feat(model): make DeepSeek thinking_enable tri-state so unset no longer force-sends thinking.type=disabled - #3398

AgentScope Version

2.0.10dev (main @ 7c198c5)

Description

DeepSeekChatModel.Parameters.thinking_enable is declared as a non-optional bool with default=False (src/agentscope/model/_deepseek/_model.py:39), and _call_api builds thinking.type unconditionally — so every request carries extra_body["thinking"] = {"type": "disabled"} even when the caller never configured anything. There is no way to express "unset, let the server default apply", which silently disables thinking on DeepSeek hybrid-reasoning models relative to a bare DeepSeek API call.

Changes (mirrors the volcengine model, whose near-identical block at model/_volcengine/_model.py:209-222 is already wrapped in the intended tri-state guard):

• thinking_enable: bool = Field(default=False, ...) → bool | None = Field(default=None, ...);
• the request-building block is wrapped in if self.parameters.thinking_enable is not None: — verbatim volcengine style, no new patterns.

Complementary to #3069 (extra_body mutation fix): that PR stopped the caller's thinking dict from being pinned by setdefault; this one stops the unset case from force-sending disabled at all. Same tri-state direction as #3191/#3192 for DashScope. Callers who relied on the old implicit default (nothing configured → thinking disabled) must now pass thinking_enable=False explicitly.

How to test

Regression cases added to the existing tests/model_deepseek_test.py:

• test_thinking_enable_unset_not_sent — fails on main ('thinking' unexpectedly found in {'thinking': {'type': 'disabled'}}), passes with this fix;
• test_thinking_enable_explicit_false_sends_disabled / test_thinking_enable_explicit_true_sends_enabled — pin the explicit behavior so it cannot regress;
• test_extra_body_is_not_mutated updated to pass thinking_enable=False explicitly, keeping its original intent (caller's extra_body is never mutated) under the new tri-state default.

python -m pytest tests/model_deepseek_test.py
# 20 passed.

MoonshotChatModel's non-k3 branch (model/_moonshot/_model.py:211-224) has the same gap — happy to extend this PR or follow up separately, maintainer's preference.

Fixes #3396

Checklist

• An issue has been created for this PR
• I have read the CONTRIBUTING.md
• Docstrings are in Google style
• Related documentation has been updated (e.g. links, examples, etc.) in documentation repository
• Code is ready for review