Default coding-agent servers to reasoning off - #7521
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 343bcfd803
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| "--reasoning", | ||
| server.reasoning, |
There was a problem hiding this comment.
When the outer CLI has been upgraded but the managed Studio environment still contains a llama-server version predating --reasoning, this unconditional default breaks every auto-started GGUF agent: studio.run explicitly supports re-executing a newer outer CLI into an older managed venv and forwards unknown options to llama-server, where the unsupported flag terminates startup before the health check succeeds. Pass this setting through a version-compatible mechanism or probe the installed llama-server before adding the argument.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in 2bbd9fc. The launcher now sets llama.cpp documented LLAMA_ARG_REASONING environment variable instead of forwarding the newer --reasoning flag. Current llama-server versions honor the env equivalent, while older managed versions safely ignore it rather than failing startup on an unknown argument. Regression coverage verifies the CLI flag is absent and the env defaults to off or carries an explicit mode.
| enable_tools: bool = _ENABLE_TOOLS_OPTION, | ||
| tool_call_healing: Optional[bool] = _TOOL_CALL_HEALING_OPTION, | ||
| tool_call_nudging: Optional[bool] = _TOOL_CALL_NUDGING_OPTION, | ||
| reasoning: Literal["on", "off", "auto"] = _REASONING_OPTION, |
There was a problem hiding this comment.
When a previous unsloth start has retained its auto-started server, rerunning with --reasoning on or --reasoning auto accepts the option but _require_studio immediately takes the attach path and never applies it, so the agent silently continues using the original off mode. The same attach path already warns for ignored sampling overrides; track an explicit reasoning override and likewise tell the user to stop and restart the server.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in 2bbd9fc. Reasoning is now kept unset until the user explicitly supplies --reasoning, so the attach path can distinguish an override from the auto-start default. Attaching with on, off, or auto now warns that the running server keeps its current mode and gives the stop-and-restart command. Omitted reasoning stays silent. Full test_start.py: 342 passed; Ruff and diff checks pass.
|
@codex review |
|
Codex Review: Didn't find any major issues. Can't wait for the next one! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
2 participants
Summary
--reasoning on|off|autotounsloth start--reasoning off--reasoning autoto restore llama.cpp model-detected behavior andonto force reasoningWhy
Some reasoning models emit a complete tool call in
reasoning_contentinstead of the structuredtool_callsfield expected by coding harnesses. The model can describe the correct edit without the harness ever executing it. This has also been reported upstream for Qwen3.5/3.6 in llama.cpp.The default only applies when
unsloth startlaunches its own server. Attaching to an existing Studio/server does not mutate or reload that server.Before / after
Before, a Qwen3.5-9B Hermes run generated the requested patch inside reasoning content, but no operation was executed.
After, a fresh Windows auto-start logged and launched llama.cpp with:
Pi then received a structured tool call, wrote and read
reasoning-proof.txtcontainingSTRUCTURED_TOOL_OK, and exited successfully.Validation
tests/test_start.py: 335 passed, 3 skippedos.nametont, which makespathlibrejectWindowsPathon LinuxReferences: