fix(ollama): skip think for non-reasoning models by leno23 · Pull Request #82445 · openclaw/openclaw

leno23 · 2026-05-16T04:49:00Z

Summary

avoid forwarding truthy native Ollama think values for models marked reasoning: false
keep top-level native Ollama think for reasoning-capable/unknown models
keep forwarding think: false so explicit thinking-off requests still work

Test Plan

node scripts/run-vitest.mjs extensions/ollama/src/stream-runtime.test.ts -- -t "non-reasoning"
pnpm check:test-types

Real behavior proof

Behavior addressed: native Ollama chat requests for a non-reasoning model no longer include top-level or options-level think, avoiding the error path triggered when such servers reject thinking for that model.

Real environment tested: local OpenClaw Ollama stream runtime on this PR branch, Node 22.18.0, with a local Ollama-compatible HTTP server that returns 400 model does not support thinking if either think field is present.

Exact steps or command run after this patch:

pnpm exec tsx .hermes-ollama-proof.ts

The temporary proof script started a local HTTP server, ran createConfiguredOllamaCompatStreamWrapper + createOllamaStreamFn for model llama3.2:latest with reasoning: false and configured params.thinking: "medium", then consumed the resulting stream.

Evidence after fix:

{"path":"/api/chat","hasTopLevelThink":false,"hasOptionThink":false,"model":"llama3.2:latest","optionsKeys":[]}
{"result":"stream completed","eventTypes":["start","text_start","text_delta","text_end","done"]}

Observed result after fix: the request reached /api/chat without think; the local Ollama-compatible server returned a successful NDJSON stream instead of the 400 rejection branch.

What was not tested: a live Ollama daemon/model was not available in this cron environment (ollama command was not installed), so the proof uses the real OpenClaw native Ollama request path against a local compatibility server rather than a downloaded Ollama model.

Risk/Notes

Low-to-medium risk: scoped to native Ollama thinking request shaping.
CI currently has unrelated main-branch node shard failures after merging latest upstream main; the Ollama focused test and test typecheck passed locally.

clawsweeper · 2026-05-16T04:52:34Z

Codex review: needs maintainer review before merge.

Summary
The PR adds an Ollama native think forwarding guard and regression tests, updates the changelog, and switches GitHub Copilot device-login fetches/tests to the SDK SSRF guard.

Reproducibility: yes. Current main source can emit truthy top-level native Ollama think for models represented as reasoning: false, and upstream Ollama rejects truthy req.Think for non-thinking models; I did not run the failing request locally.

Real behavior proof
Sufficient (live_output): The PR body supplies after-fix live output from the real OpenClaw native Ollama stream path against a rejecting loopback server, and maintainer comments add focused latest-head verification.

Next step before merge
No repair lane is needed; the patch has no blocking review findings and remaining work is normal maintainer review plus CI clearance.

Security
Cleared: The diff does not add dependencies, CI permissions, release scripts, or secret exposure, and the GitHub Copilot change replaces raw fixed-endpoint fetches with the existing SSRF guard.

Review details

Best possible solution:

Land this narrow guard, or an equivalent one, after maintainer review and CI clearance while preserving top-level native think for supported or unknown models and explicit think: false.

Do we have a high-confidence way to reproduce the issue?

Yes. Current main source can emit truthy top-level native Ollama think for models represented as reasoning: false, and upstream Ollama rejects truthy req.Think for non-thinking models; I did not run the failing request locally.

Is this the best way to solve the issue?

Yes. Suppressing only truthy think for known non-reasoning models is narrower than moving think into options, because upstream models think as a top-level ChatRequest field and Options has no think member.

What I checked:

Live PR state: GitHub REST shows this PR is open, not draft, mergeable, and currently targets main b9921e2 from head ff73e14. (ff73e14638e9)
Current main can send top-level Ollama think: Current main's native Ollama wrapper writes payloadRecord.think = think, and configured model params are resolved into top-level requestParams.think. (extensions/ollama/src/stream.ts:239, b9921e21b93d)
Current main has the model capability signal the PR uses: Ollama model definitions set reasoning from /api/show capabilities when available, so missing thinking capability becomes reasoning: false. (extensions/ollama/src/provider-models.ts:251, b9921e21b93d)
Upstream Ollama contract supports the root cause: Ollama's ChatRequest has top-level Think, and server routing rejects truthy req.Think when the loaded model lacks CapabilityThinking.
Upstream options contract does not make options.think the native fix: Upstream Options lists runtime/model options such as temperature, stop, and num_ctx runner fields, but not think; think is modeled separately on ChatRequest.
PR implementation: The PR adds shouldForwardNativeOllamaThink, preserving think: false while suppressing truthy think only when model?.reasoning === false, and applies it to configured and runtime thinking paths. (extensions/ollama/src/stream.ts:278, ff73e14638e9)

Likely related people:

steipete: Provided maintainer verification on the PR discussion and has recent Ollama SDK/barrel history plus PR-branch commits for changelog and the GitHub Copilot guarded-fetch addition. (role: recent area contributor and verifier; confidence: high; commits: 9917f3b3a10e, a3e73daa6b17, 932248107585; files: extensions/ollama/api.ts, extensions/ollama/runtime-api.ts, src/plugin-sdk/ollama.ts)
BruceMacD: Authored the earlier Ollama think=false request-shaping change in the same native stream behavior lineage. (role: introduced related runtime behavior; confidence: medium; commits: 773c57b41882; files: extensions/ollama/src/stream.ts, src/agents/ollama-stream.test.ts, src/agents/pi-embedded-runner/extra-params.ts)
yfge: Authored the Ollama thinking profile exposure work that connects model capability/profile metadata to activation behavior. (role: thinking-profile contributor; confidence: medium; commits: 7a9efc138998; files: extensions/ollama/index.ts, extensions/ollama/provider-policy-api.ts, extensions/ollama/provider-policy-api.test.ts)
Sunwoo Yu: Introduced the native Ollama /api/chat provider and stream/test surface that this PR changes. (role: native Ollama provider introducer; confidence: medium; commits: 11702290ff80; files: src/agents/ollama-stream.ts, src/agents/ollama-stream.test.ts, docs/providers/ollama.md)

Remaining risk / open question:

Latest head still has failing CI checks that should be cleared or confirmed unrelated before merge.
The proof uses OpenClaw’s native Ollama request path against a local compatibility server, not a live Ollama daemon/model.
The PR includes a GitHub Copilot SSRF-guard fix beyond the Ollama title, so maintainers should intentionally accept that extra scope.

Codex review notes: model gpt-5.5, reasoning high; reviewed against b9921e21b93d.

leno23 · 2026-05-16T06:29:22Z

Updated the PR body with real behavior proof and re-ran the proof check.

Verification:

node scripts/run-vitest.mjs extensions/ollama/src/stream-runtime.test.ts -- -t "non-reasoning"
pnpm check:test-types
local Ollama-compatible stream proof: request omitted both top-level and options-level think; stream completed successfully

GitHub Real behavior proof is now green: https://github.com/openclaw/openclaw/actions/runs/25954911080/job/76299733325

Remaining CI failures match current upstream main failures in unrelated cron/gateway/doctor/session shards, not the Ollama plugin surface touched here.

steipete · 2026-05-16T13:03:27Z

Maintainer verification for head 90a51665c53a3ad138d6caba67808c10bbe9802e:

Behavior addressed: Ollama provider no longer forwards truthy native think payloads for models marked reasoning: false, while preserving explicit think: false.
Real environment tested: local OpenClaw checkout on macOS, rebased onto origin/main at f410a95081c8f52f3d3b846b4c8f803664937eb0.
Exact steps or command run after this patch:

node scripts/run-vitest.mjs extensions/ollama/src/stream-runtime.test.ts -- -t "non-reasoning"
git diff --check
codex-review --mode branch --parallel-tests "node scripts/run-vitest.mjs extensions/ollama/src/stream-runtime.test.ts -- -t 'non-reasoning'"
Evidence after fix: focused Ollama stream runtime test file passed, 79 tests; diff whitespace check passed; codex-review reported no accepted/actionable findings.
Observed result after fix: configured and runtime thinking settings are omitted from Ollama request payloads when the resolved model declares reasoning: false.
What was not tested: no live Ollama daemon/model was available in this maintainer checkout; provider behavior is covered by the focused request-payload regression tests and the contributor's loopback rejecting-server proof.

