LLM backends

Turn a request into a reply: the one external seam of a campaign. See LLM backends.

Interface

factor_agent.llm.base.LLMBackend	Anything that completes a request.
factor_agent.llm.base.Request	One call: the role making it, the model, the prompt and the sampling parameters.
factor_agent.llm.base.Reply	The provider's answer: its text, token usage as the provider reports it and the model that answered.
factor_agent.llm.base.ProviderQuotaExhausted	The provider refuses calls because the balance or quota is exhausted; its message says so in its words.
factor_agent.llm.base.ProviderReplyMalformed	The provider answered, without an error, something that is not a reply; its message says what it lacked.
factor_agent.llm.base.AskingAgain	A backend that asks backend again when it answers something that is not a reply.
factor_agent.llm.base.make_backend	Build the backend a role's config names in spec["backend"].

Backends

factor_agent.llm.openai_compatible.OpenAICompatibleBackend	Complete a request with one chat-completions call.
factor_agent.llm.anthropic.AnthropicBackend	Complete a request with one Messages API call.
factor_agent.llm.claude_code.ClaudeCodeBackend	Complete a request with one headless Claude Code call.
factor_agent.llm.replay.ReplayBackend	Answer each request with the recorded reply to the same role, model and prompt.
factor_agent.llm.replay.ReplayExhausted	The recorded run holds no reply for a request: it ran out of calls, or the run asked something else.
factor_agent.llm.log.LLMLog	The call files of one run directory's llm_log.