[DevelopmentCats]

Fixes #23.

What I found on the Sparks

spark-1 still has the profile from the original report: vllm-helper new wrote EXTRA_ARGS="--reasoning-parser deepseek_r1" for Qwen/Qwen3-0.6B on 2026-08-25, and the vllm.service journal shows vLLM 0.27.1 serving it with reasoning_parser: 'deepseek_r1' while qwen3 was registered.

The issue's measurement ("reasoning_content is empty") was a field-name mix-up. vLLM 0.27.1 returns reasoning in reasoning and no longer sends reasoning_content. The real failure is narrower. I served Qwen3-0.6B on spark-1 with each parser and sent the same request four ways:

Parser	Thinking on	Thinking on, streamed	Thinking off	Thinking off, streamed
deepseek_r1	correct	correct	content: null, answer in reasoning	correct
qwen3	correct	correct	correct	correct

With thinking off, Qwen3's template puts an empty <think></think> in the prompt, so the model's output has no tags, and deepseek_r1 treats untagged output as reasoning.

Fix

Qwen3's hybrid cards name deepseek_r1 in the vllm serve line and qwen3 in the SGLang line. When a card's SGLang command names a reasoning parser that the installed vLLM registers under exactly that name, that parser replaces the vLLM line's, and inspect, new and the panel say so. Nothing is guessed from similar names.

Cards whose lines agree are unchanged. I checked 32 reasoning-model cards: the Qwen3 hybrid models change to qwen3, while the Thinking-2507 and Next-Thinking cards (deepseek_r1 and deepseek-r1) and GLM-4.5/4.7 (glm45 on both lines) stay as they were.

Existing profiles are not rewritten. One can be fixed with llm-helper set <profile> EXTRA_ARGS=....

The issue's other two suggestions already work: --enable-reasoning is dropped with a warning that names it, and inspect shows the result before anything is pulled.

Verification

• inspect on spark-1 against vLLM 0.27.1, v0.6.0 and this branch side by side:
  • Qwen3-0.6B and Qwen3-8B: deepseek_r1 before, qwen3 with the change, plus a changed: line saying why.
  • Qwen3-30B-A3B-Thinking-2507 and GLM-4.5: identical output from both builds.
• The live serving test above, with a memory watchdog. GLM on spark-1 stayed healthy (/health 200 before and after).
• Unit tests for reading the SGLang parser from a card (including the real Qwen3-0.6B card) and for the replacement rules. Replacing into deepseek_v3 keeps its --default-chat-template-kwargs {"thinking":true}, and replacing away from it drops that default.
• go vet (Linux and macOS), go test -race ./..., gofmt, prettier, gendocs (no diff), recipediff (unchanged).

[Copilot]

🟡 Changes recommended

Replacing a parser with deepseek_v3 or poolside_v1 omits the required thinking default, potentially making the parser a no-op and leaking raw thinking tags into content.

1 open issue

• Apply thinking defaults after selecting the final parser · New

What changed in this PR

This PR addresses stale reasoning-parser recommendations by using the parser named in a model card’s SGLang command when the installed vLLM registers that exact name.

Changes:

• Extracts SGLang reasoning-parser hints and applies them during inspection and profile creation.
• Reports parser replacements in the CLI and web panel without rewriting existing profiles.
• Adds parser-selection tests, a Qwen3 card fixture, and user documentation.

File	Description
internal/​web/​static/​models.js	Displays parser-replacement notes.
internal/​vllm/​cardparser.go	Selects registered SGLang parser replacements.
internal/​vllm/​cardparser_internal_test.go	Tests replacement rules and unchanged cases.
internal/​hf/​testdata/​README_qwen3_hybrid.md	Adds a Qwen3 hybrid card fixture.
internal/​hf/​snippet.go	Extracts parser names from SGLang commands.
internal/​hf/​snippet_internal_test.go	Tests SGLang snippet parsing.
internal/​hf/​hf.go	Includes the SGLang parser in detected flags.
internal/​hf/​hf_internal_test.go	Tests detection using the card fixture.
internal/​cli/​search.go	Prints parser-replacement explanations.
internal/​catalog/​catalog.go	Applies replacements and propagates notes.
docs/​models.md	Explains parser selection and Qwen3 behavior.
CHANGELOG.md	Records the fix and existing-profile guidance.

_{Review effort: Balanced}