Add `reasoning_style` override for inline-tag thinking blocks on OpenAI chat-completions (MiniMax M3, Qwen, GLM)

[buko]

Executive Summary

• The parses of reasoning content from Minimax M3 is broken in CodeWhale when the Minimax Token Plan is used as an 'OpenAI-comaptible' Provider.
• There is an existing issue v0.8.60: Add first-party MiniMax provider route that adds Minimax Token Plan as a supported provider. This is a step in the right direction but this issue does not address the more fundamental problem: developers using CodeWhale with OpenAI-compatible providers need to be able to override how the TUI parses reasoning content. This is especially true for developers using CodeWhale with private models that would never be added to the well-known Providers enum.
• This patch introduces a "OpenAI Reasoning Style Override". It is a configuration option that developers can use to change how Codewhale parses reasoning content with any OpenAI-compatible provider.

Problem

When MiniMax M3 is reached through CodeWhale's generic OpenAI-compatible provider, reasoning is streamed as inline <think>…</think> blocks inside delta.content instead of a separate delta.reasoning_content field. CodeWhale's parser has no inline-tag handling, so the literal tags appear in the visible chat. The same problem affects Qwen thinking models on raw vLLM/Ollama and GLM models on aggregators that don't have a GLM-specific parser. The format depends on the serving gateway, not the model identity.

Proposed solution

A new [overrides] config bucket lets the user declare which reasoning format a given protocol emits. The change to ~/.codewhale/config.toml looks like this:

[overrides.openai.protocol]
reasoning_style = "https://codewhale.dev/configuration/reasoning_style#inline_tags"

Built-in fragments (all under https://codewhale.dev/configuration/reasoning_style):

• #separate_field (default) — reasoning arrives in delta.reasoning_content or delta.reasoning. Current behavior.
• #inline_tags — reasoning arrives as <think>…</think> blocks inside delta.content. Used by M3, Qwen thinking family, GLM, and similar.
• #none — no reasoning support. Treats the content stream as plain text.

The fragment is the strategy selector; the base URL is the namespace. New strategies ship as new fragments, not new keys or new code paths. Plugins can register custom strategies by pointing the URL elsewhere.

Resolution order (highest priority wins): CLI flag → environment variable → [overrides] block → per-provider config → per-model config → hardcoded heuristic.

A typo in the URI or an unknown key is a startup error with a clear message. No silent fallbacks.

Use case

The format depends on the gateway, not the model. A user running Qwen via raw Ollama needs #inline_tags; the same model via Alibaba Cloud DashScope needs #separate_field. With this change, the user picks the right strategy per gateway with a single config line — no code change, no model-name matching, no per-model registry update.

Alternatives considered

• Per-model hardcoded matching (current approach). Doesn't scale to gateway-dependent formats; requires a code change per new model.
• Strip-only post-processing. An existing strip_thinking_tags already strips the tags, but only for the saved-session history list, not the live stream — and it silently eats the reasoning rather than routing it to a thinking cell.
• New ApiProvider variant per model family. v0.8.60: Add first-party MiniMax provider route #1310 (closed) added the first-party MiniMax route for v0.8.60, but that's orthogonal — it covers the provider route, not the inline-tag parser gap.

Impact

• Removes the blocker preventing M3 (and Qwen/GLM) from being usable through CodeWhale.
• Single config edit covers the whole class of inline-tag models. Future models that adopt inline tags need a config line, not a PR.
• Convergent with v0.8.58: Reasoning-content integrity — audit #861's four root causes on the constitution branch and fix the remainder #3016, v0.8.58: Wire reasoning-effort and token-limit params for OpenAI, Moonshot, Ollama, Atlascloud — kill the silent no-ops #3024, v0.8.58: Fix provider capability matrix — true context windows and thinking flags for Moonshot/OpenAI/Ollama; query Ollama num_ctx #3023, and Validate OpenRouter-compatible custom base_url reasoning/cache support #1978 — all of which touch the same gap of telling CodeWhale the reasoning format when a model is served through a non-native provider.

Additional context

Reproduction

export OPENAI_BASE_URL="https://api.minimax.io/v1"
export OPENAI_API_KEY="<token-plan-key>"
export DEEPSEEK_PROVIDER=openai
codewhale --provider openai --model MiniMax-M3
# send: "hello"

The visible response begins with a literal <think>…</think> block. Expected: a reasoning cell plus the actual reply.

Related CodeWhale issues

• #1310 — v0.8.60: Add first-party MiniMax provider route (closed)
• #3016 — Reasoning-content integrity audit (open)
• #3023 — Provider capability matrix for thinking flags (closed)
• #3024 — Wire reasoning-effort params for OpenAI-compatible providers (open)
• #1978 — Custom base_url reasoning/cache support (open)
• #861 — Thinking-block collapse root causes (referenced by v0.8.58: Reasoning-content integrity — audit #861's four root causes on the constitution branch and fix the remainder #3016)

External references

• QwenLM/Qwen3 discussion #1657 — Qwen team's thread on inconsistent <think> tag behavior across model sizes
• Alibaba Cloud Model Studio — deep thinking — Qwen DashScope uses delta.reasoning_content as a separate field
• Z.AI API docs — GLM emits both inline tags and reasoning_content depending on request config
• anomalyco/opencode#31999 — Upstream opencode's parallel M3 bug (PR #32152 in flight)

[github-actions]

Thanks @buko for the report.

This issue is staying open for maintainer triage. CodeWhale gets better because people bring us real edge cases from real machines, providers, regions, and workflows.

If you can add a reproduction, logs, version output, screenshots, or the provider/model involved, that makes it much easier for us to verify and harvest the fix. Maintainers may comment /lgtmi to mark recurring issue reporters as approved so this intake note is skipped next time.

[buko]

Please see the attached patch. This is a large patch but it significantly improves the parsing logic and moves Codewhale towards being a much more configurable system.

codewhale-reasoning-style-override.patch