Skip to content

Add a first-party local runtime evaluation protocol with capability and auth provenance #927

Description

@tcconnally

Problem

Helix Agent Lab uses an asynchronous local MCP server, deterministic prompt-case scoring, hidden simulator state, runtime capability descriptors, bounded logs, reproducible seeds, and explicit separation between stateful system benchmarks and prompt-only evaluations. It also keeps provider authentication inside official installed CLIs instead of turning consumer credentials into a general API.

Perseus has strong memory/context/evidence benchmarks, but its cross-runtime evaluation contract should make provider/auth mode, runtime capability, asynchronous terminal state, and benchmark-family parity equally explicit.

Proposed change

Add a reusable local runtime evaluation protocol in the appropriate Perseus benchmark/flywheel surface. Define:

  • versioned observation/action envelopes;
  • capability descriptors distinguishing persistent state, seeded memory, filesystem, tools, and feedback memory;
  • account/environment auth mode and provider provenance without reading credentials;
  • asynchronous run lifecycle and bounded artifact/log retention;
  • deterministic assertions and hidden-state scenarios;
  • separate score families for stateful system evaluations, prompt-case evaluations, and wrappers.

Acceptance criteria

  • Every run emits a self-describing manifest containing suite digest, repo revision, runtime, model/provider/version, auth mode, seed, scorer version, scope, timestamps, and artifact digests.
  • Run status distinguishes queued/running/passed/failed/cancelled/interrupted/failed_to_start.
  • Account mode never reads or returns credential contents and uses explicit provider CLI boundaries where applicable.
  • Stateful and prompt-only results cannot appear on one undifferentiated leaderboard.
  • The offline harness runs without credentials/network; live-provider mode is explicit and reports authoritative provider usage separately from estimates.
  • Cancellation, crash/restart, timeout, malformed output, and partial results are covered.

Grounding

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions