Auto

Autonomously plan and implement a feature with external LLM review.

Reviewer Configuration

Arguments: $ARGUMENTS

Check the arguments for flags:

Reviewer flags (mutually exclusive):

• --gemini → use only Gemini for reviews
• --codex → use only Codex for reviews
• --claude → use a Claude subagent (Task tool) for reviews
• --no-review → skip all review phases (3, 4, and 6)
• No flag → use both Gemini and Codex in parallel (default)

Mode flags:

• --dry-run → plan only, skip implementation and final review (stop after Phase 4)
• --skip-final → skip the final review phase (Phase 6)
• --rounds N → repeat the review-refine cycle N times (default: 1). Max: 3.

Strip all flags from arguments to get the task description.

Phase 1: Understand the Task (No Questions)

1. Explore the codebase - use Glob, Grep, Read to understand:

   • Relevant files and their structure
   • Existing patterns and conventions
   • Dependencies and interfaces

2. Make reasonable assumptions - do NOT ask clarifying questions

   • Use best judgment based on codebase context
   • Prefer simpler solutions when ambiguous
   • Follow existing patterns in the codebase

Phase 2: Write the Plan

Create a plan document following this structure:

# [Feature Name] Implementation Plan

**Goal:** [One sentence describing what this builds]

**Approach:** [2-3 sentences about the approach]

**Assumptions:** [List any assumptions made without asking the user]

---

### Task 1: [Short description]

**Files:**
- Create: `exact/path/to/file.py`
- Modify: `exact/path/to/existing.py` (lines 123-145)

**Steps:**
1. [Specific action]
2. [Specific action]

**Code:**
```language
// Include actual code, not placeholders
```

---

Guidelines:

• Exact file paths - never "somewhere in src/"
• Complete code - show the actual code
• Small tasks - 2-5 minutes of work each
• DRY, YAGNI - only what's needed

Save the plan to history/<date>-plan-<feature-name>.md (e.g. history/2026-02-15-plan-user-auth.md).

Phase 3: Consult Reviewer(s)

If --no-review: Skip to Phase 5 (Implementation).

Based on the reviewer flag from arguments:

If --gemini: Gemini only

Call mcp__consult-llm__consult_llm with:

• model: "gemini"
• prompt: See review prompt below
• files: Array including the plan file and relevant source files

If --codex: Codex only

Call mcp__consult-llm__consult_llm with:

• model: "openai"
• prompt: See review prompt below
• files: Array including the plan file and relevant source files

If --claude: Claude subagent

Use the Task tool with subagent_type: "general-purpose" and a prompt like:

Review this implementation plan. The plan is in: [plan file path]

Consider:
- Are the tasks correctly ordered and sized?
- Are there any missing steps or edge cases?
- Are the file paths and code snippets accurate?
- Any architectural concerns or better approaches?
- Is the approach consistent with the existing codebase patterns?

Read the plan file and relevant source files, then provide specific, actionable feedback. Be concise.

If no flag (default): Both Gemini and Codex in parallel

Spawn BOTH as parallel subagents (Agent tool, subagent_type: "general-purpose", model: "sonnet"). NEVER run subagents in the background — always run them in the foreground so you can process their results immediately. Each subagent prompt must include the full review prompt and file list so it can make the MCP call independently.

Gemini subagent — prompt must include:

• Call mcp__consult-llm__consult_llm with model: "gemini", prompt: the review prompt, files: [array including the plan file and relevant source files]
• Return the COMPLETE response including any [thread_id:xxx] prefix

Codex subagent — prompt must include:

• Call mcp__consult-llm__consult_llm with model: "openai", prompt: the review prompt, files: [array including the plan file and relevant source files]
• Return the COMPLETE response including any [thread_id:xxx] prefix

Review prompt (for Gemini/Codex):

Review this implementation plan. Consider:
- Are the tasks correctly ordered and sized?
- Are there any missing steps or edge cases?
- Are the file paths and code snippets accurate?
- Any architectural concerns or better approaches?
- Is the approach consistent with the existing codebase patterns?

