fix(ollama): implement _get_disable_thinking_kwargs so structured-output fallback can disable thinking - #3588

Fixes #3583

AgentScope Version

2.0.12dev (main @ 2f8c029)

Description

ChatModelBase.generate_structured_output only includes the no_think strategy in its fallback ladder when the provider overrides _get_disable_thinking_kwargs (src/agentscope/model/_base.py:509-524). Five providers implement it, but OllamaChatModel does not — so for Ollama reasoning models (qwen3, deepseek-r1, ...) whose thinking output crowds out the structured payload, the ladder is only forced → auto → none and never attempts the think=False request that Ollama fully supports (the parameter path already maps thinking_enable to the top-level think option at model/_ollama/_model.py:178-179).

Changes — implement the override, verbatim provider style:

def _get_disable_thinking_kwargs(self) -> dict:
    """Ollama uses a top-level ``think`` option."""
    return {"think": False}

Complementary to #3510 (Ollama think tri-state): that fix made the parameter tri-state; this makes the structured-output fallback actually use it. One interaction, documented correctly: _call_api writes kwargs["think"] = parameters.thinking_enable (model/_ollama/_model.py:178-179) before kwargs.update(generate_kwargs) (:181), so the no_think attempt's think=False overrides an explicit thinking_enable=True for that attempt — unlike the DashScope setdefault semantics; happy to align if maintainers prefer.

How to test

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

• test_disable_thinking_kwargs_is_think_false — fails on main ({} != {'think': False}).
• test_fallback_ladder_sends_think_false_and_parses — mocks client.chat and asserts the ladder sends think=False on the third (no_think) attempt and that the returned tool arguments are parsed and validated. On main the same test fails with [True, True, True] != [True, True, False].

python -m pytest tests/model_ollama_test.py
# 15 passed.

Real environment reproduction

Reproduced with a real Ollama server (0.40.3, Windows, CPU), Python client ollama 0.6.2, and qwen3:0.6b (Q4_K_M; 751.63M parameters; default thinking true). Request: extract {name: string, age: integer} from "Alice is 30 years old and lives in Berlin." through the generate_structured_output tool; temperature=0 (required for a deterministic repro — with the model's default sampling, one base-revision run succeeded on the second attempt).

Raw /api/chat with tools (3 runs each, identical):

Through generate_structured_output (thinking_enable=True, temperature=0):

The default thinking_enable=None configuration behaves the same (base fails, this PR succeeds), because qwen3:0.6b enables thinking by default. On Ollama the first three strategies are byte-identical (tool_choice is ignored and the none strategy still sends the tool schema), so the no_think attempt is the only request that differs.

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