openclaw - 💡(How to fix) Fix [Request]: Backport fix for Gemma 4 reasoning_content replay bug (#68704) to v2026.5.x stable [1 comments, 2 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Issue #68704 was fixed in commit 556c3e8 and merged to main, but has not been backported to stable. Users on v2026.5.2 (current stable) are still affected — after 3–5 turns with reasoning enabled, agent responses become incoherent due to prior-turn thinking blocks being re-sent to the LM Studio OpenAI-compat Gemma 4 endpoint.

Steps to reproduce

1. Set up OpenClaw with LM Studio provider, model: google/gemma-4-26b-a4b-it (or similar Gemma 4 variant).
2. Enable reasoning: /reasoning on (or leave at default with Think: high).
3. Keep LM Studio's Reasoning Parsing enabled in Inference settings.
4. Send 3–5 messages in a single session.
5. Observe agent responses become incoherent — answers reference unrelated content from earlier turns instead of the current message.

Workaround: /reasoning off (with LM Studio Reasoning Parsing still ON) prevents corruption. Agent still reasons internally; OpenClaw simply does not manage the thinking blocks.

Expected behavior

Prior-turn thinking blocks should be stripped from replay history before being sent to the LM Studio Gemma 4 endpoint, as Gemma 4's OpenAI-compat API does not accept reasoning_content in conversation history. On v2026.4.26, Gemma 4 worked correctly (reasoning blocks were not injected). Commit 556c3e8 implements this stripping policy for main; backporting to v2026.5.x stable should restore correct behavior.

Actual behavior

After 3–5 turns with reasoning enabled, agent responses become incoherent — the model pattern-matches from stale earlier turns instead of responding to the current message. The context window grows normally (21%→22%), but response content references unrelated earlier conversation content. Disabling /reasoning off stops the corruption immediately.

OpenClaw version

v2026.5.2 (8b2a6e5)

Operating system

WSL2 Ubuntu on Windows 11

Install method

npm global

Model

lmstudio/gemma-4-26b-a4b-it

Provider / routing chain

openclaw -> LM Studio (OpenAI-compat local endpoint)

Additional provider/model setup details

Provider config: LM Studio local server at http://localhost:1234/v1, api: openai-responses, reasoning: true. Model entry: { id: "google/gemma-4-26b-a4b-it", reasoning: true, input: ["text", "image"] }

Logs, screenshots, and evidence

Version comparison (cross-version control):
- v2026.4.26: Qwen 3.6 35B (LM Studio) — reasoning NOT displayed. Gemma 4 26B — reasoning NOT displayed. No context corruption on either.
- v2026.5.2: Qwen 3.6 35B — reasoning NOW displayed (new). Gemma 4 26B — reasoning still NOT displayed. Context corruption appears on Gemma 4.

The Qwen change confirms commit 2fd1e7b ("fix: normalize LM Studio binary reasoning efforts") introduced a universal normalize layer affecting all LM Studio local models — not Gemma-specific.

Related upstream fix (not yet in stable release):
- Commit 556c3e8 "fix(agents): strip Gemma reasoning from local replay" (v2026.5.4-beta.1) — adds dropReasoningFromHistory policy for isStrictOpenAiCompatible + isGemma4ModelRequiringReasoningStrip models.

Session snapshot:
- OpenClaw v2026.5.2 (8b2a6e5)
- Model: lmstudio/gemma-4-26b-a4b-it
- Think: high · elevated
- Context grew from 21% → 22% (55k–104k tokens) across observed sessions
- Compactions: 0

Impact and severity

Affected: Users running OpenClaw with LM Studio + Gemma 4 models (any variant) on v2026.5.2. Severity: High — agent produces hallucinated or context-mismatched responses, making the model unusable for multi-turn conversations. Frequency: Always reproducible after 3–5 turns with reasoning enabled. Consequence: Effectively blocks use of Gemma 4 via LM Studio until workaround (/reasoning off) is applied or 556c3e8 is released in stable.

Additional information

Last known good version: v2026.4.26 (Gemma 4 worked without context corruption, reasoning simply not displayed). First known bad version: v2026.5.2.

Root cause (supported by commit diff): Commit 2fd1e7b added normalizeLmstudioTransportReasoningCompat() which enables reasoning_effort round-trip for all LM Studio models. For Gemma 4, this causes prior-turn thinking blocks to be included in replay messages. LM Studio's OpenAI-compat Gemma 4 endpoint silently rejects or mishandles these blocks, corrupting message history.

Commit 556c3e8 (already merged to main, not yet in stable) addresses this by stripping prior-turn reasoning from Gemma 4 replay via dropReasoningFromHistory policy.

Related: #68704 (locked as resolved — fix in main but not backported to stable).

Request: Please backport 556c3e8 to a v2026.5.x patch release.