steipete · 2026-05-16T13:22:41Z

Updated maintainer verification for head c91088f580986b160ec18e46473c2f92da3816ec:

Behavior addressed: Ollama provider no longer forwards truthy native think payloads for models marked reasoning: false, while preserving explicit think: false and supported thinking models.
Real environment tested: local OpenClaw checkout on macOS, rebased onto origin/main at 1bd10cfee6a4ad608c97e47b4294cc81586406d9.
Exact steps or command run after this patch:

node scripts/check-no-raw-channel-fetch.mjs
node scripts/run-vitest.mjs extensions/github-copilot/index.test.ts
node scripts/run-vitest.mjs extensions/ollama/src/stream-runtime.test.ts -- -t "non-reasoning"
git diff --check origin/main...HEAD
codex-review --mode branch --parallel-tests "node scripts/run-vitest.mjs extensions/ollama/src/stream-runtime.test.ts -- -t 'non-reasoning'"
Evidence after fix: raw channel/plugin fetch boundary check passed; GitHub Copilot provider onboarding test file passed, 12 tests; focused Ollama stream runtime test file passed, 79 tests; diff whitespace check passed; codex-review reported no accepted/actionable findings.
Observed result after fix: configured and runtime thinking settings are omitted from Ollama request payloads when the resolved model declares reasoning: false; the unrelated latest-main raw fetch boundary failure in GitHub Copilot device login is also fixed by using the plugin SDK guarded fetch helper.
What was not tested: no live Ollama daemon/model was available in this maintainer checkout; provider behavior is covered by the focused request-payload regression tests and the contributor's loopback rejecting-server proof.

steipete · 2026-05-16T13:26:13Z

Rebased again for latest main; updated head is ff73e14638b6da206298d51f97762495f289861a.

Additional proof after rebase:

node scripts/check-no-raw-channel-fetch.mjs
node scripts/run-vitest.mjs extensions/github-copilot/index.test.ts
node scripts/run-vitest.mjs extensions/ollama/src/stream-runtime.test.ts -- -t "non-reasoning"
git diff --check origin/main...HEAD

All passed locally. Changelog conflict was resolved by preserving both the latest Telegram entry from main and this PR's Ollama entry.

steipete · 2026-05-16T14:05:14Z

Maintainer update pushed after rebasing on current main.

Behavior addressed: Ollama native chat no longer forwards truthy top-level think values for models whose runtime catalog marks reasoning: false; explicit think: false and reasoning-capable models are preserved.
Real environment tested: local OpenClaw checkout on Node via repo Vitest wrapper.
Exact steps or command run after this patch:

node scripts/run-vitest.mjs extensions/ollama/src/stream-runtime.test.ts -- -t 'non-reasoning'
/Users/steipete/Projects/agent-scripts/skills/codex-review/scripts/codex-review --mode branch --parallel-tests "node scripts/run-vitest.mjs extensions/ollama/src/stream-runtime.test.ts -- -t 'non-reasoning'"
git diff --check origin/main...HEAD
Evidence after fix: focused Ollama request-body regressions pass and assert both configured and runtime non-reasoning paths omit think.
Observed result after fix: no accepted/actionable Codex review findings; whitespace check clean.
What was not tested: live Ollama daemon/provider call.

