Stats
Actions
Tags
From casegraph-plugin
Use when the user wants to review AI-performed work on a CaseGraph case. Trigger on phrases like "case <id> をレビューして", "AI がやった作業を確認して", "review the case", "audit this case", "証跡をチェックして", or whenever a structured read-only inspection of task completion, evidence integrity, decision traceability, and event-log trustworthiness is needed.
How this skill is triggered — by the user, by Claude, or both
Slash command
/casegraph-plugin:cg-review-driverThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Run a structured, read-only audit of a case after an AI has performed work on it. Produce a markdown review report classifying findings as must-fix (🔴), should-review (🟡), or note (🟢), and list external artifacts that still require human cross-check. Never mutate the graph — the only writes allowed are the report text and, if explicitly requested, one `evidence.attached` event at the end to ...
Run a structured, read-only audit of a case after an AI has performed work on it. Produce a markdown review report classifying findings as must-fix (🔴), should-review (🟡), or note (🟢), and list external artifacts that still require human cross-check. Never mutate the graph — the only writes allowed are the report text and, if explicitly requested, one evidence.attached event at the end to record the review itself.
Use this skill after cg-workflow-driver or another AI-driven loop has produced done tasks, new decisions, evidence nodes, and patch-applied events. For live workflow orchestration use cg-workflow-driver. For AI-authored graph changes use casegraph-patch. For raw workspace reading use casegraph.
Resolve a working launcher before using the commands below:
cg --help works, use cg.pnpm exec cg --help or npx cg --help.pnpm install and pnpm build, then use pnpm cg --help.@caphtech/casegraph-cli and use either global cg or project-local pnpm exec cg.In the rest of this skill, cg ... means "use the launcher that succeeded here."
Use this skill when:
Skip it for a case that has only human edits, or for a case with fewer than a handful of nodes.
--since-revision <n> or --since-timestamp <ISO8601> to the skill to ignore history before that point. Useful on long-running cases to avoid re-reviewing everything.evidence node on the case root via cg evidence add. Default off.Out of scope in v0.1: batch review across multiple cases, diff reviews between two revisions, automated Phase F execution.
Run in order. Each phase reads from --format json and pipes through jq — never emit mutating commands. Detail in review-phases.md.
cg case show --case <id> --format json → read revision.current, case.state. Then cg case view --case <id> --format json → full nodes[] / edges[] / derived[] / validation[] arrays that later phases query. If --since-revision is set, scope all later event-log queries to events newer than that revision.cg validate --case <id> --format json (errors must be 0), cg events verify --case <id> --format json (HARD STOP if integrity fails — the graph itself is not trustworthy), then cg frontier --case <id> --format json and cg blockers --case <id> --format json.kind:decision in state:done should have metadata.result or a non-empty description. Flag naked decisions.cg events export --case <id> --format json and apply event-trail-jq.md patterns to classify patches by generator.kind, detect orphan workers, list cancel/fail transitions.| Verdict | Condition |
|---|---|
| 🔴 Red | Phase B failed (event log verify or validation errors > 0), or any kind in {task, decision} with state: done has no verifies edge, or any worker.dispatched event lacks a matching worker.finished. Short-circuit: Red results cause later phases to be reported but not trusted. |
| 🟡 Yellow | Any just-in-time evidence, naked decision, AI-authored (generator.kind in {agent, worker}) patch.applied without an accompanying evidence.attached event in the same revision range, or state: open with a non-empty frontier when the user's prompt implied the case was done. |
| 🟢 Green | No Red or Yellow findings. Only 🟢 notes and "Manual cross-check required" items remain. A Green verdict is not an automatic approval — it means the graph side passes; the human still owns Phase F. |
Emit a single markdown document. Do not split into multiple files. Shape is fixed by report-template.md. Keep it pasteable into a PR description or a chat thread.
When the user asks to record the review, attach it via:
cg evidence add --case <id> \
--id evidence_review_<ISO8601_compact> \
--title "Review report at revision <n>" \
--target <case_root_or_most_relevant_goal_id> \
--description "<full markdown report or summary + link to where the full text is stored>"
This is the only mutating command the skill ever emits, and only on explicit request.
cg case close stays with cg-workflow-driver.git log, reruns tests, opens PRs.--since-revision is set — scoping is strict so a long case stays reviewable.cg patch apply, cg task ..., cg node ..., cg edge ..., or any other mutating verb. Its entire output is read queries plus (optionally) one cg evidence add at the end.GraphPatch: casegraph-patchnpx claudepluginhub caphtech/casegraph --plugin casegraph-pluginCreates bite-sized, testable implementation plans from specs or requirements, with file structure and task decomposition. Activates before coding multi-step tasks.