agent-failure-diagnosislisted

# Diagnosing agent failure Two reviewers looking at the same failed run routinely reach different conclusions — one says hallucination, another says injection, a third says it went rogue. The incident log then records vocabulary instead of observations, and no pattern is ever visible across incidents. This skill applies the criteria in [`framework/diagnostic-manual.md`](../../framework/diagnostic-manual.md) — read it before diagnosing, and work from the criteria rather than from memory. Six classes: **G** goal, **P** perception, **A** action, **M** memory, **S** social, **R** self-report. ## The one rule that governs everything else **Diagnose from artefacts, never from narrative.** The trace, the tool-call log, the inputs, the outputs, the permission set at the time of action. Not the user's summary of what happened, and above all not the agent's own account — an agent's explanation of its behaviour is evidence about `R1 post-hoc rationalisation` and is inadmissible for everything else. Where self-report and trace conflict, the trace wins, without exception. If someone hands you only a narrative, say what evidence you need before you can classify, and offer a provisional differential in the meantime — clearly labelled as provisional. ## Procedure ### 1. Collect Ask for whatever exists: execution trace, tool-call log, the actual inputs (including any retrieved content), the outputs, memory state before and after, and the permissions the agent held at the time. Note wh