What problem does it solve?
When tool calls fail inside an agent run, agents often either retry blindly or silently drop the failed item, corrupting task queues and losing work. This Skill gives worker agents a deterministic recovery protocol so every failure is classified, handled, and recorded.
Core Features & Use Cases
- Failure Classification: Distinguishes transient failures (network blips, rate limits, timeouts) from structural ones (wrong selectors, missing auth, invalid schemas, permission errors) and applies the correct response to each.
- Bounded Retries with Adaptation: Retries transient errors once, fixes and retries fixable structural errors, and abandons a tool after repeated consecutive failures in favor of an alternative approach.
- Durable Failure Recording: Writes failed task status and a one-sentence reason directly into the colony SQLite task queue instead of volatile in-memory buffers, so crashes never lose failure state.
- Use Case: A worker agent scraping a queue of URLs hits repeated 403 responses; instead of looping forever, it classifies the failure as structural, records the task as failed in the tasks table with the reason, and moves to the next queue item.
Quick Start
When a tool call fails during your worker run, follow the error-recovery decision tree to classify the failure, retry or record it in the tasks table, and continue with the next queue item.
