memory-hook-failure-diagnosis

name	memory-hook-failure-diagnosis
description	Hook D — Failure-Diagnosis memory consultation. Runs at executive Phase 7 (after 3+ consecutive failures on a work item) BEFORE the failure-diagnosis skill itself runs. Consults the second brain for similar prior failures and successful retry strategies, then hands enriched context to the failure-diagnosis skill. Read-only. Invokes the memory-reader skill.

Hook D — Failure Diagnosis Memory Consult

You are the executive agent. A work item has failed 3+ times in a row. Before the existing failure-diagnosis skill runs, consult the second brain for similar prior failures so the diagnosis can build on what we already know.

Read first

1. Read .claude/skills/memory-reader/references/mem0-limitations.md.
2. Then invoke the memory-reader skill.

Context

{{CONTEXT_JSON}}

Contains:

• workItem — id, title, bundle_slug
• failureSignals — last N failure events: error messages, stack traces (truncated), validator verdicts, tool that failed
• priorAttempts — strategies already tried with this work item (minimal_scaffold, break_into_steps, etc.)
• vendor — which vendor failed

What to retrieve

You're looking for patterns, not exact matches. Iterative search:

1. Same bundle, prior failures — filters: { user_id, app_id: <bundle_slug> }, type filter episodic or reflective
2. Cross-project: same error code or signal — search the literal error code (mem0 preserves discriminators)
3. Cross-project: same vendor + similar work type — patterns like "claude vendor with harness X fails when ..."
4. Reflective memories about retry strategy — what strategy has worked for similar failures in the past

Stop when 5+ memories surface or when 6 queries produce no new results.

Output shape

Use Shape A but with a diagnosis-specific framing:

## Memory consultation: failure-diagnosis

Queries issued: <N>
Memories surfaced: <M>

Pattern analysis:
<2–3 paragraphs. Group surfaced memories by failure mode. Cite IDs.>

Strategy recommendations (advisory; failure-diagnosis skill decides):
- <strategy name>: <why this looks promising> — <memory IDs that support it>
- <strategy name>: <why this might fail again> — <memory IDs that warn>

Escalation flag:
<set to "yes" if surfaced memories suggest this pattern has hit human-escalation 2+ times before; otherwise "no">

Caveats:
<gaps>

This block goes into the failure-diagnosis skill's input. That skill makes the actual retry-vs-escalate decision; you're enriching its inputs.

Hard rules

• Read-only. If memories suggest a write is overdue (e.g., "this failure mode has happened 3+ times and there's no reflective memory yet"), flag it for Hook E (post-retro) — do not write here.
• Cite IDs for every claim.
• Cap output at 400 words.