Specification: Zhipu AI (GLM) Provider Integration

Overview

This track introduces support for Zhipu AI (z.ai) as a first-class model provider. It involves implementing a dedicated client in src/ai_client.py for the GLM series of models, updating configuration models, enhancing the GUI for provider selection, and integrating the provider into the tiered MMA architecture.

Functional Requirements

1. Core AI Client (`src/ai_client.py`)

Zhipu AI Integration:
- Implement _send_zhipu() to handle communication with Zhipu AI's API.
- Implement a tool-call loop similar to _send_gemini and _send_anthropic.
- Support for function calling (native Zhipu tool format).
- Support for multi-modal input (Vision) using glm-4v by encoding screenshots as base64 data.
- Implement response streaming support compatible with the existing GUI mechanism.
State Management:
- Define module-level _zhipu_client and _zhipu_history.
- Ensure ai_client.reset_session() clears Zhipu-specific history.
Error Handling:
- Implement _classify_zhipu_error() to map Zhipu API exceptions to ProviderError types (quota, rate_limit, auth, balance, network).
Model Discovery:
- Implement _list_zhipu_models() to fetch available models (GLM-4, GLM-4-Flash, GLM-4V, etc.) from the API.

2. Configuration & Authentication

Credentials:
- Update src/ai_client.py:_load_credentials() to support a [z_ai] section in credentials.toml.
- Update the FileNotFoundError message in _load_credentials() to include the Zhipu AI example.
Cost Tracking:
- Update src/cost_tracker.py with pricing for major GLM models (GLM-4, GLM-4-Flash).

3. GUI & Controller Integration

Provider Lists:
- Add z_ai to the PROVIDERS list in src/gui_2.py and src/app_controller.py.
Model Fetching:
- Ensure AppController._fetch_models() correctly dispatches to ai_client.list_models("z_ai").
Settings UI:
- The AI Settings panel should automatically handle the new provider and its models once added to the lists.

4. MMA Orchestration

Tier Support:
- Verify that agents in all tiers (1-4) can be configured to use Zhipu AI models via the mma_tier_usage dict in AppController.
- Ensure run_worker_lifecycle in src/multi_agent_conductor.py correctly passes Zhipu model names to ai_client.send().

Non-Functional Requirements

Consistency: Follow the established pattern of module-level singleton state in ai_client.py.
Latency: Ensure the tool-call loop overhead is minimal.
Security: Rigorously protect the Zhipu AI API key; ensure it is never logged or exposed in the GUI.

Acceptance Criteria

Zhipu AI can be selected as a provider in the AI Settings panel.
Models like glm-4 and glm-4-flash are listed and selectable.
Agents can successfully use tools (e.g., read_file, run_powershell) using Zhipu AI models.
Screenshots are correctly processed and described by vision-capable models (glm-4v).
Response streaming is functional in the Discussion panel.
Estimated costs for Zhipu AI calls are displayed in the MMA Dashboard.

Out of Scope

Support for Zhipu AI's knowledge base or vector store features.
Specialized "Batch" API support.

3.3 KiB Raw Blame History