Provide specific, actionable feedback. Be concise.

Maintaining conversation context with thread_id

After each MCP call, check the response for a [thread_id:xxx] prefix. Extract and store the thread ID per model (e.g., gemini_thread_id, codex_thread_id). Pass the corresponding thread_id in all subsequent calls to the same model (Phase 4 rounds, Phase 6) so the reviewer retains full context from earlier reviews without resending the plan and files.

Phase 4: Apply Improvements

After receiving feedback from both LLMs:

1. Synthesize feedback - identify common suggestions and valid concerns
2. Apply improvements automatically - update the plan with:
   • Bug fixes or edge cases both LLMs identified
   • Architectural improvements that make sense
   • Missing steps that are clearly needed
3. Discard conflicting advice - if Gemini and Codex disagree, use your judgment
4. Update the plan file with the improvements
5. Do NOT ask the user - proceed with your best judgment

Multiple Rounds (--rounds N)

If --rounds N is specified and N > 1, repeat Phases 3-4 for additional rounds:

• Round 2+: Send the refined plan back to reviewers with this prompt, passing the thread_id from the previous round. The reviewer already has context from prior rounds, so you don't need to resend files — just include the updated plan file:

  This is revision [N] of the implementation plan. Review the updated plan:
  - Were previous concerns addressed adequately?
  - Any new issues introduced by the changes?
  - Any remaining gaps or edge cases?
  - Is the plan ready for implementation?
  
  Focus on what changed. Be concise.
  

• Apply improvements from each round before proceeding to the next
• Track changes by round in the plan file under a "Revision History" section
• Stop early if reviewers indicate no further changes needed
• Update thread_id after each round — use the latest thread_id from the response

After completing all rounds (or stopping early), proceed to Phase 5.

Phase 5: Implement

If --dry-run: Skip to Phase 7 (Summary) - report the plan without implementing.

Implement the plan without further interaction:

1. Follow the plan exactly - implement each task in order
2. Commit after each logical unit - keep commits small and focused
3. If something is unclear - make a reasonable decision and note it in the commit message
4. If a task fails - attempt to fix it before moving on
5. Only stop if there's a blocking error that cannot be resolved

Implementation rules:

• Work through tasks sequentially
• Test changes when possible
• Keep commits atomic and well-documented
• Use commit messages that explain the "why"
• Commit improvements from consultations separately from original plan tasks

Phase 6: Final Review

If --no-review or --skip-final: Skip to Phase 7 (Summary).

After implementation is complete, review the changes using the same reviewer from Phase 3:

1. Generate the diff - get the full diff of all changes made

2. Consult the reviewer based on the flag:

   • --gemini or no flag: Use mcp__consult-llm__consult_llm with model: "gemini"
   • --codex: Use mcp__consult-llm__consult_llm with model: "openai"
   • --claude: Use Task tool with subagent_type: "general-purpose"

   Include git_diff (for mcp tool) or instruct the subagent to run git diff to see changes.

   Pass the thread_id from Phase 3/4 so the reviewer has full context from the plan review. The reviewer already knows the plan — the diff shows how it was implemented.

Final review prompt:

Review this implementation for bugs, issues, or improvements:
- Any obvious bugs or edge cases missed?
- Code quality issues (error handling, naming, structure)?
- Deviations from best practices?
- Security concerns?

Be concise. Only flag issues worth fixing.

3. Apply fixes automatically - if the review identifies real issues:

   • Fix bugs and edge cases
   • Improve error handling if clearly needed
   • Commit each fix separately with clear messages (not bundled with implementation commits)

4. Skip minor style suggestions - don't refactor for style alone

Phase 7: Summary

Present a final summary to the user:

## Summary

**Implemented:** [One sentence describing what was built]

**Plan improvements applied:**
- [Improvement from Gemini/Codex feedback]
- [Another improvement]

**Post-implementation fixes:**
- [Fix applied after final review, if any]

**Commits:**
- `abc1234` - [commit message]
- `def5678` - [commit message]