autonomous-fix-review

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

Autonomous Fix-Review Pipeline

Give it a finding, walk away, come back to a committed fix that passed multi-model review.

Usage

/autonomous-fix-review "Fix the null pointer in src/main.py:42"
/autonomous-fix-review docs/research/trail/findings/finding-001.yaml

7-Phase Pipeline

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

Autonomy Modes

• supervised: Human approves everything except unanimous approval
• semi-autonomous (default): Auto-retry on warnings, escalate on disagreement
• autonomous: Proceed on any non-abort consensus

Decision Matrix

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

Workflow

# 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

Constraints

• Zero API keys — all LLM calls via subscription CLIs (claude, codex, gemini)
• Worktree isolation for implementation phase
• Human interview ONLY on consensus disagreement (ESCALATE)
• Quality gate (6/6) required before commit