fix: Fix vLLM provider sending hard-coded reasoning_effort values that fail server-side validation

[idling11]

Commit Message — vLLM reasoning_effort fix (#2169)

Summary

Fix vLLM provider sending hard-coded reasoning_effort values that
fail server-side validation.

Closes: #2169

Problem

CodeWhale's apply_reasoning_effort function hard-coded invalid values
for the vLLM provider:

• "max" branch sent reasoning_effort: "max" — vLLM only supports
  none, low, medium, high
• "low"/"medium"/"high" branch always sent "high", ignoring the
  user's actual configured value

Result: 400 Bad Request on every chat completion attempt.

Fix

Two-line change in crates/tui/src/client.rs:

Branch	Before	After
medium / high	hard-coded "high"	pass through actual low / medium / high
max	hard-coded "max"	downgrade to "high" (vLLM-compatible)

Files Changed

File	Change
crates/tui/src/client.rs	+11 / -2 lines

How to Test

cargo test -p codewhale-tui -- client

Config:

[providers.vllm]
reasoning_effort = "medium"

Send any chat message — should no longer receive 400.

Greptile Summary

Fixes two bugs in apply_reasoning_effort for ApiProvider::Vllm: the low/medium/high branch was always sending "high" regardless of the user's config, and the max branch was sending "max" which vLLM rejects with a 400.

• low/medium/high branch: a new inner match now maps "low"/"minimal" → "low", "medium"/"mid" → "medium", and all others → "high", mirroring the existing Openrouter/Novita logic.
• max/xhigh branch: now sends "high" (vLLM's maximum) instead of the invalid "max" string, with a clarifying comment.

Confidence Score: 4/5

The fix is targeted and correct; both vLLM branches now send valid values. The only gap is missing test coverage for the new code paths.

The two-line bug — always sending 'high' for configurable efforts and the invalid 'max' for the top tier — is correctly patched. The inner match mirrors proven logic already used for Openrouter/Novita. No vLLM-specific unit tests accompany the change, so if someone later edits the inner match arms, there's nothing to catch a regression.

crates/tui/src/client.rs — the two new vLLM branches in apply_reasoning_effort have no dedicated tests.

Important Files Changed

Filename	Overview
crates/tui/src/client.rs	Fixes hard-coded reasoning_effort values for vLLM: low/medium/high are now passed through correctly, and 'max' is downgraded to 'high'. Logic is correct; no vLLM-specific tests were added to cover the new branches.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[apply_reasoning_effort called\nwith vLLM provider] --> B{normalized effort}
    B -- off/disabled/none --> C[chat_template_kwargs:\nenable_thinking: false]
    B -- low/minimal/medium/mid/high/empty --> D[chat_template_kwargs:\nenable_thinking: true]
    D --> E{inner match}
    E -- low or minimal --> F[reasoning_effort: 'low']
    E -- medium or mid --> G[reasoning_effort: 'medium']
    E -- high or empty or other --> H[reasoning_effort: 'high']
    B -- xhigh/max/highest --> I[chat_template_kwargs:\nenable_thinking: true]
    I --> J[reasoning_effort: 'high'\n downgraded from max]

Loading

Comments Outside Diff (1)

1. crates/tui/src/client.rs, line 936-947 (link)

   No tests added for vLLM reasoning effort mapping

   Every other provider in apply_reasoning_effort has at least one test (Deepseek, NvidiaNim, Fireworks, Openrouter each have dedicated #[test] cases), but the two vLLM branches changed in this PR have no coverage at all. Without tests, a future refactor of the inner match arms or a mistaken copy-paste could silently reintroduce the hard-coded "high" regression. A test along the lines of the existing reasoning_effort_maps_openrouter_scale_without_deepseek_max_label test that covers low → "low", medium → "medium", high → "high", and max → "high" for ApiProvider::Vllm would lock in the fix.

   

_{Reviews (1): Last reviewed commit: "fix: vLLM provider — pass through reason..." | Re-trigger Greptile}

[gemini-code-assist]

Code Review

This pull request updates the reasoning effort handling for the vLLM provider in crates/tui/src/client.rs. Instead of hardcoding "high", it now dynamically maps the user-chosen reasoning effort value ("low"/"minimal" to "low", "medium"/"mid" to "medium", and defaults to "high"). Additionally, it downgrades "max" to "high" to prevent sending an invalid value to vLLM. There are no review comments to address, and I have no additional feedback to provide.