavaImport compares in POSIX space (backslash-normalized repo keys) — on Windows the '/'-joined class file never matched a backslash-keyed repo (observed: 675 calls, zero hits). - Wildcard bindings never overwrite single imports (a member shadowing its own wildcard is honored); unresolved wildcards degrade to the existing skip floor. Rebased onto current main: the parse-worker gate this originally carried is superseded by the provider moduleConstantHeuristic architecture; only the resolver-side wildcard expansion and POSIX tolerance remain. * chore(cache): claim SCHEMA_BUMP 83 (82 taken upstream by Spring lookup facts) * style: prettier * fix(java): make wildcard static imports actually resolve route constants extractJavaModuleConstants never populated wildcardImports, so `import static a.b.C.*;` was inert: the asterisk is a sibling of scoped_identifier in tree-sitter-java, not a path segment. Record the class FQN there and stop binding the class simple name as a field. Wildcard-only files also fell through every harvest gate (the java provider heuristic regex, the parse worker emit check, and the group prepareRepo filter), so the constants never reached either layer. Expansion now resolves targets against constant-defining files only, keeping ingestion and group in parity (#2980 R4). Adds unit coverage for extraction, expansion/shadowing, the unresolved skip floor, the harvest heuristic, Windows path keys, and group <-> ingestion parity. Co-authored-by: Cursor <cursoragent@cursor.com> * feat(kotlin): fold package-star imports for route constants Kotlin `import pkg.*` now records star scopes and resolves unique top-level names after local and explicit imports, matching the Java wildcard path without accepting invalid object-star imports. Reuse a Java constant-file suffix index across expansion so wildcard materialization stays linear as both constants and importers scale. Add named-vs-star benches and CI --check gates. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(jvm): fold wildcard imports without shadowing locals or skipping same-package names Keep Java unfoldable declarations from being resurrected by static wildcards, bind only the target type's members, and prefer Kotlin same-package names over package-star imports. Move repo-wide preparation behind a language-provider hook so the shared parse phase stays language-agnostic. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#3110) - Stop treating a Java static wildcard as a type import for qualified refs. - Harvest only class and static (including on-demand) imports, not import pkg.*. - Let harvested Kotlin top-level names shadow same-package star imports. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#3110) - Fold Kotlin classifier-star imports (`import Type.*`) the same way package stars already fold. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#3110) - Rebuild the Kotlin constant index when overlay replaces a contributing file. - Document the wildcard shape in the Java pipeline e2e fixture. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: l.cx <l.cx@winning.com.cn> Co-authored-by: ChunxueLi <mecoloud@users.noreply.gitee.com> Co-authored-by: Gergő Magyar <gergomagyar@icloud.com> Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
ChunxueLi
|
678a0e11c9
|
fix(server,web): honor the grep tool contract — real regex, fileFilter, caseSensitive (#3109)
* fix(server,web): honor grep tool contract — real regex, fileFilter, caseSensitive (Patch 12)
Background
==========
The web chat's grep tool schema has always promised regex search with an
optional path-substring fileFilter and caseSensitive control, but the
GET /api/grep handler escapeRegExp()'d every pattern into a literal
substring (a ReDoS hardening from
|
||
|
ChunxueLi
|
4aa6bddd0a
|
feat(jvm): synthesize Lombok and Kotlin JVM accessor methods (#2885)
* feat(java): synthesize Lombok @Data/@Getter/@Setter accessor methods * fix(lombok): resolve class identity by AST node id, not simple name Root-cause fix for the bot review's name-ambiguity findings: 1. Cross-file collision: the owner map was rebuilt per file from result.symbols, which accumulates across the whole language group — a later Java file with the same simple class name resolved to the earlier file's class node. The map is now filled INSIDE the capture loop (per-file scope) and keyed by the class_declaration AST node id (SyntaxNode.id), which is unique by construction. 2. Same-tail nested classes (Outer.A vs Other.A): a name-keyed map overwrote one with the other; AST-node-id keys cannot collide. 3. Synthesized method ids now follow the SAME convention real nested member ids use (keyed by the class's own simple name, matching findEnclosingClassInfo().className), so call resolution can hit synthesized accessors exactly like hand-written ones. 4. Lombok semantics: setters are no longer generated for final fields (Lombok never emits those) and @Setter(AccessLevel.NONE) now suppresses setters, symmetric to the existing getter suppression. Also tightens two vacuous test loops flagged by the bot (empty-array for..of passed trivially): counts are asserted before property loops, and a new regression test pins distinct owners for same-tailed nested classes plus the real id convention for nested accessors. * feat(java): synthesize Lombok accessors via provider hook and scope dual-path Replace the worker language===Java branch with LanguageProvider.synthesizeStructureMembers, align MethodRegistry ownership through scope captures, and bump parse-cache schema to 83 so warm caches cannot replay pre-synthesis worker output. Co-authored-by: Cursor <cursoragent@cursor.com> * test(java): cover Lombok synthesis semantics, cache replay, and CI bench Add unit/integration matrices (including durable cold/warm/historical parse-cache), a permanent no-Lombok vs Lombok-heavy harness with fingerprint budgets, and a CI --check step. Document that Kotlin→Java member CALLS remains a pre-existing gap. Co-authored-by: Cursor <cursoragent@cursor.com> * refactor(lombok): drop dead state and redundant scans from accessor synthesis Collapse Lombok import provenance into one compilation-unit scan with a cached wildcard flag, remove unused planned-accessor fields and the duplicate @Data enable flag, and plan scope captures without wrapping a fake Parser.Tree. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#2885) Give each Lombok accessor a unique scope range so multi-declarator fields do not share @scope.function IDs, and type the owner map as ReadonlyMap to match the provider hook. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(bench): pin the real Lombok synthesis fingerprint (#2885) The committed baseline held a fingerprint no revision of this branch ever produced, so the CI guard failed on every push. Re-pin it to the value the synthesizer deterministically emits and correct the method count the comment claims (800 x 4 x 2 = 6400, not 12800). Co-authored-by: Cursor <cursoragent@cursor.com> * feat(kotlin): synthesize JVM accessors using shared beanspec helpers (#2885) Kotlin val/var properties now emit the same JavaBeans get/set Methods as Lombok, via jvm/beanspec + jvm/synthetic-accessors. SCHEMA_BUMP 84 invalidates warm caches that would omit those callables. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(kotlin): match kotlinc JVM accessor ABI (#2885) Emit custom getters, preserve is-prefix names, and convert synthetic graph lines to 0-based so same-name accessors resolve to the owner. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#2885) Restrict Lombok provenance to lombok/experimental FQNs and match Kotlin existing methods by exact JVM name. Co-authored-by: Cursor <cursoragent@cursor.com> * refactor(jvm): consolidate accessor synthesis (#2885) Keep language-specific discovery in Java and Kotlin adapters while centralizing owner orchestration, collision policy, graph emission, and captures. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(jvm): align accessor synthesis with compiler ABI (#2885) Match Lombok and kotlinc provenance, companion owners, and collision arity so mixed-JVM CALLS bind to the Methods compilers actually emit. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#2885) Mark Kotlin interface accessors abstract, pin the Lombok case-fold collision test, and document the non-lowercase is-prefix rule. Co-authored-by: Cursor <cursoragent@cursor.com> * Address PR review feedback (#2885) Honor explicit Getter/Setter over @Data regardless of order, and let field @Accessors replace class-level fluent/chain. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(ci): pin Kotlin scope-capture fingerprint after interface accessors (#2885) Invalidate warm parse cache so interface property Methods are not replayed as concrete. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: ChunxueLi <mecoloud@users.noreply.gitee.com> Co-authored-by: Gergő Magyar <gergomagyar@icloud.com> Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
ChunxueLi
|
19f6731c34
|
feat(java): resolve SpringContextUtil.getBeans(X.class) dynamic lookups (#2886)
* feat(java): resolve SpringContextUtil.getBeans(X.class) dynamic lookups * fix(ingestion): make Spring dynamic lookups graph-correct Capture Java and Kotlin lookups from ASTs and resolve them through scoped type bindings and transitive JVM assignability so emitted INJECTS edges are attributable, cache-safe, and production-tested. Co-authored-by: Cursor <cursoragent@cursor.com> * perf(ingestion): keep Spring lookup capture linear Reuse Java and Kotlin scope-query call nodes instead of rewalking each AST, cache DI subtype closures, and enforce linear scaling with production-path benchmarks in CI. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Gergő Magyar <gergomagyar@icloud.com> Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Gergő Magyar
|
43a842724d
|
fix: bind Razor ViewComponent names to in-repo classes (#3104)
* fix: bind Razor ViewComponent names to in-repo classes
Index Component.InvokeAsync("Name") and in-repo ViewComponent("Name")
as CALLS to workspace ViewComponent classes so impact sees real callers
instead of an empty graph. SDK types stay unresolved.
Fixes #2991
Co-authored-by: Cursor <cursoragent@cursor.com>
* chore(autofix): apply prettier + eslint fixes via /autofix command
* fix: scan Razor and C# ViewComponent names without regex holes
Use string-aware lexers so combined Name= aliases, code-block calls,
this/base helpers, and escaped @@ markup match ASP.NET instead of
emitting false or missing CALLS.
Co-authored-by: Cursor <cursoragent@cursor.com>
* chore(autofix): apply prettier + eslint fixes via /autofix command
* perf: skip Razor scans without ViewComponent tokens
Preserve the lexer correctness fixes while avoiding per-character work for
the common view that cannot contain a supported invocation.
Co-authored-by: Cursor <cursoragent@cursor.com>
* test: gate Razor ViewComponent extractor scaling in CI
Wire mixed-corpus tripwire + GITNEXUS_BENCH loader/scaling checks into the dedicated ci-tests benchmarks job so the #2991 lexer cannot regress without a wall-clock gate.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix: read Razor views through one file handle
CodeQL js/file-system-race: the size gate stat'd the path and the read
re-resolved it, so a template swapped in between could be read past the
size ceiling. Both now go through the same handle.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
|
||
|
dependabot[bot]
|
31c9d9223e
|
chore(deps): bump the codeql-action group with 3 updates (#3056)
Bumps the codeql-action group with 3 updates: [github/codeql-action/init](https://github.com/github/codeql-action), [github/codeql-action/analyze](https://github.com/github/codeql-action) and [github/codeql-action/upload-sarif](https://github.com/github/codeql-action). Updates `github/codeql-action/init` from 4.37.6 to 4.37.7 - [Release notes](https://github.com/github/codeql-action/releases) - [Changelog](https://github.com/github/codeql-action/blob/main/CHANGELOG.md) - [Commits]( |
||
|
dependabot[bot]
|
414687ad10
|
chore(deps): bump docker/setup-buildx-action from 4.2.0 to 4.3.0 (#3057)
Bumps [docker/setup-buildx-action](https://github.com/docker/setup-buildx-action) from 4.2.0 to 4.3.0.
- [Release notes](https://github.com/docker/setup-buildx-action/releases)
- [Commits](
|
||
|
dependabot[bot]
|
c056d136ad
|
chore(deps): bump actions/attest-build-provenance from 4.1.1 to 4.2.2 (#3005)
Bumps [actions/attest-build-provenance](https://github.com/actions/attest-build-provenance) from 4.1.1 to 4.2.2.
- [Release notes](https://github.com/actions/attest-build-provenance/releases)
- [Changelog](https://github.com/actions/attest-build-provenance/blob/main/RELEASE.md)
- [Commits](
|
||
|
azizur100389
|
b77d6f662b
|
fix(kotlin): resolve imports from declared packages (#2990)
Some checks failed
CodeQL / Analyze (javascript-typescript) (push) Waiting to run
CodeQL / Analyze (python) (push) Waiting to run
Gitleaks / gitleaks (push) Waiting to run
Publish / RC guard (marker + release-PR skip) (push) Blocked by required conditions
Publish / Classify release event (push) Waiting to run
Publish / ci (push) Blocked by required conditions
Publish / Publish to npm (push) Blocked by required conditions
Publish / Build & Push RC Docker images (push) Blocked by required conditions
Scorecard / Scorecard analysis (push) Waiting to run
Trivy Image Scan / Trivy (gitnexus-cli) (push) Waiting to run
Trivy Image Scan / Trivy (gitnexus-web) (push) Waiting to run
Skill copy sync / shipped skills drift guard (push) Has been cancelled
|
||
|
azizur100389
|
7f0ab16ffe
|
feat(routes): support JS data route tables (#2972)
Some checks are pending
CodeQL / Analyze (javascript-typescript) (push) Waiting to run
CodeQL / Analyze (python) (push) Waiting to run
Gitleaks / gitleaks (push) Waiting to run
Publish / RC guard (marker + release-PR skip) (push) Blocked by required conditions
Publish / Classify release event (push) Waiting to run
Publish / ci (push) Blocked by required conditions
Publish / Publish to npm (push) Blocked by required conditions
Publish / Build & Push RC Docker images (push) Blocked by required conditions
Scorecard / Scorecard analysis (push) Waiting to run
Trivy Image Scan / Trivy (gitnexus-cli) (push) Waiting to run
Trivy Image Scan / Trivy (gitnexus-web) (push) Waiting to run
|
||
|
dependabot[bot]
|
02008e0288
|
chore(deps): bump the codeql-action group with 3 updates (#2947)
Bumps the codeql-action group with 3 updates: [github/codeql-action/init](https://github.com/github/codeql-action), [github/codeql-action/analyze](https://github.com/github/codeql-action) and [github/codeql-action/upload-sarif](https://github.com/github/codeql-action). Updates `github/codeql-action/init` from 4.37.3 to 4.37.6 - [Release notes](https://github.com/github/codeql-action/releases) - [Changelog](https://github.com/github/codeql-action/blob/main/CHANGELOG.md) - [Commits]( |
||
|
dependabot[bot]
|
8c2452a4e8
|
chore(deps): bump dorny/paths-filter from 4.0.2 to 4.0.3 (#2948)
Bumps [dorny/paths-filter](https://github.com/dorny/paths-filter) from 4.0.2 to 4.0.3.
- [Release notes](https://github.com/dorny/paths-filter/releases)
- [Changelog](https://github.com/dorny/paths-filter/blob/master/CHANGELOG.md)
- [Commits](
|
||
|
Gergő Magyar
|
18bc51dfd2
|
perf(import-resolvers): index every scanning resolver, consolidate the memo, gate every registered language (#2911)
* perf(import-resolvers): build buildSuffixIndex's dirMap lazily (#2903) `buildSuffixIndex` eagerly built three maps. `dirMap` is the array-valued one — one entry per directory suffix per file, so O(files x depth) in entries and array churn — and only four call sites ever read it, all via `getFilesInDir`: `import-resolvers/{php,csharp,jvm}.ts` and `import-resolvers/configs/python.ts`. Ruby (through workspace-file-index), the TypeScript scope resolver, Vue's import-target and the include-extractor never ask a directory question, and built it anyway. Since #2880 these indexes are retained for a whole resolution pass rather than rebuilt per import, so that waste is now resident memory. Deferring it to the first `getFilesInDir` call is behaviour-identical — same key, same descending-suffix order, same per-bucket push order, same `substring(lastIndexOf('.'))` extension clamp. The builder assigns the MAP on completion, so a repeated miss cannot rebuild it. Measured on `buildSuffixIndex` alone, 32k paths, index built and `getFilesInDir` never called: C# layout, 13 segments 79,018,680 -> 66,580,488 B -15.74% Ruby layout, 11 segments 60,752,792 -> 48,656,856 B -19.91% and on the whole retained WorkspaceFileIndex the bench measures: csharp 32k 73.62 -> 61.76 MiB ruby 32k 55.26 -> 43.69 MiB When `getFilesInDir` IS called the footprint is unchanged, so the deferral is never a loss. No new retention: all five construction sites already hold both input arrays alive beside the index. The laziness is pinned structurally rather than by timing. The test's corpus is a `string[]` whose elements are accessor properties, so an indexed read is observable and the read count IS the pass count: 14 after construction, still 14 after any number of get/getInsensitive, 28 after the first `getFilesInDir`, 28 after five more. Memoizing the decision instead of the map would read 42. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B * perf(php): resolve imports from a per-run index, not a scan per import (#2901) PHP was the last language whose import resolution scanned the workspace per import. Both `resolvePhpImportTarget` and `resolvePhpImportTargetInternal` materialized two full arrays from the Set on every call, then passed `undefined` as the `index` argument — so `resolvePhpImportInternal` fell through to `suffixResolve`'s linear `findIndex`, once per extension per path part. Measured at 20,000 files: 96.40 ms per import. **Handing it the shared SuffixIndex would have moved IMPORTS edges.** All three index-fed sites answer a different question than the scan they short-circuit, each found by differential with a concrete witness: 1. `getInsensitive` — the scan leg is `allFiles.has(path)`, exact whole-path with no case-insensitive counterpart; the shared index answers a ci SUFFIX probe. 2. `getFilesInDir` — the scan is root-anchored `startsWith(nsDir + '/')`; `dirMap` is keyed on every directory SUFFIX, so a vendor copy can win. 3. `suffixResolve` — the scan's `endsWith('/' + S)` matches only a PROPER suffix; `buildSuffixIndex` indexes j=0, so a root-level `Foo.php` starts resolving `use Foo` where it returned null. 3b. the scan's `endsWith(p) || lower.endsWith(lower(p))` has a second disjunct that subsumes the first, so it is purely first-in-Set-order and case-insensitive; `get(S) || getInsensitive(S)` lets a case-exact hit anywhere beat an earlier ci hit. So this is not Ruby's #2880 shape. Both sites take `getWorkspaceFileIndex` for the memoized arrays and hand the internal resolver a PARITY `SuffixIndex` memoized on the same Set identity: `getInsensitive` disabled, `get` implementing the scan's real rule via the shared ci lookup plus one O(files) whole-path correction map, `getFilesInDir` root-anchored in Set order. no composer.json 96.40 -> 0.036 ms/import steady state with composer.json 100.19 -> 0.068 ms/import steady state Also closes PHP's last per-import traversal, in `import-resolvers/php.ts`: its namespace-directory scan ran whenever `getFilesInDir` came back EMPTY, not merely when no index was supplied — despite the comment above it claiming "only when SuffixIndex unavailable". An empty bucket is already the answer, so the scan could only confirm it, at one full pass per import whose namespace matches a PSR-4 prefix but whose directory has no direct `.php` child (measured 11 traversals for 10 imports; now 1). Moving it into the `else` is safe because the bucket is a SUPERSET of what the scan finds — a root-anchored direct child `nsDir/<x>.php` has its directory exactly equal to `nsDir`, and a directory is always one of its own suffixes, so both index shapes contain it. Nine mutations of the new code are caught, including M1 "pass the raw shared index" (the naive fix) at 23 arms. The adapter guard reads 600 instead of 1 under a defensive `new Set(allFilePaths)` — the #1918 P1 hazard the unit differential is structurally blind to. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B * perf(java): index import resolution instead of scanning per import (#2908) Java scanned the whole workspace twice per import: once for the three-tier direct match, and again INSIDE the progressive prefix-stripping loop — so a single unresolvable import cost one full pass per stripped segment. No WeakMap, no index, and it is registered in `SCOPE_RESOLVERS`, so it ran in production. This is byte-for-byte the C# shape #2878 fixed, so Java now reads the same machinery: `getWorkspaceFileIndex` for `normToRaw` + the segment-suffix index, and a Java-owned `PackageDirIndex` WeakMap over `buildPackageDirIndex(_, n => n.endsWith('.java'))` read through `firstFileDirectlyInPkgDir`. Structure mirrors C#'s `narrowContext` / `resolveDirectMatch` / `resolveByProgressiveStripping`. 20k files, 256 imports, 7-in-8 unresolvable: 8.05 -> 0.62 ms/import steady state once the index is built: 0.0036 ms/import Tie-breaks preserved, and Java's are NOT identical to C#'s: - tier 1 `break`s on the exact match, so an exact whole-path hit wins even when a suffix or directory-child hit came earlier in iteration order — hence `normToRaw.get` before `index.get`, which conflates them; - the stripping loop instead returns at the FIRST hit of `f === tailFile || f.endsWith('/' + tailFile)` and only yields its directory child after the scan completes, so the conflated `index.get` is the correct lookup THERE. Applying tier 1's exact-wins rule inside the loop is a real behaviour change (mutation M6); - `.*` wildcard stripping stays ahead of everything; - `firstFileDirectlyInPkgDir` reproduces Java's at-root/at-nested predicate exactly, including the first-`indexOf` rule — proved algebraically rather than assumed: the `atRoot` branch matches iff `dir === pathLike`, which is `D.indexOf(P) === 0 === D.length - P.length`, and the `atNested` branch's first occurrence in `f` is the first occurrence in `D` shifted by one. Six mutations are caught; a seventh (swapping the two index builds) is a true equivalence and is recorded as such. Hand-derivation also corrected four cases where the legacy code resolves and I had predicted null — including `java.util.List` reaching a local `util/List.java`, because Java has no in-repo-namespace gate like C#'s #1881. That is preserved here and filed separately as #2910; the parity test pins it so the fix is visible. The adapter guard reads 800 instead of 2 under a defensive `new Set(allFilePaths)`. Two traversals is correct: the workspace index and the package-dir index are separate WeakMaps and each iterates the Set once, the same accounting as C#. