git:20260922.c9be8be to git:20261017.89cb6dd
21 added, 34 removed. Audit A to A.
# LLM Provider Sub-module
- ## Overview
-
- Implements the provider abstraction layer for the LLM module. Each provider implements the `BaseLLMProvider` abstract class, ensuring consistent behavior across different LLM backends.
-
- ## Architecture
-
- ```
- providers/
- ├── __init__.py # Provider registry and factory (48 lines)
- ├── base_provider.py # Abstract base class defining the provider interface (255 lines)
- ├── ollama_provider.py # Ollama local model provider (330 lines)
- ├── openai_provider.py # OpenAI API provider (354 lines)
- ├── openrouter_provider.py # OpenRouter multi-model gateway (436 lines)
- └── perplexity_provider.py # Perplexity AI provider (419 lines)
- ```
-
- ## Provider Interface
-
- All providers implement `BaseLLMProvider` with these core members:
-
- - **`generate_response(messages, config) → LLMResponse`** — Text generation.
- - **`generate_stream(messages, config) → AsyncGenerator`** — Streaming text generation.
- - **`get_available_models() → List[str]`** — List models available from this provider.
- - **`health_check() → bool`** — Verify provider connectivity.
- - **`default_model` / `available_models`** — Properties declaring provider defaults.
-
- ## Provider Selection
+ Providers implement `BaseLLMProvider`: `initialize`, `validate_config`, async
+ `generate_response`, async-generator `generate_stream`, `close`, and
+ `provider_type`, `default_model`, `available_models` properties. The advertised
+ model list is not a live daemon/model readiness probe. There is no common
+ `health_check` or `get_available_models` interface; OpenRouter has its own
+ additional model-list method.
- The provider factory in `__init__.py` selects providers based on:
- 1. Explicit configuration in `input/config.yaml`
- 2. Environment variables (`OLLAMA_MODEL`, `OPENAI_API_KEY`, etc.)
- 3. Fallback chain: Ollama → OpenAI → OpenRouter → Perplexity
+ The maintained adapters are Ollama, OpenAI, OpenRouter and Perplexity. Explicit
+ provider requests remain pinned even on unavailable providers or circuit-breaker
+ failure. Generic unpinned operations retain the shared processor's recovery
+ behavior. See [parent contracts](../AGENTS.md).
- ## Parent Module
+ Ollama uses bounded standard-library HTTP chat, preserving whole messages with
+ input truncation and shifting disabled. `preflight_model` verifies stable runtime
+ >=0.35.0 and the exact model tag, and records its digest and context/seed options.
+ Generation uses one total request deadline, including runtime capability checks;
+ pipeline workers supply hard process containment. Missing SDK/CLI is irrelevant
+ to this transport. Never start/pull a daemon or change the model implicitly.
- See [llm/AGENTS.md](../AGENTS.md) for the overall LLM architecture.
+ Native completion is distinct from receiving text. Preserve finish reason,
+ usage and transport metadata in `LLMResponse`. Token-limit and unfinished replies
+ cannot satisfy complete prompt coverage or enter a completed response cache.
- **Version**: [pyproject.toml](../../../../pyproject.toml) (canonical)
+ See [provider overview](README.md), [specification](SPEC.md), and
+ [Mac runbook](../../../../docs/llm/README.md).