Add Cursor agent transcript adapter

[MoMitwalyIgniteTech]

Why

Cursor sessions were invisible to watchmen — the wishlist pencilled a Cursor adapter in as SQLite/state.vscdb reverse-engineering, but Cursor's agent now writes plain JSONL transcripts under ~/.cursor/projects/<slug>/agent-transcripts/, which makes coverage much cheaper than expected.

What changed

• New src/watchmen/adapters/cursor.py following the NAME/discover()/scan() contract:
  • prompts extracted from <user_query> tags; editor-injected context tags (<attached_files>, <plugin_info>, …) excluded from prompt mining
  • human-format <timestamp> tags (Tuesday, Apr 28, 2026, 10:15 AM (UTC+3)) parsed to UTC ISO; older transcripts carry no tags at all, so scan() falls back to file mtime to keep sessions datable
  • <manually_attached_skills> entries recorded as pseudo-Skill tool calls — Cursor inlines the skill content, so the SKILL.md-read detection in _shared alone would go blind; regular tool_use paths still go through extract_skill_from_args
  • project-dir slugs decode via the existing decode_project_dir (same dash-flattened encoding as Claude Code, minus the leading dash)
  • the format records no tokens, model names, or tool results — those columns honestly stay at defaults
• Registered in ADAPTERS; cr / Cursor / violet added to the display touchpoints (util.py, metrics.py, insights.py, pipeline.py)
• Fixture-driven tests in tests/test_adapter_cursor.py + tests/fixtures/cursor_session.jsonl (12 tests: timestamp parsing incl. noon/midnight/negative/fractional offsets, prompt extraction, skill attribution, mtime fallback, discovery walk, corrupt-file tolerance)
• test_unknown_agent_falls_through_to_raw_slug now uses windsurf as its unknown slug — cursor stopped being one
• CONTRIBUTING refreshed: the stale "Cursor stores sessions in state.vscdb, no hooks" entry replaced with the two real remaining frontiers (IDE state.vscdb chat history, Cursor 1.7+ hooks host in hooks_setup.py)

Testing

• uv run pytest tests/ — 300 passed
• uv tool run ruff check src/watchmen tests — clean
• uv build — wheel + sdist build
• Validated against a real corpus: 126 transcripts across 5 projects (WSL2), 0 parse errors → 979 prompts, 9,310 tool calls, all project dirs decoded to real paths, skill attribution populated (coderabbit-respond ×34, search-company-knowledge ×14, …); watchmen ingest + mission control render the new agent slice end-to-end

Notes

• The transcript schema is reverse-engineered and undocumented — Cursor may change it between releases. The adapter degrades gracefully (unparseable lines skipped, missing install silent).
• No privacy implications beyond the existing adapters: reads local transcripts only, same as claude_code/codex/pi.
• No release/migration/plugin-install implications; the corpus schema is untouched (agent column already generic).

🤖 Generated with Claude Code

[aktasbatuhan]

Thanks for this, genuinely clean work. I pulled it down and ran it locally and it holds up well:

• ruff clean, full suite green (529 passing on my machine, the "300" in the description is just stale)
• the two-source design is the right call: one adapter, discover() fans out to both the IDE chat DB and the agent-transcript JSONLs, and scan() dispatches on entry shape. Composer ids stay namespaced cursor/<id> and transcripts use bare UUID stems, so they can't collide in corpus.db. I checked that against the session_id PRIMARY KEY + INSERT OR REPLACE upsert and it's solid.
• the schema notes in the header are excellent, and I appreciate the honest defaults: no invented tokens/models/cost, mtime fallback so undateable transcripts still show up, half-hour UTC offsets handled even though they don't appear in real data yet.

I just approved the workflow run so CI can go. This was your first PR here so GitHub held it for manual approval, nothing on your side.

A few small, non-blocking things if you feel like it. None of these gate the merge:

1. The description still reads like a brand new file, but the rebase landed it as an extension of the existing SQLite adapter. A one-line tweak would make the history read cleaner.
2. Cross-store question: if a single Cursor agent run ever gets persisted in both the IDE chat DB and the agent-transcript JSONL, we'd surface it twice under two ids. You document the stores as disjoint surfaces, which matches what I'd expect, just flagging it in case you've seen any overlap in your own corpus.
3. _scan_transcript builds its session dict inline rather than going through _empty(). Minor drift risk if the schema columns change later, totally optional.

Once CI is green I'll merge. Thanks again for picking this up.

[aktasbatuhan]

Verified locally and CI is green across all platforms including Windows. Clean two-source design, solid docs. Merging. The optional follow-ups above can be a future PR if you want them.