openclaw-barnacle Bot added extensions: ollama triage: needs-real-behavior-proof Candidate: external PR needs after-fix proof from a real setup. size: S labels May 16, 2026

openclaw-barnacle Bot added proof: supplied External PR includes structured after-fix real behavior proof. and removed triage: needs-real-behavior-proof Candidate: external PR needs after-fix proof from a real setup. labels May 16, 2026

clawsweeper Bot added the proof: sufficient ClawSweeper judged the real behavior proof convincing. label May 16, 2026

openclaw-barnacle Bot removed the proof: sufficient ClawSweeper judged the real behavior proof convincing. label May 16, 2026

clawsweeper Bot added the proof: sufficient ClawSweeper judged the real behavior proof convincing. label May 16, 2026

clawsweeper Bot mentioned this pull request May 16, 2026

fix(ollama): propagate supportsTools, disable idle watchdog, fix thinking and keep_alive #81915

Closed

steipete force-pushed the fix/80332-ollama-think-options branch from 786cc92 to 90a5166 Compare May 16, 2026 13:03

openclaw-barnacle Bot removed the proof: sufficient ClawSweeper judged the real behavior proof convincing. label May 16, 2026

clawsweeper Bot added the proof: sufficient ClawSweeper judged the real behavior proof convincing. label May 16, 2026

steipete force-pushed the fix/80332-ollama-think-options branch from 90a5166 to c91088f Compare May 16, 2026 13:22

openclaw-barnacle Bot added the extensions: github-copilot label May 16, 2026

openclaw-barnacle Bot added size: M and removed size: S labels May 16, 2026

steipete force-pushed the fix/80332-ollama-think-options branch from c91088f to ff73e14 Compare May 16, 2026 13:26

openclaw-barnacle Bot removed the proof: sufficient ClawSweeper judged the real behavior proof convincing. label May 16, 2026

clawsweeper Bot added the proof: sufficient ClawSweeper judged the real behavior proof convincing. label May 16, 2026

leno23 and others added 4 commits May 16, 2026 14:52

fix(ollama): skip think for non-reasoning models

fa96b5c

docs: update changelog for Ollama think fix

33c41e4

test(gateway): preserve stale main labels

4a99f1f

test(gateway): align stale main label expectation

e042f0a

steipete force-pushed the fix/80332-ollama-think-options branch from ff73e14 to e042f0a Compare May 16, 2026 14:05

openclaw-barnacle Bot added size: S and removed extensions: github-copilot size: M labels May 16, 2026

steipete merged commit eebdbab into openclaw:main May 16, 2026
108 of 112 checks passed

steipete mentioned this pull request May 16, 2026

Bug: Ollama think parameter sent at top-level causes 400 for non-thinking models; should be inside options #80332

Closed

sene1337 mentioned this pull request May 16, 2026

[Bug]: switching session modelOverride to non-thinking Ollama model can silently kill the conversation (thinking 400, failover next=none) #82714

Closed

clawsweeper Bot mentioned this pull request May 20, 2026

Ollama: think:true forwarded to non-reasoning models causes HTTP 400 (regression of #69902 fix in 2026.4.23) #84496

Closed

clawsweeper Bot mentioned this pull request May 30, 2026

Fix/gateway reliability 20260421 #76125

Closed

25 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(ollama): skip think for non-reasoning models#82445

fix(ollama): skip think for non-reasoning models#82445
steipete merged 4 commits into
openclaw:mainfrom
leno23:fix/80332-ollama-think-options

leno23 commented May 16, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented May 16, 2026 •

edited

Loading

Uh oh!

leno23 commented May 16, 2026

Uh oh!

steipete commented May 16, 2026

Uh oh!

steipete commented May 16, 2026

Uh oh!

steipete commented May 16, 2026

Uh oh!

steipete commented May 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

leno23 commented May 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test Plan

Real behavior proof

Risk/Notes

Uh oh!

clawsweeper Bot commented May 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

leno23 commented May 16, 2026

Uh oh!

steipete commented May 16, 2026

Uh oh!

steipete commented May 16, 2026

Uh oh!

steipete commented May 16, 2026

Uh oh!

steipete commented May 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

leno23 commented May 16, 2026 •

edited

Loading

clawsweeper Bot commented May 16, 2026 •

edited

Loading