Skip to content

feat(benchmark): pre-bake hermes agent into dataset images for closed-book runs #259

Description

@gburachas

Problem

Closed-book benchmark runs (no network egress at task time) cannot use the Hermes agent, because
the Harbor task image installs it at runtime via curl | bash. With egress blocked the install
fails and the task cannot run.

Proposal

Pre-bake the NousResearch hermes-agent into the dataset image at build time, alongside the agents
already pre-baked there (Claude Code, Codex, OpenCode, Node). Pin the version through the existing
benchmark/agent-versions.env mechanism so runs stay reproducible, and add an "already installed"
guard so the runtime installer short-circuits instead of failing.

Scope

benchmark/ only — no library or routing code. Roughly 130 lines across three files.

Notes

This is benchmark tooling and stays Python/shell; it is unaffected by the ongoing Rust migration.

Happy to open the PR — the change is already written and tested against current main.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions