fix(ollama): send think=false for thinking models when thinking is off by BruceMacD · Pull Request #53200 · openclaw/openclaw

BruceMacD · 2026-03-23T22:30:41Z

Summary

Problem: When thinkingLevel is off, OpenClaw never sends think: false in the Ollama /api/chat request body. Ollama defaults to think: true for thinking-capable models, so they continue generating thinking tokens that the response parser discards. This wastes tokens and in the case of some models which behave badly it may produce empty responses.
Why it matters: All thinking-capable Ollama models (qwen3.5, kimi-k2.5, glm-5, etc.) would continue to think silently when thinking was off, delaying response.
What changed: (1) Wire onPayload into createOllamaStreamFn so payload wrappers can mutate the request body before serialization. (2) Add an Ollama-specific wrapper in extra-params.ts that sets top-level think: false when thinkingLevel === "off". (3) Regression tests covering the payload injection, non-Ollama passthrough, and non-off passthrough. (4) Widen test-support thinkingLevel type to use the canonical ThinkLevel union.
What did NOT change (scope boundary): Response-side parsing (buildAssistantMessage) is untouched. No changes to other providers. No new files beyond the test.

Note: @SnowSky1 has been added as a collaborator. This builds on their work in #50741 but aims to be as minimal as possible. Their PR had a bug where think was nested under options instead of being set as a top-level request body field where Ollama expects it.

Change Type (select all)

Scope (select all touched areas)

Linked Issue/PR

Closes Ollama thinking models produce empty responses — think: false not sent in API requests #46680
Closes Ollama provider does not pass think:false — thinking models return empty content #50702
Closes [Bug] Ollama thinking models: think:false not sent in API request body #50712
Related fix(ollama): send think=false when thinking is off #50741
This PR fixes a bug or regression

Root Cause / Regression History (if applicable)

Root cause: The Ollama stream path (createOllamaStreamFn) never called onPayload, so payload wrappers had no way to mutate the request body. Additionally, no wrapper existed to map thinkingLevel=off to Ollama's native think: false parameter. Other thinking-aware providers (SiliconFlow, Moonshot, Google) have a dedicated wrapper already.
Prior context: The Ollama native stream path was added in [Feature]: Add native Ollama API provider for streaming + tool calling support #11828. Provider-specific thinking wrappers were added incrementally for SiliconFlow, Moonshot/Kimi, and Google, but Ollama was never wired up.
If unknown, what was ruled out: N/A, root cause is clear.

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/agents/pi-embedded-runner/extra-params.ollama.test.ts
Scenario the test should lock in: When thinkingLevel=off and model.api=ollama, the request payload must have top-level think: false. Non-Ollama models and non-off thinking levels must not be affected.
Why this is the smallest reliable guardrail: The test exercises the wrapper through applyExtraParamsToAgent and asserts the final payload shape.
Existing test that already covers this (if any): None.

User-visible / Behavior Changes

Ollama thinking-capable models now return actual content immediately when thinking is set to off. No config changes needed.

Security Impact (required)

New permissions/capabilities? No
Secrets/tokens handling changed? No
New/changed network calls? No (same Ollama /api/chat endpoint, one additional field in body)
Command/tool execution surface changed? No
Data access scope changed? No

Repro + Verification

Environment

OS: macOS
Runtime/container: Node 22, Ollama 0.18.0+
Model/provider: qwen3.5:cloud via Ollama
Integration/channel (if any): N/A
Relevant config (redacted): thinking: "off"

Steps

Run Ollama with a thinking-capable model (e.g. qwen3.5:cloud)
Set thinking to off (/think off)
Send a message

Expected

Model returns response without thinking, user sees response

Actual (before fix)

Model does think, but this is not clear to the user

Evidence

Failing test/log before + passing after

Human Verification (required)

Verified scenarios: Confirmed via debug logging in an ollama server that think: false is received only when think is off. Ran pnpm test. Typecheck and all lint/format checks green.
Edge cases checked: Non-Ollama models unaffected (test). Thinking levels other than off unaffected (end-to-end).
What you did not verify:

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.

Compatibility / Migration

Backward compatible? Yes
Config/env changes? No
Migration needed? No

Failure Recovery (if this breaks)

How to disable/revert this change quickly: Revert the single commit; the wrapper is guarded by model.api !== "ollama" so only Ollama is affected.
Files/config to restore: src/agents/ollama-stream.ts, src/agents/pi-embedded-runner/extra-params.ts
Known bad symptoms reviewers should watch for:

Risks and Mitigations

Risk: Older Ollama versions (pre-0.18.0) may not recognize the think field.
- Mitigation: Ollama ignores unknown top-level fields; verified no error on older versions. The field is only injected when thinkingLevel=off, which is already a no-op for non-thinking models.

greptile-apps · 2026-03-23T22:34:16Z

Greptile Summary

This PR fixes a bug where Ollama thinking-capable models (e.g. qwen3.5, kimi-k2.5) continued generating reasoning tokens even when thinking was set to off, because the Ollama stream path never called onPayload and no wrapper existed to map thinkingLevel=off to Ollama's native think: false parameter.

ollama-stream.ts: One-liner that wires options?.onPayload?.(body, model) after the request body is assembled and before JSON.stringify — mutations from wrappers now reach the serialised body.
extra-params.ts: New createOllamaThinkingOffWrapper that intercepts the onPayload callback and injects think: false at the top level (not under options) when model.api === "ollama". Applied unconditionally when thinkingLevel === "off" but guarded internally to be a no-op for non-Ollama models.
extra-params.ollama.test.ts: Three regression tests — payload injection, non-Ollama passthrough, and non-off passthrough — including an explicit assertion that think is not placed under options (avoiding the bug from fix(ollama): send think=false when thinking is off #50741).
extra-params.test-support.ts: thinkingLevel parameter widened to the canonical ThinkLevel union so "off" and "xhigh" are usable in tests.
Minor type gap: OllamaChatRequest in ollama-stream.ts does not yet declare think?: boolean, so the wrapper relies on a Record<string, unknown> cast to inject the field. Functionally correct at runtime, but updating the interface would make the type a faithful model of the Ollama API.

Confidence Score: 5/5

PR is safe to merge; the fix is minimal, well-tested, and backward-compatible with older Ollama versions.
The change is small and tightly scoped: one hook call in the stream path, one wrapper function, and three regression tests that cover all the cases described in the PR. The think field is only injected when thinkingLevel === "off" and model.api === "ollama", so all other providers and thinking levels are unaffected. The only non-blocking finding is that OllamaChatRequest doesn't declare think?: boolean, which is a type-safety nicety rather than a correctness issue.
No files require special attention.

Comments Outside Diff (1)

src/agents/ollama-stream.ts, line 41-47 (link)

OllamaChatRequest missing think field

The OllamaChatRequest interface doesn't include a think?: boolean field, even though the new wrapper now injects it at the top level. The workaround — casting payload to Record<string, unknown> inside createOllamaThinkingOffWrapper — gets the job done at runtime (JavaScript objects are mutable and JSON.stringify will serialize the extra property), but it bypasses the type system.

Adding think?: boolean here would let TypeScript verify the field at the injection site and make the interface a faithful description of the Ollama /api/chat body.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agents/ollama-stream.ts
Line: 41-47

Comment:
**`OllamaChatRequest` missing `think` field**

The `OllamaChatRequest` interface doesn't include a `think?: boolean` field, even though the new wrapper now injects it at the top level. The workaround — casting `payload` to `Record<string, unknown>` inside `createOllamaThinkingOffWrapper` — gets the job done at runtime (JavaScript objects are mutable and `JSON.stringify` will serialize the extra property), but it bypasses the type system.

Adding `think?: boolean` here would let TypeScript verify the field at the injection site and make the interface a faithful description of the Ollama `/api/chat` body.



How can I resolve this? If you propose a fix, please make it concise.

Prompt To Fix All With AI

This is a comment left during a code review.
Path: src/agents/ollama-stream.ts
Line: 41-47

Comment:
**`OllamaChatRequest` missing `think` field**

The `OllamaChatRequest` interface doesn't include a `think?: boolean` field, even though the new wrapper now injects it at the top level. The workaround — casting `payload` to `Record<string, unknown>` inside `createOllamaThinkingOffWrapper` — gets the job done at runtime (JavaScript objects are mutable and `JSON.stringify` will serialize the extra property), but it bypasses the type system.

