tinyvault-claude-review

Imported from jonathanavni/tinyvault (.agents/skills/tinyvault-claude-review/SKILL.md). Install upstream with npx skills add jonathanavni/tinyvault --skill tinyvault-claude-review. Copyright stays with the author.

Independent Claude Opus 5 review

Read AGENTS.md and the session protocol. Codex retains continuity; Claude is a fresh, read-only worker. Use the user's authorized review scope.

1. Hold the candidate checkout stable. Use a separate checkout if another session is editing it. Prepare a packet outside all source worktrees from the packet template. Supply exact base/head, files, locked requirements, threat model, accepted residuals, and available verification evidence. Untracked files are included in the candidate inventory; ignored files are not. Do not include real credentials or another reviewer's findings in a blind review packet. The helper automatically embeds the reviewed checkout's root CLAUDE.md, with a saved snapshot and hash. It must exist in the candidate inventory and be nonempty. Supply other required context in the packet; the embedded project guidance does not grant Claude orchestration or write authority.

2. Run the helper from the TinyVault root:

   node scripts/claude-review.mjs --repo /absolute/review-checkout --packet /absolute/task.md --channel qa --base BASE_SHA --output /absolute/new-report-directory

   Choose plan, qa, or security. Run each required channel as a separate invocation, with its own packet and output directory. security automatically includes the versioned security methodology; if the contract requires a specific external auditor, that remains a separate gate. The default timeout is 900 seconds; adjust explicitly with --timeout-seconds (1–3600) for a genuinely larger review.

3. The helper pins claude-opus-5 at high effort with no fallback. Use the existing Claude Code login. If sandbox access to that login/network is blocked, use the host's normal approval mechanism for this exact invocation. Never read/copy credentials or enable permission bypasses. Do not silently substitute a different model. Missing CLI/authentication/model access leaves the gate pending.

4. Keep the process/session identifier and let the bounded run finish. Inspect events.jsonl for live progress, then summary.json and report.md. Exit 0 means a completed PASS review; 2 means a completed NEEDS-ATTENTION review; 1 is dispatch/validation failure; 124 is timeout; 130 is interruption. No summary, a failed summary, or a killed outer process is never a successful review. Do not rerun merely because findings are inconvenient. Check an interrupted run has stopped before retrying.

5. Read the full findings and evidence. Synthesize them under the canonical ladder; record the reviewer, candidate digest/refs, channel, report path, and disposition in the appropriate append-only register when authorized as continuity owner. A successful dispatch is not proof the candidate is safe.

The reviewer can only Read/Glob/Grep; it cannot execute tests, mutate files, launch workers, or load custom commands/hooks/MCP tools. Give it already-produced diff and test artifacts. It must report missing dynamic evidence. These are CLI tool restrictions, not an OS-level sandbox or a repository-only read boundary. The candidate digest is a before/after drift check, not an atomic lock; freeze the checkout during review. Opus 5 is checked against actual assistant events; CLI auxiliary model usage is preserved separately.