Agent-Forge is a robust multi-agent automation framework designed to generate production-grade code through a rigorous "Plan-Execute-Audit" closed loop.

Unlike traditional "fire-and-forget" AI coding tools, agent-forge introduces a Human-in-the-loop planning confirmation mechanism and an Audit-Driven acceptance workflow. Code is only accepted after passing strict static analysis and verification standards.

• Interactive Planning:
  • The Orchestrator automatically generates a detailed execution plan based on a TASK_ID (saved in .claude/plan/).
  • User Confirmation Required: Execution only begins after you verify and approve the plan.
• Strict Quality Gates:
  • An embedded Audit Agent forces a review of all artifacts.
  • Substandard code automatically enters a "Rework Loop" until it receives a PASS rating.
• Adaptive Worker:
  • The Worker Agent dynamically adopts roles such as "Backend Dev", "Frontend Dev", "QA Engineer", or "Architect" based on task requirements.
• Session-based XML Logging:
  • Logs are isolated by TASK_ID and TIMESTAMP.
  • Action Log: Tracks the system's operational timeline.
  • Audit Log: Records detailed code review feedback and scores.

1. Orchestrator (Manager & Planner):

   • Brain: Decomposes tasks, generates plan files, and requests user approval.
   • Dispatcher: Manages Sub-agent calls and file paths.
   • Constraint: Does not write code directly.

2. Worker (Adaptive Executer):

   • Polymorphic: Accepts role_definition to implement logic and write tests.
   • Responsive: Fixes code based on specific Audit feedback.

3. Audit (Quality Control):

   • Inspector: Scores outputs against Acceptance Criteria.
   • Output: Appends XML-formatted reports to the specific session log.

4. Recorder (Scribe):

   • Logger: Maintains the "System Truth" by appending all events to the XML action log.

Ensure your cc (Claude Code) environment is configured as follows:

Create the configuration directories in your home folder:

mkdir -p ~/.claude/agents
mkdir -p ~/.claude/skills

Note: During runtime, the system will automatically create .claude/plan/, .claude/audit/, and .claude/action/ in your current project root.

Save the generated Prompt files here:

• orchestrator.md: Master logic with planning and confirmation.
• worker.md: Adaptive worker prompt with role injection.
• audit.md: Audit prompt with XML append capability.
• recorder.md: Recorder prompt with XML append capability.

1. Start: Load the Orchestrator Agent.
2. Initialize:
   • The system will request a TASK_ID (e.g., Refactor_Login_Module).
3. Confirm Plan:
   • The Orchestrator generates a plan in .claude/plan/.
   • Review the plan and reply "Yes" or "Approve".
4. Auto-Execution:
   • The system begins dispatching Workers and generating logs in .claude/audit/ and .claude/action/.

All logs are stored in XML format and isolated by session:

• Execution Plan: .claude/plan/plan_{TASK_ID}_{TIME}.md
• Action Stream: .claude/action/action_{TASK_ID}_{TIME}.xml
• Audit Details: .claude/audit/audit_{TASK_ID}_{TIME}.xml