Bug Description
When using PinchBench to execute test tasks with OpenClaw agents, the agent receives "Continue where you left off. The previous model attempt failed or timed out." message instead of the actual task prompt, causing the agent to wait for user instructions instead of executing the task.
Symptoms
| Symptom |
Description |
| Agent Response |
"I don't have context about the previous session" or "Continue where you left off..." |
| Agent Behavior |
Waits for user instructions instead of executing task |
| Task Score |
0.0/1.0 (0%) |
| Workspace |
New workspace created successfully, but agent doesn't use it |
| Transcript |
First user message is "Continue where you left off..." |
Root Cause
After deep analysis, the root cause is OpenClaw CLI (openclaw agent) behavior characteristics:
1. CLI Injects Default "continue" Message
When using openclaw agent --message "task" command, OpenClaw CLI injects a default "continue" message before sending the user's message.
2. Session State Issues
Even when deleting all session history and creating a new workspace each time, OpenClaw still sends the "continue" message.
3. Not Related to Bootstrap
This is not a BOOTSTRAP.md issue - even after completing bootstrap initialization, the problem persists.
4. Not Related to Model Invocation
This is not a model timeout or fallback retry issue.
Workaround
Use OpenClaw Gateway HTTP API directly instead of CLI:
- Enable HTTP API endpoint in OpenClaw node's
~/.openclaw/openclaw.json:
{
"gateway": {
"http": {
"endpoints": {
"responses": {
"enabled": true
}
}
}
}
}
- Modify PinchBench code to use HTTP API instead of CLI
Expected Behavior
The agent should receive and execute the actual task prompt, not a "continue" message.
Environment
- PinchBench version: 2.0.0-rc2
- OpenClaw Gateway
- Task execution mode: CLI (current) vs HTTP API (workaround)
Bug Description
When using PinchBench to execute test tasks with OpenClaw agents, the agent receives "Continue where you left off. The previous model attempt failed or timed out." message instead of the actual task prompt, causing the agent to wait for user instructions instead of executing the task.
Symptoms
Root Cause
After deep analysis, the root cause is OpenClaw CLI (
openclaw agent) behavior characteristics:1. CLI Injects Default "continue" Message
When using
openclaw agent --message "task"command, OpenClaw CLI injects a default "continue" message before sending the user's message.2. Session State Issues
Even when deleting all session history and creating a new workspace each time, OpenClaw still sends the "continue" message.
3. Not Related to Bootstrap
This is not a BOOTSTRAP.md issue - even after completing bootstrap initialization, the problem persists.
4. Not Related to Model Invocation
This is not a model timeout or fallback retry issue.
Workaround
Use OpenClaw Gateway HTTP API directly instead of CLI:
~/.openclaw/openclaw.json:{ "gateway": { "http": { "endpoints": { "responses": { "enabled": true } } } } }Expected Behavior
The agent should receive and execute the actual task prompt, not a "continue" message.
Environment