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B * perf(cobol): index COPY resolution instead of two scans per statement (#2908) `cobolScopeResolver.resolveImportTarget` ran two full workspace scans per `COPY`, each calling `path.extname` + `path.basename` + `.toUpperCase()` on every entry: tier 1 over `.cpy`/`.copybook`, tier 2 over `.cbl`/`.cob`/ `.cobol`. No WeakMap, no index, and registered in `SCOPE_RESOLVERS`. Two uppercased-basename maps, one per tier, filled in a SINGLE pass over the Set and memoized on Set identity. Lookup is `copybooks.get(upper) ?? sources.get(upper) ?? null`. 20k files, 500 COPY operands: 3879-4082 -> 10.5-11.7 us/import (~350-369x) steady state once built: 0.253 us/import Tie-breaks preserved: - TIER ORDER. A `.cpy` match beats a `.cbl` match even when the source file appears EARLIER in Set-iteration order. This is the one a naive single-map rewrite silently breaks, so it gets its own fixture. - Within a tier, first in Set-iteration order wins (`if (!tier.has(...))`, mirroring the scans' first-match return). - The key is built with the identical call sequence, `basename(fp, extname(fp).toLowerCase()).toUpperCase()`, so `Foo.CPY` still keys under `FOO.CPY` rather than `FOO`. - `path` stays in the loop rather than hand-rolled `/`-slicing, so backslash handling is unchanged on every platform — pinned by a `dir\sub\BOOK.cpy` case. All six mutations are caught: collapsing the tiers, within-tier last-wins, dropping the target uppercase, dropping the extension lowercase, hand-rolled slicing, and the adapter's defensive copy. The first five are caught by the differential and are invisible to the adapter guard; the sixth is the reverse, which is the layering working as intended — the guard reads 600 instead of 1. `COBOL_SOURCE_EXTENSIONS` was being re-allocated on every call; hoisted to module scope beside `COPYBOOK_EXTENSIONS`. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B * perf(csharp): index the csproj leg's namespace-directory scan (#2902) #2878 moved C#'s no-csproj leg onto memoized indexes; the csproj leg kept a per-import full scan in `resolveCSharpImportInternal` step 3, measured at ~1.10 ms per import at 50,000 `.cs` files. **The fix the issue proposed would have moved edges.** It suggested skipping the fallback when an exhaustive index is available, on the assumption that step 2's `getFilesInDir` answers the same question. It does not: step 2's `dirMap` is keyed on segment-aligned directory suffixes, while step 3's `normalized.indexOf(dirPrefix + '/')` is an UNANCHORED substring match, so step 3 finds a strict superset — and it runs only when step 2 came back empty, so those extra hits are observable, not shadowed: dirPrefix 'ubModels' step 2 [] step 3 ['src/SubModels/Widget.cs'] dirPrefix 'rc/Models' step 2 [] step 3 src/Models/* AND vendor/mysrc/Models/* So the predicate is kept byte-for-byte and made fast instead. It depends only on the file's directory (the needle ends with `/`, so every occurrence lies wholly inside `D + '/'`), which reduces to the `package-dir-index` formula minus the anchoring leading slash. `PackageDirIndex` itself cannot be reused for the same reason — its matcher is anchored. The index is memoized on the `normalizedFileList` array identity and built lazily at the point step 3 is first reached, so BCL usings — which `continue` out at the root-namespace gate — never pay for it. Candidates come from an exact last-segment bucket when `dirPrefix` contains a slash, a last-segment key sweep when it does not, and `singleSegmentDirs` when it is empty. Positions rather than paths, merged and sorted when several directories match, so file-list order survives. App.Missing @ {App, src} 1103.0 -> 7.6 us (145x, and flat in file count: 7.3 @10k, 7.6 @50k, 8.4 @200k) App.Missing @ {App, ''} 626.7 -> 108.5 us App @ {App, ''} 1077.9 -> 2.0 us (539x) App.Ns8 @ {App, src} 0.6 -> 0.6 us (step-2 hit, untouched) `relative === ''` is preserved exactly, including the no-`projectDir` case where the needle is a bare `/` and the answer is "every `.cs` whose directory has no slash of its own" — `getFilesInDir('', '.cs')` cannot answer that over repo-relative paths, so it has its own arm. 13 of 14 mutations are caught, including M1, the naive skip-when-indexed cleanup, at 9 arms. The survivor drops the empty-prefix fast path and is a true equivalence. M9 initially survived and exposed a real corpus gap — no non-`.cs` file lived inside a directory — now covered. The remaining non-constant term is the slash-free sweep, O(distinct last segments): 456 us at 200k files on a unique-name layout, but 7.9 us on a `SrcN/Models` layout, which is how C# repos are actually laid out. Closing the unique-name case needs a character-suffix map over segments — the O(files x depth) memory shape `package-dir-index.ts` cites #2649 to avoid — so it is documented in the code as a design change rather than tuned here. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B * test(scope-resolution): assert index reuse for every registered language (#2909) Index reuse was asserted by nine hand-written per-language files, so the guarantee existed exactly for the languages someone remembered — and #2908 is the proof that is not good enough: Java and COBOL were registered, quadratic and unguarded until this branch. `resolveImportTarget` is a required member of `ScopeResolver` with one signature and 16 registrations, so "calling it N times against a stable `allFilePaths` must not traverse the set N times" is a property of the CONTRACT. `import-target-index-reuse.contract.test.ts` drives every entry of `SCOPE_RESOLVERS`, modelled on `construction-syntax-wiring.test.ts` — the established shape here for a property plus a justified inventory. Measured counts, all memoized: c 1 cobol 1 cpp 1 csharp 2 dart 1 go 1 java 2 javascript 2 kotlin 1 php 1 python 1 ruby 1 rust 0 swift 1 typescript 2 vue 2 **`KNOWN_UNINDEXED` is empty.** The audit that produced it also cleared C, C++, Rust, Swift, TypeScript, Vue and JavaScript by hand — Rust's memo lives in `qualified-call.ts::moduleIndexFor`, C's and Swift's loops are inside their WeakMap builders. The empty map stays as a mechanism: a 17th language cannot opt out silently, and the inventory arm fails when a registered resolver has no fixture. Two things the assertion had to get right: - it is `scans(200) === scans(2)`, not `scans === 1`. Per-language counts legitimately differ (C# and Java build two indexes), and comparing two counts needs no per-language expected value. - Rust legitimately scans ZERO times — it answers every leg with `allFilePaths.has(candidate)` probes — so the floor is a per-language `minimumScans`, 1 for fifteen languages and 0 for Rust with the reason on the interface. Paired with a `hitTarget` that must resolve non-null, so the property cannot pass vacuously on a resolver that stopped answering. Miss targets are distinct per import, which defeats the TS/JS/Vue per-target `resolveCache`. Also unifies the instrument. Kotlin and Python counted index BUILDS from production; the other seven count traversals of a `CountingSet`. The build counter is strictly weaker — a scan added BESIDE a reused index moves no build count, which is exactly the mutation `baselines.json` `_blind_spot` records as invisible to every timing arm — and it costs two production modules that ship in the bundle purely for tests, holding module-global state every test must `reset()`. Both guards migrate to `CountingSet`, and `languages/{kotlin,python}/index-stats.ts` plus both call sites are gone, for -59 lines of shipped source. (Mechanical note: the two `index-stats.ts` file deletions appear in the #2901 commit rather than this one. They were staged with `git rm` while a concurrent commit swept the index. The final tree is correct; only that attribution is off, and rewriting a sibling commit to move them was not worth the risk.) Coverage went up in the swap: Kotlin's old "rebuilds when the file set is a different object" arm (3 sets, 3 builds) would have PASSED under a defensive adapter copy. Its replacement fails, as do all six arms across the two files. Verified by mutation: `new Set(allFilePaths)` inserted into the kotlin, python and go adapters fails exactly those three and no others — `python: 200 imports cost 201 traversals, 2 cost 3`. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B * test(import-target): gate the four newly-indexed resolvers, retighten heap The bench covered go/csharp/dart/ruby/kotlin. The four resolvers indexed on this branch shipped unmeasured, and #2903's memory win was not locked in. **php, java and cobol join the shared corpus**, each with the two load-bearing properties the header requires: imports scale with file count, and most imports MISS so the full cascade runs (resolve rates php 36.0%, java 34.4%, cobol 36.0%). Java's miss families were measured rather than assumed, since it has no in-repo-namespace gate (#2910): `java.*` 1041 imports and `com.google.*` 1006, both resolving 0. COBOL's collide layout repeats a bookname across BOTH extension tiers, so it reaches the copybook-over-source tie-break rather than only the basename map. **`csharp_csproj` is a sixth LANGS entry**, not a new arm dimension — an entry needs five small additions and inherits all five arms and all seven gates, where a context axis would have to be threaded through `buildRepo`, `resolveAll`, `identityPass`, the report shape and every gate. `buildFiles` aliases it to `csharp`, so the two share one corpus by construction and cannot drift. Two configs (`{App, 'src'}`, `{Lib, ''}`) produce all three `dirPrefix` shapes — slashed, slash-free and empty — in five arms instead of ten: App.Ns{d} 30.6% src/Ns{d} step 2 hit App.Missing{n} 25.5% src/Missing{n} step 3, last-segment bucket Lib 14.0% (empty) step 3, singleSegmentDirs Lib.Missing{n} 12.0% Missing{n} step 3, KEY SWEEP — the one non-constant path BCL / Ghost 12.4% — root-namespace-gate control **2221 of 3200 imports reach the indexed leg**, only 12.4% `continue` out. What that arm pins is stated plainly rather than overclaimed: step 3 answers null for all 2221 here (the hits land at step 2), so it gates that leg's COST and its null answers; its positive tie-breaks stay pinned by the unit parity test. **Heap ceilings retightened.