recall-failure-lessons

Recall Failure Lessons

Call recall_failure_lessons once with compact task evidence in text. Add component only when it is already known. Do not invent an invariant, cause, or prevention action merely to make the query look complete.

Normally omit mode, top_k, and min_relevance; the tool applies its calibrated relevance threshold, authoritative lifecycle filter, pattern collapse, and dynamic zero-to-three result limit. Treat top_k only as a maximum and min_relevance as an optional stricter floor. Never lower the threshold, broaden, or retry to fill a result list.

Accept an empty result and continue. Apply returned lessons only as proposed cautions after validating them against the current task.

If later evidence in the same task clearly establishes whether the recall was applied, not applicable, contradicted, or prevented/failed to prevent recurrence, call report_memory_outcome once with the returned attempt_id, the affected returned lesson IDs, a compact evidence_code, the supported source and evidence_kind, and an observation window of 1, 7, or 30 days. Omitted optional outcome metadata uses the tool's documented defaults. Do not pause the task to manufacture an outcome or report one when it remains unknown.

Only when the tool capability is absent before invocation, pass the identical JSON on standard input to failure-memory recall once. Do not use the fallback after a timeout or ambiguous tool failure. After such a tool failure, continue without memory guidance and mention it once only when relevant. If the command is absent, continue without memory guidance and mention the installation issue once.

Do not inspect plugin files or SQLite, install anything, create temporary files, or include raw prompts, secrets, personal data, or unnecessary paths. Similarity is not proof, policy authority, or permission to merge lessons.