fix(loaders): strip git log -p commit metadata before parsing

[gonzaloserrano]

Piping git log -p (or git show -p of multiple commits) through hunk's pager mode floods stderr with hundreds of
parseLineType: Invalid firstChar / processFile: invalid rawLine warnings from @pierre/diffs.

Root cause: pager mode's looksLikePatchInput heuristic accepts the stream because it contains diff --git and @@, then forwards the whole blob to parsePatchFiles. But @pierre/diffs is a strict patch parser: every hunk-body line must start with +, -, or \. The commit <sha>, Author:, Date:, blank, and 4-space-indented commit message lines that appear between commits all fail that check, and the parser logs each rejection unconditionally (no quiet flag).

Fix in normalizePatchChangeset rather than in pager.ts so all input paths benefit (a saved git log -p fed via hunk patch file.txt would have hit the same bug). The new stripGitLogMetadata:

• Fast-paths to a no-op when no ^commit [0-9a-f]+ boundary exists, keeping the regular patch path zero-allocation.
• State-machines per line: enters "header" mode on a commit boundary, drops every following line until a patch start (diff --git , --- , +++ ).
• Preserves context lines that mention "commit" textually — they start with the diff line-type marker ( /+/-), so the ^commit regex doesn't match.

Tests cover: single-commit, multi-commit, decorated headers (refs in parens), abbreviated SHAs, --stat blocks, merge metadata, context lines mentioning "commit", trailing diff-less commit, and an end-to-end check that the stripped output round-trips through parsePatchFiles with zero Invalid firstChar warnings.

[greptile-apps]

Greptile Summary

This PR adds stripGitLogMetadata to loaders.ts to strip the per-commit envelope lines (commit <sha>, Author:, Date:, indented message body) that git log -p emits between patch hunks, so the remaining text is a clean patch stream that @pierre/diffs can parse without flooding stderr with parseLineType: Invalid firstChar warnings. The fix is placed in normalizePatchChangeset so it applies to every input path (pager mode and direct file input alike).

• stripGitLogMetadata fast-paths to a no-op when no commit <sha> boundary is detected, keeping the hot path zero-allocation for ordinary patch input.
• The state machine exits header mode on the first diff --git, ---, or +++ line and does not disturb hunk-body lines whose leading +/-/ already prevents the ^commit regex from matching.
• A comprehensive test suite covers single-commit, multi-commit, decorated refs, abbreviated SHAs, --stat blocks, merge metadata, context lines containing the word "commit", and an end-to-end round-trip through parsePatchFiles asserting zero noisy warnings.

Confidence Score: 4/5

The core logic is correct and well-tested; the only gap is that SHA-256 repositories will silently bypass stripping and reintroduce the original warning noise.

The state machine is straightforward and the test suite is thorough. The SHA regex upper bound of 40 means the fix does nothing for git repos using SHA-256 object format (64-character hashes), restoring the original bug in that case. There is also a minor duplicate regex literal in the loop body that could be unified with the named constant above it.

src/core/loaders.ts — specifically the two occurrences of {4,40} in stripGitLogMetadata.

Important Files Changed

Filename	Overview
src/core/loaders.ts	Adds stripGitLogMetadata to remove git log -p commit boundary lines before patch parsing; the SHA regex range {4,40} won't match 64-char SHA-256 hashes, and the same regex is duplicated in the loop body.
src/core/loaders.gitLog.test.ts	New test suite covering strip-metadata scenarios including single/multi-commit, decorated headers, abbreviated SHAs, --stat blocks, merge metadata, context lines, trailing diff-less commits, and an end-to-end round-trip through @pierre/diffs.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["patchText input"] --> B["replaceAll CRLF → LF"]
    B --> C["stripTerminalControl"]
    C --> D["stripGitLogMetadata"]
    D --> E{{"commit sha boundary present?"}}
    E -- No --> F["return text unchanged (zero-allocation fast-path)"]
    E -- Yes --> G["split into lines, inHeader = false"]
    G --> H["for each line"]
    H --> I{{"line matches ^commit [0-9a-f]{4,40}"}}
    I -- Yes --> J["inHeader = true, skip line"]
    J --> H
    I -- No --> K{{"inHeader?"}}
    K -- No --> L["push to output"]
    L --> H
    K -- Yes --> M{{"starts with diff --git / --- / +++"}}
    M -- No --> N["skip line (metadata)"]
    N --> H
    M -- Yes --> O["inHeader = false, push to output"]
    O --> H
    H -- done --> P["join lines → clean patch"]
    P --> Q["parsePatchFiles"]
    F --> Q

Loading

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
src/core/loaders.ts:86-96
**SHA-256 repositories not covered by regex range**

The pattern `[0-9a-f]{4,40}` caps at 40 hex characters, which covers SHA-1 hashes, but SHA-256 object hashes (used when a repo is initialised with `git init --object-format=sha256`) are 64 hex characters. For any such repo, neither the fast-path `COMMIT_BOUNDARY` test nor the per-line check in the loop will ever match, so the function returns the input unchanged and all `commit <64-char-sha>` boundary lines are fed to `@pierre/diffs`, reproducing the original warning spam. Widening both occurrences of the range to `{4,64}` covers both formats.

### Issue 2 of 2
src/core/loaders.ts:96
**Duplicate regex literal on every loop iteration**

`COMMIT_BOUNDARY` (line 86) was defined precisely for this test, but a new equivalent regex literal without the `m` flag is constructed inline here. Since each `line` from `split("\n")` is already a single string with no embedded newlines, `^` anchors to the start of the string regardless of the `m` flag, so both regexes behave identically per line. Reusing `COMMIT_BOUNDARY` directly (or extracting a module-level constant) removes the redundancy and makes future maintenance simpler.

_{Reviews (1): Last reviewed commit: "fix(loaders): strip git log -p commit me..." | Re-trigger Greptile}

[benvinegar]

Thanks!