Intended behavior
Under /reasoning on, a reasoning model's thinking should be sealed as its own block, and the final answer delivered as a separate message/block.
Actual behavior
For deepseek-v4 (api openai-completions, reasoning: true), the final answer is appended into the last 🧠 reasoning block rather than emitted as a distinct answer — the thought and the answer run together. This is independent of the delivery fixes (#95029 Discord, #95051 Telegram): once the reasoning lane delivers, the content itself is merged.
Root cause (source-level)
deepseek streams reasoning via reasoning_content deltas, then switches to the answer via content deltas, with no boundary event between them. In src/llm/providers/openai-completions.ts:
reasoning_content → thinking_delta (opens the thought)content → text_delta (the answer)thinking_end is emitted only by finishBlock, which runs once at end-of-stream (the for (const block of blocks) finishBlock(block) loop, ~L462–464) — never at the reasoning→content transition.
So the event order the subscribe handler sees is:
thinking_delta … text_delta … text_delta … [stream end] thinking_end
Because onReasoningEnd is keyed on thinking_end (src/agents/embedded-agent-subscribe.handlers.messages.ts, ~L633), it fires only after the answer has already streamed — too late. The channel keeps the reasoning block open across the answer text and merges them.
Providers that DO emit a discrete thinking_end at the transition (Anthropic, OpenAI-responses, Google, Mistral) are unaffected.
Reproduction
- Model:
deepseek/deepseek-v4-flash (api openai-completions, reasoning enabled)
/reasoning on, /think high- Prompt anything that produces visible reasoning followed by an answer
- Observe: the answer text appears inside the last thinking block
Deterministic (unit-level): a handler test that feeds thinking_delta then text_delta with no thinking_end and asserts the reasoning stream closes before the answer fails on current code (onReasoningEnd called 0 times), because nothing closes the thought at the text-lane transition.
Proposed fix
Treat the text-lane opening as the reasoning boundary for native-thinking providers: when a text delta begins while the reasoning stream is still open and we are not inside a tag-based <think> block (whose end is detected during content parsing), close the reasoning stream so the thought is sealed before the answer. PR to follow.
Environment
Reported with AI assistance (Claude); behavior verified on a live OpenClaw 2026.6.9-beta.1 instance.
Intended behavior
Under
/reasoning on, a reasoning model's thinking should be sealed as its own block, and the final answer delivered as a separate message/block.Actual behavior
For deepseek-v4 (api
openai-completions,reasoning: true), the final answer is appended into the last 🧠 reasoning block rather than emitted as a distinct answer — the thought and the answer run together. This is independent of the delivery fixes (#95029 Discord, #95051 Telegram): once the reasoning lane delivers, the content itself is merged.Root cause (source-level)
deepseek streams reasoning via
reasoning_contentdeltas, then switches to the answer viacontentdeltas, with no boundary event between them. Insrc/llm/providers/openai-completions.ts:reasoning_content→thinking_delta(opens the thought)content→text_delta(the answer)thinking_endis emitted only byfinishBlock, which runs once at end-of-stream (thefor (const block of blocks) finishBlock(block)loop, ~L462–464) — never at the reasoning→content transition.So the event order the subscribe handler sees is:
Because
onReasoningEndis keyed onthinking_end(src/agents/embedded-agent-subscribe.handlers.messages.ts, ~L633), it fires only after the answer has already streamed — too late. The channel keeps the reasoning block open across the answer text and merges them.Providers that DO emit a discrete
thinking_endat the transition (Anthropic, OpenAI-responses, Google, Mistral) are unaffected.Reproduction
deepseek/deepseek-v4-flash(apiopenai-completions, reasoning enabled)/reasoning on,/think highDeterministic (unit-level): a handler test that feeds
thinking_deltathentext_deltawith nothinking_endand asserts the reasoning stream closes before the answer fails on current code (onReasoningEndcalled 0 times), because nothing closes the thought at the text-lane transition.Proposed fix
Treat the text-lane opening as the reasoning boundary for native-thinking providers: when a text delta begins while the reasoning stream is still open and we are not inside a tag-based
<think>block (whose end is detected during content parsing), close the reasoning stream so the thought is sealed before the answer. PR to follow.Environment
Reported with AI assistance (Claude); behavior verified on a live OpenClaw 2026.6.9-beta.1 instance.