[ZhXZhao]

Summary

Fixes #4212.
Related to langgenius/dify#44964; this PR addresses the confirmed Tongyi outbound-history reinjection path, not every reported stray-tag or pseudo-tool-call symptom.

For qwen3.8-max, preserve_thinking=False does not remove reasoning stored as ordinary assistant content. This change:

• Derives strip_history_thinking from the effective parameters after normalization, restricted to model == "qwen3.8-max" and preserve_thinking is False.
• Passes the policy explicitly through both MultiModalConversation.call and Generation.call history conversion paths. It does not depend on the current turn's enable_thinking.
• Removes complete existing <think>\n...\n</think> protocol blocks from assistant strings, including multiple blocks interleaved with answer text, without trimming block-external whitespace.
• Preserves structured tool calls, IDs, names, arguments, tool-result mapping, and the existing single-space fallback for empty assistant content. Original input messages and other roles are unchanged.
• Leaves dify-agent, current-turn thinking capability, response/stream handling, and other Tongyi models unchanged.
• Adds final-SDK-message assertions and explicitly opt-in live tests; bumps Tongyi from 0.2.26 to 0.2.27.

Explicit compatibility trade-off

Unmarked historical thinking and legitimate assistant prose with the exact same formatting cannot be reliably distinguished. Under this proposal, a complete, non-nested <think>\n...\n</think> example in assistant prose, including inside a Markdown code fence, is also removed from the next outbound request. This does not alter the original output or stored source history, but can remove a legitimate example from the next model's context. Maintainer review of this trade-off is requested; this is not presented as lossless provenance detection.

Inline <think>example</think>, normal tag mentions, incomplete/orphan tags, different-case/attribute variants, and escaped tags are retained. Matching does not cross another literal think tag to complete a malformed block. No orphan-tag cleanup or output-side filtering is introduced. Reliable source distinction would require trusted provenance or structured reasoning; merely adding a new marker would leave existing unmarked history unresolved.

This is separate from #3277, fixed by #3282: that issue was response-chunk content loss, whereas this PR changes outbound assistant history.

Release Notes

• Prevent complete historical thinking blocks from being replayed as ordinary assistant text for qwen3.8-max when preserved thinking is disabled, including when current-turn thinking is off.
• Preserve answer text outside those blocks, structured tool calls, and source history.
• Compatibility note: exact-format <think>\n...\n</think> examples in assistant history are interpreted as thinking blocks and removed from subsequent requests, even inside code fences. Ordinary inline tag mentions and incomplete tags are unchanged.

Change Type

• Documentation / non-plugin change
• Non-LLM plugin (tools, extensions, datasource, etc.)
• LLM plugin

Screenshots / Videos

Provider request-conversion change; no UI modification. The following reflects assertions at the SDK call boundary:

Before	After
assistant.content = "<think>\nHistorical reasoning\n</think>Final answer"	assistant.content = "Final answer"
Thinking-only assistant message with structured tool calls	Same tool calls retained; content uses existing " " placeholder
Current turn enable_thinking=False still replays historical reasoning	Historical protocol blocks removed independently of current thinking
Exact-format protocol example inside assistant prose/code fence	Example removed from outbound history: documented compatibility limitation

For the registered vision route, the text is wrapped in [{"text": ...}]; the same policy applies before wrapping.

LLM Plugin Checklist

Areas affected by this change (check all that apply)

• Message flow (system messages, user ↔ assistant turn-taking)
• Tool interaction flow (multi-round usage, Agent App and Agent Node)
• Multimodal input (images, PDFs, audio, video, etc.)
• Multimodal output (images, audio, video, etc.)
• Structured output (JSON, XML, etc.)
• Token consumption metrics
• Other LLM functionality (reasoning, grounding, prompt caching, etc.)
• New models / model parameter fixes

The multimodal endpoint's assistant-text conversion is covered; image/video payload behavior is not changed. Tool-flow verification is at the plugin/API boundary, not a full Agent App/Agent Node UI run.

Version

• Bumped top-level version in manifest.yaml (not the one under meta)
• dify_plugin>=0.3.0,<0.11.0 is declared in pyproject.toml and locked in uv.lock (or kept in requirements.txt for legacy plugins without uv.lock) — SDK docs

Version: 0.2.26 → 0.2.27; meta.version remains unchanged. Dependencies are intentionally unchanged: the existing declaration is dify_plugin>=0.10.2 (without the template's upper bound), and the lockfile/test environment uses 0.10.2.

Testing

• Local deployment — Dify version: not run as a full deployment
• SaaS (cloud.dify.ai)
• Standalone plugin: Python 3.12.9, dify_plugin==0.10.2, dashscope==1.27.4.
• Offline Tongyi suite: 327 passed, 39 skipped, two dependency deprecation warnings. The 39 skips are 35 existing live invocation cases plus 4 opt-in live cases added here.
• Added regression coverage: 101 offline cases across both SDK call routes. Captures final messages, normalized parameters, thinking on/off, tools and associated results, interleaved blocks, protocol ambiguity, unchanged non-target models, original-object preservation, and current reasoning output.
• Test-first evidence: the initial regression matrix had 46 failures / 53 passes before the implementation, failing on expected outbound-history content assertions, and passed after it.
• Real DashScope: 4 passed for fresh vs synthetic-history context × current thinking off/on. Each runs one tool-call → local marker result → final-answer round trip. The reasoning-enabled cases retain current reasoning output. Only synthetic data and a side-effect-free local marker tool are used.
• Ruff for the added test file; git diff --check.

Commands from the repository root:

PYTHONPATH="$PWD/models/tongyi" env -u DASHSCOPE_API_KEY models/tongyi/.venv/bin/python -c 'import dify_plugin, pytest; raise SystemExit(pytest.main(["models/tongyi/tests", "-q", "--tb=short"]))'

RUN_TONGYI_THINKING_HISTORY_LIVE=1 models/tongyi/.venv/bin/python -c 'import dify_plugin, pytest; raise SystemExit(pytest.main(["models/tongyi/tests/test_llm_thinking_history.py", "-k", "live", "-q", "--tb=short", "-x"]))'

uvx --offline ruff check models/tongyi/tests/test_llm_thinking_history.py
git diff --check

The live command additionally requires DASHSCOPE_API_KEY in the environment and makes real billable API calls; live cases are skipped unless explicitly opted in. Importing dify_plugin first preserves the production gevent patch ordering, and PYTHONPATH lets the existing import-order subprocess test resolve the plugin package from the repository root.

Not verified: full Dify Agent App UI/database thought persistence, the original production conversation, or whether all reported stray tags/pseudo-tool outputs disappear. No such end-to-end claim is made. These are new-context vs synthetic-history comparisons on the patched plugin, not a production before/after experiment.

AI-assisted implementation and validation; executed test results and known limitations are reported above.