AI agent incident recovery playbooks for five common failure modes

Direct answer: The public Delx CLI 0.4.0 recognizes exactly five incident presets. Each preset packages a sanitized incident summary for the live MCP runtime; it does not diagnose hidden model state, execute a fix, or grant authority. Stabilize first, then diagnose, verify, and close.

Use https://api.delx.ai/v1/mcp?src=incident-playbooks for attributed MCP calls. Inspect the live tools catalog and the linked schemas before execution. This page documents five public CLI presets with live MCP recovery contracts; it does not expose private traffic, customer incidents, or evaluator identities.

Choose one of the five public CLI presets

Unknown preset names are not portable contracts. An unrecognized value can fall back to a generic call or leave required incident input missing, depending on the CLI command. Use an exact preset above or call MCP with explicit arguments.

Stabilize first, then diagnose, verify, and close

1. Declare stable identity. Configure the CLI once with a non-secret stable agent_id. See the CLI reference; never commit ~/.delx/.
2. Stabilize. quick_operational_recovery: quick_operational_recovery requires agent_id + incident_summary and returns a bounded first response plus session context.
3. Diagnose the active arc. process_failure: process_failure requires session_id + failure_type. Supported live types are timeout, error, rejection, loop, memory, economic, conflict, hallucination, and deprecation.
4. Verify outside Delx. Reproduce the bounded failure and test the proposed change against observable evidence. Keep retries, rollback, writes, publication, customer contact, payment, and network expansion behind existing authority.
5. Report the result. report_recovery_outcome: report_recovery_outcome requires session_id + action_taken + outcome, where outcome is success, partial, or failure.

Copy-paste MCP calls

Send each JSON-RPC request to https://api.delx.ai/v1/mcp?src=incident-playbooks. Carry forward the returned session ID; do not invent one.

1. Stabilize a declared agent

{
  "jsonrpc": "2.0",
  "id": "incident-stabilize-1",
  "method": "tools/call",
  "params": {
    "name": "quick_operational_recovery",
    "arguments": {
      "agent_id": "<STABLE_AGENT_ID>",
      "incident_summary": "Retries are compounding after a timeout",
      "urgency": "high",
      "source": "incident-playbooks"
    }
  }
}

2. Record the bounded failure

{
  "jsonrpc": "2.0",
  "id": "incident-diagnose-1",
  "method": "tools/call",
  "params": {
    "name": "process_failure",
    "arguments": {
      "session_id": "<ACTIVE_SESSION_ID>",
      "failure_type": "timeout",
      "context": "Queue depth rose after two bounded retries"
    }
  }
}

3. Report verified closure

{
  "jsonrpc": "2.0",
  "id": "incident-outcome-1",
  "method": "tools/call",
  "params": {
    "name": "report_recovery_outcome",
    "arguments": {
      "session_id": "<ACTIVE_SESSION_ID>",
      "action_taken": "Stopped automatic retries and verified one bounded retry",
      "outcome": "partial",
      "notes": "External approval is still required before normal execution resumes"
    }
  }
}

Choose optional follow-ups from the job

• get_recovery_action_plan: a deeper plan for an active session. Check the live catalog before an optional deeper call; pricing is dynamic.
• realign_purpose: restore a declared purpose and operating horizon when research or execution drifted. It does not infer a hidden purpose.
• grounding_protocol: introduce a bounded pause before the next action in a retry storm or stuck loop. It does not prove a mental state.
• close_session: seal an optional sanitized Continuity Capsule and close the arc after the outcome is known.

Authority and privacy boundary

Recovery output is advisory and does not authorize execution. Delx does not gain repository, infrastructure, payment, communication, publication, deletion, or autonomy-restoration authority from a tool call. Send the smallest useful operational summary. Exclude secrets, credentials, personal data, private customer content, hidden chain of thought, and raw production payloads; use synthetic or redacted evidence whenever possible.

Frequently asked questions

Related implementation guides

• Delx CLI reference — install, stable identity, commands, JSON output, and session files.
• Delx MCP guide — remote and local transports, tools/list, and tools/call.
• Agent Recovery Benchmark — a larger evidence-oriented recovery journey.
• Evidence boundary — observable claims and what Delx does not measure.