Skip to content

fix: improve Ollama thinking profile resolution for live-discovered reasoning models using name heuristic#94362

Closed
lizanle521 wants to merge 1 commit into
openclaw:mainfrom
lizanle521:fix-93835
Closed

fix: improve Ollama thinking profile resolution for live-discovered reasoning models using name heuristic#94362
lizanle521 wants to merge 1 commit into
openclaw:mainfrom
lizanle521:fix-93835

Conversation

@lizanle521

@lizanle521 lizanle521 commented Jun 18, 2026 •
edited
Loading

Copy link
Copy Markdown

Problem

Telegram /think menu shows only off for live-discovered Ollama thinking models (no catalog entry), but /think <level> accepts them.

Root Cause

For models discovered at runtime (not in static config/catalog), reasoning flag passed to resolveThinkingProfile is undefined. The Ollama plugin previously returned an off-only profile when reasoning was not explicitly true.

Solution

Modified resolveThinkingProfile in extensions/ollama/provider-policy-api.ts to also consider model name heuristics when reasoning is not explicitly false. We now treat a model as reasoning if either:

  • reasoning === true (catalog explicitly says it's a reasoning model)
  • isReasoningModelHeuristic(modelId) returns true (based on common reasoning model naming patterns)

This ensures that live-discovered Ollama models with reasoning capabilities (e.g. deepseek-r1, qwen3:...thinker) get the full thinking level menu, while still respecting explicit non-reasoning flags.

Fixes #93835

Real behavior proof

Setup

  • OpenClaw instance running with Telegram extension enabled.
  • Ollama provider configured and reachable.
  • Model list includes at least one reasoning model not present in the static catalog (e.g. deepseek-r1:latest or qwen3:14b).

Steps

  1. Ensure the model is discovered live (not in config.yaml model catalog).
  2. In a Telegram chat with the bot, issue /think to view the menu.
  3. Observe the available thinking levels.

Expected after fix

  • The /think menu displays multiple reasoning levels (e.g., off, low, medium, high or similar) for the live-discovered reasoning model.
  • Previously only off was shown; now the full list appears.

Optional verification

  • Run unit tests for isReasoningModelHeuristic with known reasoning model name patterns.
  • Execute listThinkingLevels(provider='ollama', model='deepseek-r1', catalog={}) and confirm it returns a non-off-only profile.

(If you have runtime logs or screenshots, attach them here.)

@openclaw-barnacle openclaw-barnacle Bot added extensions: ollama size: XS triage: needs-real-behavior-proof Candidate: external PR needs after-fix proof from a real setup. labels Jun 18, 2026
@clawsweeper

clawsweeper Bot commented Jun 18, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

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

Close as superseded: the open runtime-catalog PR is the stronger canonical fix for the same live-discovered Ollama /think menu bug, while this provider-side heuristic does not cover the reported glm-5.2:cloud case and lacks real behavior proof.

Root-cause cluster
Relationship: superseded
Canonical: #94067
Summary: The cluster centers on live-discovered Ollama reasoning metadata missing from native /think menus; this PR is a provider-side heuristic candidate, while the canonical PR fixes the caller layer by passing runtime catalog metadata into native menu resolution.

Members:

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Canonical path: Keep the remaining work on #94067 and #93835, where runtime catalog metadata remains authoritative for native /think menus.

So I’m closing this here and keeping the remaining discussion on #94067 and #93835.

Review details

Best possible solution:

Keep the remaining work on #94067 and #93835, where runtime catalog metadata remains authoritative for native /think menus.

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

Yes. The linked issue gives concrete OpenClaw 2026.6.8 Telegram/Ollama steps, and current main shows native menus can fall back to configured-only catalog metadata while Ollama discovery records reasoning: true; 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 branch is a plausible provider-policy mitigation for heuristic-matching models, but the best fix is to pass the existing runtime catalog into native /think menu and autocomplete resolution so live Ollama capability metadata stays authoritative.

Security review:

Security review cleared: The diff changes provider policy logic and imports an existing local helper; it does not touch dependencies, lockfiles, workflows, secrets, package resolution, or other supply-chain surfaces.

AGENTS.md: found and applied where relevant.

