session-titles

Overview

This skill owns the entire session title lifecycle:

1. Generation -- Stop hook extracts context from the transcript (primary request, branch, files) and calls Haiku to produce a 4-7 word active-voice title. Detects focus shifts and tracks them with a (N) prefix.
2. Feedback -- Each generated title is saved as a pending feedback entry for later scoring.
3. Rating -- Interactive workflow where an AI judge scores first, then the human confirms or corrects. Builds dual-perspective training data.
4. Evaluation -- Pattern checks (fallback, meta-language, too long, etc.) plus optional LLM judge scoring across all pending entries.
5. Evolution -- GEPA-inspired prompt mutation: reflect on failures, propose targeted changes, keep improvements.
6. Golden dataset -- Extract candidates from real sessions, curate ideal titles, run regression evals.

How It Works

Session ends
  --> Stop hook (hooks/stop.sh)
    --> scripts/generate.ts  (stdin: session_id, cwd, transcript_path)
      --> generate-core.ts
        1. extractSessionContext()  -- parse transcript JSONL
        2. evolveTitleWithContext()  -- initial title or shift detection via Haiku
        3. sanitizeTitle()          -- strip preambles, enforce length
        4. savePendingFeedback()    -- append to title-feedback/pending.jsonl

Scripts

All scripts support --help style flags. Run with bun.

Script	Purpose
generate.ts	Hook entry point. Reads JSON from stdin.
generate-core.ts	Core module: context extraction, title generation, shift detection.
generate-core.test.ts	Unit + integration tests. bun test scripts/generate-core.test.ts
schema.ts	TitleFeedback types, prompt version constants.
store.ts	JSONL persistence for pending/scored feedback.
eval-quality.ts	Pattern checks + optional --judge LLM scoring.
evolve-prompt.ts	GEPA evolution. --iterations N, --pareto-size N.
extract-candidates.ts	Pull test cases from session transcripts. --limit N, --project NAME.
run-eval.ts	Run golden dataset eval. --judge-model MODEL.
report.ts	Generate report from latest eval results. --file PATH.

Rate Title

Interactive rating workflow (invoke as /rate-title or manually):

1. AI Judge assesses first -- Score (1-5), reasoning, proposed better title.
2. Human calibrates -- Agree? Different score? Better suggestion?
3. Both perspectives saved to ~/.claude/title-feedback/scored.jsonl.

The dual-perspective data enables DSPy optimization of both the judge prompt (learn to rate like the human) and the journalist prompt (generate titles humans rate highly).

Rating Scale

Score	Meaning
5	Perfect -- specific, actionable, concise
4	Good -- minor phrasing improvements possible
3	Acceptable -- gets the gist but generic
2	Poor -- too vague or wrong focus
1	Bad -- completely off-base or misleading

See references/scoring-rubric.md for detailed criteria.

Data Layout

Runtime data (gitignored, at ~/.claude/title-feedback/):

• pending.jsonl -- written by Stop hook
• scored.jsonl -- written by /rate-title

Evaluation data (gitignored, at skills/session-titles/data/):

• candidates.jsonl -- extracted test cases
• golden.jsonl -- curated with ideal titles
• results/ -- timestamped eval outputs
• baseline-*.md, evolution-*.md -- quality reports

References

• references/scoring-rubric.md -- 5-point rating criteria
• references/adaptive-title-plan.md -- Roadmap: phases, LanceDB vectors, DSPy optimization