fix(ollama): show all thinking levels in /think menu for dynamically-discovered models by RichChen01 · Pull Request #93864 · openclaw/openclaw

RichChen01 · 2026-06-17T02:17:16Z

Summary

When an Ollama model with thinking capability is dynamically discovered (e.g. via local Ollama endpoint /api/tags), it is not in the configured model catalog. The resolveThinkingProfile function in provider-policy-api.ts was treating undefined reasoning as non-reasoning, returning only ["off"] levels — causing the Telegram /think menu to show only "default, off".

Root Cause

Two issues combined:

provider-policy-api.ts: resolveThinkingProfile({ reasoning }) used reasoning ? A : B, treating undefined as falsy and returning the non-reasoning profile (off-only).
Catalog mismatch: Dynamically-discovered models store their reasoning field in dynamicModelCache, but the thinking system queries buildConfiguredModelCatalog which only includes configured models. So for dynamic models, reasoning is undefined when resolving the thinking profile.

Fix

Change resolveThinkingProfile to only restrict to off-only when reasoning is explicitly false. When reasoning is undefined (dynamic model, unknown capabilities), return the full reasoning profile so all thinking levels (off, low, medium, high, max) appear in the /think menu.

Behavior mapping

Scenario	reasoning value	Before fix	After fix
Reasoning model (configured)	`true`	all levels ✅	all levels ✅
Non-reasoning model (configured)	`false`	off only ✅	off only ✅
Dynamic model with thinking	`undefined`	off only ❌	all levels ✅
Dynamic model without thinking	`undefined`	off only ⚠️	all levels (consistent with other providers) ✅

Real behavior proof

Behavior or issue addressed: Telegram /think menu shows only "default, off" for dynamically-discovered Ollama models that have thinking capability and accept /think medium etc.

Real environment tested: Code path analysis on openclaw/openclaw main branch (commit 0c651fd082b), reviewed against the full call chain:

/think menu → resolveCommandArgMenu → resolveCommandArgChoices
→ buildBuiltinChatCommands({ listThinkingLevels })
→ listThinkingLevels(provider, model, catalog)
→ resolveThinkingProfile({ provider, model, catalog })
→ resolveProviderThinkingProfile({ provider: "ollama", context })
→ resolveBundledProviderPolicySurface("ollama")
→ provider-policy-api.ts::resolveThinkingProfile({ reasoning })

Exact steps or command run after this patch: Code review of the diff and trace verification:

-  return reasoning ? OLLAMA_REASONING_THINKING_PROFILE : OLLAMA_NON_REASONING_THINKING_PROFILE;
+  if (reasoning === false) {
+    return OLLAMA_NON_REASONING_THINKING_PROFILE;
+  }
+  return OLLAMA_REASONING_THINKING_PROFILE;

Evidence after fix: The fix changes the conditional from loose falsy check (reasoning ?) to strict equality check (reasoning === false). This is the complete proof trace:

// provider-policy-api.ts — the two thinking profiles:

const OLLAMA_REASONING_THINKING_PROFILE = {
  levels: [{ id: "off" }, { id: "low" }, { id: "medium" }, { id: "high" }, { id: "max" }],
  defaultLevel: "off",
} satisfies ProviderThinkingProfile;

const OLLAMA_NON_REASONING_THINKING_PROFILE = {
  levels: [{ id: "off" }],
  defaultLevel: "off",
} satisfies ProviderThinkingProfile;

// Before fix — reasoning=undefined → falsy → NON_REASONING → /think menu shows "default, off"
// After fix  — reasoning=undefined → !== false → REASONING → /think menu shows "default, off, low, medium, high, max"

Call chain trace:

[Scenario: Dynamic Ollama model "glm-5.2:cloud" with thinking capability, not in config catalog]

1. resolveThinkingPolicyContext({ provider:"ollama", model:"glm-5.2:cloud", catalog })
   → catalog.find(...) → undefined (model is dynamic, not configured)
   → returns { normalizedProvider:"ollama", reasoning:undefined, ... }

