audit-agent-history

Audit Agent History

Purpose

Use this skill to understand local agent history as a storage and continuity surface, not as a prompt dump. Prefer compact inventory, metadata, and narrow keyword searches before opening any transcript content.

This skill can prepare provider coverage notes or redacted evidence packets for an assessment, but it does not score AI fluency by itself.

For AI usage assessment from provider histories, use audit-agent-history -> evaluate-ai-usage-maturity. This skill produces a redacted Provider Evidence Packet, then evaluate-ai-usage-maturity scores the workflow from that packet and durable artifacts.

Trigger Examples

• "Claude/Codex 대화내역 기반으로 만들 만한 스킬 찾아봐."
• "히스토리 구조 개선할 수 있는지 봐."
• "예전 agent가 뭘 했는지 로그에서 찾아줘."
• "Codex, Claude Code, Antigravity 기준으로 반복 패턴 뽑아줘."
• "세션/아카이브가 너무 커졌는데 정리 방향만 제안해."

Safety Rules

• Default to read-only inspection.
• Do not delete, move, archive, compress, or quarantine history unless the user explicitly asks.
• Do not print secrets, OAuth tokens, raw credentials, or long transcript excerpts.
• Treat generated summaries as hints that must be rechecked against current files or live systems before execution.

Workflow

1. Identify provider homes and likely storage roots for Codex, Claude Code, and Antigravity, plus project-local state, memory indexes, and archived session folders.
2. Build a metadata inventory first:
   • file counts
   • total sizes
   • newest and oldest modification times
   • session/archive/index layout
   • notable config and state files
3. Search by path and keyword instead of loading broad histories. Use rg, find, jq, or provider-native indexes when available.
4. Open only the minimum number of matching files needed to answer the user's question.
5. Synthesize the result into actionable buckets:
   • reusable workflow patterns
   • stale or duplicated history surfaces
   • retention or cleanup candidates
   • candidate skills, agents, hooks, or docs
   • facts that require live revalidation
   • provider coverage gaps for assessment work

Search Strategy

Prefer narrow commands and stable metadata:

• Use find for path, size, and modified-time inventory.
• Use rg for task keywords, repo names, branch names, PR numbers, and exact error text.
• Use jq only after confirming the file is JSON or JSONL.
• Open one or two matching summaries before raw logs.
• Stop reading when you have enough evidence to answer the current question.

Candidate Extraction

When looking for reusable skills or plugin features, score each pattern:

• repeatability: Has the user asked for this more than once?
• risk: Does the pattern prevent a costly false success, data leak, bad deploy, or destructive command?
• portability: Can it apply across repositories or runtimes?
• non-obviousness: Would a fresh agent likely miss it without guidance?
• automation-fit: Is it better as a skill, script, hook, agent, or plain documentation?

Common Mistakes

• Loading entire chat logs into context before building an inventory.
• Treating memory-derived facts as current proof.
• Mixing provider runtime state with source-controlled configuration.
• Recommending deletion or archive actions without explicit user intent.

Output Contract

Provider inventory:
- Codex: ...
- Claude Code: ...
- Antigravity: ...
- Hermes/other: ...

High-signal patterns:
- ...

Candidates:
- skill/plugin/doc/hook/agent/script ...

Retention or cleanup notes:
- read-only recommendation ...

Revalidation needed:
- ...