[sosidudku1]

Linear: ATO-1397.

Problem

Two providers fail with HTTP 400 out of the box (Desktop 0.0.8 / agent 0.7.0):

• Anthropic: every turn on Claude Sonnet 5.5 or Haiku 5.5 is refused because the request carries temperature, which these models no longer accept.
• DeepSeek in thinking mode: the second request of a tool loop is refused because the earlier assistant tool-call message does not carry its reasoning_content.

Cause

• model-params.ts knew two temperature exceptions (OpenAI reasoning models, Kimi K3.1). Every Claude model got the default 0.2. The run-time retry from fix(llm): send each model the temperature it accepts (Kimi K3.1 = 1), retry once on a named temperature #1008 only read "only N is allowed", not a 400 that refuses the field itself.
• fix(llm): hand a tool call's reasoning back to a local chat template (ATO-887) #1050 hands reasoning back only to a local server that renders the chat template. Cloud bodies never carried it, and the cloud context journal (the cloud prompt path) stores calls without their reasoning at all.

Change

Temperature (first commit):

• model-params.ts: Claude Opus 4.7 and later, Sonnet 5 and later, Haiku 5 and later, and every Fable and Mythos model get the forbidden rule. It is read from family and version at the start of the id, with or without a router or Bedrock prefix (anthropic/claude-sonnet-5, anthropic.claude-haiku-5-5, us.anthropic.claude-opus-5-5, dotted or dashed versions, date suffixes). Anchored at the start, so a fine-tune that only names Claude (Qwen3-8B-Claude-Sonnet-5-Distill) keeps its temperature. Older Claude models (Opus 4.6, Sonnet 4.6, Haiku 4.5, the claude-3-5-sonnet naming) keep the free rule. max_tokens is unchanged.
• allowed-temperature.ts: new parseTemperatureRefused / temperatureLessonFromError. A 400 that says temperature is deprecated, unsupported, unknown or not permitted (within 80 characters of the word) reads as "this model takes no temperature". A body that names the one allowed value keeps its existing meaning.
• openai-provider.ts: the existing one resend per completion (fix(llm): send each model the temperature it accepts (Kimi K3.1 = 1), retry once on a named temperature #1008) now covers both lessons. On a refused field the model is remembered as taking no temperature (rememberTemperatureForbidden) and the resend goes out without one. So a Claude id the static rule does not know yet recovers after one 400, unary and streamed. When the refusal is already on record and the body that failed carried no temperature, there is no resend: the same body would go out again.
• openai-build-body.ts, openai-describe-image.ts: a learned refusal also removes a temperature set through extraBody or model params, as a learned value already overrides them.
• The key check and contract probe pick this up through resolveTemperature, so they stop sending 0 to these models too.

DeepSeek reasoning (second commit):

• build-cloud-prompt.ts: the current, un-replied macro-turn's call reasoning goes back on the outgoing turns, read from the transcript. The persisted journal and the text prompt do not change. A provider that does not hand reasoning back sends the same body as before.
• openai-native-messages.ts: new reasoningRequired option. It hands reasoning back without the 2048-token local cap. Every assistant tool-call message carries the field, empty where the transcript kept none (answered turns, later calls of a batch, reasoning left out by the budget). This includes the first step of a follow-up turn, when only answered calls are in the history. A reasoningBudget option caps it, newest calls first.
• openai-build-body.ts: the reasoning handed back is held to the room the cloud window has left. The body is assembled once without any reasoning, after the extraBody / model params merge (so a max_tokens set there counts) and with the tools as they go on the wire (strict tools included). It is measured exactly as the capacity check measures it, then assembled again with the remaining room, less 10% for JSON escaping, as the reasoning budget. The cloud prompt does not count reasoning, so without this a long thinking turn would fail the capacity check instead of sending less of it.
• openai-provider.ts / openai-build-body.ts: the option is on only for DeepSeek's own host (isDeepSeekHost, *.deepseek.com). Other OpenAI-compatible services may reject the unknown field, and routers such as OpenRouter carry reasoning their own way.

Tests

• model-params.test.ts: Claude models without sampling parameters (forbidden list, older models and Claude-named fine-tunes still free, learned refusal per model id).
• allowed-temperature.test.ts: parseTemperatureRefused, temperatureLessonFromError.
• openai-build-body-temperature.test.ts: no temperature for Claude 5.5 / Fable, 0.2 kept for Sonnet 4.6, learned refusal beats extraBody and model params.
• openai-temperature-retry.test.ts: refused-temperature retry, unary and streamed. No resend when the refusal is on record and the body carried no temperature. No temperature to claude-sonnet-5-5 on the first send.
• openai-native-messages.test.ts: reasoning DeepSeek requires back (uncapped, empty field on calls without reasoning, the field kept on answered calls on the first step of a follow-up turn, reasoningBudget).
• openai-build-body.test.ts: reasoningRequired puts reasoning_content on a cloud body, leaves it off without the option, sends less reasoning instead of failing in a tight window (size pinned), and sizes it against a model-params max_tokens and against strict tools, with no capacity throw and reasoning still sent.
• openai-provider.test.ts: reasoning goes to api.deepseek.com and to no other host; isDeepSeekHost.
• cloud-context.test.ts: current-turn reasoning on the outgoing cloud turns, not in the journal or the text prompt, gone once the turn is answered. One path from session through buildPrompt to the body OpenAiProvider sends to api.deepseek.com: the turn's reasoning on the call, then an empty field on the first step of the next turn.

Not confirmed live

• Anthropic: the model ids and which models refuse temperature come from Anthropic's model documentation, and the exact 400 wording is assumed. Neither has been checked against the OpenAI-compatible endpoint with a real key.
• DeepSeek: that the field is required only on tool-call messages, and that an empty string is accepted where no reasoning was kept, is assumed from the 400 and DeepSeek's docs. Not checked with a real key, in thinking mode or not.
• DeepSeek with thinking off now also receives reasoning_content: "" on its tool-call messages. Not checked that the API accepts it there.
• Overlaps PR fix(llm): treat GPT-6 as an OpenAI reasoning model #1131 (GPT-6 as an OpenAI reasoning model) in model-params.ts; whichever merges second needs a rebase.

Not yet checked on the second Mac: lint and tests run there in one batch after approval.