Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

I observed a reasoning-leak issue in an OpenClaw + Ollama environment where internal reasoning text was not only shown in the UI, but also persisted into session .jsonl history.

The affected reply shape looked like:

[reasoning prose] </think> [final answer]

In this malformed case, the reasoning text survived visible-text sanitization and was treated as normal assistant content.

Workaround

A local workaround that worked for us was to add a sanitizer fallback with this behavior:

• find the last </think> in assistant text
• if there is non-empty text before and after it
• keep only the text after </think>

In other words, for malformed replies shaped like:

[reasoning prose] </think> [final answer]

the sanitizer returns only:

[final answer]

Suggested Fix

Strengthen assistant visible-text sanitization so malformed reasoning output is safely removed even when:

• there is only a trailing </think>
• reasoning prose appears outside a valid tagged block
• upstream output is partially malformed or mixed into a normal text block

During the same investigation, we also encountered separate Ollama memory-pressure/timeouts under concurrent multi-user use of a large model. That appeared to be an operational capacity issue, not the root cause of the reasoning leak itself.

Steps to reproduce

Run OpenClaw with an Ollama-backed reasoning-capable model.
2. Start a new assistant session.
3. Send a simple message such as:

• Hi Neo! How are you today?
• or trigger a startup/session greeting flow

4. Observe cases where the upstream assistant output becomes malformed and contains:
   • reasoning prose
   • a trailing </think>
   • then the final answer
5. Inspect:
   • the visible reply in the UI
   • the corresponding session .jsonl file

Expected behavior

• Internal reasoning should not be shown to the user.
• Internal reasoning should not be persisted as normal assistant-visible content.
• Sanitization should remain safe even when upstream output is malformed or only partially tagged.

Actual behavior

• Internal reasoning text is shown to the user.
• The leaked reasoning is also persisted into session history as assistant text.
• Config-level settings intended to keep thinking/reasoning off are not sufficient in the affected malformed-output path.

OpenClaw version

2026.4.8

Operating system

Ubuntu

Install method

docker

Model

Local Ollama

Provider / routing chain

Openclaw -> Ollama

Additional provider/model setup details

Environment

• OpenClaw runtime
• Ollama backend
• Reasoning-capable model path
• Observed during embedded assistant runs
• Affected channels in our case included web chat, and related persisted session history

Logs, screenshots, and evidence

Impact and severity

This issue is more serious than a cosmetic UI leak because the reasoning text may also be written into persisted session history.

Additional information

No response