** #2903 dropped the measured figures, leaving the 1.5x ceilings at ~1.9x — a straight revert to the old size would have passed: csharp 116,000,000 -> 98,000,000 B (measured 61.76 MiB) ruby 87,000,000 -> 69,000,000 B (measured 43.69 MiB) php new 106,000,000 B (measured 67.29 MiB) java new 154,000,000 B (measured 97.32 MiB, the largest in the file — Maven layout is 18 segments) php and java are gated because both retained NOTHING across imports at BASE and now retain the O(files x depth) suffix index — the same argument that gates C#. cobol is not: two `Map<basename, path>`, O(files) with no depth term, and its retained delta does not clear measurement noise, so a ceiling would gate nothing. `csharp_csproj` is not: same corpus, same index, a duplicate number — its one distinguishing footprint, the lazily-built `dirMap` its `getFilesInDir` forces back, is measured at +20.8% and recorded as a residual instead, because gating it would licence eager-dirMap everywhere. csharp's `depth_ratio` also fell 3.318 -> 2.31 (the no-csproj leg never asks a directory question, so the deep arm stopped paying an eager dirMap build). Budget 5 -> 3.5, restoring the file's 1.5x convention — and `_arms_note` says plainly that 3.5 does NOT lock that win in, because locking it needs ~2.9, which is 1.25x over a 1.05x spread and the kind of tightening `_triage` warns buys flake rather than signal. All five pre-existing languages are byte-identical: 25 cells x 5 fields = 125 values, 0 mismatches. The new arms were proven live by a doctored baseline (cobol ceiling 0.01, php heap 1000 B, java resolved 999) producing three correctly-worded failures and exit 1. Wall-clock 10.9 -> 26.1 s, php and csharp_csproj ~11 s of it — both cascades end in `suffixResolve`'s ~50-extension probe, and both gate the two largest wins on this branch, so neither is a candidate to drop. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B * perf(javascript): build the suffix index JS resolution never had JavaScript's `PassCache` was TypeScript's minus one field: `index`. So JS called the shared `resolveTsTarget` with `ctx.index === undefined`, and `import-resolvers/standard.ts` fell through to `suffixResolve`'s linear `findIndex` — scanning the materialized path list once per extension (~39) per path part, per import. 2000 files 6448.9 -> 28.5 us/import (TypeScript: 25.0) 8000 files 25972.6 -> 27.4 us/import (TypeScript: 27.0) Per-import scaling over 4x the files: 4.12x -> 1.09x. **Every instrument on this branch was blind to it.** `CountingSet` counts traversals of the Set; this walked the array the adapter had already materialized — the blind spot `counting-file-set.ts` documents in its own header and `baselines.json` records under `_blind_spot`. Under mutation M1, which drops `index` and reproduces the shipped defect exactly, the sixteen- language contract test stays GREEN for javascript, because the pass cache is still reused and `files.scans` reads 2 either way. Two new arms do catch it: a `suffixResolve` linear-branch counter that runs the legacy adapter first as its control (135 entries legacy, 0 now), and a mock-free behavioural assertion that a repo-root module resolves by bare specifier. Adding an index moves output, exactly as it did for PHP in #2901, so it was characterized rather than assumed — 211,200 pairs (400 corpora x 3 importers x 176 targets) plus 184 hand cases. **Two classes move and there is no third:** A null -> repo-root file (108) `require('config')` with root `config.js`. The scan tests `endsWith('/' + suffix)`, so a path with no slash has no proper suffix and was unreachable through that leg — while `./config` from the root already resolved via the exact `Set.has` branch. JS was internally inconsistent. B file -> different file (5679) `import 'app/main'` was resolving to `node_modules/dep0/lib/main.js`; the scan skipped the whole-path candidate at the 2-segment suffix and fell through to the 1-segment `/main.js`, taking the first such file in Set order. C hit -> null ZERO, and impossible: proper-suffix keys are a subset of the index's keys. Both moved classes are JS being wrong. **JS-new agrees with TypeScript on all 211,200 pairs and every corpus case, 0 disagreements** — which is the intended design, since JS delegates to the TS resolver and differed only by this field. Also swaps the single-slot `let cached: PassCache | null` in JS, TS and Vue for a module-level `WeakMap`, matching every other language. Two alternating file sets rebuilt everything on every call: 12.0 -> 1438.2 ms at 4000 files x 400 imports (120x); after, 11.0 -> 15.7 ms. This is LATENT, not live — `pipeline/run.ts:673` builds one Set per provider pass and the three are separate providers — but it is why these were the only languages that could not carry the standard distinct-set guard. They can now: the arm fails on HEAD for all three (`expected 42 to be 2`) and passes after. Six mutations caught, including a global `resolveCache` (M5), which needed a new arm — `expectDistinctFileSetsGetOwnIndex` builds two IDENTICAL corpora, so a stale answer carried between them is also the right answer. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B * refactor(ingestion): one per-file-set memo primitive, twenty-one call sites Every language that indexes its import resolution hand-rolled the same memo: declare a module-level `WeakMap` keyed on the file-set object, `get`, `if undefined` build and `set`, return. One concept, written twenty-one times, and this branch had just added five more. `import-resolvers/per-file-set.ts` exports it once: perFileSet<K extends object, T extends object>(build: (key: K) => T): (key: K) => T Two decisions, both recorded in the file. `T extends object` rather than `has`-then-`get`: `WeakMap.get` returning `undefined` cannot distinguish "not built" from "built as undefined", and the `has` form needs a cast or a non-null assertion, both banned here — the constraint makes the ambiguous case unrepresentable instead, and a future caller wanting `string | null` gets a compile error pointing at the decision. A throwing build stores nothing and runs again next call, so failures are not memoized and a half-filled index is never published — inert for these pure builders, and the safer direction. `K extends object` rather than `ReadonlySet<string>` is what lets C#'s `readonly string[]`-keyed cache share the helper. Twenty-one sites migrated across `import-resolvers/` and fifteen languages. Every existing doc comment was re-homed onto the new call rather than deleted — several record real invariants (the Set-identity contract, the #1918 pass-through rule, why Rust's memo lives on a different hook). TypeScript, JavaScript and Vue additionally had byte-identical `PassCache` interfaces and builders. `import-resolvers/pass-cache.ts` now holds the one builder, taking a single argument — every difference the three have lives in the CONSUMER (`tsconfigPaths`, the extension list), not the builder. The builder is shared, the memo deliberately is not: each adapter keeps its own `perFileSet`, hence its own index and its own `resolveCache`, because the three disagree about what a specifier resolves to and one shared cache would hand a language another language's answers. It buys no runtime reuse and the module says so — each provider pass builds its own `allFilePaths` Set, so the three are always different keys. C and C++'s `augmentedFilePaths` was a two-LEVEL memo, and needed no new abstraction: the outer memo's value is a function and a function is an object, so `perFileSet(perFileSet(...))` composes. The two instances stay one per file, and the reason is now in BOTH doc comments rather than only C++'s — cpp delegates to `resolveCImportTarget`, whose `suffixIndex` is keyed on the augmented set, so a shared memo would cross the two languages' indexes. Two sites are deliberately NOT migrated, each with the reason written at the declaration so the next sweep does not re-litigate them: - `configs/swift.ts` is a two-input memo keyed on one. `targets` is not derivable from the key; re-keying on `ctx` would force a banned non-null assertion or an unreachable fallback inside a memo builder. - `rust/qualified-call.ts` `MODULE_SCOPE_CACHE` is three inputs keyed on one, and sits ten lines below a `perFileSet` in the same file — the likeliest thing to be "fixed" by mistake. The other ten remaining `WeakMap`s are different concerns and stay: AST-node caches, worker-pool runtime state, graph metadata, mutable lazily-filled accumulators, and the C++ ADL / inline-namespace indexes, which are reassigned by explicit clear functions and epoch-stamped on read — validity rules beyond key identity that a closure over a private cache cannot express. Net −20 lines of code, +22 of the two "why not" notes. The primitive's own doc is where the cost sits: the Set-identity contract and the two design decisions are written once instead of being twenty-one implicit facts. Pure refactor: 1764 unit tests, 42 guard tests, all sixteen contract-test traversal counts unchanged (c 1, cobol 1, cpp 1, csharp 2, dart 1, go 1, java 2, javascript 2, kotlin 1, php 1, python 1, ruby 1, rust 0, swift 1, typescript 2, vue 2), 647 C/C++ tests, and every bench fingerprint unchanged. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B * test(import-target): gate every registered language, not nine of sixteen The bench pinned output fingerprints and scaling for 9 of the 16 languages in `SCOPE_RESOLVERS`. The other seven — c, cpp, javascript, python, rust, swift, typescript, vue — resolve imports in production with nothing pinning their output or their cost. JavaScript was the sharpest case: the 25,972 us/import defect fixed earlier on this branch was gated by unit tests alone. All 16 are now gated, plus the `csharp_csproj` variant: 17 entries. **The nine existing languages are byte-identical** — 234 committed values (9 x 5 arms x 5 fields, plus 9 top-level fingerprints), 0 changed, and no pre-existing budget touched. Measured both before and after the memo consolidation in |
||
|
Carter LaSalle
|
81100e2c74
|
fix(python): resolve calls through __init__.py re-exports (#2864)
* fix(python): resolve calls through `__init__.py` re-exports
A call to a name imported from a package never resolved when the package's
`__init__.py` re-exported it rather than defining it:
pkg/impl.py def target_fn(x): ...
pkg/__init__.py from pkg.impl import target_fn
caller.py from pkg import target_fn
def calls_it(): return target_fn(21) # no CALLS edge
`caller.py` gets no CALLS edge. Both IMPORTS hops are recorded, and all four
functions are extracted as nodes — only the call binding is missing. Because
`__init__.py` re-exports are how Python packages declare a public surface, this
misses a large fraction of real call edges, and the failure is silent: the
defining file looks like dead code with zero callers.
The re-export closure that should carry this already exists and is fully general
(`buildReexportClosures` — SCC over the re-export subgraph, bounded fixpoint for
cycles, transitive `via` chains). Python just never fed it: the subgraph admits
only `kind: 'reexport'` and `kind: 'wildcard'`, and Python emits neither for
`from m import x`.
Python has no dedicated re-export form. A module-level `from pkg.impl import X`
binds X locally AND publishes it as `pkg.X`, so it is both a named import and a
re-export. Emitting `kind: 'reexport'` would be wrong — that form drops the local
binding, which Python's does create. Instead add an optional `reexportsName` flag
to the `named`/`alias` variants, alongside the existing provider-specific
`importedSymbolKind` / `targetIncludesImportedName` flags, and admit flagged
imports into the closure subgraph. Languages with an explicit form keep emitting
`kind: 'reexport'` and leave the flag unset, so nothing changes for them — a
negative-control test asserts a plain named import still does not resolve.
Verified on a fixture covering the three shapes (direct, top-level-via-re-export,
function-local-via-re-export): 1 of 3 CALLS edges resolved before, 3 of 3 after.
On a 12.4k-file Python/Go/TypeScript repository: edges 294,416 -> 301,443
(+7,027) and execution flows 300 -> 813. A previously "100% orphaned" module
(`shared/db/event_writer.py`) now correctly reports its caller.
5 new finalize tests (single hop, 3-hop chain, alias keying, cycle termination,
and the negative control) plus 6 updated Python fixture shapes.
`npx tsc --noEmit` clean in both packages; full unit suite shows no regression
against baseline (remaining failures are pre-existing load-sensitive flakes in
analyzer-identity / evidence-provenance-helper / skip-git-cli / hooks, each
verified passing in isolation).
* fix(python): set reexportsName only for module-level imports
`interpretPythonImport` flagged every `from m import x` as republishing the
name, but only a module-level statement does. A `from m import X` inside a
`def` or `class` body binds locally and puts nothing in the module namespace,
so flagging it fabricates a re-export of a name no importer can reach:
# pkg/__init__.py
def loader():
from pkg.impl import InternalHelper
# caller.py
from pkg import InternalHelper # CPython: ImportError
resolved to `def:pkg.impl.InternalHelper`. Worse, with declaration-order
first-wins in the closure, a scope-blind entry could claim a name ahead of the
real module-level import and give a WRONG def for legal, running code.
`interpretImport` receives a `CaptureMatch`, which is `{name, range, text}`
with no syntax node, so the scope is not recoverable there — and it is not
recoverable downstream either: `pass3CollectImports` applies no scope filter
and `ImportEdgeDraft.fromScope` is hardcoded to the module scope. The decision
therefore moves up to `import-decomposer.ts`, which still holds the live
`import_from_statement` node, and rides down as an `@import.publishes` marker.
Computed once per statement, not once per imported name, with the existing
`findAncestorBeforeBoundary` helper.
Only `function_definition` and `class_definition` suppress publication.
`if` / `try` / `for` / `with` do NOT — Python has no block scope — so the
predicate is an ancestor walk for those two node types and nothing else.
Verified against CPython 3.11 in both directions; both are now pinned by
tests, including the counterpart control that a branch-nested import still
republishes.
Also corrects the docblock in `scope-extractor.ts` that sent this change the
wrong way. It claims pass 3 attaches imports "not to any `Scope` — finalize
reconstructs the owning scope via `provider.importOwningScope` during Phase
2". Finalize does no such thing: `importOwningScope` is declared on
`LanguageProvider` and implemented by a dozen providers, and
`grep -rnE "\.importOwningScope\b" gitnexus/src/` returns exactly one hit —
that doc comment. Nothing invokes it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rzsb6mdGtbu66BG1EaF6Zz
* fix(shared): stop guessing ambiguous and namespace re-exports; bound the via chain
Four changes to the re-export closure, all reachable only now that Python
feeds it.
1. AMBIGUOUS NAMES ARE DROPPED, NOT GUESSED. `populateFileClosure` documented
"declaration order first-wins for duplicates of the same exported name",
which is sound only where a duplicate export is illegal — two
`export { X } from …` is a TypeScript compile error, so the rule never
fires. Python has no such guarantee:
from .v1 import Client # legacy, left behind
from .v2 import Client # the actual public Client
CPython binds v2 (verified on 3.11); first-wins attributed every
`from pkg import Client` in the repo to the DEAD implementation, and
`impact("Client")` pointed at the wrong file. Last-wins is not the fix
either: for the equally common `try:`/`except ImportError:` and
`if sys.version_info` pairs exactly one branch runs, and which one is not
decidable here. Both directions are wrong on real code, so the entry is
dropped — the importer stays unresolved, which is exactly the pre-#2864
answer, and the file-level IMPORTS edge is untouched.
`collectAmbiguousReexports` runs as a PRE-PASS over data phase 0 froze,
so the poisoned set is constant across the fixpoint. That matters: a set
that grew mid-fixpoint would need retraction to propagate to files that
already inherited the name, would make `myClosure.size > before` an
unsound progress signal, and would invalidate the `|SCC| + 1` cap. As a
pre-pass the closure map stays monotone and every existing termination
argument survives unchanged. Only two flagged drafts resolving to two
DIFFERENT in-workspace files count; duplicates of one target are
harmless, and unresolvable targets never entered the closure.
Checked in both loops. Named re-exports take precedence over wildcards,
so suppressing only the named loop would hand the name to a later
`import *` and reinstate an arbitrary winner through the back door.
2. NAMESPACE-RECLASSIFIED DRAFTS ARE EXCLUDED. The admission guards tested
`draft.source.kind` while `tryFinalize` tests the post-reclassification
`draft.base.kind`. Python's `from . import logger` is emitted as `named`,
reclassified to `namespace` by `isNamespaceImport`, and was still
admitted — republishing whatever def shared the module's simple name. For
a `logger.py` holding a module-level `logger = logging.getLogger(...)`,
importers of `from pkg import logger` bound to that Variable instead of
the module. Reproduced end to end. Both predicates now take the draft and
test `base.kind`; this is a no-op for TS/Rust, whose only
`isNamespaceImport` implementation is Python's.
3. `transitiveVia` IS CAPPED AT 32. Each hop copies the inherited path, so
an unbounded chain is Theta(depth^2) in time AND retained memory, and
Theta(|SCC|^2) for a cycle whose chain tracks it. `MAX_REEXPORT_DEPTH =
100` covered this until
|
||
|
DuduPhudu
|
fa31a7d824
|
fix: close the nine follow-up review findings from #2856 (routes, receiver typing, truncation honesty) (#2899)
* fix(typescript): a type parameter shadows a declared type of the same name (W2-8)
First item of wave 2, promised to the reviewer on #2856.
`export function unwrap<Result>(value: Result): Result` names the PARAMETER, not
the `interface Result` beside it — tsc resolves both annotations to the
parameter. The type-reference capture that makes a contract answerable ("what
breaks if I remove this field?") had no notion of a parameter binding, so every
annotation mentioning `Result` inside `unwrap` minted a `USES` edge into the
interface, at the same confidence as a real consumer and indistinguishable from
one. Measured on the new fixture: `unwrap` produced TWO false edges while the
genuine consumer produced one.
Blast radius is every generic whose parameter name collides with a declared
type, and the colliding names are ordinary choices for both: `Result`, `Key`,
`Value`, `Item`, `Node`, `Options`, `Config`, `Props`, `State`, `Response`.
TWO HALVES, and the first is why upstream's fix could not reach this. #2833
introduced `bindsTypeParameter` for the CALL-receiver path, where a workspace
`class T` was answering for `<T>`. Reusing it here changed nothing at first, and
the reason is its own documented contract: `@declaration.type-parameters` was
captured for class/interface declarations ONLY, so a generic FUNCTION recorded
no parameter list and the predicate correctly returned false — absence is not
evidence. The data was missing, not the logic. So:
- TYPESCRIPT_SCOPE_QUERY now captures type parameters on `function_declaration`,
`generator_function_declaration` and `type_alias_declaration`;
- the graph bridge consults `bindsTypeParameter` before emitting `USES`.
Both are load-bearing — removing either one fails the fixture.
The fixture carries two controls, because the obvious wrong fix is to stop
emitting: a genuine consumer of the interface must still link, and a generic
whose parameter does NOT collide must still link its real reference. Both are
asserted, and the "genuine consumer" case is asserted FIRST so the absences
below it cannot pass vacuously.
SCHEMA_BUMP 53 -> 54: parse-time capture change. A warm cache replays defs with
no parameter list, so the guard reads nothing and the feature is inert while
looking implemented.
Capture fingerprint re-baselined with justification. NO NEW CAPTURE NAME —
diffing the capture-name sets against the wave-1 branch returns empty; the tag
existed and now fires on more declarations. capture_groups_fp 2338 -> 2371,
fixture_count 151 -> 152, scaling 1.06 < 1.5, and JavaScript's fingerprint does
not move at all, which is the check that this is the TS declaration rules rather
than something broader.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(analyze): close the four false-success paths in the graph-write-collapse guard (W2-6)
Second wave-2 item, promised on #2856. All four were reported; all four
reproduced by reading the code they name.
(a) A SAME-COMMIT RE-RUN REPORTED SUCCESS FOREVER. Every other meta-driven
trigger — schema fingerprint, PDG mode, runner identity, CJK segmentation,
embedding dims — has a block that forces a rebuild before the
`alreadyUpToDate` fast path. `graphWriteCollapsed` had none; `grep -rn` found
writes and no reads. So the one state meaning "most of your edges are gone"
was the one state that repaired itself only if the user happened to pass
`--force`. Now forces a full rebuild, and forcing is right rather than merely
re-running: the persisted graph disagrees with what the pipeline produced, so
an incremental pass over unchanged files would write nothing and re-stamp the
same broken index as fresh.
(b) AN INCREMENTAL RE-RUN ERASED THE STAMP. `saveMeta` is a full atomic
overwrite, and the field was spread in only when the CURRENT run had a
verdict. `undefined` meant two different things at that site — "full run, no
collapse" (a positive all-clear) and "incremental write, not comparable" (no
opinion) — so the second case silently dropped `graph-write-collapsed` from
meta.json while the edges were still missing. Now three-way: stamp on
detection, CLEAR on a healthy full run, CARRY FORWARD when there is no
verdict. That is the shape `branch: branchLabel ?? existingMeta?.branch` two
lines away had all along.
(c) THE SERVER PATH NEVER CONSUMED IT. `analyze-worker-ipc.ts` projects the field
"so a server-side caller sees the same degraded outcome the CLI does" — but
nothing read it, so the comment described an intention and every collapsed
run reported `complete` to the UI and to every API consumer. Now reports
`failed` with the counts and the remedy, matching the CLI, which prints
`Repository indexed INCOMPLETELY` and exits non-zero. A consumer that reads
"complete" will query the index and get confident wrong answers.
(d) --pdg ROWS MASKED TOTAL STRUCTURAL LOSS. `expected` counts the in-memory
graph plus the streamed STRUCTURAL manifest; the streamed PDG layers never
enter `graph.relationshipCount`. But `persisted` was `stats.edges`, a count
of EVERY `CodeRelation` row, and PDG writes into that same table. With 1,000
structural edges expected and 4,000 PDG rows persisted, losing every
structural edge still read `persisted = 4000`, cleared the ratio, and stayed
silent — on exactly the large repos `--pdg` is used for.
Worth recording that the OBVIOUS fix does not work. Padding `expected` with
the PDG rows makes the two universes match but leaves the ratio judging a
minority population: 4,000 of 5,000 still clears 0.5. I wrote that first, and
the test I wrote to prove it failed. Only comparing structural against
structural asks the question the check exists to ask, so `getLbugStats` gains
a `structuralEdges` count excluding `PDG_EDGE_TYPES`. `TAINT_PATH` is
deliberately NOT in that set — it is a whole-program Function→Function edge
persisted by the normal emit, so it is structural and stays counted on both
sides.
`index-freshness-graph-collapse.test.ts` had pinned the masking as correct
(`detectGraphWriteCollapse(1000, 4000)` → undefined, "PDG layers write into
the same table, so persisted > expected is normal"). True about the table,
and it licensed the hole. Replaced with the case that matters and a note on
why the fix is at the caller.
The new `structuralEdges` assertion in `lbug-core-adapter` is there because the
failure mode is silent: the query sits in a try/catch that yields `undefined`,
and `undefined` makes the collapse check decline to compare — so a typo in the
Cypher would throw nothing, fail nothing, and switch the guard off. Verified
against a real LadybugDB and mutation-checked by breaking the query.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(processes): make process selection insertion-order invariant (W2-5)
Third wave-2 item. Reproduced before fixing: two equal three-step flows with
`maxProcesses: 1` select `handleAlpha`; inserting the identical nodes and CALLS
edges in reverse select `handleBeta`. Same repository, same commit, a different
persisted graph — so a filesystem that enumerates differently, or an incremental
run that reorders assembly, silently changes what the tool reports.
Four sorts ranked by score or length alone and returned 0 on a tie.
`Array.prototype.sort` is stable, so a 0 preserves INPUT order, which traces
back to `graph.iterNodes()`. Under `maxProcesses` capping that decided which
`Process` and `STEP_IN_PROCESS` nodes were persisted at all. Each now falls
through to a totally-ordered, content-derived key — node id for entry points,
the joined path for traces.
WHAT IS ACTUALLY VERIFIED, stated precisely because "four fixes" would overclaim:
- the ENTRY-POINT sort is individually mutation-verified;
- the two DEDUP sorts are collectively mutation-verified;
- the TRACE-RANK tiebreak is NOT individually observable, and the source says
so. The dedup sorts already impose a total order on the list that reaches
it, so removing it alone fails nothing. Kept as defence in depth: it cannot
misbehave — it only makes an already-deterministic order explicit — and it
is what stops a change to dedup ordering from silently re-opening this.
Finding that out took two fixtures. The first (three chains, three entry points)
is separated by the entry-point sort before trace ranking is reached, so it never
exercises the trace comparator at all; the second gives ONE entry point two
equal-length branches to different terminals, which is the only shape where the
trace comparator decides. Both are kept — they gate different sites.
The invariance tests assert the INVARIANT rather than any single sort, so they
cover all four sites and any future one without needing to know where they are.
Three assertions: same selection under a cap, identical set uncapped, and
identical ORDER — the last because order is what the cap consumes, so a set-only
assertion would pass while the defect persisted.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(impact): UNKNOWN dominates a mixed candidate set, instead of reporting the known floor (W2-4)
Fourth wave-2 item. The all-UNKNOWN branch here was reasoned about carefully and
is correct — its comment even names the two ways a set can be all-UNKNOWN. The
MIXED case fell straight through it.
`RISK_ORDER` is `['LOW','MEDIUM','HIGH','CRITICAL']` and has no `UNKNOWN` entry,
so `indexOf('UNKNOWN')` is -1 and an UNKNOWN candidate can never win the reduce.
An ambiguous name with one caller-less candidate (UNKNOWN, per the round-1 fix)
beside one single-caller candidate (LOW) reported `maxRisk: 'LOW'` — a confident
floor over a set containing an interpretation nobody measured. That is the same
false-safe the all-UNKNOWN branch exists to prevent, one case over, and it
surfaced in the UI as "Max blast radius N (LOW risk)".
`maxRisk` answers "how bad could this be?", and an unresolved candidate could be
CRITICAL — so any UNKNOWN in the set makes the aggregate UNKNOWN. Narrowing it
that way would normally cost information, so the measured part travels alongside
as `knownMaxRisk`, present only when the two differ: absent on a fully-resolved
set, where it would duplicate `maxRisk`, and absent on a fully-unknown one, where
there is no measured part. A reader gets "at least LOW among what resolved, and
one interpretation could not be walked at all", which is strictly more than
either value alone. The human-readable message says the same thing.
The seed gained a mixed pair because the existing one could not reach this: both
its twins are caller-less, so it only ever exercises the all-UNKNOWN branch —
which is precisely why the gap survived a round of review. Three assertions,
both halves mutation-verified.
`eval-server.ts` needs no change: it renders `result.maxRisk ?? 'UNKNOWN'`, so it
now shows UNKNOWN where it previously showed the floor.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(routes): track ternary polarity in dispatch guards, so a selected verb cannot be inverted (W2-9)
`if ((req.method === 'GET' ? false : true) && pathname === '/api/i')` emitted
`GET /api/i` — the one method that branch guarantees the request does NOT have.
A ternary SELECTS between its arms, so a verb inside one is not reached merely
because the whole condition is truthy, but `findVerbInSubtree` descended into
both arms and returned the first verb it saw. Same inversion `!` produced before
|
||
|
Gergő Magyar
|
78ecce1b92
|
perf(import-target): index the workspace once per run for go/csharp/dart/ruby (#2898)
* perf(import-target): index the workspace once per run for go/csharp/dart/ruby Four import-target resolvers answered their lookups with a full `allFilePaths` scan per import, making resolution O(imports x files): - go (#2877): `findRootPackageFiles` / `findAllFilesInPkgDir`, the latter once per path segment on the GOPATH fallback. Most Go imports are external, so the whole cascade ran to completion before returning null. - csharp (#2878): the no-csproj leg took the raw Set past the memoized index the csproj leg was already using - up to eight passes for a four-segment `using`. - dart (#2879): one scan per candidate path, and for an external package both candidates miss, so both always ran to completion. - ruby (#2880): a complete `buildSuffixIndex` rebuilt and discarded per `require` - every require paid to index every file in the repo. Each now reads an index memoized on the `allFilePaths` Set identity, the shape `getPythonFileIndex` (#1918) and csharp's own `getWorkspaceFileIndex` (#1881) already used. Two shared modules back them: - `workspace-file-index.ts`: normalized list + `SuffixIndex` + a normalized->raw map, for csharp and ruby. - `package-dir-index.ts`: "which files live directly inside a directory ending with <path>", for go and csharp. Candidates are bucketed by the directory's last segment rather than by indexing every directory suffix, which would cost O(files x depth) entries at kernel scale (#2649). Behaviour is unchanged, including the tie-breaks that are expressed only through Set-iteration order and `indexOf` positions: the go root leg stays sorted and its package leg stays unsorted, the first-occurrence rule that excludes a directory nested inside a same-named directory is preserved, csharp's whole-path match still beats an earlier suffix match, and dart still tries `lib/<rel>` fully before bare `<rel>` and matches raw paths. Verified two ways. `import-target-index-parity.test.ts` keeps verbatim copies of the pre-change implementations and diffs against them over a deterministic corpus plus hand-built layouts for each tie-break; six mutations of the new code were confirmed to fail it. Separately, the bench corpus produces byte-identical fingerprints against the pre-change resolvers at both 400 and 1600 files. `bench/import-target/measure.mjs` gates both arms in CI: per-language output fingerprints, a scaling budget (measured 0.98-1.12 here, 3.32-4.10 against the pre-change scans), and the corpus shape, so the corpus cannot be shrunk below the size the scaling arm needs and still print PASS. Closes #2877 Closes #2878 Closes #2879 Closes #2880 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BCptYZWRgnnJ821rzebQyj * perf(import-target): cover kotlin, and add depth + absolute-cost arms #2872 landed the same index hoist for Kotlin while this branch was open. Fold it into the shared measures so all five resolvers are gated on one corpus, and adopt the two arms that PR's review proved a scaling ratio alone cannot carry. - `bench/import-target/measure.mjs` gains a kotlin arm: a Gradle-shaped corpus with per-module source roots over one package namespace, `.kt` and `.kts` stems, a nested same-name package directory, and a share of wildcard `.*` imports so the package fan-out tier — the only tier whose output is order-bearing — is inside the fingerprint. - `depth_ratio`: deep arm at a FIXED file count with ~6x the path components. `scaling_ratio` divides the file count out, so it is scale-invariant and structurally cannot see a cost that grows with path depth instead, and `buildSuffixIndex` (C#, Ruby) and Kotlin's `suffixByStem` each emit one entry per component. Measured: go 0.98, dart 0.88 (depth-free indexes), ruby 1.48, kotlin 2.20, csharp 3.45 — which is why the budget is per language. One global budget would have to sit at 5.0 and would let Dart go 0.88 -> 4.9 unnoticed. - `small_ms_ceiling`: an absolute bound at 4x the measured arm, because a constant-factor regression that grows both scale arms equally passes every ratio. - The deep arm must resolve exactly what the small arm resolves. Padding was supposed to change depth and nothing else; a deep arm that stopped resolving would be timing the null path. The five fingerprints are unchanged by this commit - verified against the previous baseline before rewriting it, so adding the kotlin arm and the deep scale did not perturb the four languages' output. Kotlin joins the Set-iteration counter in `import-target-index-parity.test.ts` too. Its own guard (`kotlin-import-index-reuse.test.ts`) counts index BUILDS, which a scan added beside a reused index does not move. That counter is also the only DETERMINISTIC guard against a reintroduced scan, and this commit documents why rather than pretending otherwise: a full workspace scan on 1-in-32 imports was measured to pass every timing arm here (dart, 1.458 scaling against a 1.8 budget, 1.736 ms against a 4 ms ceiling) while the counter reads 14 instead of 1. Tightening the ceilings toward the noise floor to chase that case would only buy flaky CI. `bench/kotlin-import-target/` stays: it fingerprints both file-set iteration orders and probes the four-tier cascade shape by shape, neither of which this corpus does. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BCptYZWRgnnJ821rzebQyj * perf(import-target): merge matching package dirs in one pass `filesDirectlyInPkgDir` re-spread its accumulator once per matching directory, costing O(files x dirs^2) copies per import. On a Go monorepo where many services carry the same package directory (`svcN/internal/pkg`, which Go's GOPATH cascade queries by two-segment tail) that made the index SLOWER than the scan it replaced: 13.4x at 1600 matching directories. Append into one array and sort once. Measured against a verbatim copy of the pre-change scan, output byte-identical at every k: k=200 1400 files old 0.126 ms was 0.169 ms now 0.042 ms k=800 5600 files old 0.457 ms was 3.002 ms now 0.185 ms k=1600 11200 files old 0.960 ms was 12.890 ms now 0.232 ms The index now beats the scan by 2.5-4.1x on this shape instead of losing to it by up to 13x. Also drop the min-`ord` comparison in `firstFileDirectlyInPkgDir`: the build loop appends a directory to its last-segment bucket the moment it accepts that directory's first file, so bucket order already IS ascending first-file-`ord` order and the first hit is the minimum. Differentially verified at 0 divergences. The invariant, and the build-loop edits that would silently break it, are now recorded at the early return. Type the index containers as deeply readonly so Go's deliberate `[...rootFiles].sort()` copy is compile-enforced rather than comment-enforced, and correct the header's claim that a polyglot repo "never pays" -- only the stored index is per-language. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013rLy781K8E3rRHFGEe1khJ * test(import-target): guard index reuse at the adapter boundary `workspace-file-index.ts` documented the hazard as "a defensive `new Set(allFilePaths)` in an ADAPTER" -- the bug #1918 shipped -- and named the unit parity test as the guard. It is not: that test imports the resolvers directly, while production reaches them through `<lang>ScopeResolver.resolveImportTarget`. Inserting the copy at `go/scope-resolver.ts:31`, `csharp:35`, `dart:189` and `ruby:268` left the parity test 28/28 green and `measure.mjs --check` PASS in all four cases. Kotlin and Python already had adapter-level guards; go/csharp/dart/ruby had none. Add `test/integration/<lang>-import-index-reuse.test.ts` for the four, mirroring the Kotlin/Python precedent: resolve through the scope resolver, assert the file set is traversed once (twice for C#, which builds two indexes), and pair every count with a result assertion so a count of 1 cannot be the count of an adapter that resolves nothing. Each was proven to fail under the copy it exists to catch: go expected 600 to be 1 dart expected 600 to be 1 ruby expected 400 to be 1 csharp expected 600 to be 2 `CountingSet` moves to `test/helpers/counting-file-set.ts` and now counts `forEach`, `values`, `keys` and `entries` as well as `[Symbol.iterator]`. It missed a rescan spelled `allFilePaths.forEach(...)` entirely; with the overrides that mutation reads 14 instead of 1. Four fixtures that pinned the guard next door, each now shown to kill its mutation: - the Dart "matched RAW" case used a forward-slash target, so the basename bucket missed before the raw comparison was reached and it asserted `null === null`. A positive twin carrying the backslash in the TARGET catches both half-mutations. - no C# or Ruby target addressed the corpus's `win\dir\thing` file, so deleting the backslash normalization in `workspace-file-index.ts` passed both gates. Now 4 failures. - `normToRaw`'s first-wins rule had no normalization twin in any corpus. - the Go nested-package fixture was decided by the `endsWith` half and never reached the first-occurrence branch its title names; addressing the directory as a single segment makes it reach it. The parity test's own docstring no longer claims the scan count is a complete census -- it names the three materialized arrays it cannot see. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013rLy781K8E3rRHFGEe1khJ * test(import-target): assert every scale, add collide and retained-heap arms Three holes in the gate this PR ships as its own proof. 1. `--check` computed three fingerprints per language, stored all three, and compared one. `DEEP_PAD = 16 -> 0` deleted the entire depth arm and still printed PASS, because depth padding is count-neutral by design so no asserted number moved. Assert `fingerprint` per scale, and assert `deep.fingerprint !== small.fingerprint` so the padding's EFFECT is pinned, not just its output. 2. The corpus minted per-index directory names (`src/pkg${d}`, `src/Ns${d}`, `lib/feature${d}`), so max last-segment bucket and max matching dirs were both 1 -- and bucket cardinality is the only non-constant term the index has. The `dirCount > 1` merge branch had never executed in any arm. Add a `collide` arm on shared-leaf layouts with identical files/imports/resolved counts; it reaches 9,269 multi-directory merges per run, up to 34 directories at once. Go and C#/Dart legitimately score above the linear budget there and get their own; Ruby and Kotlin stay at 1.8 because their keyed maps are collision-immune and that immunity is the assertion. 3. No arm measured memory, while the C# no-csproj leg newly retains an O(files x depth) suffix index. Add a retained-heap arm on the `bench/cfg` pattern, including its loud failure when `--expose-gc` is missing rather than a silent skip. Measured at 32k files: csharp 73.62 MiB, ruby 55.26 MiB. Ceiling is 1.5x, NOT the 4x the timing arms use -- the measurement is byte-stable to 0.00085% across processes, so 4x would be throwing away the gate. `_arms_note` records why, so nobody harmonises it back. `depth_ratio`, added by this PR, flaked ~1-in-20: go peaked at 1.748 and dart at 2.043 against a 1.6 budget, both ratios of two sub-3 ms minima. Fixed at the estimator, not the threshold -- REPS 5 -> 15, matching `bench/cfg`, `schema-pairs` and `callable-value-flow` (5 was the lowest in the repo; the sibling `kotlin-import-target` uses 7, which was not enough here). 22/22 PASS, every arm now at 70-78% of its budget with a <=1.26x swing. No budget was widened; the distributions are recorded in `_arms_note` so the headroom is visibly earned. Three copies of the same overclaim corrected: the parity test NARROWS the 1-in-32 blind spot, it does not close it -- it watches the Set while the resolvers hold materialized arrays. `_floor` no longer claims its ratios "match" the issues' (different corpora, both quadratic). The step moves to the END of the benchmarks job and runs with `--expose-gc`. A failing step aborts every step after it (#2895), so the newest, least-proven gate must not sit ahead of eight established ones. All five output fingerprints are byte-identical to before this session -- the proof that every change here was behaviour-preserving. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013rLy781K8E3rRHFGEe1khJ * docs(import-target): point the reuse contract at the guard that guards it `workspace-file-index.ts` told callers the unit parity test guards the adapter-copy hazard. It does not -- it never crosses the adapter. Name both layers and say which catches what: the per-language `test/integration/<lang>-import-index-reuse.test.ts` files at the adapter boundary, the parity test for a rescan reintroduced inside a resolver. The C# namespace-dir index comment named `findDirectChild`, which this PR deleted; it feeds `firstFileDirectlyInPkgDir` now. Drop `GoResolveContext`, dead since the legacy call-resolution DAG was removed in #942 -- zero importers, and `gitnexus`'s package.json declares no `main`, `exports` or `types`, so it is not a published surface. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013rLy781K8E3rRHFGEe1khJ * refactor(import-target): quality pass over the review-round changes Four cleanup lanes (reuse / simplification / efficiency / altitude) over the previous four commits. No behaviour change anywhere: all 25 bench cells (5 languages x 5 arms) are byte-identical on files, imports, resolved, distinct_outcomes and fingerprint, re-verified after each individual edit. **Restores a fast path the last commit lost.** Fixing the O(k^2) accumulator made the SINGLE-directory case — the overwhelmingly common one — copy the bucket where the original aliased it: measured 1.11x slower at 4 files/dir rising to 1.72x at 128. Holding the first bucket by reference and promoting to an accumulator only when a second directory appears is 0.65-0.97x of the previous code at dirCount=1 and parity at dirCount=64. 176-case differential, 0 divergences. **`sortedRootFiles` accessor.** `rootFiles` was the only index container read directly from outside the module. `readonly` is erased at runtime and `Array.isArray` widens it back, so the copy rule now lives with the code that owns the invariant instead of at the call site. No `Object.freeze`: V8's PACKED_FROZEN_ELEMENTS read cost lands on the hot `matchingDirs` path. **One shared arm for the four reuse guards.** The distinct-file-set test was copy-pasted four ways, 33-38 identical lines each, and this repo's own helpers (`mini-repo.ts`, `scope-model.ts`) document extracting at the SECOND verbatim consumer. `expectDistinctFileSetsGetOwnIndex` takes what actually varies; its `expected` type excludes `null` so the pairing rule cannot be reinstated as a hole. The per-language first and third arms stay duplicated on purpose — corpora and payload shapes genuinely differ. Re-proven: all four still fail under an adapter-inserted `new Set(allFilePaths)`. **Bench.** `dirsFor` shared by the two functions that must agree on directory fan-out (they mint and address the same files). `SCALES` derived from the arm table, so a future arm cannot be measured, printed and silently never asserted. Five timing checks with one shape collapsed to a table — the trailing sentence had already drifted into four wordings. `uniqueTarget`/`collideTarget` as flat functions, mirroring the `uniqueDir`/`collideDir` split rather than nesting a second axis four ternaries deep. One `identityPass` replaces two untimed full resolution passes per cell: -371 ms median. **CI step moved back where it belongs.** It was parked last "until #2895 lands", but that reasoning was backwards twice over: the flake that motivated it was fixed at the estimator in the previous commit, and #2895's own audit measured the last slot as executing zero times in 13 runs. It sits with the other resolver-index guards; #2899 carries the `if: !cancelled()` that fixes step masking for every step at once. Filed rather than fixed here: #2908 (java and cobol still scan the workspace per import, same shape as #2877-#2880, neither memoized), #2909 (make index reuse a contract test over SCOPE_RESOLVERS on one instrument). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Co58j4au9JLf8dmJpwdF9B --------- Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Gergő Magyar
|
f7a1c28291
|
Merge branch 'main' into fix/skill-evolution-gate | ||
|
glier
|
c6b24162d9
|
perf(kotlin): index import resolution instead of scanning per import (#2872)
Some checks are pending
CodeQL / Analyze (javascript-typescript) (push) Waiting to run
CodeQL / Analyze (python) (push) Waiting to run
Gitleaks / gitleaks (push) Waiting to run
Publish / Classify release event (push) Waiting to run
Publish / RC guard (marker + release-PR skip) (push) Blocked by required conditions
Publish / ci (push) Blocked by required conditions
Publish / Publish to npm (push) Blocked by required conditions
Publish / Build & Push RC Docker images (push) Blocked by required conditions
Scorecard / Scorecard analysis (push) Waiting to run
Skill copy sync / shipped skills drift guard (push) Waiting to run
Trivy Image Scan / Trivy (gitnexus-cli) (push) Waiting to run
Trivy Image Scan / Trivy (gitnexus-web) (push) Waiting to run
* perf(kotlin): index import resolution instead of scanning per import
`resolveKotlinImportTarget` walked the entire workspace on every import.
