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
Step 1: Identify changed files, filter non-code, split by size
      |
      v
Step 2: Generate session UUID, record commit hash + timestamp
      |
      v
Step 3: LINT PASS
        ruff/mypy (Python) or eslint/tsc (JS/TS)
        --> haiku lint-filter strips warnings
        --> sonnet fixer fixes errors only
      |
      v
Step 4: REVIEW PASSES (3 mandatory + 1 conditional)
        |
        |- Pass 1: Code correctness (syntax, logic, race conditions)
        |- Pass 2: Functional correctness (does it do what it should?)
        |- Pass 3: Adversarial review (injection, exhaustion, corruption)
        |- Pass 4: Security audit (only if files touch auth/network/crypto/etc.)
        |
        Each pass: LLM Externalizer review --> sonnet fixer
        Files >250KB: opus[1m] agent instead of LLM Externalizer
        Files >500KB: skipped
      |
      v
Step 5: Commit fixes with [rechecker: skip] marker
      |
      v
Step 6: finalize-session.py
        --> count tokens, write history.jsonl, move reports
      |
      v
Done. Summary reported to user.

Components

Component	What It Does
/recheck skill	The full pipeline — lint, 3+1 review passes, commit, finalize
sonnet-code-fixer	Sonnet agent that fixes reported issues using Serena MCP
lint-filter	Haiku agent that strips warnings from lint output, keeping only errors
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)
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

Each recheck run gets a unique UUID. A session record is appended to .rechecker/history.jsonl:

{"uuid":"a1b2c3d4e5f6","commit":"abc123...","started":"...","finished":"...","files_reviewed":8,"issues_found":3,"issues_fixed":3,"commit_made":true,"tokens":{...}}

Reports are stored in .rechecker/reports/<uuid>/ (gitignored).

---

Plugin Structure

rechecker-plugin/
├── .claude-plugin/
│   └── plugin.json              # Plugin manifest
├── hooks/
│   └── hooks.json               # StopFailure error logging
├── agents/
│   ├── sonnet-code-fixer.md     # Sonnet fixer (Serena MCP + Grepika)
│   └── lint-filter.md           # Haiku lint output filter
├── rules/
│   └── recheck-after-commit.md  # Auto-trigger rule
├── skills/
│   └── recheck/
│       └── SKILL.md             # /recheck pipeline
├── scripts/
│   ├── finalize-session.py      # Token counting + history + cleanup
│   ├── count-tokens.py          # JSONL transcript parser
│   ├── log-stop-failure.py      # API error logger
│   └── publish.py               # Dev: lint, bump, tag, release
└── README.md

---

Safety

Concern	How It's Handled
Recursion	Commits include [rechecker: skip] marker; rule and skill both check it
Token waste	Lint filtered by haiku; reviews via external LLM; finalize automated by script
Code deletion	All review prompts include "NEVER suggest removing code" rules
Fixer scope	"Fix ONLY what is reported. When in doubt, SKIP."
Cleanup	Reports moved to .rechecker/ (gitignored); temp files removed
Blocking	Runs inline — no background processes, no worktrees, no orphaned branches

---

Cross-Platform Support

Platform	Status
macOS	Supported
Linux	Supported
WSL	Supported
Windows	Untested

Python 3.12+ required on all platforms.

---

License

MIT — see LICENSE

Author

Emasoft — GitHub