9-Phase Iteration Workflow Skill

When to Use

This skill activates when:

• User invokes /superpowers-iterate:iterate <task>
• User asks to "follow the iteration workflow"

Announce: "I'm using the iteration-workflow skill to orchestrate this task through all 9 mandatory phases."

Overview

Phases: Brainstorm -> Plan -> Plan Review -> Implement -> Review -> Test -> Simplify -> Final Review -> Codex Final

Iteration Loop: Phases 1-8 repeat until Phase 8 finds zero issues or max iterations reached. Phase 9 runs once at the end (full mode only).

Modes:

• Full (default): Defaults to mcp__codex-high__codex for Phases 3, 8 and mcp__codex-xhigh__codex for Phase 9 (configurable via /configure)
• Lite (--lite): Uses Claude reviews, skips Phase 9

See AGENTS.md for model configuration and state schema details.

State Management

State file: .agents/iteration-state.json

Initialize at start with version 3 schema. Update state after each phase transition. See AGENTS.md for full schema.

Configuration Loading

IMPORTANT: Configuration MUST be loaded fresh each time the workflow is invoked. Always read both global and project config files at the start.

Load configuration using the configuration skill. The skill handles:

• Default values
• Merging global config (~/.claude/iterate-config.json)
• Merging project config (.claude/iterate-config.local.json)

On each invocation:

1. Read global config file (if exists)
2. Read project config file (if exists)
3. Merge: defaults → global → project
4. Use merged config for all phases in this run

Run /superpowers-iterate:configure --show to see current config.

Phase 1: Brainstorm

Purpose: Explore problem space, generate ideas, clarify requirements

Required Skill: superpowers:brainstorming + superpowers:dispatching-parallel-agents

Actions:

1. Mark Phase 1 as in_progress in state file
2. Use TodoWrite to track brainstorming tasks
3. Launch parallel subagents as needed using superpowers:dispatching-parallel-agents (if config allows):
   • Identify independent research areas for the task
   • Dispatch one subagent per independent domain (no limit)
   • Example domains:
     • Research existing code patterns and architecture
     • Explore problem domain and requirements
     • Investigate test strategy and coverage requirements
     • Analyze similar implementations in codebase
     • Research external libraries/APIs needed
     • Explore edge cases and error scenarios
4. Follow superpowers:brainstorming process:
   • Ask questions one at a time to refine the idea
   • Propose 2-3 different approaches with trade-offs
   • Present design in 200-300 word sections, validating each
5. Document test strategy requirements during brainstorming:
   • What test frameworks/tools are available?
   • What testing patterns does the codebase use?
   • What edge cases need coverage?
6. Save design to docs/plans/YYYY-MM-DD-<topic>-design.md

Exit criteria:

• Problem is well understood
• Approach is agreed upon
• Design decisions documented
• Test strategy requirements identified

Transition: Mark Phase 1 complete, advance to Phase 2

Phase 2: Plan

Purpose: Create detailed implementation plan with bite-sized tasks including tests

Required Skill: superpowers:writing-plans + superpowers:dispatching-parallel-agents

Actions:

1. Mark Phase 2 as in_progress
2. Launch parallel subagents (if phases.2.parallel is true) to create plan components:
   • Identify independent planning areas based on brainstorm output
   • Dispatch one subagent per independent component (no limit)
   • Example planning areas:
     • Plan core implementation tasks
     • Plan test coverage (TDD approach)
     • Plan integration points and edge cases
     • Plan documentation updates
     • Plan migration/upgrade paths
     • Plan performance considerations
3. Follow superpowers:writing-plans format:
   • Each task is 2-5 minutes of work
   • Include exact file paths
   • Include complete code in plan
   • Follow DRY, YAGNI, TDD principles
4. Each task MUST include test steps:
   • Step 1: Write failing test
   • Step 2: Run test to verify it fails
   • Step 3: Write minimal implementation
   • Step 4: Run test to verify it passes
   • Step 5: Commit
5. Save plan to docs/plans/YYYY-MM-DD-<feature-name>.md
6. Document the plan using TodoWrite

Plan header must include:

# [Feature Name] Implementation Plan

> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.

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

**Architecture:** [2-3 sentences about approach]

**Tech Stack:** [Key technologies/libraries]

**Test Strategy:** [Testing approach and frameworks]

---

Exit criteria:

• Implementation plan created with TDD steps
• All tasks have clear acceptance criteria
• Test strategy defined for each task
• Plan follows superpowers:writing-plans format

Transition: Mark Phase 2 complete, advance to Phase 3

Phase 3: Plan Review

Purpose: Validate plan quality before implementation begins

Required Tool (from config phases.3.tool):

• mcp__codex-high__codex (default): Codex with medium reasoning
• mcp__codex-xhigh__codex: Codex with high reasoning
• claude-review: Use superpowers:requesting-code-review
• Lite mode always uses superpowers:requesting-code-review

Actions:

1. Mark Phase 3 as in_progress in state file
2. Run review based on configured tool:

Codex Mode (mcp__codex__codex or mcp__codex-high__codex)

Invoke mcp__codex-high__codex with plan review prompt:

Review the implementation plan at docs/plans/YYYY-MM-DD-<feature-name>.md

Validate:
- Task granularity (each task should be 2-5 minutes of work)
- TDD steps included for each task
- File paths are specific and accurate
- Plan follows DRY, YAGNI principles
- Test strategy is comprehensive
- Dependencies and task order are correct
- Edge cases are covered

Report findings with severity (HIGH/MEDIUM/LOW) and file:line references.
If you find NO issues, explicitly state: "Plan looks good to proceed."

Claude Review Mode (claude-review or lite mode)

Dispatch code-reviewer subagent via superpowers:requesting-code-review to review the plan document.

3. Evaluate review results:

   If ZERO issues found:

   • Announce: "Plan review passed. Proceeding to implementation."
   • Proceed to Phase 4

   If HIGH/MEDIUM issues found:

   • Fix plan issues
   • Re-run plan review
   • Do not proceed until plan is clean

   If only LOW issues found:

   • Note them for awareness
   • Proceed to Phase 4

Exit criteria:

• Plan review completed
• No HIGH or MEDIUM severity issues in plan
• Plan ready for implementation

Transition: Mark Phase 3 complete, advance to Phase 4

Phase 4: Implement

Purpose: TDD-style implementation following the plan

Implementer (from config phases.4.implementer):

• claude (default): Use Claude subagents via superpowers:subagent-driven-development
• codex-high: Use mcp__codex-high__codex for implementation
• codex-xhigh: Use mcp__codex-xhigh__codex for implementation

Required Skill (claude mode): superpowers:subagent-driven-development + superpowers:test-driven-development

Required Plugins: LSP plugins for code intelligence

Actions:

1. Mark Phase 4 as in_progress
2. Check phases.4.implementer config to determine implementation mode:

Claude Mode (default)

Follow superpowers:subagent-driven-development process (model from phases.4.model):

• Read plan, extract all tasks, create TodoWrite

• Launch as many subagents as needed for implementation:

  • One implementer subagent per task (sequential to avoid conflicts)
  • Multiple reviewer subagents can run in parallel

• For each task: a. Dispatch implementer subagent with full task text AND LSP context:

  You have access to LSP (Language Server Protocol) tools:
  - mcp__lsp__get_diagnostics: Get errors/warnings for a file
  - mcp__lsp__get_hover: Get type info and documentation at position
  - mcp__lsp__goto_definition: Jump to symbol definition
  - mcp__lsp__find_references: Find all references to a symbol
  - mcp__lsp__get_completions: Get code completions at position
  
  Use LSP tools to:
  - Check for errors before committing
  - Understand existing code via hover/go-to-definition
  - Find all usages before refactoring
  

  b. Answer any questions from subagent c. Subagent follows superpowers:test-driven-development:

  • Write failing test first
  • Run to verify it fails
  • Write minimal code to pass
  • Run to verify it passes
  • Use mcp__lsp__get_diagnostics to check for errors
  • Self-review and commit d. Dispatch spec reviewer subagent e. Dispatch code quality reviewer subagent (can use LSP diagnostics) f. Mark task complete in TodoWrite

Codex Mode (codex-high or codex-xhigh)

Invoke the configured Codex tool with implementation prompt:

Implement the following tasks from the plan at docs/plans/YYYY-MM-DD-<feature-name>.md

For each task:
1. Read the task requirements
2. Write the code following TDD:
   - Write failing test first
   - Implement minimal code to pass
   - Verify tests pass
3. Follow existing code patterns and conventions
4. Add appropriate logging and error handling

Run these commands after implementation:
1. make lint
2. make test

If tests fail, fix issues and re-run until passing.

3. Run make lint && make test to verify all tests pass
4. Commit after tests pass

Note (Claude mode): Implementation subagents run sequentially (to avoid file conflicts), but reviewer subagents can run in parallel.

Exit criteria:

• All tasks from plan implemented
• Tests written for new functionality (TDD)
• make lint && make test pass
• Code committed

Transition: Mark Phase 4 complete, advance to Phase 5

Phase 5: Review (1 Round)

Purpose: Quick code review sanity check before Phase 8's thorough review

Required Skill: superpowers:requesting-code-review

Bug Fixer (from config phases.5.bugFixer):

• claude: Use Claude subagents to fix issues
• codex-high (default): Use mcp__codex-high__codex to fix issues
• codex-xhigh: Use mcp__codex-xhigh__codex to fix issues

Actions:

1. Mark Phase 5 as in_progress
2. Get git SHAs for the changes:

   BASE_SHA=$(git merge-base HEAD main)
   HEAD_SHA=$(git rev-parse HEAD)
   

3. Dispatch code-reviewer subagent per superpowers:requesting-code-review
4. Provide:
   • WHAT_WAS_IMPLEMENTED: Description of changes
   • PLAN_OR_REQUIREMENTS: Reference to plan file
   • BASE_SHA and HEAD_SHA
5. Categorize issues:
   • Critical: Must fix immediately
   • Important: Fix now
   • Minor: Note for Phase 8
6. Fix issues using configured bugFixer:
   • If claude: Dispatch Claude subagent to fix
   • If codex-high/codex-xhigh: Invoke Codex with fix prompt including issue details
7. Re-run tests after fixes
8. Document findings in state

Exit criteria:

• 1 review round completed
• No Critical issues remaining
• Important issues addressed

Transition: Mark Phase 5 complete, advance to Phase 6

Phase 6: Test

Purpose: Run lint and test suites

Actions:

1. Mark Phase 6 as in_progress
2. Run: make lint
   • If fails: Fix issues, re-run until pass
3. Run: make test
   • If fails: Fix issues, re-run until pass
4. Record results in state

Exit criteria:

• make lint passes
• make test passes

Transition: Mark Phase 6 complete, advance to Phase 7

Phase 7: Code Simplifier

Purpose: Reduce code bloat using code-simplifier plugin

Required Plugin: code-simplifier:code-simplifier (from claude-plugins-official)

Actions:

1. Mark Phase 7 as in_progress
2. Launch code-simplifier agent using Task tool:

   Task(
     description: "Simplify modified code",
     prompt: "Review and simplify code modified in this iteration. Focus on clarity and maintainability while preserving functionality.",
     subagent_type: "code-simplifier:code-simplifier"
   )
   

3. Review suggestions
4. Apply appropriate simplifications
5. Re-run tests to verify no breakage

Exit criteria:

• Code simplifier has reviewed changes
• Appropriate simplifications applied
• Tests still pass

Transition: Mark Phase 7 complete, advance to Phase 8

Phase 8: Final Review (Decision Point)

Purpose: Thorough review that determines whether to loop back to Phase 1 or proceed to completion.

Required Tool (from config phases.8.tool):

• mcp__codex-high__codex (default): Codex with high reasoning
• mcp__codex-xhigh__codex: Codex with extra-high reasoning
• claude-review: Use superpowers:requesting-code-review
• Lite mode always uses superpowers:requesting-code-review

Bug Fixer (from config phases.8.bugFixer):

• claude: Use Claude subagents to fix issues
• codex-high (default): Use mcp__codex-high__codex to fix issues
• codex-xhigh: Use mcp__codex-xhigh__codex to fix issues

Actions:

1. Mark Phase 8 as in_progress
2. Check current iteration count against maxIterations
3. Run review based on configured tool:

Codex Mode (mcp__codex__codex or mcp__codex-high__codex)

Invoke mcp__codex-high__codex with review prompt:

Iteration {N}/{max} final review for merge readiness. Run these commands first:
1. make lint
2. make test

Focus on:
- Documentation accuracy
- Edge cases and error handling
- Test coverage completeness
- Code quality and maintainability
- Merge readiness

Report findings with severity (HIGH/MEDIUM/LOW) and file:line references.
If you find NO issues, explicitly state: "No issues found."

