openclaw - 💡(How to fix) Fix [Bug]: Agents respond they are working on a request but then fail to perform any actions. [1 comments, 1 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Summary

All agents intermittently (now consistently) return a single “working on it” style message, then do no real work. No active sessions/tasks are created, no files/posts are produced, and no completion follows.

Example bad response:

“No, not yet. I’m doing the real update now and will send you the completion count as soon as it finishes.”

Version / Environment

• OpenClaw version: 2026.4.5 (3e72c03)
• Host OS: macOS
• Channel: Discord (multi-account)
• Agents: 8 configured
• Default model route: minimax-portal/MiniMax-M2.7
• Gateway target: ws://127.0.0.1:18789

What I expected

When asked to perform work, the agent should either:

1. execute immediately (with tool activity), or
2. create a real task/run and report status tied to that run.

What happened instead

• Agent sends one short “I’m working on it” type response
• Typing/progress stops
• No real execution loop/task is active
• Re-asking yields similar non-executing status text

Reproduction (high level)

1. Send work request in Discord thread/channel
2. Agent replies with short acknowledgement (“working on it” style)
3. No further execution/output appears
4. Repeats across multiple agents

Diagnostics already run

• openclaw tasks maintenance --apply
• Result: 0 reconcile / 0 cleanup / 0 prune
• Task health:
• 0 queued
• 0 running
• many historical audit warnings
• openclaw channels status --probe previously reported gateway timeout (ws://127.0.0.1:18789)
• Port check showed node listening on 18789 while probe timed out:
• 127.0.0.1:18789 (LISTEN)
• [::1]:18789 (LISTEN)
• Restarted server/gateway and cleaned rogue service
• Disabled/adjusted thread binding behavior (spawnAcpSessions)
• Added/removing guardrail prompts in AGENTS; issue still reproduces

Relevant config notes

• Discord uses groupPolicy: "allowlist" with account-level guild/channel/user allowlists
• Multi-agent setup with separate workspaces
• agents.defaults.maxConcurrent and subagent concurrency were tuned during testing

Impact

Severe: all agents eventually become non-functional for real task execution.

Attachments

I can provide redacted outputs for:

• openclaw status --all
• openclaw channels status --probe
• openclaw tasks audit --json
• recent log tail around a failed request

Steps to reproduce

1. Send work request in Discord thread/channel
2. Agent replies with short acknowledgement (“working on it” style)
3. No further execution/output appears
4. Repeats across multiple agents

Expected behavior

What I expected

When asked to perform work, the agent should either:

1. execute immediately (with tool activity), or
2. create a real task/run and report status tied to that run.

Actual behavior

What happened instead

• Agent sends one short “I’m working on it” type response
• Typing/progress stops
• No real execution loop/task is active
• Re-asking yields similar non-executing status text

OpenClaw version

2026.4.5

Operating system

MacOS 12.7.4

Install method

npm global

Model

minimax-portal/miniMaxM2.7

Provider / routing chain

home network

Additional provider/model setup details

openai-codex

Logs, screenshots, and evidence

Impact and severity

Severe: all agents eventually become non-functional for real task execution. This is intermittent but can't recover some of them at all. No work is being performed.

Additional information

No response