Skip to content

[Bug]: Empty responses from OpenAI-compatible models with reasoning #81058

@varoudis

Description

@varoudis

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Bug Report: Empty responses from OpenAI-compatible models with reasoning

I'm seeing empty responses ("⚠️ Agent couldn't generate a response") from self-hosted OpenAI-compatible models with reasoning enabled. Sessions run successfully on the backend but deliver nothing to the user.

Root cause:

Commit f66b1d1 (May 10) — "fix: strip OpenAI-compatible replay reasoning" — changed reasoning stripping from Gemma4-only to ALL openai-completions models by default. Any model using openai-completions API now has historical reasoning blocks stripped on replay unless the provider explicitly opts out via dropReasoningFromHistory: false.

Commit e7277b4 (May 8) — "refactor(agents): preserve raw reasoning stream and push formatting to edge" — keeps reasoning deltas raw internally instead of pre-formatting them.

The interaction issue:

Models that output reasoning/thinking content in a raw format get their content mangled when replayed. The core sends the raw reasoning stream, but the replay sanitizer (via the new default) strips it. On some backends this causes the content block parser to fail — the model output gets placed entirely inside a thinking block instead of proper thinking + text separation. The result is an empty user-facing response.

Workaround:

Set "dropReasoningFromHistory": false in the provider config. This matches what DeepSeek and ZAI already do:

"my-provider": {
"api": "openai-completions",
"dropReasoningFromHistory": false,
...
}

Suggested fix:

The blanket openai-completions default stripping is too broad. Either:

  1. Only apply dropReasoningFromHistory: true to models known to fail on replayed reasoning (e.g., Gemma 4), or
  2. Let individual provider plugins declare their own default instead of inheriting a universal one

Anyone else hitting empty responses from self-hosted reasoning models since the May releases?

Steps to reproduce

Models that output reasoning/thinking content in a raw format get their content mangled when replayed. The core sends the raw reasoning stream, but the replay sanitizer (via the new default) strips it. On some backends this causes the content block parser to fail — the model output gets placed entirely inside a thinking block instead of proper thinking + text separation. The result is an empty user-facing response.

Expected behavior

not an empty response.

Actual behavior

Agent couldn't generate a response. Note: some tool actions may have already been executed — please verify before retrying.

OpenClaw version

2026.5.7

Operating system

macos

Install method

npm

Model

kimi 2.6

Provider / routing chain

sglang

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workingregressionBehavior that previously worked and now fails

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions