3 commits

Author	SHA1	Message	Date
Gergő Magyar	2be508e796	fix(mcp): stop scaling the detect_changes query with the diff's hunk count (#2915) (#2930) ... * fix(mcp): map diff hunks to symbols without per-hunk OR conditions (#2915) `detect_changes` folded one `(n.startLine <= $hunkEndI AND n.endLine >= $hunkStartI)` pair per diff hunk into a single WHERE clause, one query per changed file. A machine-generated file (cache JSON, lockfile, golden fixture) diffs at thousands of hunks with `-U0`, and the expression tree that produces overflows LadybugDB's recursive evaluator copy on a TaskScheduler worker thread: a bare SIGBUS with no error output where secondary threads get 512 KB of stack (macOS), a swallowed 30s query timeout where they get more (Linux), which the CLI then printed as "No changes detected." with exit 0. Coalesce each file's hunks into sorted, disjoint ranges and run the overlap test in JS instead. Only ranges that overlap or abut are merged, so the union covers exactly the lines the raw hunks covered. Query text and parameters are now identical whether a file changed in 1 place or 100,000, and files are queried in batches of 100 rather than one full node scan each. Reproduced on Linux by running the engine with macOS-sized (512 KB) thread stacks: 2,500 hunks passed, 3,333 and 4,000 segfaulted — matching the reporter's macOS threshold table. After the change the same repo maps a 100,001-hunk diff in 2.1s with no crash. Also fixes a line-base mismatch the rewrite exposed: graph rows are 0-based (#2377) while git hunk lines are 1-based, so the raw comparison shifted every symbol one line up. An edit to a symbol's LAST line reported nothing changed — a one-line function whose body was edited was invisible to the pre-commit gate. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * fix(cli): say when a detect_changes result is partial (#2915) When a graph query fails, `detect_changes` swallows the error, sets `partial: true` and leaves the counts at zero (#2283). The CLI formatter never read that flag, so a degraded run printed "No changes detected." and exited 0 — the pre-commit safety gate reporting a clean bill of health for a check that did not complete. Print the partial note in both the empty and non-empty branches. Also restore the `Symbol` placeholder for rows whose label came back as an empty string: the changed-symbol mapping now keeps `''` instead of dropping it to undefined, so the formatter needs `||`, not `??`. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * refactor(mcp): bound the hunk→symbol query and simplify the overlap helpers (#2915) Cleanup pass over the #2915 fix. No change to which symbols detect_changes reports, except that a node matched by two changed paths is now reported once. * Push a per-file [lo, hi] span into the query. Coalesced ranges are sorted and disjoint, so a file's whole touched span is free, and the engine can drop the symbols outside it instead of shipping every row in the file across the native boundary. Measured on a 400-file batch against a 25k-node index: 546ms/13,870 rows before, 84ms/1,555 rows after, identical kept set. Depth stays constant (two comparisons per file, not per hunk), so #2915 cannot come back — the JS test still rejects symbols landing in the gaps between hunks. The struct-list parameter was verified against @ladybugdb/core 0.18.3 and 0.19.1. * Convert hunks into the graph's 0-based space once, at the point they are grouped, with the existing `toZeroBasedLine`. Every comparison downstream is then base-neutral, and `toDisplayLine` goes back to being what its doc says it is: an MCP response-boundary converter, not a filter input. * Deduplicate matched nodes by id. `ENDS WITH` is a plain string suffix, so a diff touching both `README.md` and `pkg/README.md` counted the same node twice (169 duplicates in 13,870 rows on a real 400-file diff). Pre-existing, free to fix now that the rows are shaped in one place. * Drop the positional `?? sym[N]` row fallbacks in this block. `executeParameterized` returns `getAll()` rows, which are alias-keyed objects, so the fallbacks were dead — and they coupled the mapping to RETURN column order, which is what made adding a column a renumbering exercise. * Build the path→hunks map in one pass, so "every value is coalesced" holds at every point rather than being repaired by a second loop. Simplify `coalesceHunks` (the length<2 branch and the sort tiebreaker changed nothing) and state `hunksOverlapRange` as a standard half-open lower bound. * Document `partial` in the detect_changes tool description. The CLI now prints it, but the MCP client — the main consumer of the pre-commit gate — was getting the flag as an undocumented raw key. * Tests: pin the query text as identical for a 1-hunk and a 3,000-hunk diff (replacing a magic length bound), pin the 0-based bounds parameter, pin the dedup, and fold two near-identical row mocks into one helper. Temp dirs now come from the shared pool helper, whose cleanup is per-directory and Windows-lock aware. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * feat(mcp): bound and batch the hunk→symbol query, and anchor its path match (#2915) Follow-up review pass on the #2915 fix, implementing every remaining finding. * Push a per-file `[lo, hi]` span into the query. Coalesced hunks are sorted and disjoint, so a file's touched span is free, and the engine drops the symbols outside it instead of shipping every row in the file across the native boundary. Measured on a 25k-node index, 400-file batch: 546ms/13,870 rows before, 84ms/1,555 after, identical kept set. Depth stays constant (two comparisons per file, not per hunk), so #2915 cannot return. The struct-list parameter was probed against @ladybugdb/core 0.18.3 and 0.19.1 first; the index-subscript form `$paths[i]` does not parse. * Anchor the path match: `n.filePath = b.path OR n.filePath ENDS WITH b.suffix` where suffix is the path with a leading separator. A bare `ENDS WITH` is a plain string suffix, so a diff touching `lib/a.py` also reported a symbol from an indexed `src/mylib/a.py` — a file the diff never touched. This is the form `explain` already uses. Pinned by an integration test against a real engine (it fails 3/3 with the un-anchored predicate). * Run batches a few at a time. `executeParameterized` checks a connection out of the 8-connection per-repo pool for the duration of a query, so parallel calls never share one — the same reason ~15 other queries in this file already run under `Promise.all`. `allSettled`, so one failed batch degrades the result to `partial` instead of discarding the batches that succeeded beside it. * Deduplicate matched nodes by id, and count `changed_files` as distinct paths: a path can appear twice in one diff (a rename reported alongside an edit). * Cap the listed symbols at 1,000 with `symbols_truncated: {listed, total}`. A repo-wide diff otherwise puts an unbounded array in one MCP payload — the CLI has `--limit`, an MCP client has nothing. Counts are never capped, so the risk level and the CLI's "... and N more" still see the true total. * Extract `chunk` / `mapBatches` / `LBUG_QUERY_BATCH_SIZE` into `core/lbug/query-batch.ts`. Every query built from a caller-sized array has this ceiling; the shape now has one name and the measured batch size is recorded where it is defined rather than in three constants under three names. * Move hunk grouping and the 0-based conversion into `coalesceHunksByPath`, at the parse boundary. `parseDiffHunks` stays faithful to git (1-based, like the `@@` headers it reads), consumers compare graph-native values, and the conversion is unit-testable instead of living in the backend. * Document `partial` and `symbols_truncated` in the detect_changes tool description — the MCP client is the main consumer of the pre-commit gate and was getting both as undocumented raw keys. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * refactor(core): batch every remaining repo-sized query list (#2915) `detect_changes` was not the only place building query text from a caller-sized array. `core/wiki/graph-queries.ts` interpolated the whole file list of a module into four `IN [...]` literals, growing the query with the repo — flat breadth rather than the nested depth that crashed #2915, but the same unbounded shape, and the one the repo's own `DELETE_FILES_CHUNK_SIZE` precedent already chunks elsewhere. All four now run one query per batch and merge in JS. The membership arms need care, and each is documented where it happens: * `getIntraModuleCallEdges` batches the caller arm only. A per-batch callee arm would drop a call from batch 0 to batch 2, both inside the module, so that predicate moves to JS against the whole set. Results are now sorted: the single-query form had no ORDER BY, and batch order would hand the entire 30-edge window `formatCallEdges` keeps to the first 100 files (#2787). * `getInterModuleCallEdges` keeps the SAME batch list in its `NOT` arm. That is sound — a file outside the module is outside every batch — and it preserves the null handling: `NOT null IN [...]` is null, so the original dropped edges to a node with no filePath, where a JS-only `!has(undefined)` would admit them. ORDER BY and LIMIT move to JS because a per-batch limit would cut rows before the cross-batch membership filter ran. * `getProcessesForFiles` keeps `LIMIT` inside the batch: `stepCount DESC, id` is a total order, so a process in the global top-N is in its own batch's top-N. Also adopt the shared `chunk()` at the hand-rolled slice loops in `lbug-adapter.ts`, `embeddings/http-client.ts` and `run-analyze.ts`. The loops whose index fed a progress callback or an error message use `chunk(...).entries()`, which removes the `i / SIZE` and `Math.floor(i / SIZE)` arithmetic rather than reproducing it. No batch size changed. One trap that survived tsc and is worth naming: after renaming a loop variable away from `chunk`, a leftover `chunk.length` silently resolved to the imported FUNCTION's arity, reporting `chunkSize: 1` for a 200-path batch. Only `lbug-query-importers-batch`'s exact-value assertion caught it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * refactor: name the line-base conversions and share the symbol line (#2915) The 0-based-graph vs 1-based-elsewhere rule was open-coded in five places with the reasoning living only in comments — the same rule that, applied by hand and skipped once, hid every last-line edit from `detect_changes`. * Add `toOneBasedLine` beside `toZeroBasedLine` in `ingestion/utils/line-base.ts` so the module owns both directions, and adopt it at the four CFG/PDG join sites in `pdg-impact.ts` and the two in `local-backend.ts`. This is NOT `line-display.ts`'s `toDisplayLine`, which is documented as a response boundary converter with an `undefined` passthrough; the joins need arithmetic, and the guards that produce `Number.NaN` for an absent line are kept verbatim. * `http-route-extractor.ts` probed graph spans with a bare `line - 1` and a 20-line comment. It calls `toZeroBasedLine` now; the `?? pick(line)` fallback arm is untouched, so which node is picked cannot change (the clamp differs only for a negative line, which no emitter can produce). * Extract `formatSymbolLine`: `detect-changes-format.ts` and `eval-server.ts` rendered the same `type name → filePath` line. One behavior note — the two were not byte-identical, and eval-server had no placeholder on `name`, so a definition with an empty name rendered the literal `undefined` and now renders `?`. Both `definitions[]` shapes set name from a graph row, so this is unreachable in practice, and printing `undefined` into LLM-facing output is the bug, not the intent. `||` (not `??`) in the placeholders is deliberate and documented: a node label can come back as an empty string and still needs the placeholder. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * perf(wiki): bind the module file list instead of splicing it into the query (#2915) The wiki's four `IN [...]` sites interpolated every file of a module into the query text, so the text grew with the repo — the shape that overflowed LadybugDB's recursive evaluator copy in `detect_changes`. The previous commit chunked them, which worked but cost real complexity: the callee arm had to leave Cypher and be re-implemented in JS, DISTINCT had to be re-established across batches, and ORDER BY/LIMIT had to move to JS so a per-batch window could not cut rows the cross-batch filter still needed. Binding the list as a parameter removes the reason for all of it. The text is constant at any list length, and measured against a real index a bound list is ~3x faster than the equivalent literal (5,000 items: 139ms vs 459ms; 20,000: 598ms vs 1,686ms). Every predicate goes back into Cypher, including the `NOT ... IN` arms whose null handling is load-bearing — `NOT null IN [...]` is null, so a callee with no filePath is dropped by the engine, where a JS membership test would have admitted it. Verified on this repo's own index: a 2,000-path bound list returns 14,856 rows in 877ms. Also collapses the per-process step query into one grouped `p.id IN $ids` fetch — 105ms to 13ms for 20 processes — and drops `fileListLiteral`, `callEdgeKey`, `compareProcessHeaders` and the batching loops with it. `compareStrings` was a byte-identical re-roll of `compareCodeUnits` (src/lib/utils.ts), including its #2787 rationale; it now calls the shared one. Intra-module edges are sorted where the original had no ORDER BY: `formatCallEdges` keeps only the first 30, and an unordered cut keeps a different subset per machine. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * refactor(core): one home for batching, and a backstop for the shape that crashed (#2915) `chunk` moves to `src/lib/utils.ts`, the repo's generic-utility home: it is an array helper, and leaving it in `core/lbug/query-batch.ts` made an HTTP embedding client import batching from the graph-DB namespace. `query-batch.ts` keeps what is actually about queries — the measured `LBUG_QUERY_BATCH_SIZE`, the concurrency helper, and the ceiling — and now documents the preference the wiki change proved: bind the list as a parameter first, chunk only when you cannot. `mapBatches` becomes `mapConcurrent`: nothing about it is batch-specific, and it now has non-query callers. Its body is a per-item try/catch plus `Promise.all`, so ordering comes from the primitive rather than from unwrapping a settled union. The wave barrier stays — measured against a rolling window it is 538ms vs 532ms on a 1,000-file diff, whose per-batch times spread only 1.35x. Adopted at the loops that were still hand-rolled: `file-hash.ts`, `cluster-enricher.ts` (its progress callback now accumulates `batch.length` instead of clamping an index), `filesystem-walker.ts` and `language-config.ts` (wave scheduling with `allSettled`, which is exactly `mapConcurrent`). Deliberately not adopted, each for a stated reason: the analyzer-identity probe runs as a standalone `node -e` script with no module resolution; the embedding sub-batch loop slices two parallel arrays and breaks early; `walkRepositoryPaths` reports progress from inside each wave, which `mapConcurrent` cannot express. `warnIfQueryTextUnbounded` is the backstop: #2915 died in native code with no message, and a query built by concatenating a caller-sized list is the shape that gets there. Wired at both execution chokepoints (`pool-adapter`'s `executeParameterized`, `lbug-adapter`'s `executePrepared`/`streamQuery`; their `executeQuery` siblings delegate and are covered once). It never throws — a long query the engine can actually run must not start failing on a heuristic — and it is deliberately absent from the raw write path, where a node's `content` is inlined and a large source file would warn legitimately. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * refactor(mcp): name the path-match rule, and key detect_changes by node id (#2915) * `path-predicate.ts` names the three ways a caller's path can match a stored `filePath` — `exact`, `pathSuffix`, `fragment` — instead of each call site copying whichever idiom its neighbour used. A bare `ENDS WITH` is a plain string suffix, which is how a diff touching `lib/a.ts` came to report a symbol from `src/mylib/a.ts`; the loose `CONTAINS` sites are loose ON PURPOSE (a user hint of `src/mcp` should match a directory fragment), and naming the modes is what lets a call site choose rather than inherit. * `detectChanges` kept four structures over one row set — an array, a dedup Set, an id list and an id→name Map — that had to stay in sync by hand. One id-keyed Map is all of them; insertion order is preserved, so every output is byte-identical. * `symbols_truncated: {listed, total}` becomes `truncated: true`, the key `explain`/`pdg_query`/`trace` already use. The true total was always in `summary.changed_count`, so the nested object said nothing the existing vocabulary could not. * `GraphLineRange` is now a distinct type from `DiffHunk`: they carry the same two fields in different bases, and mixing them IS #2377. The name means a 1-based hunk cannot reach `hunksOverlapRange` without a conversion between. * `coalesceHunksByPath` accumulates raw ranges and coalesces once per path rather than re-sorting on every occurrence. * `chunk` adopted at this file's own five loops — the point of extracting it — including two locals named `chunk` that shadowed the import. That shadowing is not cosmetic: it is how a leftover `chunk.length` silently became the function's arity earlier in this branch. One bug caught by the real-engine integration test and worth naming: Cypher comments are `//`, not `--`. A `--` comment inside the query string made LadybugDB reject the whole query at PREPARE, which `detect_changes` swallows into `partial` and renders as "No changes detected." Every mocked unit test passed. Prose stays out of query strings. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * refactor(test): share the git-repo bootstrap, and move the shared formatter out (#2915) `formatSymbolLine` lived in `detect-changes-format.ts` but is rendered by `eval-server`'s query formatter too, so a `query` formatter imported from a `detect_changes` module. It moves to `src/cli/format-symbol.ts`; both callers import it from there. The `||`-not-`??` fallbacks stay documented — a node label can come back as an empty string and still needs its placeholder. `test/helpers/temp-git-repo.ts` gives `initGitRepo(dir, identity?)` and `commitAll(dir, message)` to the ~10 test files that hand-rolled the same `git init -q` + two `git config` + `add -A` + `commit` sequence. It takes a directory and never owns one, matching `temp-dir-pool.ts`'s split of lifecycle from seeding; the identity is a parameter because the existing consumers genuinely disagree about it, and each keeps exactly what it configured. Four files stay hand-rolled for stated reasons — pinned author dates for a deterministic digest, remote handling, `--allow-empty`, and the `-c key=value` form that never persists to the repo. Test trims: the `formatSymbolLine` fallback cases collapse into one `it.each` table (the case pinning that BOTH consumers emit the helper's exact line stays — no table row can express it); two `line-base` cases that were compositions of their neighbours go; and `detect-changes-path-anchoring` runs its `detect_changes` call once in `beforeAll` instead of three times, keeping the three named failures. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019d85r7TrMYjWbTUT3pXccS * perf(mcp): filter the batched hunk query before the engine materialises (#2915) `UNWIND $bounds AS b MATCH (n) WHERE …b…` compiles to a CROSS_PRODUCT whose build side is a RESULT_COLLECTOR over the whole filtered node table: only the `n`-only predicates get pushed below the accumulate, so neither the anchored path match nor the [lo, hi] span could reduce the scan. Measured at 1M nodes: +242 MB for one batch and +922 MB for the four concurrent ones, paid even for a one-file diff — and at a 268 MB buffer pool the query died with `Buffer manager exception` where the old per-file query completed, landing in `partial:true` + `changed_count:0`, the #2915 false clean by another route. Adding the batch-wide, `b`-free disjunction as a redundant leading conjunct lets the planner push it below the accumulate: EXPLAIN now shows it as FILTER[2] directly under SCAN_NODE_TABLE[0]. It is a provable superset of the correlated predicate, so it cannot drop a row the correlated filter keeps. 10x less memory, ~20% faster, identical result sets. Also in detect_changes: - Sort rows on (filePath, startLine, id) before the 1000-symbol cut. The cut was slicing engine row order — measured 5 distinct orders across 8 runs on one connection, the #2787 class this branch fixes 200 lines away in the wiki. - Chunk `symIds`, the one caller-sized list left unbatched: 500k ids measured 4.0 GB RSS. Binding keeps the query TEXT constant, which is all the unbounded guard measures, while the bound VALUE stayed repo-sized. - Prefer exact path equality and widen to the anchored suffix only for paths that matched nothing, so a root README.md stops reporting pkg/*/README.md. - Report `risk_level:'unknown'` rather than 'low' when a query was swallowed. A degraded pre-commit gate must not read as an all-clear. - Pass --no-ext-diff --src-prefix=a/ --dst-prefix=b/. `diff.noprefix` in a user's gitconfig makes git emit `+++ f.py`, which parseDiffHunks cannot match, so every run printed "No changes detected." and exited 0 before any query ran. A diff that parses to zero files now raises `partial` instead of the clean branch. - `labels(n)`, not `labels(n)[0]`: labels() returns a scalar string here, so the subscript was always '' and `type` never carried a label. - Validate IMPACT_MAX_CHUNKS. The chunk() adoption turned an entry condition into an exit condition, so a non-numeric value ran every chunk instead of none. - Record why four-way concurrency is safe here, and scope the arm64 sequential comment to the query it was written for (#496). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv * fix(cli): fail the detect_changes gate instead of exiting 0 when it degrades (#2915) The secondary half of #2915 was that a swallowed query failure printed "No changes detected." and exited 0, so a shell pre-commit gate passed on a broken analysis. This branch added the PARTIAL text. It did not change the exit status, so `gitnexus detect-changes && git commit` still proceeded. `detectChangesCommand` passed a STRING to `output()`, and `output()` sets a failing code only for an OBJECT carrying `error` — under a comment calling itself "the one place that keeps scripted callers honest". A string never matches, so this command opted itself out of the only mechanism the file provides. It was broader than `partial`: the formatter also renders a backend `{error}` payload as text, so hard failures exited 0 too. Fixed narrowly in `detectChangesCommand`, following the object-first shape `checkCommand` already uses, rather than widening `output()`'s shared contract — every one of its other seven callers already passes an object and is unaffected. One code for both `error` and `partial`: `&&` only distinguishes zero from non-zero, and a softer code for `partial` would invite `|| [ $? -eq 2 ]` exemptions that reopen exactly this hole. `truncated` deliberately stays exit 0 — only the listing is capped, while the counts and risk are computed over the full set, so the verdict is sound and failing on it would fire on every large-but-healthy diff. Also wires `truncated` through the formatter, which this branch had left as a producer-only flag while `partial` went end to end, with the note in both locales and no count of its own so the existing "... and N more" line stays the sole numeric report. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv * fix(wiki): restore step order and symbol labels, and cut the edge list in Cypher (#2915) Found by running the queries against a real engine, which nothing did before: this branch's regrouped `withSteps` returned step traces OUT OF ORDER. `ORDER BY pid, r.step` combined with `WHERE p.id IN $ids` silently drops the second sort key — `proc_1_incrementalupdate` came back 2,7,1,3,4,5,6. `ORDER BY step` alone is correct, and so was the pre-branch per-process query, so this was introduced by the batching. `formatProcesses` prints "${s.step}. ${s.name}", so every module and overview page was getting scrambled execution traces. The mocked suite passed 112/112 before and after. `labels(x)[0]` is always the empty string: labels() returns a scalar string and the subscript is 1-based over its characters ([1] is "F"). `prompts.ts` renders "${s.name} (${s.type})", so all 5,027 exported symbols reached the LLM as "name ()". `getIntraModuleCallEdges` shipped every edge to use 30 — measured 18,299 rows and 851 ms with all 2,079 paths bound, against 30 rows and 94 ms with ORDER BY + LIMIT in Cypher, which the sibling `getInterModuleCallEdges` twenty lines below already did. The determinism fix (#2787) was right; the placement was not. `compareCallEdges` goes with it — it was intransitive when a name was null or empty, so `Array.sort` was input-permutation dependent, i.e. the nondeterminism it was added to remove. Deletes the positional row ABI this branch newly documented. The vendor declaration is `getAll(): Promise<Record<string, LbugValue>[]>` — string keys only — and `row[0]` probes back `undefined`; the same PR deleted ~30 identical fallbacks from local-backend.ts. They were already stale here: `withSteps` prepends `p.id AS pid`, so `toProcessStep` was reading the pre-branch layout. Rows are now typed by alias, so renaming an `AS` is a compile error. `??` for `||` so a step of 0 or an empty label keeps its own value. Tests: a real-engine integration suite covering all seven exported queries (PREPARE included — the trap that shipped a `--` comment on this branch), and the four holes that let the ordering bug through — a vacuous order assertion, a LIMIT never reached by a 2-edge fixture, a fake that returned rows pre-ordered and ignored ORDER BY, and a hardcoded `type: 'Function'` that hid labels(). The step-ordering fixture is empirically sized: 2 processes never reproduced the bug, ~400 step edges was intermittent, 710 (20 processes x 26-45 steps) hit 11 of 11 runs. Seeded descending and interleaved so no grouping looks sorted by accident. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv * refactor: put the shared helpers where their callers are, and make their contracts true (#2915) `mapConcurrent` moves to lib/utils.ts beside chunk(). Nothing about it is query-specific and it already had filesystem callers, while its docstring justified concurrency safety through the per-repo connection pool — an argument that does not apply to fs.readFile. This is the precondition the branch's own commit message stated ("it now has non-query callers") and then did not apply. LBUG_QUERY_BATCH_SIZE and warnIfQueryTextUnbounded genuinely are query-specific and stay. `pathMatch`/`PathMatchMode` deleted: zero callers, and none of the three sites its docstring cited were migrated, so the tree carried the abstraction and the copies it was written to replace. `pathSuffixOf` stays and the module now documents the anchoring rule it actually implements. Contracts that were not true: - QUERY_TEXT_CEILING_BYTES was compared against `cypher.length` — UTF-16 code units, not bytes — so non-ASCII query text was undercounted and the reported KB was wrong. Buffer.byteLength now, behind a `length * 3 <= ceiling` early return so only text over ~21 KB pays for the count. - chunk(items, NaN) returned [[]], against a docstring promising never to return an empty slice, and mapConcurrent's Math.max(1, NaN) propagated it — which would have resolved [] for non-empty input with no error, read as "no results" by every call site. - GraphLineRange claimed a 1-based hunk could not reach hunksOverlapRange without a conversion, but it was structurally identical to DiffHunk so tsc accepted one with no diagnostic, and coalesceHunks<T extends GraphLineRange> actively laundered the base while its accumulator was still DiffHunk[]. The useless generic is gone and a one-line phantom on each interface makes the claim real; a bare {startLine, endLine} literal still satisfies both, so no construction site needs a cast. Pure deletions no longer vanish. A -U0 deletion emits `+N,0`, which parseDiffHunks dropped, so the file survived with no hunks, no query ran, and detect_changes reported `changed_files:1, changed_count:0, risk_level:'low'` — "No changes detected." for a commit that deleted a function. A unified diff spells an empty range as the line before it, so the anchor is line N alone: a symbol containing the deleted text also contains N, while extending to N+1 would claim a symbol that merely starts after the gap — the widening coalesceHunks guarantees it never does. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv * docs: say that a partial or truncated detect_changes is not a clean gate (#2915) The gate itself now fails loudly, but the instructions every agent reads still described a zero as a result. Fixed at the source: AGENTS.md's gitnexus block is generated from a template in cli/ai-context.ts and injected into every user's repo, so the sentence goes there and AGENTS.md/CLAUDE.md are regenerated through the real code path (which also picks up a pre-existing `analyze --index-only` drift the committed docs were behind). That block is under a test-enforced size cap with 30 characters of headroom, so the 144-character clause was paid for in the same currency: the header exhortation, which the Always Do list restates as MUSTs with commands, and a verbatim repeat of the detect-changes command in the regression-compare example. 3549 of 3552. Worth noting for whoever adds the next line — #2899 replaced an absolute cap with a 0.65 ratio to let "a legitimate clause fit without ceremony", but set the ratio flush against the block's then-current size, so it is a ratchet with no ratchet. The canonical block does not make the skills redundant: three of the four install channels ship skills without touching AGENTS.md, --skip-agents-md does the same in-repo, and a user-trimmed gitnexus:keep block legitimately has no Always Do section — in those repos the skill file is the only carrier. Precedent agrees: the risk:UNKNOWN rule is deliberately carried in both places. So one sentence each in gitnexus-work (the commit gate), gitnexus-impact-analysis (beside the UNKNOWN paragraph) and gitnexus-refactoring, whose post-hoc "verify only expected files changed" is the worst of the three because a degraded result makes it vacuously pass. gitnexus-taint-analysis is left alone: its audience is always inside this repo, where the canonical block loads. All copies mirrored to npm, plugin and cursor. The cursor copies are condensed checklists rather than byte-mirrors, so they carry the equivalent note placed where it governs every detect_changes line in the file — and nothing tests that, since standard skills are fragment-checked rather than byte-compared. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv * refactor: break the seven small import cycles gitnexus check reports (#2915) `check` reported 11 cycles. Five are paths inside a single 257-file strongly connected component in core/ingestion (call-extractors / cfg visitors / utils/ast-helpers), with a second 26-file component behind it — fixing those paths would only make check print different ones, so both are left for their own PR. This closes the seven that are genuinely separable, taking the graph from 9 strongly connected components to 2. Six of the seven were one value import plus one `import type` edge. tsconfig sets neither verbatimModuleSyntax nor isolatedModules, so those edges erase entirely — the cut is a graph and readability change with no emitted-JS difference. Each moved type went to a leaf module, with a re-export left behind only where an importer outside the change actually needed it: - cli/ai-context <-> cli/skill-gen: GeneratedSkillInfo -> cli/generated-skill.ts. One importer, no package export surface, so a clean move with no re-export. - cli/analyze-config <-> cli/analyze (+core/run-analyze): AnalyzeOptions -> cli/analyze-options.ts. Re-export kept because a test imports it from analyze.js. run-analyze needed no edit — cutting the one type edge collapses the 3-file component into a DAG. Its own same-named AnalyzeOptions is a different interface and was deliberately not merged. - ingestion/import-resolvers/types <-> ingestion/language-config: type-only in BOTH directions, so it had no runtime existence at all. ImportConfigs has no importers outside the pair and is the return type of loadImportConfigs, so it moved into language-config. Side effect worth having: the shared resolver types module no longer names a single language, which is an AGENTS.md rule for core/ingestion shared pipeline code. - ingestion/di-extractors barrel <-> spring: DiResolver and the two match types -> di-extractors/types.ts, following the import-resolvers/types.ts precedent. - scope-resolution/walkers <-> workspace-index: WorkspaceResolutionIndex -> workspace-index-types.ts. Re-export is load-bearing — 9 src importers, 4 test files, and a dynamic import() at contract/scope-resolver.ts. Moving the value isClassLike instead was rejected: ~15 value importers, and it is documented as a pair with isShapeLike. - server/analyze-worker <-> analyze-worker-core: the WorkerMessage protocol -> analyze-worker-protocol.ts, a declarations-only leaf. storage/branch-index <-> storage/repo-manager was the one genuine two-way runtime cycle: branch-index called getStoragePaths/loadMeta, repo-manager used branchSlug/BRANCHES_DIR. branch-index's header conceded the cycle and argued it was ESM-safe because neither side calls across at module-evaluation time — a guarantee resting on call ordering rather than structure. Folding resolveBranchPlacement back the other way does not help, because BranchSummary.stats is typed RepoMeta['stats'], so RepoMeta had to move either way. Extracted storage/repo-meta.ts, a leaf importing only fs and path, holding the metadata read primitives; repo-manager re-exports the public names so all 54 RepoMeta and 50 loadMeta importers are untouched. The moved block diffs byte-identical against HEAD. Verified beyond typecheck, because the worker entrypoint is the risky part and nothing in the suite forks it: emitted analyze-worker.js still contains exactly one runtime import, and forking the real worker over IPC boots it through entry -> core -> protocol -> terminal-claim. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv * refactor: apply the reuse, simplification, efficiency and altitude cleanups (#2915) The one that mattered: the degradation exit code was fixed at the wrong depth. `output()` has never inspected `partial` — it tests `error` only — so putting the check in `detectChangesCommand` left every other tool exiting 0 on a degraded run. `partial` is cross-tool vocabulary: query (enrichmentDegraded || ftsPartial), impact (!traversalComplete, perSymbolEnrichmentCapped) and the mode:'pdg' envelope all emit it. A truncated impact traversal returns a short caller set and an under-ranked risk, then exits 0 — so `gitnexus impact … && <edit>` proceeds, in the tool AGENTS.md makes a MUST gate before every edit. The justification also cited checkCommand as precedent, but checkCommand passes STRINGS too — it was the second command already hand-rolling around this gap, while output()'s docstring called itself "the one place that keeps scripted callers honest". output() now takes an optional renderer and fails on error OR partial; two hand-rolled sites go away and three tools are covered instead of one. truncated stays exit 0 (only the listing is capped) and checkCommand's cycleCount policy stays put. Efficiency, all re-measured on the 25k-node index: - The process lookup was chunked with LBUG_QUERY_BATCH_SIZE, calibrated for the opposite query shape — that constant is for a whole-node-table scan where more items amortise the scan, while this is an `id IN $ids` probe where round trips dominate. 20k ids: 617ms at 100, 261ms at 1000. New LBUG_ID_PROBE_BATCH_SIZE, documented against its sibling so they cannot be re-merged. This also settles the older "chunking this query is a regression" measurement — that was chunk=100. - The sort comparator re-coerced fields ChangedSymbolRow already types, O(n log n) redundant conversions (+31-38%). Row shape probed directly: alias-keyed, no positional keys, numeric columns are JS numbers. - exactlyMatchedPaths built two throwaway arrays; one loop instead (40k rows 11.4ms -> 4.5ms). - The integration fixture seeded 710 step edges one round trip at a time; one UNWIND instead. File wall time 6.91s -> 3.63s. Fixture size unchanged — its docstring records the threshold below which the bug stops reproducing, and the mutation check still fails 3/3 when ORDER BY step is reverted. Reuse and simplification: - CALL_EDGE_LIMIT existed in four places; its own docstring predicted the drift it then caused. prompts.ts owns it now — it is a zero-import leaf so the direction cannot cycle, and had graph-queries.ts owned it the four suites that vi.mock that module would have left slice(0, undefined), silently returning every edge in exactly the tests meant to police the cap. - Six dead positional row fallbacks survived the rewrite in the loop this branch re-indented, in the same PR that deleted the identical ABI from graph-queries.ts. - Two test files independently modelled the same labels() scalar-string quirk. Deleted the wiki one — the file's own new header says semantics belong in the real-engine test — and kept projectTypeColumn, the only instrument that can see the bug for the detect_changes query. - makeRepo onto the shared git bootstrap (the eleventh copy of the sequence the helper was extracted to own), the duplicate diff-args unwrapper merged into test/helpers, hand-rolled comparators onto compareCodeUnits, real-timer sleeps replaced by wave-released promises with a strengthened per-wave assertion. - Re-exports trimmed to what is actually imported, a cross-reference this branch invalidated by moving mapConcurrent, and a "~20% faster" claim that does not survive at real index sizes (1-9%; the 10x memory win does). Also adds the drift guard the new doc text lacked: fragment coverage for the partial/truncated paragraph in every skill copy and in the managed AGENTS.md / CLAUDE.md block. Falsifiability checked — none of those fragments exist at the merge base. Not done here, deliberately: 27 live labels(x)[0] projections remain across impact/context/query/trace and MCP resources, with four load-bearing workarounds that have begun depending on each other and one that fabricates rather than degrades. That is a semantic change to five agent-facing tools and wants its own PR, scoped to delete the workarounds too. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv * fix(cli): restore the detect-changes subcommand in the regression example (#2915) Caught by the gitnexus-check bot on the PR. The regression-review fallback in the injected mandate rendered as `--scope compare --base-ref "main" --repo .` with no command, so anyone copying it invokes the runner with an option as its first argument. Self-inflicted, and by exactly the mechanism flagged when it landed: the block is under a test-enforced size cap (#856) that had 30 characters of headroom, so adding the partial/truncated clause required paying for it, and the 38-character "repeat" that was dropped turned out to be the subcommand rather than a repeat. Paid for the restoration out of the clause instead — both parentheticals are gone, since `partial` and `truncated` are already defined in the tool description this text points at. Block is back under the cap at 3548/3552. Notably the cap has now been raised four times (2700 -> 2900 -> 2950, then 0.55 -> 0.65) each with the argument that the new line is load-bearing, and it has now also caused a user-facing defect. It is not functioning as a budget. Left at 0.65 here rather than making it five: moving the threshold to fit one's own text is how it got here. Worth restructuring separately. The fragment guard added a commit ago caught the rewording immediately, which is what it is for; its fragments now pin the two policy claims rather than the prose around them, since that prose is what gets re-trimmed under the cap. Also verified and NOT changed: the bot's other error, that detect_changes compares 1-based hunks against 0-based graph lines. `bounds` is built from `coalesceHunksByPath`, which applies `toZeroBasedLine` to both ends at the grouping boundary, and both a mocked and a real-engine test pin an edit landing on a symbol's last line. The bot read `parseDiffHunks` in isolation. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv * fix(core): reject a fractional chunk size, and stop the truncation note overclaiming (#2915) All five from the gitnexus-check bot's pass on the previous push; two were introduced by the cleanup round that preceded it. `chunk` guarded with `Number.isFinite`, which admits a fractional size — and that one does not fail, it DUPLICATES. `slice` truncates its indices while `i` does not, so size 1.5 yields slice(0, 1.5) = items 0-1 then slice(1.5, 3) = items 1-2, putting item 1 in two batches; a caller batching a query would send it twice. A size is a count, so `Number.isInteger`. Unreachable today (every caller passes a constant) but the guard existed precisely for the unreachable case, and the NaN half of it was already there. `mapConcurrent`'s per-item degradation contract had a hole: `onError` is caller-supplied and was invoked outside a try, so a throwing reporter rejected `settle`, rejected the whole `Promise.all` wave, and discarded the neighbouring successes the function exists to preserve. Reporting a failure must not become one. The CLI truncation note asserted "the counts and risk level still cover all of them", which is true only when `truncated` fires alone — with `partial` the counts are summed from the batches that succeeded. It now varies: a distinct string when both flags are set, saying the counts are a lower bound. This is the same claim already corrected in the tool description; the CLI text still had the old one. The di-extractors contract docstring claimed the barrel re-exports everything from it. That stopped being true when the re-export was trimmed to what is actually imported, one commit earlier. The real-engine wiki test claimed to prepare "every exported query" and omitted `getInterModuleEdgesForOverview`, which `generateOverview` calls. Added — it aggregates in JS over `getInterFileCallEdges` rather than issuing its own Cypher, so the note says why it is in a prepare test. Verified and NOT changed: the bot's other error, that detect_changes compares 1-based hunks against 0-based graph lines. `bounds` is built from `coalesceHunksByPath`, which converts both ends at the grouping boundary (storage/git.ts), and two tests pin an edit landing on a symbol's last line. The remaining seven findings are changed-symbol heads-ups with no signature change; their callers' suites are green. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv * fix(mcp): make the IMPACT_MAX_CHUNKS fallback actually fire (#2915) The validation added earlier this branch used `Number.parseInt`, which takes the numeric PREFIX: '1.5' parses to 1, satisfies `Number.isInteger`, and silently caps enrichment after a single 100-item batch — the opposite of the fallback the comment beside it promised. `Number` instead, so a fractional value is rejected and falls back to 10. The emptiness check is load-bearing rather than defensive: `Number('')` is 0 and 0 is a legitimate value here (enrich nothing), so an UNSET variable would otherwise mean "enrich nothing" rather than "use the default". Behaviour table, old vs new: '1.5' 1 -> 10 (the bug), and undefined/''/' '/ '10junk'/'-2'/'all' -> 10, '0' -> 0, '3' -> 3, ' 5 ' -> 5 all unchanged. So the only case that moves is the reported one. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0139ycbqnAorkJGQaZQXUNuv --------- Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-12 14:51:17 +01:00
Gergő Magyar	740f0a4e57	fix(skills): publish gitnexus-plan artifacts on macOS without an interpreter (#2905) (#2922) ... * fix(skills): anchor gitnexus-plan safe writer on macOS (#2905) The safe generated-plan writer refused to run on anything but Linux. `requireDescriptorAnchoring()` hard-gated `process.platform !== 'linux'` because every name it resolves went through `/proc/self/fd/<fd>/<child>`, and publication went through `renameat2(RENAME_NOREPLACE)`. macOS has neither, so `write-plan` and `read-plan` failed on every input and `snapshot` failed whenever a materialized path was absent. Node cannot perform openat-style directory-relative resolution on macOS at all: `node:fs` exposes no dir_fd parameter, and `fcntl(F_GETPATH)` is a snapshot string that XNU reconstructs from the name cache, so using it would reintroduce the exact race this helper exists to prevent. Python does expose the *at() family via dir_fd, and macOS has renameatx_np with RENAME_EXCL, so the anchoring borrows the interpreter the writer already spawns for renameat2. Anchoring now goes through a backend with two implementations. The Linux one keeps the original expressions, flags, ordering and error strings. The Darwin one runs each operation in the integrity-checked python3: it re-walks the chain from the repository root with O_DIRECTORY|O_NOFOLLOW, asserting the caller's recorded device, inode and mode at every level before acting. A chain that fails that assertion reports a dedicated anchoring errno and never ENOENT, so a moved parent cannot be read as an absent file. Node holds an open descriptor on every chain element for the anchor's lifetime, which pins the inodes so their numbers cannot be recycled between spawns, and that coupling is re-checked on the way into every request rather than left implicit. A filesystem that answers ENOTSUP to RENAME_EXCL is a refusal, never a fallback to a replacing rename. Every other platform is still refused. The suite had silently skipped on every non-Linux runner, so it is now gated on linux-or-darwin and registered in the cross-platform test list, which puts it on the macos-latest CI matrix. Disclosed rather than papered over: operations that must hand Node a file descriptor are anchored in the helper and then opened lexically with O_NOFOLLOW and identity-compared. A racer can force a mismatch, which aborts, or land on the inode the anchored walk already found, which is harmless. A perfect ABA inside that window is impossible on Linux and detected in all but its narrowest form on macOS. The reference doc says so. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NtGZN6YSNU8chALSnKYHBp * test(skills): normalize the anchoring-gate fixture repo on Windows The two capability-gate tests are the only ones in this file that run on Windows, and both failed there: `createBaseRepo` returned the path `os.tmpdir()` gave it, which on Windows is the 8.3 short form (C:\Users\RUNNER~1\...). `assertRepository` compares fs.realpathSync of the caller's path against the realpath of `git rev-parse --show-toplevel`, and plain realpathSync does not expand short names while git always reports the long form, so the helper rejected its own fixture with "--repo must be the Git worktree root" before either platform gate was reached. Resolve the fixture with the native resolver, which returns the canonical long path. No-op on platforms where the two already agree. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NtGZN6YSNU8chALSnKYHBp * test(skills): skip the darwin backend gate on Windows Spoofing process.platform does not spoof fs.constants. Windows Node defines no O_DIRECTORY, so a darwin-spoofed run there refuses at the anchoring-flag check and returns that message instead of ever reaching the python3-backend branch the test exists to cover. Skip it on win32 rather than loosening the regex, which would also let a macOS run pass on the wrong message. The sibling test still asserts the Windows refusal on Windows. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NtGZN6YSNU8chALSnKYHBp * refactor(skills): tighten the macOS anchoring backend Quality pass over the Darwin backend. No behaviour change was intended on the success paths; the guarantees are the same or stronger. Structural: - openChildRead now proves identity inside the backend instead of by comment. It was returning a raw descriptor from a lexical open, with the "callers always compare against the preceding anchored stat" invariant enforced across four call sites in prose — and since the Linux predicate is a literal `return true`, a fifth caller that forgot would have been an unanchored open on macOS that Linux CI could not see. It routes through darwinAdoptAnchoredFile, which already did open-then-compare-then-close-on-mismatch for createChild. - recordAnchoredAbsence shares one prefix walk per snapshot instead of re-walking from the repository root for every absent cited path. With three absent paths under a three-deep prefix that is 12 helper spawns down to 6 and 12 retained descriptors down to 4. citedPaths is caller-supplied and unbounded, so the descriptor retention was the real problem; the cache is now the sole close owner. This does change Linux descriptor lifetime — prefixes stay open for the snapshot rather than only the tail, deduplicated across paths. - assertRepository and the sibling realpath comparisons use realpathSync.native. Windows hands back 8.3 short names that plain realpathSync preserves while git reports the long form, so `snapshot`, which is not platform-gated, could reject a worktree root by quoting that same directory back at the user. The fixture workaround that papered over this for the new gate tests is gone. Efficiency, all measured at ~13.5ms per helper spawn: - consume the identity mkdir already computed rather than re-stat it - act on renameNoReplace's return value rather than spending two stats re-deriving what it already reported - drop a duplicate anchored stat taken twice in a row in movePathToVault - import ctypes only where it is used; 19 of 20 spawns never touch it Simplification: pins folded into the descriptors the handle already carried, an unreachable refreshAnchorTail branch and the dead darwinHardenedOpen mode parameter removed, the four copies of the spawn options collapsed, the spawn-and-parse shared between the probe and the request path, the unreachable launch-path fallback and a redundant memo deleted, and the helper's dispatch made a real elif chain with leaf name and mode validated at one chokepoint rather than per operation. The two chain encodings were left alone deliberately: merging them would have grown triple fields on Linux for no Linux benefit and changed the Linux validatePlanParent comparison. The double re-stamp that motivated the merge is contained in one named helper with the hazard documented. Rejected candidate interpreters now say which dir_fd operations were missing instead of producing a generic refusal. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NtGZN6YSNU8chALSnKYHBp * refactor(skills): publish plans with link(2) and drop the interpreter The macOS backend spawned python3 for two jobs: openat-style resolution, which Node cannot do, and a no-replace rename. Only the first is actually unavoidable, and the second was carrying the whole dependency. link(2) is a no-replace publish. It is atomic, it fails EEXIST when the destination name is taken, and it refuses a symlinked destination without following it — the same guarantee renameat2(RENAME_NOREPLACE) and renameatx_np(RENAME_EXCL) give, reachable from plain fs.linkSync. The published file is the same inode as the verified temporary, so the downstream identity checks hold by construction rather than by argument. That removes the interpreter from Linux entirely, since /proc already did the resolving there, and it removes ctypes, libSystem, RENAME_EXCL and the ENOTSUP handling from macOS. Deleted with them: the trusted-executable validation, the held-descriptor exec and its two-tier probe, the capability probe, the JSON request protocol, and both embedded Python programs. The helper drops from 3047 to 2327 lines. macOS keeps the part that genuinely cannot be done in Node, and now does it without a subprocess: a lexical O_NOFOLLOW walk that holds an open descriptor on every directory in the chain and re-proves the chain either side of every step. Pinning is load-bearing — an open descriptor keeps its inode number from being recycled, which is what makes the recorded identities trustworthy across steps. The guarantees are no longer symmetric and the docs say so plainly. /dev/fd/<fd> is a devfs node, not a magic link: opening it works, resolving through it does not, open("/dev/fd/<fd>/child") returns ENOENT and realpath returns /dev/fd/<fd> — measured on macOS 26 rather than inferred. So Linux makes a parent swap impossible while macOS detects one and aborts. Also fixes the writer on 9p mounts, where renameat2(RENAME_NOREPLACE) returns EINVAL and publication failed every time; link(2) succeeds there. Tests 174 -> 154: dropped 29 fixtures that drove the deleted Python program directly, added coverage for the link publish, for a macOS parent swap caught through the pinned chain, and for a spoofed-darwin round trip that asserts no /proc path reaches the hooks, which the portable backend now makes runnable on Linux CI. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NtGZN6YSNU8chALSnKYHBp * fix(skills): drop O_NOFOLLOW_ANY, guard trailing slashes, handle link edge cases macOS CI rejected our hardened directory open with EINVAL on 30 tests. The flag O_NOFOLLOW_ANY was ORed into every open on the theory that XNU ignores unrecognized open bits, so it would be inert where unsupported. That theory is wrong, at least combined with O_DIRECTORY. The Python design never hit it because the walk ran inside the interpreter; once Node did the opening, every Darwin directory open went through it. Removed rather than probed. The per-component O_NOFOLLOW walk is what delivers the guarantee, and cap-std — the closest reference implementation of this problem — has not adopted O_NOFOLLOW_ANY either. A fixture now pins the exact flags of every directory open under a spoofed darwin, so the next failure names the flag instead of printing a stack trace. With the flag gone the two backends' directory open became identical, so it is no longer a platform concern at all. Three findings from researching the prior art, all now covered: Trailing slashes. CVE-2026-39822 escaped Go's os.Root because open(fd, path, O_NOFOLLOW) follows symlinks when the path ends in "/". It reproduces here: with docs a symlink, opening "docs" is ENOTDIR but "docs/" succeeds into the attacker's directory, and path.join preserves the slash. We were safe only by construction, and only for repo-derived names — the generated temporary and vault artifact names never passed through the validator. The guard now sits at anchoredChild, the single place a name becomes a path, so it holds for every caller. link() can lie on NFS. Per link(2) BUGS, the return code may be wrong if the server creates the link then dies before replying; open(2) NOTES gives the remedy, which is to stat the source and treat a link count of 2 as success. Implemented, with the man-page reasoning in the comment so it is not later removed as paranoia. Filesystems without hard links now fail loudly. EPERM, ENOTSUP and EMLINK say so and refuse to fall back to a replacing rename. Git falls back and accepts losing collision detection because its objects are content addressed; that reasoning does not transfer to a named plan destination. Durability was already correct — the temporary is fsynced before publication and the parent directory immediately after — but the comment now records why the parent fsync is required for link as it was for rename, and the honest limitation that fsync is not a write barrier on macOS while F_FULLFSYNC, which Node cannot reach, is. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NtGZN6YSNU8chALSnKYHBp * refactor(skills): shrink the anchoring seam and fix two CI breaks Four quality reviews over the pure-Node writer. Two real breaks, one drift that had already happened, and a seam that was sized for a design we deleted. The macOS round-trip fixture asserted that every observed path started with join(repo, 'docs/plans'). Reproduced on Linux by handing the helper a repo reached through a symlink, which is the shape macOS gives us via /var to /private/var: assertRepository realpaths the repo, so the handle builds paths from the resolved form while the fixture holds the form it passed in, and the prefix can never match. The assertion now proves the same thing without depending on the prefix — a lexical resolution always contains a docs/plans segment and /proc/self/fd/<fd>/<name> never does. Two publish fixtures sat in the capability-gate describe, the one block deliberately not skipped on unsupported platforms, while this PR added the file to the Windows matrix. They test link(2), not the gate, so they moved to SAFE_WRITE_FIXTURES. validatePlanParent restated verifyLexicalChain's loop without the try/catch that converts ENOENT and ENOTDIR into the parity message, so a raw errno could escape a function with a dozen call sites. It was masked on Darwin only because parentStillResolves catches first. It now calls the helpers, which also removes a second full chain walk per call there. openVerifiedFile adds O_NONBLOCK so a FIFO swapped in at the target name cannot wedge the process on open, and only Darwin was calling it. The operations are now shared, so Linux gets it by construction rather than by a per-backend decision. The backend is five methods rather than ten. The platform difference is two things — how a name becomes a path, and what guard wraps an operation — so the five operations became shared functions over a `verified` hook that is run() on Linux and the pinned-plus-lexical sandwich on Darwin. openChildRead always runs the identity adoption, so that proof is structural rather than a comment about what callers must remember. Selecting the backend is a registry that throws on an unknown platform instead of a ternary defaulting to Linux, which surfaced seven dead bindings that ran before the capability gate and made win32 report the registry error instead of the refusal. Snapshot capture no longer re-walks a prefix per record: 36,018 lstats to 6,384 and 162ms to 130ms on 2,000 dirty files across 100 directories, with a byte-identical global_dirty_digest. Absence anchoring is now bounded at 4096 pinned directories and refuses rather than evicting, because closing a cached descriptor would break the pinned chain of a guard already recorded — the inode-recycling hole the pins exist to close. The test suite no longer cache-busts its imports. That existed for the memoized python3 descriptor, the file's only mutable module binding, which is gone; the suite drops from 10.0s to 8.2s. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NtGZN6YSNU8chALSnKYHBp --------- Co-authored-by: Gergo Magyar <gergomagyar0@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-11 12:24:53 +01:00
Gergő Magyar	8b5057f325	feat(skills): GitNexus Engineering Tool Kits (#2566) ... * feat(skills): add ce-plan — GitNexus+PDG implementation-planning skill Adds .claude/skills/ce-plan: a planning-only skill that builds implementation-ready plans from GitNexus graph navigation (query/context/ impact/trace), bounded statement-level PDG slices (pdg_query, impact mode:pdg, explain), and targeted source verification, with a context ledger to prevent repeated reads and a machine-readable implementation context pack (stable contract for a future ce-implement). Whitelisted in .gitignore and registered in AGENTS.md and CLAUDE.md outside the auto-managed gitnexus block. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(skills): apply ce-plan validation findings (tool contract, consistency, conventions) Tool contract: impact mode:'pdg' shape now includes the schema-required direction param; CDG branch sense documented as the result 'label' field (reason is cypher/raw-edge only); explain caveats corrected to its real false-negative classes (cross-function TAINT_PATH is modeled). Consistency: PDG slice homed in working memory (ledger keeps one-liners); depth knob defined and category-overrides-baseline ordering stated; call_depth (consumed by nothing) and content-hash bookkeeping dropped; Never section folded into Hard rules; Phase 3 deduplicated to a pointer; allowed-repeat escalations defined; budget/discard accounting clarified; verification-commands gathering added to Phase 4; open_questions added to the context pack. From scenario runs: plans now pin the verified-at HEAD commit and index freshness in a header, tag claims [verified]/[graph]/[inferred]/[assumed], quote load-bearing tool output, prefer pre-hook-carrying npm scripts, and support an out:<path> destination override; output path defined as the Phase 1 target repo root. Conventions: AGENTS.md 1.9.0 / CLAUDE.md 1.4.0 changelog rows + metadata bumps; future ce-implement qualified as future. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): rename ce-plan → gitnexus-plan; add cross-CLI (Codex) entrypoints Renames the skill dir, frontmatter, output filename convention, plan H1 (GitNexus Engineering Plan), the future executor handle (gitnexus-implement), the .gitignore whitelist entry, and all AGENTS.md/CLAUDE.md references. Follows the pr-swarm-review cross-CLI pattern: SKILL.md is the canonical CLI-neutral spec, AGENTS.md § Engineering planning is the Codex/any-agent entrypoint, and the README documents the optional user-level ~/.codex/prompts/gitnexus-plan.md slash command plus an invocation matrix. Skill prose de-branded from Claude Code (agent-neutral verification layer). Also fixes two post-review README contradictions: the anti-reread claim now names the ledger's allowed escalations, and 'read-only by contract' is now 'planning-only' (the skill writes exactly one repo file — the plan); the scope-creep rule and template §12 now agree on where deferred follow-ups land. Drops the stale plugin-collision limitation. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(skills): document Codex user-level install path for gitnexus-plan Codex discovers SKILL.md skills from ~/.agents/skills (same path the other gitnexus-* skills install to); README now documents the cp install plus the optional ~/.codex/prompts slash-command file, with the prompt body preferring the repo copy and falling back to the user-level install. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): gitnexus-plan freshness gate + active PDG-layer refresh Freshness is now a Phase 1 gate, not advisory: under the default freshness:strict, a stale index is refreshed once per planning session via node .gitnexus/run.cjs analyze --index-only (appending --pdg when the task will reach the PDG phase), then the context resource is re-read. A missing PDG layer likewise triggers the one permitted --index-only --pdg refresh and re-probe instead of a passive recommendation. freshness:accept (or a failed/impractical refresh) preserves the old behavior: plan on the stale graph, source-weighted, labelled in the plan header. --index-only is the load-bearing flag choice — it suppresses all file generation, so the planning-only contract holds (only the .gitnexus store changes). Ledger gains an index_refresh record; plan header states fresh / refreshed / refresh-skipped-with-reason. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): gitnexus-plan runner build check before freshness refresh When the target repo builds the analyzer from its own source (bin → dist/ mapping, as gitnexus/ does), the Phase 1 freshness gate now verifies dist/ is current before running the analyze refresh — rebuilding via the package's build script when any analyzer source file is newer than the built entrypoint — and prefers that freshly built CLI. Otherwise a stale dist re-indexes with outdated extraction logic and the 'fresh' index lies. Rebuilds are recorded in the ledger's index_refresh; the PDG-phase refresh inherits the same check. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): add gitnexus-work executor and gitnexus-lfg pipeline gitnexus-work executes a gitnexus-plan as verified atomic commits: consumes the §11 implementation_context pack, drift-checks the plan's evidence pin against HEAD, re-verifies assumptions before relying on them, runs impact before every symbol edit and detect_changes before every commit (repo mandates), builds tests from the plan's scenarios, and routes structural drift back to gitnexus-plan Deepen mode instead of coding around it. gitnexus-lfg is a thin orchestrator: gitnexus-plan → blocking user gate (deepen / proceed / stop, deepen loops allowed) → gitnexus-work → review via the existing gitnexus-pr-review skill (open PR, else branch diff vs default). One bounded fix cycle for review findings; never pushes or opens a PR on its own. gitnexus-plan gains a Deepen mode (re-run freshness gate, escalate to depth:deep, re-verify graph/inferred/assumed claims toward verified, rewrite the same file); its 'future gitnexus-implement' placeholder is retired in favor of gitnexus-work. Registered via .gitignore whitelists, AGENTS.md 1.10.0 (section renamed to Engineering planning & execution), CLAUDE.md 1.5.0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(skills): apply cross-skill review findings to the gitnexus skill family Two P1s: gitnexus-plan Deepen mode now re-anchors before re-pinning (diffs the old evidence pin over every [verified]-claim file and re-reads or downgrades before the header moves — moving the pin without this laundered stale claims as verified); the index-refresh budget is stated once in Phase 1 (one --index-only refresh plus at most one Phase 3 --pdg upgrade per session, Deepen = its own session) with ledger and pdg-slice deferring to it. Contract fixes: gitnexus-work's drift check now covers every file the pack cites (not just files_to_modify) and parses the full pack incl. primary/related symbols and acceptance_criteria (walked in Phase 4 alongside §13); a pre-completed check skips §7 steps already landed and Deepen gains a reconcile-execution-state step, closing the mid-execution route-back loop; pack assumptions must name what to check and how. lfg: Lane 4 passes the merge-base to detect_changes compare (two-dot diff misattributes upstream commits when default advanced), branch-diff is the stated normal case, oversized review findings route to the plan gate instead of overflowing direct mode, the one-fix-cycle cap is explicit on re-run, and headless runs end at the plan gate with the plan as deliverable. work: blank mode narrowed to *gitnexus-plan*.md with a re-execution guard, direct-mode discipline spelled out, branch meaningfulness defined against the plan slug, and the plan document is committed as the branch's docs commit (review diff includes it). Planning-only contract now names the dist/ rebuild as the second permitted state change; Phase 5.1 names the four claim tags; stale AGENTS.md anchors fixed. Known latent issue left untouched: gitnexus/gitnexus-pr-review pairs a three-dot example with a two-dot detect_changes compare — that skill is also shipped by the plugin, so fixing it here would drift the copies; lfg compensates by passing the merge-base. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): ship the engineering skill family with the gitnexus package npm i -g gitnexus users now get gitnexus-plan / gitnexus-work / gitnexus-lfg: the three skills are added to gitnexus/skills/ in directory form (SKILL.md + references/), which installSkillsTo already enumerates dynamically and copies recursively to every editor target (~/.agents/skills for Codex, Cursor, OpenCode, Qoder, ...) on gitnexus setup — uninstall enumerates the same root, so removal stays clean. The Claude Code plugin channel (gitnexus-claude-plugin/skills/) carries the same copies plus the standard per-skill mcp.json. Global-install support in the skill text: gitnexus-plan Phase 1 now resolves the analyzer runner explicitly — node .gitnexus/run.cjs analyze when the project has a runner, else gitnexus analyze (installed CLI), else npx gitnexus analyze — and all analyze mentions route through it, satisfying the skills-steering policy (#1939/#1945) which sweeps the plugin copies. New drift guard test/unit/shipped-skills-sync.test.ts asserts the npm and plugin copies stay byte-identical to the canonical .claude/skills/ family (plugin = canonical + mcp.json), same discipline as run.cjs ↔ resolve-invocation.ts. skills-steering + shipped-skills-sync: 11/11 green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(eval): workflow_bench — measure the skill workflow's token savings Benchmarks gitnexus-plan → gitnexus-work against a baseline agent (--disallowedTools Skill) on identical tasks, in fresh detached worktrees, using real headless Claude Code sessions; every number comes from the CLI's --output-format json usage report (field names validated against a live 2.1.207 session). Reports per-arm medians (input/cache/output tokens, cost, wall time, turns), a savings row, and resolve status from a per-task verify command — savings on failed tasks are flagged, not celebrated. Per-task setup hook prepares fresh worktrees (deps); --permission-mode bypassPermissions (default) lets sessions run unattended in the throwaway trees. Free-model support: --base-url/--auth-token/--model route headless sessions through any Anthropic-compatible endpoint; free-model.litellm.yaml is a ready litellm-proxy template for OpenRouter :free variants or local Ollama, so benchmarking burns no paid tokens (README documents rate limits and the small-model skill-following caveat). Harness validated end-to-end with a stub CLI (worktree lifecycle, both arms, plan→work chaining, verify, aggregation, report) and 4 pytest units for the pure aggregation/savings/report helpers. AGENTS.md 1.11.0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(eval): record first workflow_bench calibration run Trivial-task calibration (add -V alias): both arms resolved; workflow arm ~4.3x baseline cost — the documented overhead-dominated regime, recorded so the regime boundary is empirical rather than asserted. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(eval): workflow_bench scenario matrix — arm variants, task classes, churn Ground-base measurement across scenarios: tasks.scenarios.yaml spans four labeled classes (trivial → investigation-bug → investigation-feature → cross-module) with deterministic verifies (prescribed test files). New arms: workflow_direct (gitnexus-work direct mode — the middle option that locates the routing boundary lfg's gate and work's triage encode) and baseline_nomcp (no skills AND no graph tools — separates workflow-discipline value from GitNexus-tool value; off by default). Records now carry task class and diff churn (files/+ins/−del vs the starting commit) as an over-engineering proxy; the report renders a class column and per-arm savings rows vs baseline. 5 pytest units + stub-CLI e2e of the full three-arm matrix. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(eval): record workflow_bench ground base; fix churn measurement bias Ground base (3 classes x 3 arms, n=1/cell): every arm resolved every task — pass/fail quality saturates at this difficulty, making the comparison pure cost. Full plan→work never amortized its ~$9-11 fixed cost on tasks a baseline finishes in ≤35 turns (−211% to −333% cost); workflow_direct sits near baseline (−15% to −55%, once faster wall) with more test coverage. Routing implication recorded: direct mode/plain agent below this scale, full workflow for cross-module / multi-session / plan-as-deliverable work. The cross-module cell and multi-run variance are the next measurements. Churn fix: git add --intent-to-add -A before diffing (arms that never commit no longer undercount new files) and :(exclude)docs/plans (the committed plan doc no longer inflates workflow churn); this run's churn numbers predate the fix and are omitted from the recorded table. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * perf(skills): cost-optimize the workflow from measured ground base Every optimization targets a measured fixed-cost component (eval/workflow_bench ground base: workflow arm −211% to −333% vs baseline, all tasks resolved): - Plan form is category-priced: compact form (core sections w/ § anchors preserved, ≤80 lines excl. pack, mini-pack subset of the context pack) for narrow/default categories; the full 13 sections only for deep work (refactor/security/performance/concurrency/architecture). A compact plan outgrowing its cap reclassifies to full rather than overflowing. - Freshness gate is category-priced: compact categories default to accept (source-weighted, refresh only when a graph claim becomes load-bearing); strict stays the default for full-plan categories — the rebuild+re-index was the largest single fixed cost. - Turn economy: per-category tool-call budgets (~10 to ~45; architecture uncapped); budget exhaustion routes open questions to §12 instead of more digging. - gitnexus-work fast path: HEAD == evidence pin → skip all citation re-reading (the pin's entire point); mini-pack fields tolerated. - lfg Lane 1 boundary triage: tasks below the measured ~35-turn boundary get offered gitnexus-work direct mode before the plan lane is spent. Copies re-synced (npm skills/, plugin, ~/.agents); steering + sync guards green. Re-measurement of the workflow arm follows to verify the numbers actually improve. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(eval): record optimization re-measurement — inv-bug workflow cell −20% cost Same task, same conditions, post-830a0459 skills: $14.56→$11.70 (−20%), 83→72 turns, cache_read −24%; verified in-transcript that the compact form, turn budget, and skipped rebuild/re-index all fired. Wall +15% from a work- session test-debugging tail (n=1 variance). Regime unchanged (~3.5x baseline on this class) — routing rule stands. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(eval): per-arm clone isolation — worktree ref-namespace leak contaminated an arm The cross-module workflow_direct cell reported an impossible 28-turn solve with churn byte-identical to the workflow arm: git worktree add shares the repo's ref namespace, so the workflow arm's slug branch (created by gitnexus-work Phase 2) survived worktree removal and the direct arm found and adopted the completed work. Arms now get isolated git clone --shared copies (object store via alternates, refs clone-local — agent branches and stashes die with the clone; origin/<ref> fallback for non-default refs). Leaked branch deleted; baseline arm verified clean (0 branch references in its transcript); cell marked invalidated pending re-run. Records the valid cross-module cells: workflow $18.32 vs baseline $18.03 (premium −1.6%, vs −211%..−333% on smaller classes) — fixed costs amortize at this scale, with a less destructive diff and a plan artifact as bonus; resolve rate still tied. Churn fingerprinting is what caught the contamination — noted in the README as an integrity check. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(eval): complete cross-module cell — direct mode wins 47% cost / 56% wall Clean clone-isolated re-run: workflow_direct resolved the hardest class at $9.53/52 turns/15m vs $18.03/98/34m baseline and $18.32/107/37m full workflow. The measured story across all four classes: the execution discipline (gitnexus-work) is the consistent sweet spot and delivers real token savings on hard tasks; the planning pass buys its artifact, not same-session savings. Resolve rate tied everywhere (n=1/cell caveat). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(eval): add trajectory-gated skill evolution (#2431) - Pair prompt candidates with incumbent workflow arms - Gate promotions on pinned-model quality and efficiency - Expire router evidence and document its lifecycle * fix(eval): allow pr-review skill candidates * feat(skills): rename and generalize GitNexus review * feat(eval): external-comparator and review arms for workflow_bench - ce_workflow / ce_workflow_direct: compound-engineering ce-plan/ce-work arms prompted with the same structure as the gitnexus arms - review / ce_review: gitnexus-review vs ce-code-review on an identical diff applied by the task's setup - plan handoff is snapshot-based: committed example plans in docs/plans/ tie on clone mtimes and broke the name-glob pick (executed a stale plan) - verify output tail is recorded per run and the final working-tree patch is kept, so failed rows are diagnosable after the clone is destroyed Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(skills,eval): address #2431 review — data-safe rename migration, fail-closed bench evidence - setup: never delete a legacy renamed skill dir — the installer cannot prove ownership (users customize or hand-write skills under these names); warn with the path instead, and the test now asserts survival - workflow_bench: fail closed when a session's --output-format json report is empty, malformed, or missing usage fields — an exit-0 shell with no parseable usage no longer counts as measured evidence (5 parametrized regression tests) - workflow_bench: document the trust model prominently (task setup/verify are shell-executed, sessions run bypassPermissions with the parent env, candidate overlays are prompt injection surface) in README + docstring - free-model.litellm.yaml: master_key from LITELLM_MASTER_KEY env instead of a static token; loopback-binding warning - ci: run the eval workflow_bench pytest suite on ubuntu (pytest+pyyaml only — no full eval stack) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(eval): demand observed foreground verification in headless work-arm prompts In a headless -p session there is no later turn: a work arm backgrounded its slow test run, scheduled wakeups that can never fire, and reported done while two of its tests failed. All four work-arm prompts (both skill families, symmetric) now require verification output to be observed inside the session. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): ask plan depth up front instead of offering deepen afterwards gitnexus-plan Phase 0 now asks one blocking question in interactive sessions — quick / standard / deep, mapped onto the existing depth/form/ freshness knobs — when the invocation carries no explicit depth signal. Explicit knobs and headless runs skip the question (category posture unchanged, so benchmarks and automation behave as before). gitnexus-lfg's plan gate slims to proceed/stop: depth was already the user's up-front choice, so deepening is no longer offered by default — an explicit deepen request at the gate and executor route-backs still run Deepen mode, which remains the mechanism for strengthening an existing plan document. All shipped copies resynced (npm skills/, Claude plugin); AGENTS.md 1.13.0 and CLAUDE.md 1.7.0 pointers updated, including the analyzer's regenerated index-stats block at this branch's head. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): taint pass, expert lenses, and post-work index refresh gitnexus-review gains a PDG-backed taint-and-dependence pass (explain + pdg_query, --pdg folded into the stale refresh on trust-boundary diffs) and an Expert lenses section: domain reviewers derived from the graph's clusters plus four cross-cutting lenses (architectural fit, language conformance per the repo's own contract, Definition of Done, simplicity), dispatched once after the evidence-gathering steps and scaled to the diff. gitnexus-work Phase 4 now refreshes the knowledge graph after the DoD walk via the resolved-runner ladder with analyze --index-only, so the lfg review lane and later sessions query the finished work without dirtying the tree. lfg's threshold-governance paragraph moves to its README; eval citations are tagged as measured in the GitNexus repo. All shipped copies re-synced. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(cli): remove legacy gitnexus-pr-review on uninstall; cover the rename migration uninstall's removal set now includes LEGACY_SKILL_DIR_NAMES derived from RENAMED_SKILL_DIRS, so a pre-rename install is cleaned up instead of orphaned. The rename warning gains behavioral coverage (fires with a legacy dir present, silent without), and shipped-skills-sync asserts legacy names stay absent from every shipped tree. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(eval): metric provenance, error-kind rows, skill-invocation verification, gate noise floor The promotion gate defaults to cost_usd (the only metric that includes subagent spend); token metrics carry an explicit main-loop-only warning in the report and promotion.json. Rows are classified by error_kind (session-error / verify-failed / infra-error), excluded from efficiency medians, and the gate requires equal valid-run counts. Each session's transcript is scanned for the expected Skill invocation and fails closed on a verified miss; a one-run resolution edge no longer promotes (noise floor). Per-run timeouts and setup failures record an infra-error row instead of aborting the sweep. Overlays touching skills no candidate arm exercises are rejected up front. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: fix skill routing paths, version headers, and skill rosters Routing tables point at the tracked direct skill paths (matching the post-#2434 generator output), AGENTS.md/CLAUDE.md headers match their latest changelog rows, the 1.12.0 row describes what the migration actually does, package/cursor READMEs list the full shipped skill roster, and the swarm READMEs describe /gitnexus-review's expert lenses instead of calling it single-agent. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * ci: drift-guard workflow for skill copies; pin eval pip deps; track docs/plans ci.yml ignores '**.md', so an md-only skill edit would merge without the shipped-skills-sync test running — skill-sync.yml triggers exactly on the guarded trees. The eval job's pip install is version-pinned, and docs/plans/ is unignored so gitnexus-plan output can be committed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ci): keep the runner-invocation literal in gitnexus-review; add concurrency block to skill-sync skills-steering requires skills with a stale-index hint to carry the exact 'node .gitnexus/run.cjs analyze' form — restore it with the fallback ladder as a parenthetical instead of replacing it. skill-sync.yml gains the top-level concurrency block the workflow-convention check enforces. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): token-economy guidance for expert lenses Merge lenses that ground in the same material into one reviewer, and use cheaper model/effort tiers for mechanical lenses where the harness offers them, reserving the strongest engine for adversarial judgment. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test(eval): isolate transcript home on Windows Ensure workflow_bench transcript tests set USERPROFILE alongside HOME so Path.home() resolves to the temporary test home on Windows. * docs(skills): fold PR #2522 execution learnings into review/work/plan Eight incident-backed hardenings from running the full skill cycle (review -> plan -> work, 28-finding fix series) on PR #2522: gitnexus-review: - Expert lenses execute the code under review on candidate failing shapes (empirical probe outranks source reading — every HIGH the language lenses found came from a probe, not a read). - Step 7 re-runs the exact CI check for refreshed baselines/fingerprints (a stale committed artifact is invisible in the diff; caught a red benchmarks arm). - Step 8 treats version/invalidation constants as review surface (INCREMENTAL_SCHEMA_VERSION class recurred verbatim from #2494). gitnexus-work: - Step 4 proves regression tests discriminate against the pre-fix tree. - Step 5 rebuilds executed build output before every verification run (parse workers load dist/; a correct fix 'failed' until rebuilt). - Step 6 makes stage -> detect_changes -> commit one unbroken sequence. gitnexus-plan: - Phase 0 seeded-evidence mode: plan FROM a completed review's verified findings instead of re-running the graph ladder. - Template §7: fingerprint/golden-guarded output rebaselines once, at the series tip. All distribution copies resynced; shipped-skills-sync + skills-steering 24/24 locally. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(eval): close the skill-evolution loop with an automated proposer driver workflow_bench.evolve adds the three arrows the README described as manual: a proposer session that turns loser trajectories (results.jsonl rows, transcripts, patches, the learning queue) into ONE bounded candidate overlay, a driver that iterates propose -> paired benchmark -> deterministic gate up to --generations, and an --apply step that copies a promoted overlay onto the canonical skills and shipped mirrors as a working-tree diff. The trust boundary is unchanged: overlays re-validate through candidate_overlay_files before any benchmark or apply consumes them, and committing, CI, and the PR merge stay human. learnings.jsonl is gitignored: it is machine-local evidence, like the session transcripts it complements. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(skills): route live-task friction into the evolution learning queue Each family skill gains a short 'Skill feedback' section: on friction with the skill's own instructions, append one JSON line to eval/workflow_bench/learnings.jsonl (GitNexus repo only) — never self-edit the skill from a live task. The proposer in workflow_bench.evolve consumes the queue as hints; a learning reaches a shipped skill only by beating the incumbent on the paired benchmark. All shipped mirrors re-copied byte- identical. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * ci(tests): run the evolve helper tests in the eval pytest job test_evolve.py needs only pytest+pyyaml, same as the harness tests the job already runs — without this line the new module had no CI coverage. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(ci): comment-triggered GitNexus review agent for PRs '@gitnexus review' from a maintainer (OWNER/MEMBER/COLLABORATOR; the action re-validates write access) runs the repo's gitnexus-review skill headlessly against the PR and posts the review as a sticky comment — remote triggering with no local setup. Read-only by construction: contents: read token, Write/Edit and web tools disallowed, Bash allowlisted to git reads and the gitnexus CLI; analyze parses PR code with tree-sitter, never executes it. Requires the ANTHROPIC_API_KEY repository secret; activates once the file is on the default branch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(ci): dispatch lane + existing OAuth secret for the review agent Align with claude.yml: same action pin and the CLAUDE_CODE_OAUTH_TOKEN secret the repo already carries — no new secret to configure. Add a workflow_dispatch lane (PR number input) so the agent can be triggered from the Actions UI and tested before the issue_comment trigger reaches the default branch. Allowlist gh pr view/diff and gh api, which the review skill uses to pin PR SHAs. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ci): close a fork-PR RCE vector in the review agent's tool allowlist A live headless run of the exact workflow session against PR #2431 (66 turns, full gitnexus-review pass) surfaced a real HIGH-severity confused deputy: .gitnexus/ is gitignored, not blocked — a fork PR can commit its own .gitnexus/run.cjs, issue_comment checks out PR-head content, and the skill's runner ladder tries 'node .gitnexus/run.cjs analyze' first. That would execute fork-controlled JS inside a job holding CLAUDE_CODE_OAUTH_TOKEN and a write-scoped GITHUB_TOKEN — the opposite of the 'PR code is read, never executed' claim in the workflow's own header. Fix: drop the run.cjs allowlist entry so analyze always resolves through npx gitnexus (npm registry, not the checked-out tree); the skill's documented fallback mode covers the resulting graceful degradation. Also drop 'gh api' (not read-only — accepts -X POST/PATCH/DELETE) and downgrade pull-requests: write to read (comment posting only needs issues: write; the prompt already forbids formal review submission). Same session flagged a latent evolve.py bug: select_evidence's cost sort used dict.get's missing-key default, which doesn't cover an explicit JSON null in a foreign --seed-results row and crashes proposer setup with TypeError. Guarded with 'or 0.0' and added a regression test. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: harden PR review and evolution trust boundaries * ci: follow workflow concurrency convention * fix(eval): make terminating error paths explicit * fix: unblock hardened review runtime checks * test: make containment canaries deterministic * test: expose Claude canary tool failures * fix: adapt clean shell environment for Claude * fix(eval): accept the runner's transcript source key in evidence preflight The proposer evidence preflight required transcript-artifact metadata to be exactly {path, sha256, bytes}, but the runner stamps a fourth provenance key (source=parent-captured-stream-json). Any --seed-results or generation>=2 run therefore aborted with SandboxError before proposing or promoting. Pin the producer literal as PARENT_EVENT_STREAM_SOURCE and validate it in the metadata check, and round-trip real producer output through sum_sessions into the preflight so the schema can't drift again. * fix(eval): treat an unmeasured session cost as unavailable, not $0 well_formed validated only the nested usage block, so an otherwise-successful session missing total_cost_usd was recorded as cost_usd=0.0 — and cost_usd is the default promotion metric (lower wins), so a cost-less session scored as free and could win promotion it never earned. Extract cost via measured_cost() (None on absent/garbage, a measured 0.0 preserved), propagate None through sum_sessions/aggregate/savings/report, and have the gate refuse to rank on a metric that was not measured on every run in both arms. * fix(eval): warn when ranking on the main-loop-only num_turns metric num_turns comes from the CLI's top-level usage (main-loop session only), like output_tokens, but selecting it emitted no metric_warning — so a subagent-heavy candidate could look artificially efficient. Add num_turns to MAIN_LOOP_ONLY_METRICS and broaden the warning to cover turns. * fix(eval): fail closed when an overlay adds a file with no committed base An overlay adding a new .md under gitnexus-{plan,work} passes the structural overlay checks but has no committed base for committed_destination_base_digests to bind against, so it raised an uncaught ValueError that crashed the evolve driver (and runner --candidate-overlay) mid-run. Catch it at both call sites: evolve reports NOT PROMOTED and exits, runner routes it through parser.error. * feat(eval): circuit-break the runner sweep on a systemic outage A sustained upstream outage used to pay out every remaining --timeout window one session at a time. Track consecutive session/infra/cleanup failures via a pure systemic_outage_streak helper; after --outage-streak (default 5) in a row, stop the sweep, still write report.md/promotion.json from partial evidence, and exit non-zero so evolve.py halts instead of proposing from truncated evidence. A task's own resolved=False never trips the breaker. * fix(cli): report a dirty working tree as stale in gitnexus status status --json (and the human output) computed up-to-date from commit + runner identity + completeness only, so a repo with uncommitted source changes at a matching HEAD was reported up-to-date while analyze would still re-index it. A graph-backed agent gating on that JSON could skip re-analysis on a stale graph. Extract analyze's dirty-tree check into a shared isWorkingTreeDirty() in storage/git and fold it into the status freshness decision. * fix(ci): use single-slash deny globs in the review agent's disallowedTools github.workspace already expands to an absolute path, so Read(/${{ github.workspace }}/**) and Read(//proc/**),(//sys/**),(//dev/**) produced double-slash patterns that a normalizing matcher may not match — silently no-opping the deny layer. Not exploitable (the allowlist is the primary control and never grants those paths), but the globs should be well-formed. Update the pinned test strings. * ci: install gitnexus-shared with npm ci from the committed lockfile The gitnexus-shared build floated its deps via npm install in three workflows (skill-sync, ci-tests, and — most importantly — the release publish.yml) while every other install step uses npm ci. The lockfile is committed and in sync, so switch all three to npm ci for reproducible, locked installs. * test(cli): make the shipped-skills drift guard reject symlinks listFilesRecursive walked with readdirSync and snapshotDir read with readFileSync, both of which follow symlinks — so a mirror file symlinked to the canonical tree passed the byte-compare (and a symlinked mirror dir would be followed too). Reject a symlinked root via lstat and any symlinked entry via Dirent.isSymbolicLink, with negative tests (skipped on Windows). * test(eval): guard the candidate-skill vs mirror-root coverage invariant MIRROR_SKILL_ROOTS omits the Cursor tree, safe only because no candidate skill is cursor-shipped. Pin that invariant: every CANDIDATE_SKILLS entry must exist under canonical + every mirror root and must not ship to Cursor, so adding a cursor-shipped skill to the candidate set (the PR #2488 asymmetric-sync class) fails loudly instead of syncing three of four trees. * docs(ci): describe the review agent's staged post-merge rollout The DoD asked for a dry-run or triggered run before merge, but an issue_comment (or newly added workflow_dispatch) workflow only ever executes the default-branch copy, so it cannot be exercised from the PR that introduces it. Reword the DoD and the activation checklist to a staged rollout: merge registered-but-disabled, validate same-repo and fork execution post-merge, then enable the variable. * fix: pin plugin skill mcp.json to the release version via #2445 tooling The ten plugin skill mcp.json launched `npx -y gitnexus@latest mcp` on every skill connect — non-reproducible and a supply-chain surface, and (unlike the persisted setup config) never pinned. Extend sync-plugin-manifests.mjs with an mcp surface kind that stamps the gitnexus@<version> launch arg, pin all ten to 1.6.9 now, and keep them byte-identical so the drift guard stays green. The release lifecycle + publish.yml --check now re-stamp them like the four manifest surfaces; only READMEs stay on @latest as docs. * test(eval): prove the proposer's built-in file tools are confined The real-Claude canary only exercised Bash + MCP, so it proved process/MCP containment but not that the proposer's built-in file tools stay inside their mounts. Add a canary over the exact PROPOSER_ALLOWED_TOOLS surface and the same read-only /evidence mount as run_proposer (allowlist extracted to a shared constant so it can't drift): Read reaches /evidence, a Write into the read-only evidence mount is denied, and a Write lands in the output tree. * fix(eval): apply the candidate overlay after task setup for fair arms The candidate overlay was applied before the task's untrusted setup ran, so setup could observe candidate prose and the incumbent/candidate arms started from different pre-overlay state. Reorder within the sandbox: capture the base (pre-overlay) skill digest, run setup against the base skills, verify setup did not tamper them, then apply the overlay and capture the post-overlay digest the model must preserve. apply_candidate_overlay stages path-specific overlay files, so setup's uncommitted changes stay out of the baseline and churn is unchanged. Graph freshness for the review arm is handled by the status dirty-tree fix plus the review skill's stale-triggered re-index, not by reordering the cached per-task-sha graph materialization (which is mechanically blocked). * test(eval): end-to-end containment proof of the autonomous proposer Drives the real run_proposer through bubblewrap with a deterministic scripted model (no paid API): it reads the read-only evidence bundle and writes a candidate gitnexus-plan skill edit plus a rationale into the sandbox output tree; run_proposer enforces the trust boundary and copies only the validated overlay + proposal out. This exercises the autonomous-proposal stage of the self-evolution loop end-to-end in the eval/containment CI job (the gate and apply stages are covered by test_workflow_bench_evolution and test_promotion_apply). Env-gated on GITNEXUS_REQUIRE_CLAUDE_CANARY, so it runs only where the pinned Claude binary and user namespaces are available. * fix(eval): let the proposer author its overlay via Bash Running the end-to-end proposer canary in the containment CI job surfaced a real bug: run_proposer starts the session with --bare, which hard-disables the Write/Edit tools ("Write exists but is not enabled in this context"), yet allowlisted Edit/Write and omitted Bash. The proposer therefore had no working way to write its candidate overlay — the self-evolution loop could never produce a candidate. The sandbox settings already pre-authorize Bash (autoAllowBashIfSandboxed) and confine writes to workspace/tmp/home, so switch PROPOSER_ALLOWED_TOOLS to Read/Grep/Glob/Bash and tell the proposer to author files with Bash. The end-to-end test now drives the real run_proposer through bubblewrap and asserts a validated overlay + proposal are produced (this also replaces the earlier file-tool canary, whose Write/Edit premise was moot). * test(eval): author the proposer overlay with newline-free Bash content The nested shell-sandbox prefix mangles embedded newlines, so the multi-line overlay content never landed. Use single-line content for the deterministic proposer canary. * test(eval): drop the unverifiable end-to-end proposer canary The scripted proposer overlay never materialized in the containment job across runs, and the model tool-result content is not visible in CI logs, so the test cannot be finalized without an environment where the sandbox can actually run. Keep the verified production fix (Bash-authoring in run_proposer); the proposer sandbox/containment stays covered by the existing Bash+MCP and process-tree canaries. * test(cli): drop run-analyze.ts from the windowsHide spawn-family list U7 moved run-analyze.ts's only child_process call (the git status --porcelain dirty check) into storage/git.ts (already covered by this test, with windowsHide). run-analyze.ts no longer imports a spawn-family function, so the windowsHide-regression test's 'must have >=1 spawn call' invariant failed for it. Remove it from SRC_FILES. --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Zander Raycraft <zanderjraycraft@gmail.com> Co-authored-by: Azizur Rahman <azizur100389@gmail.com>	2026-07-19 15:07:24 +01:00