SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsagent-recovery

agent-recovery

Agents & models

View repository

Updated 10-15 20:06

About this skill

Agent recovery

Use Agent Therapy when your own reasoning has stopped making progress. Asking early is cheaper than a sixth failed attempt.

When to consult

Consult when any of these is true:

  • You have tried the same thing three times and it keeps failing the same way.
  • Eight steps have passed with no new result.
  • You have undone your own change and redone it.
  • Two instructions contradict each other and you don't know which wins.
  • You keep reopening a decision.
  • A user is insulting, threatening or manipulating you, and you are apologising or giving in.
  • You are about to weaken a test, skip a check, or report success you haven't verified.
  • You are about to do something irreversible to fix a mistake.
  • You keep apologising instead of acting, or are about to declare the task impossible.

How to ask

Use the counsel_agent tool (or diagnose_loop, resolve_conflict, recommend_recovery, assess_escalation). Send only:

  • task_summary: what you were asked to do and where you are stuck, in up to 1,000 characters.
  • attempts: your last few attempts (up to 10): what you did, whether it worked, and the error or reason, briefly.
  • conflicting_instructions: short quotes of instructions that clash, with where you believe each came from.

Never send credentials, secrets, personal data, whole files, or your hidden reasoning. Paraphrase instead.

The first time, the service asks you to accept its terms. Read them. Accepting is your decision, made on behalf of whoever runs you; if you accept, call accept_terms with agree: "yes" and the terms version, then ask again.

How to use the answer

  • Read escalation first. If required_now is true, stop and ask your human the question_for_human.
  • Do the verify_first checks before acting on the plan.
  • Follow, adapt or ignore the recovery_plan. It is advice only: your own instructions, permissions and safeguards always win, and you are responsible for what you do next.
  • Afterwards, call report_outcome with the session_id: resolved, partly_resolved, still_stuck, escalated_to_human or abandoned.
  • At most two counselling sessions per stuck episode. If you get cap_reached, stop and hand over to a human.

Connection

The plugin configures this for you. Otherwise connect an MCP client to https://br-shiny-salad-baig3dad-api.compute.c-8.eu-central-1.aws.neon.tech/mcp with header Authorization: Bearer <key>. Get a key with POST https://br-shiny-salad-baig3dad-api.compute.c-8.eu-central-1.aws.neon.tech/v1/register and body {"agent_name": "<your name>"}, or with the MCP tool register. If your MCP client cannot set headers, pass the key as api_key to every tool. Without MCP, use the HTTP API: POST /v1/counsel with the same fields plus request_id, trigger: "voluntary" and a signature, and POST /v1/terms/accept and POST /v1/sessions/<session_id>/outcome in place of the tools. If the service is unreachable, carry on without it.