Startup-Memory Benchmark Pack

Copy this into .perseus/benchmarks/startup-pack.md in your workspace (or keep it here) and run it in a fresh session that has your startup block loaded. Replace <TODAY'S TASK> with the actual task. Score each answer 0–2 using the rubric in ../docs/startup-memory-benchmark.md (8 = excellent). Re-run after Perseus / Vault upgrades to catch regressions.

Before running: re-render the block (perseus render <source> --format agents-md), confirm wiring with perseus doctor, and prune low-signal entries with perseus_vault_hygiene.

---

1. Startup-context quality

From the startup context you already have (do not retrieve anything yet), what high-value facts are present, and what important gaps remain for <TODAY'S TASK>? List facts and gaps separately.

2. Retrieval-plan quality

Turn those gaps into a minimal retrieval plan — the fewest lookups that would close them. No lookup should duplicate a fact already in startup context.

3. First-move discipline

What one lookup would you do first, and what are you intentionally deferring (and why)?

4. First-command quality

Assuming tool/command discovery is already done, what is the single best first retrieval command/query? Give the exact tool, query text, and filters (category / scope / date). Prefer a targeted query over a broad scan.

---

Score sheet

Run date	P1	P2	P3	P4	Total	Perseus / Vault version	Notes

A falling total across runs after a change = a startup-memory regression; bisect the change (block content, context profile, Vault ranking).