[halindrome]

Summary

search_code matched file_pattern against the file basename only. Any pattern containing a directory, such as src/cli/*.c or apps/web/src/**/*.ts, silently returned 0 results, while search_graph matches the same pattern against the full repo-relative path.

Closes #2836

Change (src/mcp/mcp.c)

• New portable matcher, search_code_path_glob_match. It runs in linear time and doesn't use fnmatch, which Windows lacks.
  • * and ** match across /, and **/ also matches zero directories.
  • ? matches one character.
  • A pattern with no wildcards is a substring match, the same as search_graph.
• write_scoped_filelist: a file_pattern containing / is matched against the repo-relative path on every platform. Windows separators are normalised first. Patterns without / keep the existing basename behaviour.
• Windows: / patterns are scanned in process from the already-filtered list, not by the PowerShell -like post-filter. cbm_search_code_file_pattern_can_prefilter already rejects /, so nothing is filtered twice.
• When a search with a / pattern returns no results, the response carries a hint, in both tree and JSON output. The hint explains that the pattern must match the whole repo-relative path and points at path_filter.
• The schema now describes file_pattern ("Glob on file name; with '/', on repo-relative path (** = any dirs).") and path_filter ("Regex on repo-relative path.").

Acceptance criteria

1. A file_pattern containing / is matched against the repo-relative path, with ** crossing directories (including zero directories).
2. A file_pattern without / keeps the current basename semantics.
3. The behaviour is the same on POSIX and Windows for indexed (scoped) searches.
4. A / pattern that yields 0 results returns a hint pointing at path_filter.
5. The tool schema describes both file_pattern and path_filter.
6. Existing mcp suite tests stay green.

Known limits

• Unscoped searches (an unindexed root) still use find -name, so a / pattern there returns 0. The hint still fires.
• * crosses /, and matching is case-sensitive. Both are consistent with search_graph and with the existing basename match.

Testing

• New search_code_file_pattern_matches_repo_relative_path_issue2836 in tests/test_mcp.c, with fixture files src/main.c and src/cli/watcher.c:
  • src/cli/*.c → 1 match
  • src/**/*.c → 2
  • **/cli/*.c → 1
  • src/cli/ → 1
  • watcher.c → 1
  • *.c → 2
  • lib/*.c → 0, with the hint
• Red-first check: with the original mcp.c, the new test fails, because the four directory patterns return 0 and there is no hint.
• scripts/test.sh --suites mcp: 397 passed, 4 skipped, 0 failed.
• The Windows path was changed by reading only; it hasn't been run.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CDDUoWEZMNZ9fLvzp2pPGT

[halindrome]

QA round 1 for this change. It ran on the fork PR (halindrome#7) before the commits were squashed into 5a666f2, so any commit SHAs below refer to the pre-squash history. The QA was performed by Claude Opus 5.5 (Anthropic) in Claude Code.

QA Round 1

Scope: origin/main..HEAD on fix/search-code-file-pattern-path (closes #2836). The panel had 3 lenses: contract-security, regression-edges and test-quality. All 3 landed.

Contract Verification

#	Criterion	Verdict	Evidence
1	A file_pattern containing / is matched against the repo-relative path, with ** crossing directories (zero included), consistent with cbm_glob_to_like	✅ pass	search_code_path_glob_match (src/mcp/mcp.c:14826) is applied in write_scoped_filelist (mcp.c:16503-16511). */** cross /, **/ matches zero dirs, and a pattern with no wildcard is a substring match, the same as search_where_basic (#200). Test cases: src/cli/*.c, src/**/*.c, **/cli/*.c, src/cli/. Minor divergences that do not break the criterion: the match is case-sensitive where LIKE ignores ASCII case, and it is slightly stricter around /**.
2	A file_pattern without / keeps basename semantics	✅ pass	The basename fnmatch (mcp.c:16519) and the Windows prefilter are unchanged when !path_glob. Test cases: watcher.c, *.c.
3	Behaviour is the same on POSIX and Windows for scoped searches	✅ pass	The path glob runs in C before the #ifdef split, after separator normalisation. On Windows a / pattern is kept on the in-process scan (mcp.c:17536-17538), and the test is not platform-skipped. Not run on Windows in this review.
4	A / pattern yielding 0 results returns a path_filter hint	✅ pass	metrics.hint_path_filter (mcp.c:17710) is rendered in both the JSON and toon outputs, and the lib/*.c case asserts it. See Finding 1 for the wording.
5	The tool schema describes file_pattern and path_filter	✅ pass	mcp.c:732-734.
6	Existing mcp suite stays green; new tests are red without the change	⚠ partial	Red-without-change was confirmed by reading the code: basename fnmatch returns 0 for every / case and never emits the hint. Green was not verified, because lenses do not run the suite. The CI result is the evidence for this half.
7	Accepted limit: unscoped searches still use find -name	✅ pass	cbm_search_code_build_grep_cmd's unscoped branch is unchanged.

Finding 1 — path_filter hint claims file_pattern "matched nothing" when files matched but the search text did not

[claude:contract-security|regression-edges|test-quality] · minor · confirmed · src/mcp/mcp.c:17710

• Problem: metrics.hint_path_filter = file_pattern && strchr(file_pattern, '/') && gm_count == 0; is keyed on grep matches, not on whether the glob selected any files. SEARCH_CODE_PATH_HINT (mcp.c:14816) tells the caller the file_pattern "matched nothing" and points at path_filter. With a correct glob such as src/cli/*.c and a needle that does not occur, the caller is told to rewrite a correct filter.
• Expected: Fire the hint only when the / glob selected no files (scoped && scoped_written == 0, plus the unscoped find -name case), or reword it so it does not claim the glob matched nothing.
• Impact: A misleading diagnostic that wastes an agent round-trip. Results are never wrong, and the AC wording is literally met. About a one-line fix. All three lenses found this independently.

Finding 2 — Hint rendering in the default tree/toon output is untested

[claude:test-quality] · minor · confirmed · tests/test_mcp.c:11724-11735

• Problem: Every case in search_code_file_pattern_matches_repo_relative_path_issue2836 sends "format":"json". The new hint branch in assemble_search_output_toon (mcp.c:15479-15481) handles the default output format and is never exercised, so deleting it leaves the suite green.
• Expected: One tree-format case in the existing loop.
• Impact: Low. The branch is a 3-line mirror of the JSON one.

Summary

Severity	Confirmed
Critical	0
Major	0
Minor	2

No blocking findings. Contract: 6 criteria pass. 1 is partial, because suite-green is left to CI.

Observations

None.

---

SAST review not applicable

Target default has security_stage: false in the qa-cycle config. No CI
security stage is wired for this target, so no SAST/SCA delta is computed.

---

QA performed by Claude Code (claude-opus-5-5), manager + 3-lens panel

Dispositions (orchestrator, Step 3B)

• Fixed (minor): the hint said the / pattern "matched nothing", even when it had selected files that simply didn't contain the search text. The wording now reads: "no results; a file_pattern with '/' must match the whole repo-relative path (e.g. src/**/*.c); to scope by folder you can also use path_filter …". That is accurate in both cases. The hint still keys on zero matches, because the unscoped find path has no file count to key on.
• Not fixed (carried forward, minor): no test covers the hint in the tree (toon) output format. Both renderers read the same metrics.hint_path_filter flag, and the JSON branch is tested.

Verification

CC=/usr/bin/clang CXX=/usr/bin/clang++ SDKROOT=… scripts/test.sh --suites mcp → 397 passed, 4 skipped, 0 failed (exit 0). The existing test asserts the hint is present without pinning its wording, so it stays green. This is narrower than CI's default leg; the local default scripts/test.sh stops at contract Step 0t, which unmodified main also fails in this environment.

QA-Fix-Commit: 73ff216

---

Posted by the developer account (QA-agent token unavailable). Review by Claude Opus 5.5 (Anthropic) in Claude Code via qa-panel.

[halindrome]

QA round 2 for this change. It ran on the fork PR (halindrome#7) before the commits were squashed into 5a666f2, so any commit SHAs below refer to the pre-squash history. The QA was performed by Claude Opus 5.5 (Anthropic) in Claude Code.

QA Round 2

Scope: origin/main..HEAD on fix/search-code-file-pattern-path, at 73ff216f. That includes the round-1 fix commit 73ff216f, which rewords the path_filter hint. The PR closes #2836. The panel had 3 lenses: contract-security, regression-edges and test-quality. All 3 landed, and all ran on claude-opus-5-5 with CMM + context-mode navigation. The working tree was unchanged across the panel.

