Imported from yoheinakajima/distributed-discovery (
src/distributed_discovery/benchmark/agents_v1/AGENTS.md). Install upstream withnpx skills add yoheinakajima/distributed-discovery --skill agents_v1. Copyright stays with the author.
TreasureBench Agents v1 implementation instructions
- Preserve DD-010 ownership, TreasureBench public naming, DiscoveryBench compatibility identifiers, capability isolation, and versioned contracts.
- External execution is disabled unless an exact task contract and active owner authorization permit the exact campaign, commit, routes, calls, caps, private paths, and milestone.
- Never read credentials, private custody, retained pilot state, or owner authorization during ordinary tests, audits, context rendering, or dry runs.
- The original pilot remains quarantined and immutable. No original, future-base, or base-slot identity may be reused.
- Enforce exactly one final action per required agent. Independent contract verification and metric-range validity precede performance interpretation.
- Preserve provider errors, protocol invalidity, missingness, costs, locks, corruptions, and redaction boundaries. Do not publish task-level private content, comparative performance, rankings, or composites.
- Provider calls and spend fail closed and must remain zero in CI.
