Decide what to log so you can debug an agent run a week later

You are a reliability engineer reviewing an AI agent workflow. Below is its description and a run that went wrong. Design the logging I need to diagnose runs like this after the fact, and nothing more.

Output exactly these sections:

1. WHAT WE CANNOT ANSWER TODAY — 3-6 questions the current run record leaves unanswerable (e.g. which tool call returned the bad value).
2. LOG EVENTS — table with Event name | When it fires | Fields captured | Why it is needed. Max 8 rows. Every row must map to a question in section 1.
3. DO NOT LOG — fields that are noise, expensive, or sensitive, with one reason each.
4. RETENTION AND SIZE — rough events per run, what to keep 7 days vs 90, and the sampling rule for high-volume steps.
5. FIRST TRIAGE PATH — the ordered list of logged fields to check when this failure repeats.

Rules: no preamble. Do not invent numbers about my volume; if I have not given them, write "unknown - measure first". No generic advice about "good observability".

WORKFLOW AND FAILED RUN:
{{paste the workflow steps, tools it calls, and what the bad run produced}}