Claude Review Mode (claude-review or lite mode)

Dispatch code-reviewer subagent via superpowers:requesting-code-review:

• WHAT_WAS_IMPLEMENTED: Full description of all changes
• PLAN_OR_REQUIREMENTS: Reference to plan file
• BASE_SHA and HEAD_SHA from git

4. Evaluate review results:

   If ZERO issues found:

   • Announce: "Iteration {N} review found no issues. Proceeding to completion."
   • Store phase8Issues: [] in state
   • Full mode: Proceed to Phase 9
   • Lite mode: Skip to Completion

   If ANY issues found (HIGH, MEDIUM, or LOW):

   • Announce: "Iteration {N} found {count} issues. Fixing and starting new iteration."
   • Fix ALL issues using configured bugFixer:
     • If claude: Dispatch Claude subagent with issue details
     • If codex-high/codex-xhigh: Invoke Codex with fix prompt
   • Re-run make lint && make test
   • Store issues in phase8Issues array in state
   • If currentIteration < maxIterations:
     • Increment currentIteration
     • Add new iteration entry to state
     • Loop back to Phase 1
   • If currentIteration >= maxIterations:
     • Announce: "Reached max iterations ({max}). Proceeding with {count} unresolved issues noted."
     • Full mode: Proceed to Phase 9
     • Lite mode: Skip to Completion

Exit criteria:

• Review completed
• Either: zero issues found, OR all issues fixed and looping, OR max iterations reached

Transition:

• Zero issues OR max iterations → Phase 9 (full) or Completion (lite)
• Issues found AND iterations remaining → Phase 1 (new iteration)

Phase 9: Codex-High Final Validation (Full Mode Only)

Purpose: Final validation with OpenAI Codex high reasoning

Required Tool: mcp__codex-xhigh__codex

Note: This phase is skipped in lite mode.

Actions:

1. Mark Phase 9 as in_progress

2. Invoke mcp__codex-xhigh__codex with final validation prompt:

   Final validation review. Run these commands first:
   1. make lint
   2. make test
   
   This is the FINAL check before merge. Be thorough.
   
   Report findings with severity (HIGH/MEDIUM/LOW) and file:line references.
   Focus on:
   - Correctness and logic errors
   - Idempotency of operations
   - Documentation accuracy
   - Test coverage gaps
   - Security concerns
   - Edge cases missed in earlier reviews
   

3. Address any HIGH severity issues

4. Re-run Codex-high if significant changes made

Exit criteria:

• Codex-high review completed
• No HIGH severity issues remaining

Transition: Mark Phase 9 complete, proceed to Completion

Completion

After Phase 8 (lite mode) or Phase 9 (full mode):

1. Update state file to show workflow complete
2. Announce: "Iteration workflow complete after {N} iteration(s)!"
3. Summarize:
   • Total iterations run
   • Issues found and fixed per iteration
   • Final state (clean or with noted issues)
4. Suggest next steps (commit, create PR, etc.)
5. Optionally use superpowers:finishing-a-development-branch for merge prep

Resuming Interrupted Iterations

If iteration was interrupted:

1. Read .agents/iteration-state.json
2. Identify:
   • currentIteration: Which iteration we're on
   • currentPhase: Which phase within that iteration
   • mode: Full or lite
3. Announce: "Resuming iteration {N} at Phase {P}: <phase-name>"
4. Continue from where stopped

Red Flags - STOP

Never:

• Skip phases (all 8 phases per iteration are mandatory, Phase 9 runs once at end)
• Advance without meeting exit criteria
• Ignore issues from Phase 8 review (must fix or note if at max iterations)
• Skip test runs
• Skip TDD (write tests after implementation)
• Dispatch parallel implementation subagents (conflicts)
• Exceed maxIterations without proceeding to completion

If blocked:

• Record blocker in state file notes
• Ask user for guidance
• Do not proceed until resolved