llm-cli-gateway: Skill for Claude Code

Retrospective Walk

Guide a change review as a narrative with evidence. Prefer this when the user needs the story of a change set, not just hunk-level inspection.

Inputs

Accept these fields when the user supplies them, and infer conservative defaults when they do not:

{
  "scope": "diff | commit-range | worktree | job-id | episode-ref",
  "target": "diff text, HEAD~5..HEAD, path, job id, or episode reference",
  "mode": "guided | agent-driven | summary-only",
  "capture_comments": true,
  "models_for_why": [],
  "include_prior_evidence": true
}

Default mode to guided for a human in the loop, agent-driven for automated review, and summary-only only when explicitly requested or when interaction is not possible. Default capture_comments to true.

Scope Resolution

Resolve the target into immutable evidence before analysis:

• diff: use the supplied diff text. If the target is empty, capture git diff --no-ext-diff and git diff --cached --no-ext-diff.
• commit-range: capture git diff --stat <range>, git diff --name-status <range>, and per-commit metadata from git log --format=fuller <range>.
• worktree: capture status, dirty file list, staged diff, and unstaged diff. Do not stash, reset, or mutate the target worktree. If deeper model/tool analysis may write files, create a disposable isolated worktree or analyze a read-only evidence packet instead.
• job-id: fetch llm_job_status and llm_job_result; include cli, correlationId, sessionId, timestamps, exit status, captured output, and the bounded normalized progress snapshot actually returned. Page progress by afterProgressSeq when its event history matters, and record its capability and dropped count. Do not treat privacy-safe progress messages as raw reasoning, tool arguments, paths, or output evidence. Do not invent an output digest when the result has none. Personal Agent Config Kit deliberately withholds compiled context, provider output/error, and native handles from durable history; after a restart its job result may therefore be an explicit withheld marker rather than reconstructable output.
• episode-ref: resolve through the caller's available episode/DAG context. If no episode tool is available, record the unresolved reference and continue from any linked diff, job, receipt, or review artifacts the user supplied.