rechecker-plugin

rechecker-plugin

A Claude Code plugin that reviews and fixes code after git commits using a blocking /recheck skill with multiple review passes.

---

The Problem

Claude Code often introduces bugs, incomplete implementations, or subtle regressions when making code changes. Without this plugin, you'd manually ask Claude to "check your changes for errors, then fix them" — sometimes repeating this multiple times before all issues are resolved.

The Solution

This plugin adds a /recheck skill that runs a structured review-fix pipeline on committed code. A rule tells Claude to invoke it automatically after significant commits (5+ files or 50KB+). The pipeline runs inline and blocking — no worktrees, no background processes.

---

Installation

From the Emasoft Marketplace (recommended)

claude plugin marketplace add emasoft-plugins
claude plugin install rechecker-plugin@emasoft-plugins

From GitHub (direct)

claude plugin install --source github Emasoft/rechecker-plugin

Requirements

• Python 3.12+ on PATH
• git
• LLM Externalizer plugin — reviews are sent to external LLMs (grok/gemini via OpenRouter)

  claude plugin install llm-externalizer@emasoft-plugins
  

---

Usage

Automatic Mode (default)

A rule instructs Claude to run /recheck after any commit with 5 or more files or 50 KB+ total size. Commits with [rechecker: skip] in the message are ignored (recursion guard).

Manual Mode

Run the review pipeline on demand:

/rechecker-plugin:recheck

---

How It Works

You commit code
      |
      v
Claude checks the rule: >=5 files or >=50KB?
      |
      v (yes)
/recheck skill runs (blocking)
      |
      v
triage.py (one script does all mechanical work)
  |- Recursion guard check
  |- Detect changed files, filter non-code, classify by size
  |- Split into groups (max 10 files each, by extension family)
  |- Run linters (ruff/mypy/eslint/tsc/shellcheck/etc.)
  |- Filter lint errors (Python, no haiku agent needed)
  |- Detect security-relevant groups
  |- Write per-group JSON files + token snapshot
  |- Output compact manifest with ---GROUP:id--- markers
      |
      v
Orchestrator reads manifest, dispatches:
  |- Lint fix: sonnet-code-fixer per group with errors
  |- Pass 1-3: one code_task call with GROUP markers → per-group reports
  |- Pass 4 (security): only security-relevant groups
  |- Large files (>250KB): opus agent per file
  |- Each pass: sonnet-code-fixer for groups with issues
      |
      v
Commit fixes with [rechecker: skip] marker
      |
      v
finalize-session.py → count tokens, write history, move reports
      |
      v
Done. Summary reported to user.

Components

Component	What It Does
/recheck skill	The full pipeline — triage, lint, 3+1 review passes, commit, finalize
triage.py	Detects files, lints, classifies, splits into groups, outputs manifest
sonnet-code-fixer	Sonnet agent that fixes reported issues using Serena MCP
LLM Externalizer	External LLM (grok/gemini) reviews code for bugs — not Claude tokens
finalize-session.py	Automates: token counting, session history, report cleanup
count-tokens.py	Parses JSONL transcripts for per-model token breakdown (snapshot/delta/transcripts modes)
log-subagent-tokens.py	SubagentStop hook — logs isolated token usage per subagent/worktree
log-stop-failure.py	Logs API errors (rate limits, server errors, auth failures)
recheck-after-commit rule	Tells Claude when to trigger /recheck

Review Pass Details

Pass	Focus	Checks
Lint	Static analysis	Syntax errors, type errors, import errors
1	Code correctness	Logic bugs, off-by-one, race conditions, deprecated APIs
2	Functional correctness	Edge cases, return values, async handling, API contracts
3	Adversarial	Injection, resource exhaustion, state corruption, type confusion
4	Security (conditional)	SQL/XSS/command injection, auth, secrets, crypto

File Size Routing

Size	Reviewer
<=250KB	LLM Externalizer (cheap, fast)
250-500KB	Opus[1m] agent (large context)
>500KB	Skipped

Session Tracking