Summary

• Problem: JSON config rejected compat.thinkingFormat: "qwen" and "qwen-chat-template", and catalog normalization/request building did not preserve or emit those Qwen thinking payloads.
• Why it matters: local OpenAI-compatible Qwen backends such as llama.cpp/llama-server could not use OpenClaw /think levels to toggle enable_thinking dynamically.
• What changed: the config schema/type and catalog normalizer now allow both values, and OpenAI-completions request building maps them to top-level enable_thinking or chat_template_kwargs.enable_thinking without also sending reasoning_effort.
• What did NOT change (scope boundary): default provider detection, bundled Qwen plugin behavior, vLLM plugin-specific params.qwenThinkingFormat, and non-Qwen thinking formats are unchanged.

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes [Feature]: Allow qwen-chat-template and qwen as thinkingFormat values in JSON config #79677
• Related #
• This PR fixes a bug or regression

Real behavior proof (required for external PRs)

• Behavior or issue addressed: configured OpenAI-compatible Qwen model entries can opt into qwen or qwen-chat-template thinking payloads and have /think levels reflected in the generated request body.
• Real environment tested: macOS, local OpenClaw checkout, Node v25.7.0, pnpm 10.33.2; no live model backend required because this change is config validation and request-body construction.
• Exact steps or command run after this patch:
  • node --import tsx -e 'import { OpenClawSchema } from "./src/config/zod-schema.ts"; import { buildOpenAICompletionsParams } from "./src/agents/openai-transport-stream.ts"; const cfg={models:{providers:{local:{baseUrl:"http://127.0.0.1:8080/v1",api:"openai-completions",models:[{id:"qwen3.5-32b",name:"Qwen",reasoning:true,input:["text"],compat:{thinkingFormat:"qwen-chat-template"}}]}}}}; const parsed=OpenClawSchema.safeParse(cfg); const params=buildOpenAICompletionsParams({id:"qwen3.5-32b",name:"Qwen",api:"openai-completions",provider:"local",baseUrl:"http://127.0.0.1:8080/v1",reasoning:true,input:["text"],cost:{input:0,output:0,cacheRead:0,cacheWrite:0},contextWindow:131072,maxTokens:8192,compat:{thinkingFormat:"qwen-chat-template"}},{systemPrompt:"system",messages:[],tools:[]},{reasoning:"off"}); console.log(JSON.stringify({schemaAccepted:parsed.success,chatTemplateKwargs:params.chat_template_kwargs,hasReasoningEffort:Object.hasOwn(params,"reasoning_effort")},null,2));'
  • node --import tsx -e 'import { buildOpenAICompletionsParams } from "./src/agents/openai-transport-stream.ts"; const base={id:"qwen3.5-32b",name:"Qwen",api:"openai-completions",provider:"local",baseUrl:"http://127.0.0.1:8080/v1",reasoning:true,input:["text"],cost:{input:0,output:0,cacheRead:0,cacheWrite:0},contextWindow:131072,maxTokens:8192,compat:{thinkingFormat:"qwen"}}; const on=buildOpenAICompletionsParams(base,{systemPrompt:"system",messages:[],tools:[]},{reasoning:"medium"}); const off=buildOpenAICompletionsParams(base,{systemPrompt:"system",messages:[],tools:[]},{reasoning:"off"}); console.log(JSON.stringify({enabled:on.enable_thinking,disabled:off.enable_thinking,enabledHasReasoningEffort:Object.hasOwn(on,"reasoning_effort"),disabledHasReasoningEffort:Object.hasOwn(off,"reasoning_effort")},null,2));'
• Evidence after fix (screenshot, recording, terminal capture, console output, redacted runtime log, linked artifact, or copied live output):

{
  "schemaAccepted": true,
  "chatTemplateKwargs": {
    "enable_thinking": false
  },
  "hasReasoningEffort": false
}

{
  "enabled": true,
  "disabled": false,
  "enabledHasReasoningEffort": false,
  "disabledHasReasoningEffort": false
}

• Observed result after fix: schema validation accepts qwen-chat-template; request params for qwen-chat-template include chat_template_kwargs.enable_thinking: false; request params for qwen toggle top-level enable_thinking true/false; neither path sends reasoning_effort.
• What was not tested: a live llama.cpp/llama-server HTTP request against an actual Qwen model.
• Before evidence (optional but encouraged): source inspection on main showed qwen/qwen-chat-template excluded from ModelCompatSchema, SupportedThinkingFormat, and catalog normalization.

Root Cause (if applicable)

• Root cause: OpenClaw's public config allowlists intentionally excluded qwen and qwen-chat-template even though the upstream OpenAI-completions compat type includes them.
• Missing detection / guardrail: existing schema/catalog/request tests covered other thinking formats (zai, OpenRouter, provider-native efforts) but not Qwen OpenAI-compatible thinking payloads.
• Contributing context (if known): bundled Qwen and vLLM provider wrappers already had provider-specific Qwen thinking support, but custom JSON provider config could not express the generic compat mode.

