本文へスキップ
Kitbay
スキル

agent-introspection-debugging

MIT2026年11月17日 更新305,644GitHubスター

A skill from the ECC collection for an agent that keeps failing at a task, such as looping on the same tool call, running out of context or drifting from the goal.

スキルの内容

A skill from the ECC collection for an agent that keeps failing at a task, such as looping on the same tool call, running out of context or drifting from the goal. Instead of retrying, the agent records the failure (error, recent tool calls, goal, signs of context pressure, assumptions about the environment), matches it to a known pattern like a retry storm, a wrong working directory or a wrong hypothesis, and then takes the smallest reversible step that tests the diagnosis. It finishes with a short report of the cause, the action taken and whether the run recovered, partly recovered or is blocked, and escalates risky cases to a person.

できること

  • Capture the state of a failing agent run
  • Match the failure to common loop, context or environment causes
  • Choose a small, reversible recovery step
  • Write an introspection report on the outcome