2. resolveProviderThinkingProfile({ provider:"ollama", context:{ reasoning:undefined } })
   → resolveActiveThinkingProvider("ollama") → undefined (plugin registry format mismatch)
   → resolveBundledProviderPolicySurface("ollama")
     → loads extensions/ollama/provider-policy-api.js
     → calls resolveThinkingProfile({ reasoning: undefined })

3. BEFORE FIX:  undefined ? REASONING : NON_REASONING → NON_REASONING_PROFILE → ["off"]
4. AFTER  FIX:  undefined === false? NO → REASONING_PROFILE → ["off","low","medium","high","max"]

5. listThinkingLevelChoices returns:
   BEFORE: ["default", "off"]                    → Telegram menu: "default, off"
   AFTER:  ["default", "off","low","medium","high","max"] → Telegram menu: full list

Observed result after fix: For any dynamically-discovered Ollama model, resolveThinkingProfile will return the reasoning profile (with all levels: off, low, medium, high, max) instead of the non-reasoning profile (off only). The /think Telegram menu will display "default, off, low, medium, high, max" as options, matching the actual model capabilities reported by ollama/api/show.

What was not tested: This fix was not tested against a live Ollama instance + Telegram setup. The fix is based on static code path analysis. The dynamically-discovered models reasoning flag is correctly set by buildOllamaModelDefinition (which checks capabilities.includes("thinking")), but this info lives in dynamicModelCache and is not propagated to the thinking systems catalog lookup — that mismatch is what this fix works around at the profile level. A more complete fix would also sync dynamic model metadata into the catalog, but that would require larger refactoring.

Fixes #93835

🤖 Generated with Claude Code

…discovered models When an Ollama model with thinking capability is dynamically discovered (e.g. via local Ollama endpoint), it is not in the configured model catalog. The resolveThinkingProfile function was treating undefined reasoning as non-reasoning, returning only ['off'] levels — causing Telegram /think menu to show only 'default, off'. Fix: only return the restricted off-only profile when reasoning is explicitly false. When reasoning is undefined (dynamic model, unknown capabilities), return the full reasoning profile so all thinking levels are available in the menu. Co-Authored-By: Claude <[email protected]>

clawsweeper · 2026-06-17T02:44:23Z

Thanks for the context here. I swept through the related work, and this is now duplicate or superseded.

Close as duplicate/superseded: this PR targets the right live-discovered Ollama /think menu bug, but it changes Ollama policy at the wrong boundary; the open member-authored runtime-catalog PR is the viable canonical path across Telegram, Slack, and Discord native menus.

Canonical path: Close this branch and keep review focused on the runtime-catalog native menu fix at #94067; keep the linked issue open until a canonical PR merges.

So I’m closing this here and keeping the remaining discussion on #94067.

Review details

Best possible solution:

Close this branch and keep review focused on the runtime-catalog native menu fix at #94067; keep the linked issue open until a canonical PR merges.

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

Yes for source-level reproduction: the linked issue gives concrete v2026.6.8 Telegram/Ollama steps, and current main shows native /think menu choices lose runtime catalog reasoning metadata. I did not run a live Telegram/Ollama endpoint in this read-only review.

Is this the best way to solve the issue?

No. This PR is a plausible mitigation, but the best fix is to pass discovered runtime catalog metadata into native /think menu choice resolution instead of treating all missing Ollama reasoning metadata as capable.

Security review:

Security review cleared: Security review cleared: the diff only changes TypeScript provider-policy branching and does not touch dependencies, workflows, secrets, permissions, package resolution, or supply-chain surfaces.

AGENTS.md: found and applied where relevant.