Adding `think?: boolean` here would let TypeScript verify the field at the injection site and make the interface a faithful description of the Ollama `/api/chat` body.

```suggestion
interface OllamaChatRequest {
  model: string;
  messages: OllamaChatMessage[];
  stream: boolean;
  tools?: OllamaTool[];
  options?: Record<string, unknown>;
  think?: boolean;
}
```

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "fix(ollama): send think=false for thinki..." | Re-trigger Greptile}

Ollama thinking-capable models default to think=true when the parameter is absent. When OpenClaw has thinking set to off, the request never included think=false, so models continued generating thinking tokens that were then discarded by the response parser, producing empty responses. Wire onPayload into the Ollama stream path so payload wrappers can mutate the request body, and add an Ollama-specific wrapper that sets top-level think=false when thinkingLevel is off. Fixes openclaw#46680, openclaw#50702, openclaw#50712 Co-Authored-By: SnowSky1 <126348592+snowsky1@users.noreply.github.com>

)

steipete · 2026-03-27T23:49:53Z

Landed via temp rebase onto main.

Gate: scoped Ollama payload verification plus repo gate checks; pnpm check / pnpm build still hit unrelated existing main failures in src/agents/compaction* and src/agents/skills*
Land commit: 4fb8c01
Merge commit: adb78fa

Thanks @BruceMacD!

)

openclaw-barnacle Bot added agents Agent runtime and tooling size: S labels Mar 23, 2026

BruceMacD force-pushed the brucemacd/empty-response-thinking-false branch from d46af03 to 3838686 Compare March 23, 2026 22:50

BruceMacD mentioned this pull request Mar 26, 2026

Ollama models with thinking enabled cause infinite loop - think:false not passed #55119

Closed

BruceMacD and others added 2 commits March 27, 2026 23:40

fix: note ollama think=false landing (openclaw#53200) (thanks @BruceMacD

4fb8c01

)

steipete force-pushed the brucemacd/empty-response-thinking-false branch from 3838686 to 4fb8c01 Compare March 27, 2026 23:49

steipete merged commit adb78fa into openclaw:main Mar 27, 2026
9 checks passed

openclaw-barnacle Bot added the maintainer Maintainer-authored PR label Mar 27, 2026

This was referenced Mar 27, 2026

fix(ollama): send think param for thinking models #46747

Closed

fix: forward Ollama think flag when thinking is disabled #45773

Closed

johnkhagler pushed a commit to johnkhagler/openclaw that referenced this pull request Mar 29, 2026

fix: note ollama think=false landing (openclaw#53200) (thanks @BruceMacD

eac962f

)

alexcode-cc pushed a commit to alexcode-cc/clawdbot that referenced this pull request Mar 30, 2026

fix: note ollama think=false landing (openclaw#53200) (thanks @BruceMacD

f523913

)

lovewanwan pushed a commit to lovewanwan/openclaw that referenced this pull request Apr 28, 2026

fix: note ollama think=false landing (openclaw#53200) (thanks @BruceMacD

5cec733

)

ogt-redknie pushed a commit to ogt-redknie/OPENX that referenced this pull request May 2, 2026

fix: note ollama think=false landing (openclaw#53200) (thanks @BruceMacD

731baaa

)

github-actions Bot pushed a commit to Desicool/openclaw that referenced this pull request May 9, 2026

fix: note ollama think=false landing (openclaw#53200) (thanks @BruceMacD

ce93207

)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(ollama): send think=false for thinking models when thinking is off#53200

fix(ollama): send think=false for thinking models when thinking is off#53200
steipete merged 2 commits intoopenclaw:mainfrom
BruceMacD:brucemacd/empty-response-thinking-false

BruceMacD commented Mar 23, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot commented Mar 23, 2026 •

edited

Loading

Comments Outside Diff (1)

Uh oh!

Uh oh!

steipete commented Mar 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

BruceMacD commented Mar 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Change Type (select all)

Scope (select all touched areas)

Linked Issue/PR

Root Cause / Regression History (if applicable)

Regression Test Plan (if applicable)

User-visible / Behavior Changes

Security Impact (required)

Repro + Verification

Environment

Steps

Expected

Actual (before fix)

Evidence

Human Verification (required)

Review Conversations

Compatibility / Migration

Failure Recovery (if this breaks)

Risks and Mitigations

Uh oh!

greptile-apps Bot commented Mar 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 5/5

Comments Outside Diff (1)

Uh oh!

Uh oh!

steipete commented Mar 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

BruceMacD commented Mar 23, 2026 •

edited

Loading

greptile-apps Bot commented Mar 23, 2026 •

edited

Loading