Skip to content

inspection.md

Latest commit

 

History

History
78 lines (63 loc) · 3.06 KB

inspection.md

File metadata and controls

78 lines (63 loc) · 3.06 KB
Raw

llm — Module Inspection

Source Files

File Role
teaagent/llm/__init__.py Re-exports public symbols
teaagent/llm/_types.py All dataclasses, protocols, exceptions
teaagent/llm/_adapters.py ClaudeAdapter, OpenAICompatibleAdapter, GeminiAdapter, WorkersAIAdapter
teaagent/llm/_config.py ProviderConfig registry, cost estimation
teaagent/llm/_extract.py Content/delta extraction helpers per provider
teaagent/llm/_retry.py LLMRetryConfig, _call_with_retry with exponential backoff
teaagent/llm/_sse.py SSE/NDJSON stream parser (consume_sse_json_chunks)
teaagent/llm/_transport.py UrllibHTTPTransport (stdlib-only HTTP), SSL context from env

Key Exports (_types.py)

Exception hierarchy

LLMAdapterError (RuntimeError)
  ├── LLMConfigurationError  — bad config, missing env var
  ├── LLMHTTPError           — HTTP-level error (has .status_code)
  ├── LLMProviderError       — provider returned error JSON
  └── LLMResponseFormatError — unexpected response shape

Core dataclasses

  • LLMMessage(role, content) — frozen
  • LLMToolDefinition(name, description, input_schema) — frozen
  • LLMToolCall(tool_name, tool_input, call_id='') — frozen
  • LLMSafetyBlock(blocked, category?, detail='') — frozen
  • LLMRequest(messages, model?, system?, max_tokens, temperature, stream, on_chunk?, tools, response_format?) — frozen
  • LLMResponse(provider, model, content, raw, input_tokens, output_tokens, tool_calls, safety?) — frozen; has estimated_cost_cents property
  • ProviderConfig(name, api_key_env, default_model, base_url, ...) — frozen; resolved_api_key(), resolved_model(), resolved_base_url()

Protocols

  • HTTPTransport — post_json(url, headers, payload, *, timeout) -> dict
  • LLMAdapter — complete(request: LLMRequest) -> LLMResponse

Dependencies

llm/_adapters.py
  ├── teaagent.llm._extract
  ├── teaagent.llm._retry
  ├── teaagent.llm._sse
  ├── teaagent.llm._transport
  └── teaagent.llm._types

llm/_types.py
  └── (lazy) teaagent.llm._config  [estimated_cost_cents property]

All adapters use stdlib urllib only — no requests/httpx dependency.

Entry Points

  1. runner/_core.py — constructs adapter from ProviderConfig, calls adapter.complete(request) in agent loop
  2. chat_agent.py — uses adapter for chat completions
  3. intent.py — uses adapter for intent classification

Call Graph

runner._core: agent iteration
  adapter.complete(LLMRequest)
    ├── _prepare_payload(request, model)
    ├── transport.post_json(url, headers, payload, timeout)
    │     └── UrllibHTTPTransport: urllib.request.urlopen
    ├── _extract_*_content(response)
    ├── _extract_*_tool_calls(response)
    └── LLMResponse(...)

[streaming path]
  adapter._complete_streaming(request, model, payload)
    └── consume_sse_json_chunks(iter_lines, on_data)
          └── on_data(chunk) → request.on_chunk(text)