manual_slop/conductor/tracks/zhipu_integration_20260308/plan.md

Implementation Plan: Zhipu AI (GLM) Provider Integration

Phase 1: Core Client Implementation

• Task: Define Zhipu AI state and initialize client in src/ai_client.py.
  • Import zhipuai (ensure it's added to requirements.txt if missing).
  • Add _zhipu_client and _zhipu_history module-level variables.
  • Implement _classify_zhipu_error(exc) for structured error mapping.
• Task: Implement _send_zhipu tool-call loop.
  • Implement function/tool definition conversion to Zhipu format.
  • Implement the core chat completion API call with streaming support.
  • Implement tool result packaging and the recursion loop (up to MAX_TOOL_ROUNDS).
  • Integrate _reread_file_items for context refresh after tool rounds.
• Task: Implement Multi-modal (Vision) support.
  • Add logic to _send_zhipu to process screenshot inputs for glm-4v by encoding them as base64.
• Task: Implement model discovery.
  • Implement _list_zhipu_models() using the Zhipu AI models list API.
• Task: Update ai_client.py utility functions.
  • Update send() dispatcher to route to _send_zhipu.
  • Update reset_session() to clear _zhipu_history.
• Task: Conductor - User Manual Verification 'Phase 1: Core Client Implementation' (Protocol in workflow.md)

Phase 2: Configuration & Credentials

• Task: Update credential loading in src/ai_client.py.
  • Update _load_credentials() to include z_ai in the error message and loading logic.
• Task: Add cost tracking for Zhipu models in src/cost_tracker.py.
  • Add regex patterns and rates for glm-4 and glm-4-flash to MODEL_PRICING.
• Task: Verify that credentials.toml works with the new provider section.
• Task: Conductor - User Manual Verification 'Phase 2: Configuration & Credentials' (Protocol in workflow.md)

Phase 3: GUI & Controller Integration

• Task: Register z_ai as a provider.
  • Add z_ai to PROVIDERS in src/gui_2.py.
  • Add z_ai to PROVIDERS in src/app_controller.py.
• Task: Ensure model fetching works in the GUI.
  • Verify that clicking "Fetch Models" in the AI Settings panel correctly populates Zhipu models when the provider is selected.
• Task: Conductor - User Manual Verification 'Phase 3: GUI & Controller Integration' (Protocol in workflow.md)

Phase 4: MMA Integration & Final Verification

• Task: Verify MMA compatibility.
  • Test a simple multi-agent workflow where a Tier 3 worker is configured to use glm-4-flash.
  • Verify that tool calls, history, and context injection work as expected within the tiered architecture.
• Task: Run full regression suite.
  • Ensure adding Zhipu hasn't introduced side effects for existing providers.
• Task: Conductor - User Manual Verification 'Phase 4: MMA Integration & Final Verification' (Protocol in workflow.md)