recheck

Overview

Automated code review and fix pipeline. A triage script handles all mechanical work and outputs a compact manifest with pre-split file groups. The orchestrator only reads group metadata and dispatches agents — never individual file paths.

Prerequisites

Copy this checklist and track your progress:

• Git repository with at least one commit
• ${CLAUDE_PLUGIN_ROOT} set (plugin installed)
• LLM Externalizer MCP available (for review passes)
• reports/ and reports_dev/ listed in the project's .gitignore (add them if missing — reports contain private data)

Reports Location

Every artifact this plugin writes lives under one canonical root — $MAIN_ROOT/reports/rechecker/ — even when the run starts inside a linked worktree (always the main-repo root, never the worktree's own ./reports/). Per-session reviews/fixes/lint captures land in $MAIN_ROOT/reports/rechecker/<local-ts+tz>-<uuid>/; sub-agent output goes to reports/rechecker/sonnet-code-fixer/ and reports/rechecker/lint-filter/; the StopFailure hook writes to reports/rechecker/stop-failure/. Runtime state that is not a report (progress, index, batch lists, commit message) stays under .rechecker/. /reports/ and /reports_dev/ must both be in .gitignore. Full rule: ~/.claude/rules/agent-reports-location.md.

Instructions

1. Run triage — detects files, lints, classifies, splits into groups, outputs manifest:

   python3 "${CLAUDE_PLUGIN_ROOT}/scripts/triage.py"
   

   Exit 3 = skip (recursion guard or no files). Exit 0 = read JSON manifest from stdout. The manifest contains: session, files_total, grouped_input_files_paths (array with GROUP markers for LLM Externalizer), security_grouped_input_files_paths (same but only security-relevant groups), groups[], lint, security_pass. Each group has: id, group_file, report_file, fixes_file, lint_errors_file, review_with, security_relevant.

2. Fix lint errors — for each group where lint_errors_file is not null, spawn rechecker-plugin:sonnet-code-fixer with the group's lint_errors_file and group_file. The agent reads only its own group's files and errors.

3. Review passes — for each pass (see review-passes), dispatch reviews using the manifest's pre-built arrays. If grouped_input_files_paths is non-empty, pass it as input_files_paths to one LLM Externalizer code_task call — the ---GROUP:id--- markers produce per-group reports automatically. For review_with: "opus" groups: spawn opus agent per group, pass group_file in the prompt. After each pass, for groups with issues, spawn rechecker-plugin:sonnet-code-fixer with group_file + the group's report, writing to fixes_file.

   • Pass 1: correctness — use grouped_input_files_paths
   • Pass 2: functional — use grouped_input_files_paths
   • Pass 3: adversarial — use grouped_input_files_paths
   • Pass 4: security — use security_grouped_input_files_paths (only security-relevant groups). Skip if empty.

4. Commit fixes — if any files changed, stage only fixed files and commit:

   git add <fixed-files>
   git commit -m "fix: apply rechecker fixes [rechecker: skip]"
   

5. Finalize — use manifest.session values:

   python3 "${CLAUDE_PLUGIN_ROOT}/scripts/finalize-session.py" \
       --uuid "<session.uuid>" --commit "<session.commit>" --start "<session.started>" \
       --report-dir "<session.report_dir>" --snapshot "<session.snapshot_path>" \
       --files-reviewed <files_total> --issues-found <N> --issues-fixed <N> [--commit-made]
   

Output

--- Recheck: <UUID> (commit <hash>) ---
Files: <N> reviewed (<M> groups) | Lint: <status> | Security: <skipped/triggered>
Pass 1-3: <N issues fixed / clean>
Commit: <yes (hash) / no fixes needed>
Tokens: <total> (input/output/cache breakdown)
Reports: $MAIN_ROOT/reports/rechecker/<SESSION_TS>-<UUID>/
---

Error Handling

All stages fail-fast. If triage or any review/fix step fails, the pipeline aborts. No partial commits.

Examples

/recheck

Resources

• review-passes — review instructions per pass
  • Shared Rules (append to ALL pass instructions)
  • Large File Instructions (>250KB, opus agent)
  • Pass 1 — Code correctness
  • Pass 2 — Functional correctness
  • Pass 3 — Adversarial review
  • Pass 4 — Security audit (conditional)