Contract Verification

#	Criterion	Verdict	Evidence
1	A file_pattern containing / is matched against the repo-relative path, with ** crossing directories (zero included), consistent with cbm_glob_to_like	✅ pass	search_code_path_glob_match (src/mcp/mcp.c ~14826) is applied in write_scoped_filelist when the pattern contains /. */** cross /, **/ matches zero dirs, and a pattern with no wildcard is a substring match, the same as search_where_basic. It is equal to or slightly stricter than cbm_glob_to_like (src/store/store.c:6910) around /**. Test cases: src/cli/*.c, src/**/*.c, **/cli/*.c, src/cli/.
2	A file_pattern without / keeps basename semantics	✅ pass	The basename fnmatch is gated only on !path_glob. The Windows prefilter (cbm_search_code_file_pattern_can_prefilter) is unchanged. Test cases: watcher.c, *.c.
3	Behaviour is the same on POSIX and Windows for scoped searches	✅ pass	The path glob runs before the #ifdef split, after separator normalisation. On Windows, handle_search_code keeps / patterns on the in-process scan. The test is not platform-skipped. Verified by reading; not run on Windows.
4	A / pattern yielding 0 results returns a path_filter hint	✅ pass	metrics.hint_path_filter is rendered in both the JSON (mcp.c:15722) and toon (mcp.c:15480) outputs, with no duplicate hint key. After the round-1 rewording, SEARCH_CODE_PATH_HINT (mcp.c:14816) is accurate whether the glob selected no files or the selected files lacked the text. The lib/*.c case asserts use path_filter is present, and the other cases assert it is absent.
5	The tool schema describes file_pattern and path_filter	✅ pass	mcp.c ~732-734.
6	Existing mcp suite stays green; new tests are red without the change	✅ pass	Red without the change: confirmed by reading the code. Basename fnmatch rejects every / case and the old code emits no hint. Green: lenses do not run the suite. The orchestrator ran scripts/test.sh --suites mcp at this head: 397 passed, 4 skipped, 0 failed.
7	Accepted limit: unscoped searches still use find -name	✅ pass	The unscoped branch of cbm_search_code_build_grep_cmd is unchanged. A / pattern there yields 0 results and the hint fires, which is the documented limit.

Summary

Severity	Confirmed
Critical	0
Major	0
Minor	0

No new findings. The round-1 fix (73ff216f) was reviewed as part of the full diff and introduced no regressions. All 7 contract criteria pass.

Observations

• TOON-format path_filter hint branch has no test (known, deferred from round 1 Finding 2) · minor · tests/test_mcp.c:11729-11733. Every case sends "format":"json", so the hint in assemble_search_output_toon (mcp.c:15480), the default format, is not exercised. This is low risk because it mirrors the JSON branch, which reads the same flag. It was raised again by test-quality and is listed here as carried forward, not as a new finding.

---

SAST review not applicable

Target default has security_stage: false in the qa-cycle config. No CI
security stage is wired for this target, so no SAST/SCA delta is computed.

---

QA performed by Claude Code (claude-opus-5-5), manager + 3-lens panel

Cycle outcome: round 2 is clean. There are no findings, all contract criteria pass, and the round-1 fix is reviewed. Carried forward: no tree/toon-format test for the hint. Not approved: no QA-agent token is available, and approval is the human reviewer's call.

---

Posted by the developer account (QA-agent token unavailable). Review by Claude Opus 5.5 (Anthropic) in Claude Code via qa-panel.

[github-actions]

Thanks for opening this — it has been seen, and it is queued.

This note is automated, but it is not a brush-off: it exists so you know where your PR stands instead of having to guess from silence.

Current review status: working through a backlog. 0.9.1-rc.1 is out, so the release freeze that held reviews is over — but it left a large queue of open pull requests behind it, and we are reading through them oldest-first. The background is in discussion #1144.

What that means for this PR, concretely:

• It will not be closed for inactivity. No stale bot touches pull requests here.
• It may still sit a while before a human reads it. That is on us, not on you.
• Older PRs are read first, so a recent one is not being skipped — it is behind a queue.

Things that will genuinely speed it up whenever review does happen:

• Keep it rebased on main — the tree is moving quickly right now, and a conflicting branch cannot be reviewed as the diff you intended.
• Get CI green, or say which failures you believe are pre-existing.
• Keep the change to one claim. Bundled features and refactors get split before they get merged, which costs you a round trip.
• Every commit needs a sign-off (git commit -s) — CI enforces DCO.

If this fixes a bug, a reproduction we can run is worth more than a description of the symptom.

Thanks for contributing, and sorry in advance for the wait.

[DeusData]

Thank you, Shane! The directory-aware file_pattern matching is a careful fix for #2836: it mirrors search_graph's path semantics (including ** spanning zero directories and the no-wildcard substring rule), keeps basename matching for patterns without '/', and routes '/' patterns through one portable matcher on every platform. The zero-result hint and the two short schema descriptions should save agents a wasted round trip.

One maintainer change so CI covers the current tree: we merged current main into your branch (merge commit 6f63e38, signed off). Main had moved on in src/mcp/mcp.c with an unrelated change, so the earlier green no longer covered the combined tree; the merge was clean and changes nothing in your diff. A fresh CI run is on its way, and we will follow up here once it is through. Please pull before pushing again so your branch keeps this commit.

[DeusData]

Merged as 30f1fbf - thank you, Shane! A search_code file_pattern that contains '/' now matches the repo-relative path with the same semantics as search_graph, patterns without '/' keep matching the basename, and a zero-result search says why. Thank you for the careful matcher, the tests, and for filing #2836 first.