Skip to content

A model's reasoning is a block of its own — opaque, provider-bound, replayed verbatim to the provider that made it and to no other — and a reasoning model gets the output cap it asks for

Decision 658 · accepted · 2026-09-20

Superseded in part

In part by 880.

Both providers demand it back. DeepSeek's thinking mode is all-or-nothing: once one assistant message in the history carried reasoning, one that omits it fails the whole request with a 400, which refused the second request of every tool-using conversation. Anthropic signs its thinking blocks and demands them back on a tool-use turn. Troupe.LLM.Reasoning is a fourth content block (provider, text, signature, redacted), captured off both streams — reasoning_content or reasoning as a sibling of content; thinking, signature_delta and redacted_thinking as blocks — logged in llm_response.message like any block, and handed back only by the adapter whose provider produced it: OpenAI as reasoning_content on the assistant message, Anthropic as thinking / redacted_thinking blocks and only when the request has thinking enabled, since a thinking block is illegal otherwise. Message.text/1 and tool_uses/1 never see it, so a parent's summary and a client's prose do not fill with thinking; a reasoning block from a provider this build has no adapter for replays as :unknown and is carried, never sent. Live, it is llm_delta kind: "reasoning", which the TUI folds and ACP takes as agent_thought_chunk. The cap: a model's reasoning_effort (from its models: entry, through Config.target/2 onto the request) makes the OpenAI adapter send max_completion_tokens and reasoning_effort in place of max_tokens, which a reasoning model rejects; a 400 that names the other field is answered once by sending the request again with it, so nobody has to configure what the provider will say. Anthropic takes a budget rather than a level, so the level becomes thinking.budget_tokens (minimal 1k … xhigh 32k, or a number) and max_tokens is raised to hold it rather than the request failing. The effort is a per-model setting; an agent's Definition has no field for it until a profile wants to think harder than its model's default.

Governs: apps/troupe_core/lib/troupe/llm/message.ex, apps/troupe_core/lib/troupe/llm/providers/anthropic.ex, apps/troupe_core/lib/troupe/llm/providers/openai.ex, apps/troupe_core/lib/troupe/llm/request.ex, clients/gui/packages/client/src/transcript.ts