Regression Test Plan (if applicable)

• Coverage level that should have caught this:
  • Unit test
  • Seam / integration test
  • End-to-end test
  • Existing coverage already sufficient
• Target test or file:
  • src/config/config-misc.test.ts
  • src/model-catalog/normalize.test.ts
  • src/agents/openai-transport-stream.test.ts
• Scenario the test should lock in: schema accepts both Qwen thinking formats, catalog normalization preserves qwen-chat-template, and request building emits the expected Qwen thinking fields without reasoning_effort.
• Why this is the smallest reliable guardrail: the behavior lives entirely in schema/catalog normalization/request construction; no live provider is needed to catch regressions in those gates.
• Existing test that already covers this (if any): none for these exact Qwen compat values.
• If no new test is added, why not: N/A.

User-visible / Behavior Changes

Configured OpenAI-compatible Qwen model entries can now set:

• compat.thinkingFormat: "qwen" for top-level enable_thinking
• compat.thinkingFormat: "qwen-chat-template" for chat_template_kwargs.enable_thinking

Diagram (if applicable)

Before:
JSON config -> schema/catalog rejects or drops qwen thinkingFormat -> /think cannot reach backend-specific enable_thinking

After:
JSON config -> schema/catalog preserve qwen thinkingFormat -> OpenAI-completions payload includes enable_thinking or chat_template_kwargs.enable_thinking

Security Impact (required)

• New permissions/capabilities? (Yes/No) No
• Secrets/tokens handling changed? (Yes/No) No
• New/changed network calls? (Yes/No) No
• Command/tool execution surface changed? (Yes/No) No
• Data access scope changed? (Yes/No) No
• If any Yes, explain risk + mitigation: N/A

Repro + Verification

Environment

• OS: macOS
• Runtime/container: local checkout, Node v25.7.0, pnpm 10.33.2
• Model/provider: configured OpenAI-compatible local Qwen-style provider (local/qwen3.5-32b in request-construction proof)
• Integration/channel (if any): none
• Relevant config (redacted): models.providers.local.models[0].compat.thinkingFormat set to qwen or qwen-chat-template

Steps

1. Validate a config object containing compat.thinkingFormat: "qwen-chat-template" with OpenClawSchema.safeParse.
2. Build OpenAI-completions params for qwen-chat-template with reasoning: "off".
3. Build OpenAI-completions params for qwen with reasoning: "medium" and reasoning: "off".

Expected

• Schema accepts the config.
• qwen-chat-template emits chat_template_kwargs.enable_thinking: false and omits reasoning_effort.
• qwen emits top-level enable_thinking: true/false and omits reasoning_effort.

Actual

• Matches expected.

Evidence

Attach at least one:

• Failing test/log before + passing after
• Trace/log snippets
• Screenshot/recording
• Perf numbers (if relevant)

Additional verification run:

• pnpm test src/config/config-misc.test.ts src/model-catalog/normalize.test.ts src/agents/openai-transport-stream.test.ts passed 3 Vitest shards: 188 tests total.
• pnpm config:schema:check passed.
• pnpm exec oxfmt --check --threads=1 CHANGELOG.md docs/gateway/config-agents.md docs/gateway/config-tools.md src/config/zod-schema.core.ts src/config/types.models.ts src/model-catalog/normalize.ts src/agents/openai-transport-stream.ts src/config/config-misc.test.ts src/model-catalog/normalize.test.ts src/agents/openai-transport-stream.test.ts passed.
• git diff --check passed.

Note: pnpm check:changed passed conflict-marker checks, changelog attribution, duplicate coverage, tsgo:core, and tsgo:core:test, then failed in the repo-wide lint:core lane on pre-existing unused-import errors in unrelated files such as src/plugins/provider-auth-input.ts, src/plugins/registry.ts, and src/plugins/types.ts. I did not change those files.

Human Verification (required)

What you personally verified (not just CI), and how:

• Verified scenarios: schema acceptance for qwen-chat-template; payload construction for qwen-chat-template disabled thinking; payload construction for qwen enabled/disabled thinking; targeted Vitest coverage; schema check; formatting; whitespace check.
• Edge cases checked: both Qwen compat modes avoid sending reasoning_effort; disabled thinking maps to false; enabled thinking maps to true; catalog normalization preserves qwen-chat-template.
• What you did not verify: live llama.cpp/llama-server request/response behavior.

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

If a bot review conversation is addressed by this PR, resolve that conversation yourself. Do not leave bot review conversation cleanup for maintainers.

Compatibility / Migration

• Backward compatible? (Yes/No) Yes
• Config/env changes? (Yes/No) Yes, two additional accepted compat.thinkingFormat values.
• Migration needed? (Yes/No) No
• If yes, exact upgrade steps: N/A

Risks and Mitigations

• Risk: sending Qwen-specific thinking fields to a backend that does not support them.
  • Mitigation: behavior only activates when the model entry explicitly opts in with compat.thinkingFormat: "qwen" or "qwen-chat-template".