manual_slop/conductor/tracks/mma_utilization_refinement_20260226/spec.md

Specification: MMA Utilization Refinement

Overview

Refine the Multi-Model Architecture (MMA) implementation within the Conductor framework to ensure clear role segregation, proper tool permissions, and improved observability for sub-agents.

Goals

• Enforce Tier 1 as the track creator and Tier 2 as the track executor.
• Restore and fix segregated skills (mma-tier1 through mma-tier4).
• Provide Tier 3 & 4 with direct file I/O tools to reduce Tier 2 context bloat.
• Implement AST-based "Skeleton Views" for Tier 3 context injection.
• Create a non-polluting verbose log/feed for sub-agent operations.
• Remove "Context Amnesia" from Tier 2 while maintaining it for Tiers 3 & 4.

Functional Requirements

1. Skill Refinement:
   • Update mma-tier1-orchestrator to focus on /conductor:setup and /conductor:newTrack.
   • Update mma-tier2-tech-lead to manage /conductor:implement. It must maintain persistent context for the duration of a track session (no amnesia).
   • Update mma-tier3-worker and mma-tier4-qa to be stateless (Context Amnesia) but equipped with read_file, write_file, and codebase exploration tools.
2. AST Extraction (Skeleton Views):
   • Enhance mcp_client.py (or a dedicated utility) to generate Python skeletons (signatures and docstrings) using tree-sitter.
   • Update mma_exec.py to utilize these skeletons for modules NOT being actively worked on by Tier 3.
3. Observability:
   • Ensure sub-agent reasoning and tool calls are logged to a dedicated log file (e.g., logs/mma_subagents.log) or separate shell to avoid polluting the main session history.
4. Workflow Update:
   • Update conductor/workflow.md to reflect the new tier responsibilities and tool access rules.

Acceptance Criteria

• Tier 1 can successfully initialize a track.
• Tier 2 can delegate a coding task to Tier 3.
• Tier 3 receives a "Skeleton View" of relevant dependencies instead of full files.
• Tier 3 can write files back to the project.
• Tier 4 can analyze logs and provide summaries.
• Sub-agent verbose output is captured in a dedicated log.
• Tier 2 context remains focused on the high-level plan, not implementation details.