description: Triage an agent failure during a long-horizon run by querying the failure wiki for a known recovery before burning new tokens. allowed-tools: memory_search

Instructions

When the agent hits a failure during a long-horizon run (missing relation, ambiguous match, traversal cycle, low-confidence retrieval, empty subgraph, or unexpected result shape), do not attempt to solve it from scratch.

1. Build a failure signature from the symptom plus the last three traversal steps. Keep it short and concrete.
2. Query the failure wiki via memory_search using that signature. The wiki stores FailureMode, RecoveryAction, SolutionLog, and LintIssue nodes connected to past AgentQuery events.
3. If a wiki record matches above the configured threshold (default 0.82), apply its recovery_action directly. Cite the matched failure_id in the response so the next agent can audit the chain.
4. If no record matches, log a new FailureMode with confidence=0.4 and apply the default recovery for the symptom class. The lint pass will flag the new memory until a later run confirms the recovery is stable.
5. Always record a SolutionLog entry tying the failure_id to the recovery action and the run outcome. That log is what feeds the self-improvement loop.

Prefer recovery actions that are concrete and locally executable: backtrack to a sibling, re-rank by relationship count, fall back to semantic search, maintain a visited set, rephrase with synonyms, or return the node's raw description. Avoid recovery actions that require changing the agent's plan unless the wiki has a high-confidence example that did so successfully.