[Atom112]

Summary

Real-world report: DeepSeek V4.1 Flash via AIO thinks for tens of minutes, re-reads dozens of files, never starts acting — the thinking:action ratio explodes. Diagnosis: a documented model trait (thinking ON at HIGH effort by default) amplified by two genuine harness defects, both fixed here.

Defect 1 (the amplifier): reasoning_content never round-tripped

DeepSeek's thinking-mode docs make it a documented tool-call contract: in tool-carrying conversations every assistant message must carry its reasoning_content back — it is the model's cross-round working memory ("让模型继续之前的思考"). AIO dropped it on both paths:

• in-turn: append_assistant_message built {role, content, tool_calls} only — each round's thinking was discarded (RoundResult.reasoning: "kept only for debugging")
• history: finish_message_for_api never inserted it, though the DB has a reasoning column per message

Every round started thinking from scratch → re-read files to re-establish understanding → the re-read loop behind the report.

Fix: a new deepseek provider plugin (claims api.deepseek.com; everything else delegates to the fallback) + two trait additions:

• roundtrip_reasoning_content (default false — OpenAI's strict message schema rejects the unknown field, so the contract must stay provider-scoped)
• adjust_chat_body (default no-op), applied unconditionally in stream_one_round after generic body assembly

The flag gates both paths: in-turn rounds (main loop computes it once per turn; subagent loops wired identically) and history replay (the FE now passes message.reasoning through buildApiMessages; chat-mode without tools stays unflagged — the server ignores the field there per the docs).

Defect 2: the reasoning slider was a no-op for DeepSeek

Thinking is ON at HIGH effort by default; the only documented off-switch is {"thinking":{"type":"disabled"}} — which AIO never sent. The plugin's adjust_chat_body override now:

Slider	Sent to DeepSeek
Off	thinking: disabled — the slider's "off" is finally real (behavior change: no more chain-of-thought at Off; move to Low+ when thinking is wanted)
Low / Medium	reasoning_effort: low (DeepSeek's real ladder is low/high/max; upstream collapses medium→high — Low keeps three settings distinguishable)
High	reasoning_effort: high

Subagent rounds (which never inherit the slider) default to thinking disabled — executor role, not high-effort reasoning.

Model-side facts (not "fixed", documented)

• Default thinking = ON at high effort is official behavior (docs); the community even maintains dsh-repeat-guard for the looping trait
• The correct lever to reduce thinking is now the slider's Low level

Verification

• provider +4 tests (host claim/look-alike guards, effort-ladder mapping incl. off→disabled, URL delegation), round_artifacts +1 (round-trip: flag on / flag off / empty reasoning), messages +1 (history round-trip: role-scoped, flag-gated, empty-absent)
• Full suite 469 green; npm run verify fully green (87 FE tests, gates rgba=173 / any=77 ratchets held — file-size included, lesson from fix(harness): pwsh object-output P0s - formatter loss on exit, rc-0 masking, CLIXML stderr #241 applied)
• Docs: chat-and-agent zh/en reasoning section updated with the DeepSeek mapping