Skip to content

[Bug]: OpenClaw 2026.3.8 regression: empty agent payloads + websocket 1006 closures with local LM Studio backend #42270

@SamaritanOC

Description

@SamaritanOC

Bug type

Regression (worked before, now fails)

Summary

Summary

After upgrading from OpenClaw 2026.3.7 to 2026.3.8, agent responses intermittently fail.

Observed failure signatures include:
• stopReason=stop with empty payload
• stopReason=length with empty payload
• stopReason=error with "LLM request timed out."
• gateway websocket abnormal closure (1006)

This occurs with minimal prompts and with thinking disabled.

Rollback to OpenClaw 2026.3.7 restores normal behavior.

Build tested:
OpenClaw 2026.3.8 (3caab92)


Environment

OS
Linux (Parrot OS based environment)

Installation method
npm global install

npm install -g openclaw

Gateway
systemd user service

openclaw-gateway.service

Custom local LLM provider
LM Studio

Endpoint (local OpenAI-compatible API)
http://localhost:1234/v1

Available models
qwen3-8b
text-embedding-nomic-embed-text-v1.5

Sandbox mode
off

Agents involved
seer
admin
whois


Expected Behavior

Minimal agent prompt should return a response.

Command
openclaw agent --agent seer --thinking off --json --message "Reply with exactly OK"

Expected output
OK


Actual Behavior

Agent responses intermittently fail with empty payloads or timeouts.

Example outputs

Case 1
stopReason = stop
payload = ""

Case 2
stopReason = length
payload = ""

Case 3
stopReason = error
payload = "LLM request timed out."

Gateway logs during failing turns show:

gateway closed (1006 abnormal closure)
No reply from agent

Question

Is there a known regression in OpenClaw 2026.3.8 affecting:

• gateway websocket transport
• agent reply dispatch
• session routing
• embedded fallback execution

when running with OpenAI-compatible local providers such as LM Studio?

Steps to reproduce

Minimal Reproduction

Install OpenClaw 2026.3.8

npm install -g openclaw@2026.3.8

Restart gateway

systemctl --user restart openclaw-gateway.service

Run minimal test

openclaw agent
--agent seer
--thinking off
--json
--message "Reply with exactly OK"

Observed result

stopReason = stop
payload = ""


Gateway Behavior

During failing turns the gateway logs show:

gateway closed (1006 abnormal closure)
No reply from agent

Later fallback may produce

LLM request timed out

The failure pattern was observed for both

seer
admin


Backend Verification

Direct LM Studio requests succeed.

Example test

curl http://localhost:1234/v1/chat/completions

Payload

{
"model":"qwen3-8b",
"messages":[{"role":"user","content":"Reply with exactly BACKEND_OK"}],
"temperature":0
}

Response

BACKEND_OK

This indicates the LLM backend itself is reachable and responding correctly.


Regression Confirmation

Rollback procedure

npm install -g openclaw@2026.3.7
systemctl --user restart openclaw-gateway.service

Running the same test command
openclaw agent --agent seer --thinking off --json --message "Reply with exactly OK"

Returns
OK

This confirms the behavior appears only after upgrading to 2026.3.8.


Workspace Delegation Context

This installation includes a small runtime patch modifying workspace inheritance for spawned agents.

Patch locations
dist/reply-.js
dist/compact-
.js

Purpose
Ensure spawned agents inherit their own workspace directory rather than the caller workspace.

This patch was used successfully with OpenClaw 2026.3.7 and delegation behavior continued to work during the 3.8 test before the agent liveness failures occurred.

Workspace isolation does not appear to be the cause of the regression.


Diagnostic Clue From Logs

During failing runs the gateway repeatedly logs:
gateway closed (1006 abnormal closure)

followed by
No reply from agent

and eventually
LLM request timed out

This suggests the agent reply pipeline may be losing the websocket transport before the LLM response is delivered.


Evidence Bundle

A full diagnostic bundle was captured during testing.

Artifacts include

commands.sh
openclaw-version-3.8.txt
openclaw-config-validate-3.8.txt
openclaw-doctor-3.8.txt
openclaw-health-3.8.json
gateway-status-3.8.txt
journal-gateway-window.txt
seer-main.json
seer-fresh.json
admin-phase2.json
lmstudio-direct-success.json
openclaw-model-provider-summary-sanitized.json
session-id-note.txt

Rollback confirmation artifacts

post-rollback-version.txt
post-rollback-gateway.txt
post-rollback-config-validate.txt

Selected artifacts are attached to this issue.


Workaround

Remain pinned to OpenClaw 2026.3.7.
Rollback restores stable operation.


Expected behavior

Minimal agent prompt should return a response.

Command
openclaw agent --agent seer --thinking off --json --message "Reply with exactly OK"

Expected output
OK

Actual behavior

Actual Behavior

Agent responses intermittently fail with empty payloads or timeouts.

Example outputs

Case 1
stopReason = stop
payload = ""

Case 2
stopReason = length
payload = ""

Case 3
stopReason = error
payload = "LLM request timed out."

Gateway logs during failing turns show:

gateway closed (1006 abnormal closure)
No reply from agent

OpenClaw version

2026.3.8

Operating system

Parrot OS

Install method

npm install -g openclaw@2026.3.8

Model

LMStudio/qwen3-8b

Provider / routing chain

CLI request → OpenClaw gateway (systemd user service: openclaw-gateway.service) → OpenAI-compatible local provider endpoint (LM Studio) → model qwen3-8b

Config file / key location

/home/linc/.openclaw/openclaw.json

Additional provider/model setup details

Notes:

  • Provider endpoint is local
  • Sandbox mode is off
  • Failures occur in the OpenClaw gateway/agent reply path even though direct LM Studio calls succeed
  • Observed failure pattern on 2026.3.8: websocket 1006 abnormal closure → "No reply from agent" → occasional fallback timeout

Relevant areas:

  • provider/model configuration in openclaw.json
  • gateway managed by user systemd unit:
    /home/linc/.config/systemd/user/openclaw-gateway.service
  • gateway override:
    /home/linc/.config/systemd/user/openclaw-gateway.service.d/override.conf

Custom runtime context:

  • local runtime patch applied in OpenClaw dist bundle to preserve per-agent workspace inheritance
  • patch locations:
    dist/reply-.js
    dist/compact-
    .js

Important:

  • redact any tokens, secrets, or auth material before attaching config excerpts
  • this patch was already working on 2026.3.7 and delegation still worked on 2026.3.8, so it does not appear to be the direct trigger of the regression

Logs, screenshots, and evidence

Impact and severity

Openclaw with agents does not function.

Additional information

session-id-note.txt
seer-main.json
seer-fresh.json
post-rollback-version.txt
post-rollback-gateway.txt
post-rollback-config-validate.txt
openclaw-version-3.8.txt
openclaw-model-provider-summary-sanitized.json
openclaw-health-3.8.json
openclaw-doctor-3.8.txt
openclaw-config-validate-3.8.txt
lmstudio-direct-success.json
gateway-status-3.8.txt
commands.sh
admin-phase2.json

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