Skip to content

MULTI_AGENT.md

Latest commit

 

History

History
98 lines (68 loc) · 2.48 KB

MULTI_AGENT.md

File metadata and controls

98 lines (68 loc) · 2.48 KB

Multi-Agent Ready Architecture

This project keeps deterministic diagnosis as the execution backbone. Planner Agent and Critic Agent can now run as live Bailian-backed LLM agents without adding autonomous code execution loops.

Principle

Multi-agent components are responsible for:

  • repository understanding
  • triage
  • planning
  • critique

Deterministic tools remain responsible for:

  • scanning
  • rule evaluation
  • hard verification
  • policy enforcement

Current Components

  • Governor: orchestrates the whole diagnosis pass and bounded planning pass
  • Tool Runner: runs the deterministic scanning and rule engine stack
  • Scanner Agent: converts raw artifacts into hotspot-oriented repo inventory
  • Planner Agent: turns findings into bounded action steps
  • Critic Agent: checks plan scope and policy risks
  • Policy Engine: enforces protected path and file-scope guardrails

Current Outputs

In addition to the original diagnosis artifacts, the tool now writes:

  • artifacts/repo_inventory.json
  • artifacts/triage.json
  • artifacts/action_plan.json
  • artifacts/critic_review.json
  • artifacts/planner_agent.json
  • artifacts/critic_agent.json
  • artifacts/model_routing.json

Current Model Routing Blueprint

  • governor / planner / triage / critic: PLANNER_MODEL, fallback DASHSCOPE_MODEL, default deepseek-v3.2
  • refactor / test fix / small bug fix: CODER_MODEL, fallback DASHSCOPE_MODEL, default qwen3-coder-flash
  • cheap summaries / PR copy / scan explanation: DASHSCOPE_MODEL, default qwen3.5-flash
  • vector recall: text-embedding-v4
  • rerank: qwen3-rerank

Provider environment variables are:

  • DASHSCOPE_API_KEY
  • DASHSCOPE_BASE_URL
  • DASHSCOPE_MODEL
  • CODER_MODEL
  • PLANNER_MODEL
  • ENABLE_LIVE_AGENTS

Legacy compatibility:

  • BAILIAN_BASE_URL is still accepted as an alias for DASHSCOPE_BASE_URL
  • BAILIAN_MODEL is still accepted as an alias for DASHSCOPE_MODEL

Default routing source:

  • config/agent_models.json

Priority:

  • internal fallback
  • config file
  • environment variables

Config Self-Check

Use:

python main.py --check-llm-config

The check reports:

  • resolved provider config
  • live agent runtime status
  • planner / critic / coder / summary model health

If --output is also provided, the tool writes:

  • artifacts/llm_health_check.json

Phase Boundary

Planner and Critic may call a live LLM provider. Tool Runner and Policy Engine remain deterministic. No agent is allowed to execute code changes automatically.