fix: disable thinking on Moonshot calls so kimi-k2.6 populates content

[jasonm4130]

Summary

• Disable Moonshot's thinking mode for kimi-k2.6 by passing extra_body={"thinking": {"type": "disabled"}} on the OpenAI-compatible call
• Gated on "moonshot" in base_url so other OpenAI-compatible providers (real OpenAI, Together, Groq) aren't affected — they'd 400 on the flag
• Add 2 mock-based tests covering the moonshot path and the non-moonshot path

Why

kimi-k2.6 is a reasoning model. Its OpenAI-compatible response splits across two fields:

resp.choices[0].message.content            # the structured answer
resp.choices[0].message.reasoning_content  # chain-of-thought

With max_completion_tokens=8192, large chunks (the default chunk_size=20 regularly produces 50k+ token user messages) cause reasoning to consume the entire output budget. content comes back empty, _parse_llm_json("{}") returns {"nodes":[],...}, and the chunk silently contributes nothing.

Reproduction

162-file repo (~125k words, mixed code/docs/images), default settings:

	input tokens	output tokens	nodes	edges
Before	453,282	71,361	2	1
After	446,890	40,523	138	160

The 71k output tokens before were almost entirely reasoning_content that graphify ignores. Empty content, empty graph, full bill.

Test plan

• test_kimi_call_disables_thinking — verifies the flag is sent on a moonshot URL
• test_non_moonshot_call_does_not_disable_thinking — verifies the flag is not sent on api.openai.com
• Full existing test suite (443 tests) passes with zero regressions

[Qodo-Free-For-OSS]

Hi, tests/test_llm.py patches openai.OpenAI, but CI does not install the optional openai dependency, so the patch target import will raise ModuleNotFoundError and fail the test suite.

Severity: action required | Category: reliability

How to fix: Install or skip openai tests

Agent prompt to fix - you can give this to your LLM of choice:

Issue description

CI does not install the optional openai dependency, but tests/test_llm.py patches openai.OpenAI, which requires importing openai. This causes ModuleNotFoundError and fails CI.

Issue Context

openai is only declared in the optional kimi extra, while CI installs .[mcp,pdf,watch].

Fix Focus Areas

Choose one:

• Install openai in CI by adding the kimi extra (or all) to the editable install.
• Or, make tests/test_llm.py skip when openai is not installed (e.g., pytest.importorskip("openai") or conditional skip using importlib.util.find_spec).

References:

• .github/workflows/ci.yml[24-32]
• pyproject.toml[44-55]
• tests/test_llm.py[1-57]

We noticed a couple of other issues in this PR as well - happy to share if helpful.

---

Found by Qodo code review