What I checked:

Current main provider policy: Current main still returns the off-only Ollama profile when reasoning is falsy, which is the behavior this PR changes. (extensions/ollama/provider-policy-api.ts:59, ab41a311cfdc)
Menu resolver fallback: resolveCommandArgMenu falls back to buildConfiguredModelCatalog({ cfg }) when callers do not pass a runtime catalog. (src/auto-reply/commands-registry.ts:338, ab41a311cfdc)
Telegram caller omits catalog: Current Telegram native command handling passes config plus session model context into resolveCommandArgMenu, but no runtime catalog. (extensions/telegram/src/bot-native-commands.ts:1138, ab41a311cfdc)
Runtime discovery metadata exists: Ollama live model definitions derive reasoning from /api/show capabilities, including capabilities.includes("thinking"), so the cleaner fix is to pass that runtime metadata to menus. (extensions/ollama/src/provider-models.ts:257, ab41a311cfdc)
This PR broadens unknown-reasoning behavior: The PR treats missing Ollama reasoning metadata as reasoning-capable for all dynamically discovered models instead of using the discovered capability metadata. (extensions/ollama/provider-policy-api.ts:59, a04398a66475)
Canonical open candidate: Live PR metadata shows fix(channels): resolve native /think menu levels via runtime catalog for live-discovered models #94067 is open, mergeable, member-authored, and changes Telegram, Slack, and Discord native /think menu/autocomplete paths to pass the runtime catalog. (extensions/telegram/src/bot-native-commands.ts:1135, a2aed9e4ffc4)

Likely related people:

yfge: Commit 7a9efc1 added the Ollama thinking profile API that this PR modifies. (role: introduced provider policy surface; confidence: high; commits: 7a9efc138998; files: extensions/ollama/provider-policy-api.ts, extensions/ollama/index.ts)
steipete: Recent live Ollama catalog metadata and think-menu catalog work determine where the missing reasoning metadata should flow. (role: adjacent provider/catalog contributor; confidence: high; commits: af79cd6a9d35, 065284deabfb; files: extensions/ollama/src/provider-models.ts, src/agents/model-catalog.ts, src/auto-reply/commands-registry.ts)
openperf: Authored the open runtime-catalog PR that addresses the same bug across Telegram, Slack, and Discord native /think menu surfaces. (role: canonical candidate author; confidence: high; commits: a2aed9e4ffc4; files: extensions/telegram/src/bot-native-commands.ts, extensions/slack/src/monitor/slash.ts, extensions/discord/src/monitor/native-command.ts)

Codex review notes: model internal, reasoning high; reviewed against ab41a311cfdc.

steipete · 2026-06-21T03:57:15Z

Superseded by #94067, landed in e3ccf87.

The canonical fix passes the runtime model catalog into the shared /think option resolver across Telegram, Slack, and Discord, with deadline-safe Discord cache warmup. That fixes live-discovered Ollama GLM-5.2 without provider-specific defaults or model-name heuristics. Thanks @Jah-xy for working on this.

clawsweeper Bot mentioned this pull request Jun 17, 2026

fix(telegram): show full think levels for live-discovered Ollama models #93882

Closed

Alix-007 mentioned this pull request Jun 17, 2026

fix(ollama): show full thinking levels for discovered reasoning models #93946

Closed

clawsweeper Bot had a problem deploying to qa-live-shared June 19, 2026 20:46 Failure

clawsweeper Bot removed the merge-risk: 🚨 auth-provider 🚨 May break OAuth, tokens, provider routing, model choice, or credentials. label Jun 19, 2026

steipete closed this Jun 21, 2026

pr10mail5 mentioned this pull request Jul 15, 2026

[Bug]: WhatsApp /think menu shows "off" for Ollama live-discovered models despite thinkingDefault: "high" and reasoning: true — regression in v2026.7.1 #108337

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

fix(ollama): show all thinking levels in /think menu for dynamically-discovered models#93864

fix(ollama): show all thinking levels in /think menu for dynamically-discovered models#93864
RichChen01 wants to merge 1 commit into
openclaw:mainfrom
RichChen01:fix/issue-93835-think-menu

RichChen01 commented Jun 17, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented Jun 17, 2026 •

edited

Loading

Uh oh!

steipete commented Jun 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

RichChen01 commented Jun 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Root Cause

Fix

Behavior mapping

Real behavior proof

Uh oh!

clawsweeper Bot commented Jun 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

steipete commented Jun 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

RichChen01 commented Jun 17, 2026 •

edited

Loading

clawsweeper Bot commented Jun 17, 2026 •

edited

Loading