[bhavsarpratik]

What

Follow-up to #139. It makes commit work-type labels more accurate for commits without a conventional prefix, and closes the gaps that PR listed.

• More evidence for the model. For each commit, the local model now also sees:

  • the first body line that isn't a trailer;
  • a content-free line counting changed files by kind, plus total added and deleted lines, for example Changed files: 3 code, 1 tests (+120 / -30 lines).

  File paths and names never leave the Git scan. The question version moves from c1 to c2, so existing labels are redone once.

• Conventional prefixes are taken as written. A lowercase Conventional Commits type is the label, with no model call and no queue slot. The types are feat, fix, docs, style, refactor, perf, test, build, ci and chore. The model still handles capitalised prefixes (Fix:), package-style prefixes (testing: or config: in Go or Linux history), and revert:.

• MCP git_stats. It now returns the estimated split by work type, in the summary and in each day row. Keys are a fixed set and values are counts only.

• Edge cases handled:

  • Common Git and Gerrit trailers are skipped: any …-by:, Change-Id, Cc, and Fixes: or Refs: followed by an id.
  • Renames count under their new name, and quoted non-ASCII paths are decoded.
  • When a commit is too big to list its files in time, it is labeled from its message alone, and the model is told "Changed files: not listed." rather than "none".

• Tests. A failed rev-list in the backfill walk now has a test. The commit-type tests move to tests/test_git_commit_types.py to keep tests/test_git_delivery.py within its size budget.

Accuracy

The reference is each commit's own conventional-commit type, with the prefix stripped from the subject. It's a proxy for free-form messages, and the only labeled reference available.

The data is 664 of my own non-merge commits from the last 180 days, across 3 repos, labeled by token-meter-gemma. Commits are split by hash. Variants were chosen on 120 dev commits and scored once on 200 held-out commits, using both option orders.

Model input	Held-out agreement	With a body (167)	No body (33)
Subject only (c1, shipped in #139)	57.5%	61%	39%
Subject + first body line + file kinds (c2)	72.0%	73%	67%

On the dev split:

Tried	Result	Shipped
File names on top	65% with or without them	No
Prior correction	Under 1 point gain	No
Body line only	64.2%	Superseded by body + kinds
File kinds only	65.0%	Superseded by body + kinds

The no-body row is small (about ±16 points), so it shows direction only. The scratch data and the harness read private Git history, so both were deleted; only these numbers were kept. specs/WORK_INSIGHTS.md documents the method and its caveats.

Privacy: needs owner sign-off

This widens what Work insights may read. The rule in specs/AGENTS.md (and root AGENTS.md) now also allows:

• the first body line (200 characters);
• content-free counts of changed file kinds and lines.

Both sit alongside the commit subject. They stay in the in-memory queue and are sent only to the loopback model. Only labels persist. SECURITY, ARCHITECTURE, USER_GUIDE, README and the Git tooltip are updated to match. Tests check that no path, file name or body text reaches the classifier arguments or any stored file.

Validation

• Full suite on 622683d: python3 scripts/run-tests exits 0, with 2587 tests passed and 21 skipped. scripts/check-page, py_compile, git diff --check, and the design-system, agent-workflow and size-budget contracts all pass.
• Gates: one tester and two reviewers (privacy/persistence lens and correctness/evaluation lens) over five rounds, ending with no open findings. The tester mutation-checked every new guard. Each of these mutations is caught by a test:
  • remove a rename-resolution step;
  • narrow the trailer rule;
  • remove the fallback;
  • remove the prefix generation fence or the off/pause guard;
  • make prefixes case-insensitive.
• Real-model run through the production path on 12 free-form commits: all 12 labeled, mean confidence 0.88.
• Installed runtime: ./scripts/install reports installed commit 622683d. Both LaunchAgents are running, /health and /menubar return OK, and source and runtime match for every changed runtime file. Live /git-delivery shows c2 relabeling under way: 108 of 343 commits after the first scan.

Known limits (reviewed and accepted)

• Accuracy figures come from one author on one machine with a proxy reference, so real free-form messages may score lower.
• A few plain-English lines are dropped as trailers, such as Fixes: defaced banner or Fixes: 3 flaky tests. The model then gets the next body line.
• File kinds can be wrong for renames with nested literal braces (dir/{a{1} => b}/x.py). This affects only the content-free count.
• After updating, the chart shows earlier labels as not labeled yet until they are redone, at up to 200 commits per 5-minute scan. Prefixed commits are done without model calls.

🤖 Generated with Claude Code

[bhavsarpratik]

Owner sign-off on the privacy rule change: Pratik approved widening the Work insights rule so the local model may also read each pushed commit's first body line and content-free counts of the kinds of files and lines it changed. As before, these are sent only to the loopback model and never stored; paths and file names never leave the Git scan.

— Posted by Pratik's agent