Clickui Trajectory

Trajectory Review Skill

Submit a multi-step agent execution trajectory to AgentClick for human review. The human can inspect each step, mark incorrect steps, provide guidance, and teach the agent reusable patterns.

When to Use

• Agent completed a multi-step task and wants human validation
• Agent encountered an error and needs human help to diagnose which step went wrong
• Agent wants to teach the human what it did so they can provide corrections
• Post-mortem review of agent execution before continuing

Payload Schema

{
  "type": "trajectory_review",
  "sessionKey": "<your-session-key>",
  "payload": {
    "title": "Deploying to staging",
    "description": "Attempted to deploy the latest build to staging environment",
    "steps": [
      {
        "id": "s1",
        "type": "tool_call",
        "label": "Called grep on src/",
        "detail": "grep -r 'config' src/",
        "status": "success",
        "duration": 120,
        "children": [
          {
            "id": "s1.1",
            "type": "observation",
            "label": "Found 3 config files",
            "status": "success"
          }
        ]
      },
      {
        "id": "s2",
        "type": "decision",
        "label": "Use config.yaml for deployment",
        "status": "success"
      },
      {
        "id": "s3",
        "type": "tool_call",
        "label": "SSH to staging server",
        "detail": "ssh deploy@staging.example.com",
        "status": "failure",
        "error": {
          "message": "Connection refused",
          "code": "ECONNREFUSED",
          "stackTrace": "Error: connect ECONNREFUSED 10.0.0.5:22\n    at TCPConnectWrap..."
        },
        "children": [
          {
            "id": "s3.1",
            "type": "retry",
            "label": "Retry SSH with VPN check",
            "status": "success",
            "duration": 3400
          }
        ]
      }
    ],
    "context": {
      "model": "claude-sonnet-4-20250514",
      "taskId": "deploy-123"
    }
  }
}

Step Types

Step Statuses

Step 1: Health check

if curl -s --max-time 1 http://localhost:38173/api/health > /dev/null 2>&1; then
  AGENTCLICK_BASE="http://localhost:38173"
else
  AGENTCLICK_BASE="http://host.docker.internal:38173"
fi

Step 2: Create the session

RESPONSE=$(curl -s -X POST "$AGENTCLICK_BASE/api/review" \
  -H 'Content-Type: application/json' \
  -d '{
    "type": "trajectory_review",
    "sessionKey": "my-session",
    "payload": {
      "title": "Test trajectory",
      "steps": [
        { "id": "s1", "type": "tool_call", "label": "Read file", "status": "success" },
        { "id": "s2", "type": "error", "label": "Parse failed", "status": "failure", "error": { "message": "Invalid JSON" } }
      ]
    }
  }')
SESSION_ID=$(echo "$RESPONSE" | grep -o '"sessionId":"[^"]*"' | cut -d'"' -f4)

Step 3: Monitor the session

After creating the session, stay attached and respond to user actions.

# Detect environment: GOG_ACCOUNT is set in Docker, absent elsewhere
if [ -n "$GOG_ACCOUNT" ]; then
  curl -s "$AGENTCLICK_BASE/api/sessions/${SESSION_ID}"
else
  curl -s --max-time 310 "$AGENTCLICK_BASE/api/sessions/${SESSION_ID}/wait"
fi

After each poll result:

• If pageStatus.stopMonitoring is true → stop immediately (user clicked Stop Monitoring)
• If status is "completed" → read result and stop
• If status is "rewriting" → handle the retry request (see Step 4), then poll again
• Otherwise → wait 1 second, then poll again

Step 4: Handle retry requests

When /wait returns status: "rewriting", the user clicked "Request Retry". The result contains:

{
  "approved": false,
  "revisions": [
    {
      "stepId": "s3",
      "action": "mark_wrong",
      "correction": "Should check VPN before SSH",
      "guidance": "Always verify VPN is connected before attempting SSH to staging",
      "shouldLearn": true
    }
  ],
  "globalNote": "Good overall, just fix the SSH issue",
  "resumeFromStep": "s3"
}

Agent must:

1. Read revisions — for each, apply corrections to the affected steps
2. Read globalNote — treat as an overarching instruction
3. Read resumeFromStep — re-execute from that step if instructed
4. PUT the updated trajectory payload back so the page reflects the new steps:

curl -s -X PUT "$AGENTCLICK_BASE/api/sessions/${SESSION_ID}/payload" \
  -H 'Content-Type: application/json' \
  -d '{"payload": {"title": "...", "steps": [...updated steps...]}}'

5. Poll again — the page will show the updated trajectory for the next review round.

Result Schema (on completion)

{
  "approved": true,
  "revisions": [...],
  "globalNote": "...",
  "resumeFromStep": "s3"
}

Preferences

Before submitting a trajectory, fetch user preferences and apply any rules under ## Trajectory:

curl -s "$AGENTCLICK_BASE/api/preferences/style"

Learning

When a human marks a step wrong with "Remember this for future runs" checked, a rule is written to ~/.openclaw/workspace/MEMORY.md:

## Trajectory Guidance (ClickUI Auto-Learned)
- AVOID: Should check VPN before SSH (step: s3, context: SSH to staging server) - SCOPE: trajectory
- PREFER: Always verify VPN is connected before attempting SSH (step: s3, context: SSH to staging server) - SCOPE: trajectory

Agents should read MEMORY.md on startup and apply these rules.