feat(git): extend agent-provenance detection channels

[RaghavChamadiya]

What

Extends the deterministic agent-provenance classifier in the git indexer with three new AI-authorship detection channels. All are zero-LLM, local-history only, and layered onto the existing precision-first tier model without changing the shared commit-record format.

1. git-ai authorship notes (refs/notes/ai)

Reads the git-ai standard v3.0.0 authorship notes attached to commits. The note's sessions.<id>.agent_id.tool metadata resolves to the dominant agent for that commit (git_ai_note channel, tier 2). A tie between distinct tools stays unlabelled.

Read out of band via a ref-gated pass: repos without the ref do a single for-each-ref and add no git pass, so the common case pays nothing. Only repos that actually use git-ai walk the notes.

2. Assisted-by / Generated-by attribution trailers

Matches the Fedora/OpenInfra Assisted-by: and Apache Generated-by: commit trailers. The trailer key is a standardized AI marker, but the agent identity comes from the value, so these fire only when the value maps to a recognized agent. A bare human name in a misused trailer never mints a phantom agent. Generated-by is tier 2, Assisted-by tier 3.

3. Cursor commit-local markers + registry refresh

Adds the Cursor HTML-comment marker and "Generated with Cursor" footer, plus the gemini-code-assist bot login.

Validation

Ran the classifier over two real histories:

Repo	Commits	Attributed	Notable
This repo (false-positive check)	569	5 (claude 4, copilot 1)	all via genuine Co-authored-by; 0 git-ai notes; no false positives
git-ai project (true-positive)	4041	1652 (41%)	335 commits carry real ai notes; new git_ai_note channel attributes 326, the rest deferring correctly to higher-precedence bot-identity channels

Agent breakdown on the git-ai repo: claude 1286, devin 206, codex 107, copilot 29, cursor 24. Note tools present: codex 107, claude 196, devin 11, cursor 21.

Tests

Adds coverage for the note parser (single tool, alias mapping, dominant-tool, tie ambiguity, malformed input), the two trailer channels, the bare-human-name precision guard, Cursor markers, tier precedence, and the ref-gated notes loader (absent ref never spawns the walk). Full git-indexer suites pass.

[repowise-bot]

✅ Health: 7.7 (unchanged)

📋 At a glance
1 file changed health · 3 hotspots touched · 5 new findings introduced · 5 co-change pairs left out.

Files & modules (2)

• packages (2 files)
  • .../git_indexer/agent_provenance.py
  • .../ingestion/git_commit_index.py
• tests (1 file)
  • .../ingestion/test_agent_provenance.py

⚠️ Change risk: moderate _{(riskier than 51% of this repo's commits · raw 9.0/10)}
This change's risk is driven by:

• more lines added than baseline
• more scattered than baseline

File	Score	Δ	Why
.../git_indexer/agent_provenance.py	6.9 → 6.4	▼ -0.5	🔻 introduced complex method, bumpy road, error handling

💡 .../git_indexer/agent_provenance.py: Reduce cyclomatic complexity. Decompose the function along its conditional axes; if it dispatches on a type tag, replace the if/elif ladder with polymorphism or a lookup table.

🔎 More signals (2)

🔥 Hotspots touched (3)

• .../git_indexer/agent_provenance.py — 1 commits/90d, 4 dependents · primary owner: Raghav Chamadiya (100%)
• .../ingestion/test_agent_provenance.py — 1 commits/90d, 0 dependents · primary owner: Raghav Chamadiya (100%)
• .../ingestion/git_commit_index.py — 6 commits/90d, 5 dependents · primary owner: Raghav Chamadiya (100%)

🔗 Hidden coupling (1 file)

• .../ingestion/git_commit_index.py co-changes with these files (not in this PR):
  • .../git_indexer/indexer.py (3× — 🟢 routine)
  • .../git_indexer/records.py (3× — 🟢 routine)
  • .../persistence/models.py (3× — 🟢 routine)
  • .../git_indexer/commit_rows.py (2× — 🟢 routine)
  • tests/unit/test_git_commit_rows.py (2× — 🟢 routine)

---

📊 Full report · ⭐ Star Repowise · 📥 Install bot · Last updated 2026-07-08 11:37 UTC
_{Silence on a single PR with [skip repowise] in the title · Per-repo toggle on repowise.dev/settings?tab=bot}