fix(run_agent): gate reasoning_content re-feed to DeepSeek/Kimi only

[leon7609]

What does this PR do?

Gates step 3 of _copy_reasoning_content_for_api to providers that actually need the reasoning → reasoning_content promotion (DeepSeek / Kimi).

Today step 3 promotes msg["reasoning"] to api_msg["reasoning_content"] for any assistant message that carries reasoning text in the unified reasoning field — regardless of provider. This is harmful for at least two cases:

• Qwen on vLLM (and any compatible custom provider): Qwen's official multi-turn spec explicitly forbids re-feeding reasoning_content from prior turns. In practice this causes (a) context balloon across turns and (b) thinking-mode death loops where the model re-thinks its own prior reasoning back at itself. It's the dominant failure mode on long Qwen-on-vLLM sessions.
• OpenRouter / Anthropic / OpenAI: these carry reasoning continuity via the separate reasoning_details field (unaffected by this helper). They never relied on step 3's promotion path; dropping it for them is a no-op.

The fix is a one-condition addition: gate step 3 on needs_thinking_pad (already computed at line 8782 via _needs_thinking_reasoning_pad()). DeepSeek / Kimi keep their existing behaviour because they're true under that flag; everyone else falls through to step 5 and has reasoning_content popped from the API replay message.

Related Issue

No existing issue. Diagnosed from a Qwen3.6-27B-FP8 vLLM session that exhibited progressive context inflation and thinking loops; tracing showed the prior turn's <think> text was being re-fed via reasoning_content.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• run_agent.py — _copy_reasoning_content_for_api, step 3:

  - if isinstance(normalized_reasoning, str) and normalized_reasoning:
  + if needs_thinking_pad and isinstance(normalized_reasoning, str) and normalized_reasoning:
        api_msg["reasoning_content"] = normalized_reasoning
        return

  Plus a comment block above it documenting the rationale.
• tests/run_agent/test_reasoning_content_provider_gate.py — 6 new unit tests:
  • Qwen vLLM: reasoning is not promoted.
  • OpenRouter / MiniMax: same.
  • DeepSeek: regression check — reasoning is still promoted.
  • Kimi (Moonshot): regression check — same.
  • Step 1 (explicit reasoning_content) is unaffected by the gate.

The existing tests in tests/run_agent/test_deepseek_reasoning_content_echo.py continue to pass — DeepSeek/Kimi paths are unchanged.

How to Test

1. Run a multi-turn conversation against a Qwen-on-vLLM endpoint with thinking enabled.
2. Before this fix: observe the prior turn's reasoning_content re-appearing in the next request payload; on long sessions you'll see context inflate and the model re-reasoning over its own thoughts.
3. After this fix: reasoning_content is only present on DeepSeek/Kimi turns; Qwen multi-turn behaviour matches the upstream Qwen spec.

Or run the unit tests:

pytest tests/run_agent/test_reasoning_content_provider_gate.py tests/run_agent/test_deepseek_reasoning_content_echo.py -v

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(run_agent): ...)
• I searched for existing PRs — recent reasoning-content work has been DeepSeek/Kimi-specific ([Bug]: DeepSeek V4 Flash Discord session poisoned when tool-call assistant lacks reasoning_content #15250, [Bug] _copy_reasoning_content_for_api: cross-provider reasoning promotion leaks stale content to DeepSeek/Kimi #15748, fix: use non-empty reasoning_content placeholder for DeepSeek V4 Pro thinking mode #17341); none gates step 3
• My PR contains only changes related to this fix
• I've run pytest tests/ -q and the changes don't introduce new failures (vs. main baseline)
• I've added tests for the change (6 cases)
• I've tested on my platform: macOS 15 (Apple Silicon), Qwen 3.6 27B FP8 served via vLLM

Documentation & Housekeeping

• No public API surface changed
• DeepSeek / Kimi / OpenRouter / Anthropic / OpenAI behaviour unchanged

[leon7609]

Closing as superseded: upstream implemented the equivalent gate natively in 2b3a4f0 ("fix(agent): strip stale reasoning_content when falling back to a strict provider", #50480), first shipped in v0.18.0 (v2026.7.1).

Step 3 of copy_reasoning_content_for_api (now in agent/agent_runtime_helpers.py) gates the reasoning → reasoning_content promotion on needs_thinking_pad and pops the field for strict providers otherwise — semantically identical to this PR for both the promote (DeepSeek/Kimi) and drop (Qwen-on-vLLM / custom OpenAI-compatible) cases. Verified on a production deployment running v0.18.0 with a Qwen fallback: no reasoning re-feed, no regression.

Thanks!