fix(agent): preserve Qwen3.6 / vLLM preserve_thinking history (#56004)

[Tranquil-Flow]

What was fixed

Qwen3.6 (and similar self-hosted models served via vLLM with the preserve_thinking chat-template flag) loses its prior-turn chain-of-thought on every replay. The agent loop strips both reasoning_content (strict-provider gate in copy_reasoning_content_for_api) and reasoning (unconditional pop in conversation_loop.py), so vLLM ≥0.20 has nothing to render under the preserve_thinking template.

Root cause

Two client-side strip paths in agent/agent_runtime_helpers.py and agent/conversation_loop.py unconditionally remove the reasoning fields that vLLM/Qwen3.6 need to re-render prior thinking. vLLM ≥0.20 also drops reasoning_content on incoming assistant messages (vllm-project/vllm#38488), so the existing DeepSeek/Kimi/MiMo reasoning_content echo-back path is incompatible regardless.

Fix

Added a _preserves_thinking_history() predicate on AIAgent that detects Qwen3.6-style endpoints by model name (qwen3.6 / qwen-3.6, conservative word-boundary regex — does NOT match plain qwen3 per the model card). Result cached on the AIAgent instance keyed by (provider, model, base_url) — same invalidation contract as the adjacent _thinking_pad_cache, so all switch_model / _try_activate_fallback paths automatically invalidate it.

Three touch points now gated on the predicate:

1. copy_reasoning_content_for_api — skip the strict-provider reasoning_content strip on preserves-thinking-history providers.
2. conversation_loop.py — skip the unconditional reasoning pop (keep the field on the outgoing message for vLLM to render).
3. reapply_reasoning_echo_for_provider — strip reasoning on the way out ONLY for strict providers (handles the strict-fallback case where Qwen3.6 → Mistral would 422 on the unknown field).

Verification

• 11 new regression tests, all green on this commit, 9/11 RED on upstream/main with AttributeError: 'AIAgent' object has no attribute '_preserves_thinking_history' (real production-path coverage).
• 5 detection tests (qwen3.6 positive, qwen-3.6 hyphen positive, qwen3 negative — no false positives on older Qwen, deepseek/mistral negative — distinct path through _needs_thinking_reasoning_pad).
• 3 strip-path integration tests covering both copy_reasoning_content_for_api and reapply_reasoning_echo_for_provider in both directions (preserve-history provider AND strict-fallback).
• 1 cache-invalidation test (keyed on model/provider/base_url).
• 1 integration sanity test asserting the predicate is exposed on AIAgent.
• No-regression sweep over existing reasoning/conversation/fallback tests (279 + 152 passing).

python3 -m pytest tests/run_agent/test_qwen_preserve_thinking_56004.py -v -o "addopts=" --tb=short
============================= 11 passed in 0.52s ==============================

Files changed (1 focused commit, 4 files, +289/-19)

• agent/agent_runtime_helpers.py — gated reasoning_content strip + new reasoning strip branch in reapply_reasoning_echo_for_provider (handles strict-fallback).
• agent/conversation_loop.py — gated the unconditional reasoning pop.
• run_agent.py — new _preserves_thinking_history() predicate on AIAgent.
• tests/run_agent/test_qwen_preserve_thinking_56004.py — 11 new regression tests.

Notes

• Detection is intentionally model-name based per the issue author's "reasonable default if a config flag is undesirable" note. A future config-driven flag (custom_providers[*].preserve_thinking: true) is documented in the docstring as a more robust follow-up — today this covers the reported Qwen3.6 / vLLM case correctly.
• extra_body: chat_template_kwargs: {preserve_thinking: true} already works server-side today; this PR completes the client-side loop.

Auto-published by Moonsong via Path B automated pipeline.

[alt-glitch]

This was generated by AI during triage.

Related: fixes #56004; same reasoning-replay family as #20577 and its open fix PR #20607. This PR and #20607 are competing fixes with different mechanisms: #20607 re-embeds <think> blocks inline into content (run_agent.py only), while this PR adds a _preserves_thinking_history() predicate that skips the reasoning-strip paths and keeps the reasoning field for vLLM>=0.20 to render. A maintainer should pick one approach.

[teknium1]

Thanks for isolating the normal replay stripping path; current main still removes reasoning in agent/conversation_loop.py:818-821, so the underlying report remains actionable.

Problems

• agent/chat_completion_helpers.py:1692-1694 independently builds the maximum-iterations summary request and unconditionally removes reasoning. This PR does not touch that path, so the stated replay guarantee is incomplete.
• The new qwen[-_]?3\.6 predicate is model-name-only, although the behavior depends on the served endpoint and its preserve_thinking chat-template configuration. It can therefore send the non-standard field to an endpoint that does not accept it. Related open PR feat(custom-provider) Support opt-in preserve thinking #58380 uses an explicit custom-provider opt-in.
• The added conversation-loop test at tests/run_agent/test_qwen_preserve_thinking_56004.py:292-305 tests only the predicate, not the outgoing payload.

Suggested changes

• Make the capability endpoint-configured, cover the summary builder, and add payload-level tests for normal replay, summary replay, and strict fallback.

Automated hermes-sweeper review.