AI Orchestration Feedback Loop

AI Orchestration Feedback Loop

Workflow

Standard:      Plan → Validate(AI-1) → Review(AI-2) → Synthesize → Implement → Review → Done
Co-Implement:  Plan → Validate → Review → Synthesize → Core(Claude) → Aux(Gemini) → Integrate → Review → Done

CLI Patterns

Always use timeout: 600000 for all AI commands.

Model & CLI References

IMPORTANT: For available models and CLI options, refer to the required skills:

• Codex: See codex-plugin:codex-cli skill for models, reasoning effort levels, and CLI options
• Gemini: See gemini-plugin:gemini-cli skill for models, output formats, and CLI options

When asking user for model selection in Phase 0, present options based on the current skill documentation.

Phase 0: Pre-flight

mkdir -p .ai-orchestration

Ask user via AskUserQuestion with 4 questions:

Question 1: AI Participation Mode

Header: "Mode"

Question 2: Role Assignment per Phase

Header: "Roles"

If Custom selected, ask follow-up:

Question 3: Model Selection

Header: "Models"

First, load the required skills to get current model lists:

1. Load codex-plugin:codex-cli → get Codex models and reasoning effort levels
2. Load gemini-plugin:gemini-cli → get Gemini models

Then present options:

Question 4: Analysis Focus

Header: "Focus"

Question 5: Gemini Co-Implementation

Header: "Co-Impl"

Question 6: Gemini Generation Scope (if Co-Implementation enabled)

Header: "Gen Scope"

Question 7: Integration Review Mode (if Co-Implementation enabled)

Header: "Review Mode"

Save to .ai-orchestration/​config.md:

# AI Orchestration Config
## Mode: [selected mode]
## Roles
- Planning: [AI]
- Validation 1: [AI]
- Validation 2: [AI or Skip]
- Implementation (Core): Claude
- Implementation (Auxiliary): [Disabled | Gemini]
- Code Review: [AI(s)]
## Models
- Codex: [model] (reasoning: [level])
- Gemini: [model]
## Focus: [focus area]
## Co-Implementation
- Enabled: [yes/​no]
- Mode: [Disabled | Documentation Only | Boilerplate Only | Full]
- Documentation Scope: [api-docs, inline-comments, readme, jsdoc]
- Boilerplate Scope: [utilities, configs, interfaces, test-scaffolds]
- Review Mode: [review-first | auto-integrate | strict-review]

Phase 1: Planning

Executor: Based on config (default: Claude)

Create .ai-orchestration/​plan.md with: Objective, Approach, Steps, Risk Assessment, Validation Focus Areas

Phase 2: First Validation

Executor: Based on config Validation 1 setting

Detailed prompts: See prompt-templates.md for Security/​Performance/​Architecture focused prompts

Phase 3: Second Validation

Executor: Based on config Validation 2 setting (Skip if Dual-AI or config says Skip)

Detailed prompts: See prompt-templates.md for Innovation/​UX focused prompts

Phase 4: Synthesis

Read validation results. Create .ai-orchestration/​phase4_synthesis.md:

# Synthesis
## Consensus Points
## Divergence Analysis
## Prioritized Actions (P0/​P1/​P2)
## Revised Plan
## User Decisions Needed

For synthesis methodology: See synthesis-guide.md

Present to user via AskUserQuestion: Proceed / Address issues / Request more validation

Phase 5a: Core Implementation

Executor: Claude (always)

Implement core business logic using Edit/​Write/​Read tools.

Save .ai-orchestration/​implementation.md with: Implemented By, Changes Made, Issues Addressed, Testing Notes

If Co-Implementation enabled → Create handoff spec for Phase 5b:

Save .ai-orchestration/​phase5b_handoff.md:

# Gemini Co-Implementation Handoff
## Implementation Summary
[Link to implementation.md]
## Files Created/​Modified
[List of files]
## Generation Tasks
### Task 1: [Documentation/​Boilerplate]
**Type**: [api-docs | inline-comments | readme | utilities | configs | interfaces]
**Target Files**: [list]
**Code Context**:
\`\`\`[language]
[relevant snippets for context]
\`\`\`
**Requirements**:
- [specific requirements]

Phase 5b: Auxiliary Generation (Gemini)

Executor: Gemini (if Co-Implementation enabled) Skip if: Co-Implementation disabled in config

Detailed prompts: See co-implementation-guide.md for handoff format and prompts

Generate auxiliary code based on handoff specification:

gemini -m MODEL -p "Generate auxiliary code per handoff spec:
$(cat .ai-orchestration/​phase5b_handoff.md)

[Use prompt from co-implementation-guide.md based on scope]"

Output Format (FILE: marker system):

FILE: path/​to/​file.ext
---
[generated content]
---
FILE: next/​file.ext
---
[content]
---

Save to .ai-orchestration/​phase5b_gemini_output.md

Phase 5c: Integration

Executor: Claude Skip if: Co-Implementation disabled

1. Parse Gemini output (FILE: markers)
2. Validate syntax and conflicts
3. Apply Review Mode:
   • Review-first → Show to user, ask approval
   • Auto-integrate → Integrate if valid
   • Strict → Ask per file
4. Integrate approved code via Edit/​Write
5. Handle Revision (max 2 attempts):
   • If rejected → Request Gemini revision or Claude fallback

Save .ai-orchestration/​phase5c_integration.md with: Files Integrated, Review Decisions, Revisions Made

Phase 6: Code Review

Executor: Based on config Code Review setting

Detailed prompts: See prompt-templates.md

Phase 7: Final Assessment

• Both PASS/​APPROVE → Complete
• FAIL/​REJECT → Fix and re-validate
• REQUEST CHANGES → Apply and iterate

Save iterations to .ai-orchestration/​iterations.md.

Context Files

.ai-orchestration/
├── config.md
├── plan.md
├── phase2_*.md
├── phase3_*.md (Triple-AI only)
├── phase4_synthesis.md
├── implementation.md
├── phase5b_handoff.md          # Co-Implementation handoff spec
├── phase5b_gemini_output.md    # Gemini generated code
├── phase5c_integration.md      # Integration log
├── phase6a_codex_review.md
├── phase6b_gemini_review.md
└── iterations.md

Error Handling

References

• Prompt Templates: prompt-templates.md - Detailed prompts for each focus area
• Workflow Patterns: workflow-patterns.md - Security-First, Architecture Decision, Rapid Iteration patterns
• Synthesis Guide: synthesis-guide.md - Divergence analysis, priority matrix, resolution patterns
• Co-Implementation Guide: co-implementation-guide.md - Handoff format, output format, integration process