2.2 KiB
2.2 KiB
Specification: MMA Tiered Architecture Verification
Overview
This track aims to review and verify the implementation of the 4-Tier Hierarchical Multi-Model Architecture (MMA) within the Conductor framework. It will confirm that Conductor operates as a Tier 2 Tech Lead/Orchestrator and can successfully delegate tasks to Tier 3 (Workers) and Tier 4 (QA/Utility) sub-agents. A key part of this track is investigating whether this hierarchy should be enforced via a single centralized skill or through separate role-based sub-agent definitions.
Functional Requirements
- Skill Review: Analyze
mma-orchestrator/SKILL.mdandMMA_Supportdocs to ensure they correctly mandate Tier 2 behavior for Conductor. - Delegation Verification:
- Verify Conductor (Tier 2) can spawn Tier 3 sub-agents for heavy coding tasks using
.\scripts un_subagent.ps1. - Verify Tier 3/4 sub-agents can be spawned for error analysis/compression.
- Verify Conductor (Tier 2) can spawn Tier 3 sub-agents for heavy coding tasks using
- Architectural Investigation: Evaluate the pros/cons of a centralized
mma-orchestratorskill vs. independent role-based sub-agents. Determine the best way to define sub-agent roles. - Test Track Creation: Implement a "Mock Implementation" track that demonstrates the full tiered delegation flow (Tier 2 -> Tier 3 -> Tier 4).
- Automated Testing: Create
pytestcases to verify the IPC and script execution flow of the tiered sub-agents.
Non-Functional Requirements
- Traceability: All sub-agent invocations must be clearly logged in the session.
- Context Efficiency: Ensure sub-agent delegation effectively prevents token bloat in the main Conductor context.
Acceptance Criteria
- Analysis report comparing centralized skill vs. role-based sub-agents.
- A functional test track (
mma_verification_mock) that executes a full tiered delegation sequence. - Traceable logs confirming sub-agent spawning and task completion.
- Pytest suite verifying the sub-agent infrastructure and interaction logic.
- Plan alignment: The test track's
plan.mdexplicitly includes delegation steps.
Out of Scope
- Implementing a full production-ready multi-model backend.
- Modifying the core
gemini-clibinary.