carrasco-review

Carrasco Review

The harness reviews code by technology and saves the result under .harness/reviews/. This skill makes that review rigorous: it enforces the project's established standards as hard requirements and uses aggressive subagents (carrascos) — one per chunk of changed files — that respect the host workspace's architecture. Everything is configured via .harness.config.json (reviewAggressiveness).

Required Start

Announce: I'm running the carrasco-review skill.

Configuration (.harness.config.json → reviewAggressiveness)

• enabled — master switch. Defaults to false: this skill only runs the automated Stop-hook gate when a project opts in via .harness.config.json. Invoking the skill directly (e.g. "carrasco review") always works regardless of this flag.
• level — standard (default, current calibrated behavior) | strict | carrasco (uncompromising — every finding treated as blocking).
• chunking — maxFilesPerChunk, maxLinesPerChunk, byTopic. Chunking only kicks in when the change set exceeds a limit; otherwise all files are one chunk.
• carrasco — redTeamParallel (dispatch chunks in parallel), requireReproducibleTrigger, focusCategories, severityThreshold (BLOCK if any finding ≥ this).
• standards — autoDetect (read CLAUDE.md/AGENTS.md + neighboring code) and paths (authoritative standards/architecture docs to enforce).
• reportOutput — saveToHarness, format (markdown | json | both).

Procedure

1. Plan. Run the harness CLI review plan (see Running the CLI below). It writes .harness/reviews/<feature>/plan.json, one prompt file per chunk under prompts/, and an empty responses/ directory. Read plan.json to get the feature name and the chunk list. If it reports "No changed files", stop.

2. Dispatch the carrascos. For each chunk, read its prompt at .harness/reviews/<feature>/prompts/<chunk-id>.md and dispatch a superpowers-prepared:carrasco subagent with that prompt as its task.

   • Dispatch all chunks in a SINGLE message with multiple parallel Agent tool calls when carrasco.redTeamParallel is true (the default). Run sequentially only if it is false.
   • Context isolation: construct each subagent's prompt from the chunk prompt file ONLY. Never forward this session's history or other chunks' results.
   • Append to each subagent prompt: "You are a focused subagent. Do NOT invoke any skills from the superpowers-prepared plugin. Do NOT use the Skill tool. Return your full report including the <!-- REVIEWER_DECISION --> JSON block."

3. Collect. Write each subagent's returned text verbatim to .harness/reviews/<feature>/responses/<chunk-id>.txt (the chunk id matches the prompt filename).

4. Aggregate. Run the harness CLI review aggregate --feature <feature>. It parses every response, merges findings, decides the overall verdict against severityThreshold, and saves the consolidated report + decision.json. Exit code: 0 APPROVE, 2 BLOCK, 3 NEEDS_HUMAN_REVIEW.

5. Report & act.

   • Show the verdict and the report path (.harness/reviews/<feature>/carrasco-review.md).
   • On BLOCK: fix findings ASI-first (start from the report's "Fix First" entry), then re-review with review recheck (step 6 below) — never re-run from step 1. Repeat until no finding reaches the threshold.
   • On NEEDS_HUMAN_REVIEW: surface the findings to your human partner for a decision; do not silently approve.

6. Recheck (after fixing findings) — scoped, not a full re-plan. Run review recheck --feature <feature> (add --note "<one line: what you changed, plus any new follow-up ask from your human partner>"; add --chunks id-a,id-b only if you want to target specific chunks instead of the default of every chunk that didn't approve). This does not recompute the whole plan:

   • It reads the last decision's per-chunk verdicts and, for each chunk that didn't approve, writes a new prompt containing only that chunk's prior findings, the diff of just that chunk's files since the last review (the fix — not the original diff again), and your note.
   • Chunks that already approved are left untouched — their prompt and response files are not regenerated.
   • Dispatch one superpowers-prepared:carrasco subagent per chunk printed by recheck, exactly as in step 2, then write each response to the same responses/<chunk-id>.txt path.
   • Run review aggregate --feature <feature> again — it merges the fresh verdicts for rechecked chunks with the untouched ones automatically, no extra flag needed.

Running the CLI

Use the harness CLI the same way as harness-verify, with the plugin-root env var for your harness:

• Windows (PowerShell): npx tsx "$( $env:CLAUDE_PLUGIN_ROOT, $env:QWEN_PLUGIN_ROOT, $env:CURSOR_PLUGIN_ROOT, $env:CODEX_PLUGIN_ROOT | Where-Object { $_ } | Select-Object -First 1 )\tools\harness\cli.ts" review <plan|recheck|aggregate|gate-status> [--feature <name>] [--base <sha>] [--chunks <id,id>] [--note "<text>"] [--root <project>]
• Linux/macOS: npx tsx "${CLAUDE_PLUGIN_ROOT:-...}/tools/harness/cli.ts" review <subcommand> [...]

If no plugin-root env var is set, resolve from the superpowers-prepared plugin directory (the parent of hooks/).

Hard Rules

• Dispatch the carrascos from the planner's per-chunk prompts — do not hand-write the review criteria. The rigor (level, focus categories, standards enforcement, severity policy) is baked into those prompts from the config.
• Do not approve while the aggregate verdict is BLOCK. A passing run requires zero findings at or above severityThreshold.
• Do not edit files inside a carrasco subagent — carrascos review and report only. Fixes happen in the main session after aggregation.
• The Stop-hook gate's verdict is tied to the exact change set (a whole-diff fingerprint). If you change code after a passing review, the gate treats it as stale.
• After a BLOCK, do not re-run review plan from scratch and do not re-dispatch carrascos for chunks that already approved. Use review recheck (step 6) so only the chunks that had findings are re-reviewed, with a lean prompt (their prior findings + the fix diff + your note) instead of the full original chunk context.