Its four tiers — exact/suffix, directory child, package fan-out and
progressive prefix strip — each ran `for (const raw of allFilePaths)` with a
`replace(/\\/g, '/')` and several string scans per entry, and they are tried
in cascade, so one unresolved import cost two to four full passes.
Across a repository with tens of thousands of Kotlin files that is
O(imports x files): on the order of 10^10 string operations on a single
thread. It does not look like a hot loop from the outside - analyze sits at
exactly 1.00 core with a completely flat heap and emits nothing for hours,
because every allocation is a short-lived string and nothing accumulates to
hint at progress. Small repositories hide it entirely: at a few hundred files
each pass is free.
Three maps, built once per `allFilePaths` Set and memoized on its identity,
make each tier O(1): stem -> path for the exact tier, every component-suffix
of the stem for the suffix tier, and directory -> direct children for both the
fan-out and the first-child fallback. Cost becomes O(files) once plus O(1) per
import. This mirrors the existing Python index (`getPythonFileIndex`), down to
the WeakMap keying and the build counter.
Semantics are unchanged, including the parts the scans expressed only through
iteration order:
- an exact match anywhere beats a suffix match found earlier, because the
scan returned on the first exact hit but merely remembered the first
suffix hit;
- "first match" stays first in set-iteration order, so both stem maps keep
the earliest path inserted for a key;
- a directory-name match still honours the scan's `startsWith`-then-`indexOf`
rule, which only ever considered the FIRST occurrence of `/dir/`. A path
like `data/src/main/kotlin/com/example/data/Repo.kt` is therefore still
NOT a child of `data`. That is arguably wrong, but fixing it here would
silently move edges in every Kotlin repository; it belongs in its own
change with its own fixtures.
That claim is gated, not asserted. `bench/kotlin-import-target` fingerprints
every `fromFile | targetRaw -> result` triple over an exhaustive branch matrix
plus a deterministic fuzz, each file set resolved in BOTH iteration orders
because that is the only place the tie-breaks above are expressed. The
committed baseline is the value the PRE-INDEX implementation produces: both
implementations print
5ad605c179081505705ff7698a09dbdbdc4831080af6d9fdec5499cc6bce28ee over the same
20074 cases, 11612 of them non-null, and anyone can re-run it by pointing the
harness's module specifier at the old file.
Its second arm is the scaling ratio, `(t_large/t_small)/(1600/400)` over a
synthetic Kotlin monorepo whose imports are ~40% unresolvable — only a miss
drives all four tiers, which is where the scan was worst. The index measures
0.99 (8.0 ms / 31.7 ms); the implementation it replaces measures 3.737
(2207.8 ms / 33003.5 ms) on that same corpus, so the budget of 1.6 separates
them by a wide margin. Take the absolute times as an order of magnitude only
(~276x, ~1041x): the floor arm was run once cold because best-of-seven against
a quadratic implementation costs minutes, while the index arm is the usual
best-of-seven. The ratios are the comparable pair. Both arms run in the
existing always-on `benchmarks (GITNEXUS_BENCH)` job, next to the C++ guard
from #2788 and the Python one from #1918.
Two unit-level guards sit alongside it: a parity test pinning the curated
cases, and an integration test asserting the index is built once across many
imports — the adapter must pass the Set through, since a defensive copy would
hand a fresh WeakMap key per call and restore the old behaviour (the same trap
Python hit in PR #1918).
Two other providers have the same defect and are left alone here, having no
repository at hand to verify a change against:
- `go/import-target.ts`: `findRootPackageFiles` and `findAllFilesInPkgDir`
scan unmemoized, and the GOPATH fallback calls the latter once per path
segment but the last, so a single import can trigger several full passes;
- `dart/import-target.ts`: the `package:` branch scans once per candidate
path — `lib/<rel>` and bare `<rel>` — and `resolveRelative` scans again in
its suffix fallback, also unmemoized.
`csharp/import-target.ts` is a partial case worth noting: it already builds a
memoized `getWorkspaceFileIndex`, but that is reached only when a `.csproj` is
found; the no-csproj path hands the raw Set to `resolveDirectMatch` and
`resolveByProgressiveStripping`, which scan past it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* test(kotlin): close the blind axes in the import-resolution gate
Review of #2872 found the weak part was the gate, not the resolver: four
plausible follow-up mutations passed `--check` with a byte-identical
fingerprint, `cases` AND `non_null`. Each is now caught, and each was
re-checked against the mutation it exists to stop.
- The hashed record carried `order | fromFile | targetRaw | result` but not
the FILE SET, so a corpus edit that swapped the workspace under a case
while leaving its result string alone was invisible. Leaving the resolver
untouched and editing only the corpus, two documented load-bearing cases
could be gutted — the "exact beats an earlier suffix" case losing its
competing file, the repeated-directory negative case losing its file
entirely — with the gate green. The file set is now part of the record, and
that same edit now moves the fingerprint.
- The corpus capped path depth at 8 components and packages at 16 files,
which are precisely the two axes the loops this change added run on. It now
carries 11- and 13-component paths, queries against suffix keys deeper than
seven segments, a 40-file package, and a fuzz that spans both. Verified:
capping suffix-key depth at 7, skipping the `dirChildren` suffix loop above
depth 8, and capping a bucket at 17 entries each now move the fingerprint,
where all three previously passed.
- `non_null` was reported but never asserted; it is asserted beside `cases`.
That closes only the "resolves nothing at all" hole — it stayed 11612 under
all three code mutations above and under the corpus edit — so it is a
companion to the two fixes above, not a substitute for either.
- A ratio cannot see a constant factor, and a file-count ratio cannot see a
depth cost. `--check` now also asserts a DEPTH ratio (file count fixed,
paths 24 components against 8) and an absolute ceiling on the small arm: a
full workspace scan reintroduced on 1-in-32 imports scores 1.490, inside
the scaling budget, while running 2.8x slower.
The baseline is re-derived, not adjusted: the pre-index implementation and the
index both print
ebf1790bf1d42dad483a51f2cbdeb2351e493b9e8236e4eedeef592dd81e2c5c over the new
20106-case corpus, 13256 of them non-null.
Both test suites were shown to be non-load-bearing and now are:
- the parity test's repeated-directory case put `data` at the LEADING
segment, so the `startsWith` guard fired and the `indexOf` rule its own
comment describes was never reached — a resolver with that check relaxed to
`>= 0` passed all 18 cases. A mid-path case now pins it, and a backslash
fan-out case pins `norm.lastIndexOf` against `raw.lastIndexOf`, which was
also bench-only. Both mutations now fail the unit suite.
- the index-reuse test discarded all 200 return values, so a build count of 1
was equally true of an adapter that had stopped resolving anything. It now
asserts results, and its docstring premise is corrected: every one of its
imports hit the tier-1 suffix lookup and none reached the fan-out it
claimed to exercise. Half now genuinely do. The `undefined as never` casts
and the `?.` are gone — both trailing parameters are optional and the
member is required.
Resolver changes, all output-identical against the differential above:
- `dirChildren` buckets are frozen once built. `findKotlinPackageFiles` hands
a bucket straight out of the index, and the `readonly string[]` return type
does not survive the caller: the finalize pass normalizes with
`Array.isArray(t) ? t : [t]`, and `isArray`'s `arg is any[]` predicate
widens the true branch, so `tsc --strict` accepts a `.sort()` there. A
downstream sort would permanently reorder the cached bucket and flip the
first-child tier for every later import in the run.
- `stripped` is computed only after tier 1 misses, with `lastIndexOf`/`slice`
instead of `split`/`slice`/`join`. Measured -20% small arm, -21% large arm.
- `KOTLIN_EXTENSIONS` now comes from the existing `import-resolvers/jvm.ts`
export instead of a fourth inlined copy.
- A note on why the shared `buildSuffixIndex` is not reused, with the four
probes that diverge, and the measured basename-bucket comparison — the one
place this was less documented than the Python precedent it follows, and
the question the Go/Dart/C# follow-ups will each face.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Gergő Magyar <gergomagyar@icloud.com>
|
||
|
dependabot[bot]
|
29929b7488
|
chore(deps): bump docker/login-action from 4.4.0 to 4.6.0 (#2851)
Bumps [docker/login-action](https://github.com/docker/login-action) from 4.4.0 to 4.6.0.
- [Release notes](https://github.com/docker/login-action/releases)
- [Commits](
|
||
|
dependabot[bot]
|
911fdb1ae1
|
chore(deps): bump ossf/scorecard-action from 2.4.3 to 2.4.4 (#2852)
Bumps [ossf/scorecard-action](https://github.com/ossf/scorecard-action) from 2.4.3 to 2.4.4.
- [Release notes](https://github.com/ossf/scorecard-action/releases)
- [Changelog](https://github.com/ossf/scorecard-action/blob/main/RELEASE.md)
- [Commits](
|