Spec 19: skill recommender — 'you've manually run X 47 times, want a skill?'

[0bserver07]

Goal

Detect repetitive patterns the user manually re-runs across sessions and prompt to convert them into auto-skills via the existing skills generate machinery.

Why now

skills generate (v0.7.0) mines the store for patterns and writes SKILL.md files. The discovery is reactive — user has to remember to run it. This is the proactive surface.

Schema

None. Read-side over messages + message_tool_mart. Cache recommendations in a JSON file under ~/.stackunderflow/cache/skill_recommendations.json (existing TieredCache pattern).

User-visible surface

• CLI: stackunderflow recommend skills [--project P] [--threshold 5] [--format text|json] — list of "you ran X N times, here's the skill that would replace it".
• MCP tool: recommend_skills(project?, threshold?).
• Meta-agent tool: same. The active-surfacing flow (Spec 27) will use this.
• UI: notification badge on the Overview tab when there are pending recommendations.

Implementation plan

1. New service stackunderflow/services/skill_recommender.py:
   • mine_repeat_patterns(conn, project=None, threshold=5, window_days=30) — returns list of {pattern, occurrences, sessions, suggested_skill_template}.
   • Reuses skill_synth.py pattern-mining helpers (canonical-test-command, always-runs-X-after-Y, etc.); difference is the gate (occurrence threshold) and the surface (recommendation, not generation).
2. CLI command + MCP tool + meta-agent tool.
3. Optional: tie the "accept" flow to a skills generate --pattern <id> that writes the actual SKILL.md.

Tests

• Synthetic store with 7 repeats of pytest tests/ -q → recommender suggests a canonical-test-command skill.
• Threshold gate (4 occurrences below threshold of 5 → no recommendation).
• Per-project scope (default-on; --scope user opts into cross-project).
• Cache TTL.

Hard parts

• Avoid re-recommending patterns the user already has skills for. Check <project>/.claude/skills/auto-*/ and ~/.claude/skills/ before surfacing.
• Don't recommend skills for patterns where the agent has already failed (use spec 22 outcome data when available; for now, ignore failed sessions in pattern mining).

Out of scope

• Auto-applying recommendations without user confirmation.
• LLM-assisted skill text generation (v2 — once meta-agent + Ollama are stable).

Dependencies

• None blocking. Spec 27 (active surfacing) will use this.

Estimated effort

Size M — single agent, ~1 hr.

Hard rules

• DO NOT touch versions / CHANGELOG headings.
• Branch: feat/skill-recommender off main.