メニュー
Run autonomous multi-model code review on a finding. Invokes claude/codex/gemini via reviewd, evaluates consensus, and auto-commits or escalates to human.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use this skill when the user wants to transition to the next workflow step, wrap up a session cleanly, prepare a handoff, or says things like "next step", "continue the workflow", "wrap up", "hand off", "move to the next phase", "what's next in embrace", or "prepare for /clear". Also triggers when the current octo:embrace phase appears complete and no explicit next action is set.
Spawn a dotfiles team for changes spanning chezmoi templates, mise config, and system validation. Use when adding tools that need both mise config and shell config changes.
Use when diagnosing brew failures in mde update, auditing brew-owned packages, resolving brew-vs-mise ownership conflicts, or troubleshooting Homebrew on macOS. Key files: src/mde/maintain/update.py, src/mde/validate/brew.py, configs/mde-tool-ownership.json. Keywords: brew doctor, brew upgrade, cask sudo, tap repair, formula conflict, Brewfile.
This skill should be used when the user asks to manage AI agent configurations (CLAUDE.md, AGENTS.md, permissions, MCP settings, skills) across multiple tools and machines using chezmoi templates. Use when syncing .claude/ settings across machines, templating CLAUDE.md for different environments, distributing agent instructions to Claude Code/Codex/Copilot from one source, managing AI agent permissions via chezmoi data, backing up Claude Code config with chezmoi, or setting up a shared skills directory.
This skill should be used when the user asks to "add a tool to mise", "install a new CLI tool", "choose a mise backend", "search the mise registry", "detect tool drift", or mentions mise tool installation, backend selection, GIT_TERMINAL_PROMPT enforcement, or adding new dependencies to mise config.
Use when setting up Doppler as secrets source of truth, syncing Doppler to fnox/Keychain, bulk importing/exporting secrets, validating Doppler vs fnox parity, or troubleshooting Doppler CLI issues in this mise+chezmoi dev environment.
| name | autonomous-fix-review |
| description | Run autonomous multi-model code review on a finding. Invokes claude/codex/gemini via reviewd, evaluates consensus, and auto-commits or escalates to human. |
| user-invocable | true |
| argument-hint | <finding-file-or-description> |
| context | fork |
Give it a finding, walk away, come back to a committed fix that passed multi-model review.
/autonomous-fix-review "Fix the null pointer in src/main.py:42"
/autonomous-fix-review docs/research/trail/findings/finding-001.yaml
| Phase | Action | Tool |
|---|---|---|
| 1. Research | Gather context (docs, code) | /find-docs, /context7-mcp |
| 2. Implement | Dispatch coder in worktree with TDD | subagent-driven-development |
| 3. Review | Invoke claude + codex + gemini | reviewd.invoke_cli() |
| 4. Consensus | Evaluate 75% threshold | mde.consensus.evaluate_consensus() |
| 5. Decision | Apply autonomy mode | ConsensusGate.apply_autonomy() |
| 6. Quality gate | Run lint + test + validate | uv run mde-py quality |
| 7. Commit | Commit on feature branch or escalate | git commit |
| Consensus | Supervised | Semi-Autonomous | Autonomous |
|---|---|---|---|
| >= 90% approve, no criticals | PROCEED | PROCEED | PROCEED |
| >= 75% approve, no criticals | ESCALATE | RETRY | PROCEED |
| Any critical findings | ESCALATE | ESCALATE | ESCALATE |
| < 75% approve | ESCALATE | ESCALATE | ESCALATE |
| All reject | ABORT | ABORT | ABORT |
# Step 1: Parse the finding
finding=$(echo "$ARGUMENTS")
# Step 2: Research context
# Invoke /find-docs and /context7-mcp for the affected code area
# Step 3: Implement fix (if not already done)
# Dispatch coder subagent in worktree with TDD
# Step 4-5: Run multi-model review + consensus
uv run mde-py review "$finding" --autonomy semi-autonomous
# Step 6: Quality gate (auto-runs unless --skip-quality)
uv run mde-py quality
# Step 7: Based on decision output:
# PROCEED → commit on feature branch
# ESCALATE → present findings to human for decision
# RETRY → re-implement and re-review (up to max_retries)
# ABORT → stop, finding may be invalid