fix: reasoning models reject tool_choice=required; bump to 1.0.2 (closes #503, #505)

[0xallam]

Fixes #503 and #505.

Issue

Every scan against a reasoning-capable model fails immediately with provider-side 400s:

• Anthropic Claude with thinking: "Thinking may not be enabled when tool_choice forces tool use"
• DeepSeek V4 thinking: "Thinking mode does not support this tool_choice"

Affects anthropic/claude-sonnet-4-6, anthropic/claude-opus-4-7, DeepSeek V4 thinking, and any future reasoning-capable provider.

Root cause

strix/core/inputs.py:make_model_settings() unconditionally sets tool_choice="required" AND attaches Reasoning(effort=...). This combination is rejected provider-side. Quoting Anthropic's docs:

"Tool use with thinking only supports tool_choice: {"type": "auto"} (the default) or tool_choice: {"type": "none"}. Using tool_choice: {"type": "any"} … will result in an error because these options force tool use, which is incompatible with extended thinking."

LiteLLM has declined to normalize this client-side (BerriAI/litellm#8883). LangChain and pydantic-ai both shipped client-side fixes for the same constraint (langchain#35544, pydantic-ai#3611).

Fixes

• `make_model_settings()` — drop `tool_choice="required"` when reasoning is enabled. Use SDK default (`None` → auto). Lifecycle-tool termination is still enforced by `_finish_tool_use_behavior` in the agent factory.
• `STRIX_REASONING_EFFORT="none"` crash — previously crashed with `AttributeError: 'NoneType' object has no attribute 'get'` because the literal string `"none"` was truthy and made it into LiteLLM's reasoning path. Now treated as "do not attach `Reasoning(...)`" — restores the documented escape hatch.
• System prompt — reinforced that text-only turns terminate the scan, since the SDK considers a text-only response the final output when `tool_choice` isn't forced. One new explicit line near existing strong language.
• Runner defensive check — when a non-interactive scan ends without `finish_scan` being called, log a clear error pointing at the text-ended turn. Replaces silent half-scans.

Behavior matrix after the fix

`STRIX_REASONING_EFFORT`	`tool_choice`	`reasoning` attached
(unset → default `high`)	`None` (auto)	yes
`low` / `medium` / `high` / `xhigh`	`None` (auto)	yes
`"none"`	`"required"`	no (literal `"none"` is now treated as off)
`None` (Python)	`"required"`	no

Provider impact

Provider	Before	After
Anthropic Claude + thinking	400 immediately	works
DeepSeek V4 thinking	400 immediately	works
OpenAI o-series	works	works (loses `required` safety net, but follows instructions reliably)
OpenAI gpt-5.x non-reasoning	works	works
Any model with reasoning off	works	works (unchanged)

Version

Bumped to 1.0.2. After merge → tag `v1.0.2` → CI binaries + PyPI publish.

Docker image stays at `:1.0.0` (unaffected).

[greptile-apps]

Greptile Summary

This PR fixes provider-side 400 errors that occurred when tool_choice=\"required\" was combined with an active reasoning/thinking configuration — a combination explicitly rejected by Anthropic and DeepSeek. It also fixes a crash introduced by passing the literal string \"none\" as a reasoning effort.

• make_model_settings() now skips tool_choice=\"required\" and skips attaching Reasoning(...) whenever reasoning_effort is None or \"none\", restoring the documented escape hatch and unblocking Claude + DeepSeek thinking models.
• runner.py gains a defensive post-loop check that logs a clear error when a non-interactive scan returns without finish_scan having been called, surfacing the text-only-turn failure mode introduced by removing the forced tool_choice.
• System prompt gets one reinforcing line making the "text-only turn = immediate silent termination" contract explicit for reasoning models that no longer have tool_choice forcing them into tool use.

Confidence Score: 5/5

Safe to merge — the changes are well-scoped and directly fix documented provider incompatibilities without altering unrelated paths.

The use_reasoning guard is logically correct for all values in the ReasoningEffort literal. The runner defensive check only fires in non-interactive mode and only logs — it does not alter control flow. No existing non-reasoning paths are changed.

No files require special attention.

Important Files Changed

Filename	Overview
strix/core/inputs.py	Core fix: adds use_reasoning guard to drop tool_choice=required and skip attaching Reasoning(effort=...) when reasoning_effort is None or the literal string none. Logic is correct and handles both documented edge cases.
strix/core/runner.py	Defensive check added post-loop to detect when finish_scan was not called; logs a clear error. Handles str and dict final_output shapes.
strix/agents/prompts/system_prompt.jinja	Adds one explicit line reinforcing that text-only turns immediately end the scan without a report; fits naturally with the adjacent autonomous-behavior rules.
pyproject.toml	Version bump from 1.0.1 to 1.0.2; reflected in uv.lock as expected.

_{Reviews (2): Last reviewed commit: "Also handle dict final_output in scan-co..." | Re-trigger Greptile}

[0xallam]

@greptile