Agent Session Pattern Miner

Agent Session Pattern Miner

Use this skill to turn local agent history into a ranked skill backlog without exposing raw chat logs by default.

Workflow

1. Define scope:
   • Confirm whether to inspect Codex, Claude, or both.
   • Use default local sources unless the user gives explicit alternate paths.
   • Treat local chats as private evidence; aggregate first, quote only when necessary.
2. Search first:
   • List existing repo skills and manifests.
   • List installed local skills if the user asks whether a candidate already exists.
   • Mark each finding as new skill, enhance existing skill, reference-only, or not worth packaging.
3. Mine local records:
   • Run the helper when available:

python3 skills/agent-session-pattern-miner/scripts/mine_agent_sessions.py --source both

• Use --source codex or --source claude for a narrower pass.
• Use --format json when another script or test needs structured output.
• Do not use --include-examples unless raw excerpts are necessary and safe to show.

4. Normalize findings into workflow candidates:
   • Trigger phrase or task shape.
   • Repeat count and session spread.
   • Affected repos or project families.
   • Existing skill overlap.
   • Risk level, human gates, and verification needs.
5. Rank by leverage:
   • High frequency across repos beats one large session.
   • Workflows with repeated verification or release risk outrank content-only variants.
   • Enhance an existing skill when the trigger and procedure already match.
   • Create a new skill only when the trigger, workflow, gates, and done-when differ.
6. Report:
   • Start with the top 3-7 candidates.
   • For each, state action, why now, suggested skill name, and first implementation slice.
   • Separate evidence from interpretation.
   • Say when the scan is partial or a source was unavailable.

Candidate Rules

• new skill: repeated task shape, clear trigger, reusable workflow, and a distinct done-when.
• enhance existing skill: same trigger and workflow family, but missing source adapter, gate, test, or helper.
• reference-only: useful background that does not need a separate invocation surface.
• not worth packaging: one-off task, project-specific detail, or unstable behavior that has not been manually validated.

Local Sources

• Codex user history: ~/.codex/history.jsonl.
• Codex thread titles: ~/.codex/session_index.jsonl.
• Codex rollout metadata: ~/.codex/sessions/**/rollout-*.jsonl.
• Claude Code project logs: ~/.claude/projects/**/*.jsonl.
• Claude Code task records: ~/.claude/tasks/**/*.json.

When a source format is unknown, inspect a small sample first and update the extraction logic only after confirming the actual fields.

Safety

• Redact or omit secrets, tokens, cookies, keys, and private customer data.
• Aggregate by default; do not paste raw chat transcripts.
• Do not create or edit skills unless the user asks to implement the candidates.
• Do not treat historical logs as current truth; verify live repo, GitHub, runtime, and test state before claiming anything landed.

Failure Modes

• If a history path is missing, report it as unavailable and continue with remaining sources.
• If records cannot be parsed, report the file count and parse error class without dumping file contents.
• If candidate evidence is dominated by one repo, label it project-specific until another repo confirms reuse.
• If an existing skill already covers the workflow, recommend enhancement instead of a duplicate.

Done When

• The report identifies which workflows should become skills, which existing skills should be enhanced, and which findings should be ignored.
• Each recommended action has evidence counts, source scope, risk level, and a concrete first implementation slice.
• Raw private conversation text is excluded unless the user explicitly requested and it was redacted.

Resources

• scripts/mine_agent_sessions.py: local Codex/Claude aggregate miner.
• references/pattern-manifest.json: lightweight Agent Patterns index metadata.