What I checked:

  • Repository policy applied: Root and extension policies were read; provider routing/fallback behavior is compatibility-sensitive and extension public artifacts should stay lightweight and owned by plugin boundaries. (AGENTS.md:34, f18ff7551edc)
  • Current PR uses name heuristic for unknown reasoning: The PR computes effective reasoning from reasoning === true || isReasoningModelHeuristic(modelId), so live-discovered models without catalog metadata depend on regex name matching. (extensions/ollama/provider-policy-api.ts:66, 3f839fd679c9)
  • Heuristic misses the reported model: Current main's Ollama heuristic only matches r1, reasoning, think, or reason, which does not match the linked issue's glm-5.2:cloud model id. (extensions/ollama/src/provider-models.ts:241, f18ff7551edc)
  • Runtime capability metadata already exists: Ollama discovery maps /api/show capabilities to reasoning, using capabilities.includes("thinking") when capabilities are available. (extensions/ollama/src/provider-models.ts:257, f18ff7551edc)
  • Current menu fallback explains the bug: Native command menu resolution uses a supplied catalog when present, otherwise it falls back to buildConfiguredModelCatalog, which can omit live-discovered Ollama metadata. (src/auto-reply/commands-registry.ts:338, f18ff7551edc)
  • Current Telegram caller omits runtime catalog: Current main resolves provider/model context for native commands but calls resolveCommandArgMenu without a runtime catalog. (extensions/telegram/src/bot-native-commands.ts:1138, f18ff7551edc)

Likely related people:

  • openperf: Authored the open canonical runtime-catalog PR that targets the same live-discovered Ollama /think menu bug across Telegram, Slack, and Discord native surfaces. (role: likely follow-up owner; 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)
  • steipete: Authored the merged Ollama live catalog metadata work that supplies reasoning: true from capability discovery, which the native menu path should consume. (role: adjacent provider/catalog contributor; confidence: high; commits: af79cd6a9d35, 2bfe4dfb703b; files: extensions/ollama/index.ts, extensions/ollama/src/provider-models.ts, src/agents/model-catalog.ts)
  • yfge: Commit history shows this author added the Ollama provider-policy thinking profile surface whose undefined-reasoning behavior is implicated by this PR. (role: introduced provider policy surface; confidence: medium; commits: 7a9efc138998; files: extensions/ollama/provider-policy-api.ts, extensions/ollama/index.ts)
  • thewilloftheshadow: Commit history ties the shared dynamic argument-menu framework used by /think choices to this author, making them relevant for the caller-layer behavior. (role: introduced dynamic command menu framework; confidence: medium; commits: 74bc5bfd7cae; files: src/auto-reply/commands-registry.ts)

Codex review notes: model internal, reasoning high; reviewed against f18ff7551edc.

@clawsweeper clawsweeper Bot added rating: 🧂 unranked krab Not merge-ready due to missing proof or serious correctness/safety concerns. status: 📣 needs proof The PR needs real behavior proof before ClawSweeper can clear the contributor ask. mantis: telegram-visible-proof Mantis should capture Telegram visible proof. P2 Normal backlog priority with limited blast radius. labels Jun 18, 2026
@clawsweeper clawsweeper Bot added the merge-risk: 🚨 compatibility 🚨 May break existing users, config, migrations, defaults, or upgrade paths. label Jun 18, 2026
This was referenced Jun 19, 2026
Closed
Closed
Closed
Merged
@steipete

steipete commented Jun 21, 2026

Copy link
Copy Markdown
Contributor

Superseded by #94067, landed in e3ccf87.

The canonical fix uses the discovered model's declared reasoning capability through the runtime catalog across Telegram, Slack, and Discord. This avoids a GLM/model-name heuristic and keeps the behavior correct for any live-discovered reasoning model. Thanks @lizanle521 for the investigation and patch.

@steipete steipete closed this Jun 21, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

extensions: ollama mantis: telegram-visible-proof Mantis should capture Telegram visible proof. merge-risk: 🚨 compatibility 🚨 May break existing users, config, migrations, defaults, or upgrade paths. P2 Normal backlog priority with limited blast radius. rating: 🧂 unranked krab Not merge-ready due to missing proof or serious correctness/safety concerns. size: XS status: 📣 needs proof The PR needs real behavior proof before ClawSweeper can clear the contributor ask. triage: needs-real-behavior-proof Candidate: external PR needs after-fix proof from a real setup.

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

[Bug]: Telegram /think menu shows only off for live-discovered Ollama thinking models, but /think <level> accepts them

2 participants

@lizanle521 @steipete