Hermes Native Laboratory
The Hermes laboratory is an opt-in target that starts a native Hermes Agent child process against an ephemeral copy of an approved home. It is validated against:
| Component | Baseline |
|---|---|
| Hermes Agent tag | v2026.8.13 |
| Package version | 0.20.1 |
| Commit | f80f453ae0679347e38abc917c7f94f717bf96c5 |
What the laboratory enforces
Section titled “What the laboratory enforces”- A dedicated checkout and interpreter.
- A clean home or an explicitly selected, sanitized context.
- A closed manifest of copied files.
- No target tools or MCP servers.
- No custom system or developer prompt layers, profiles, prefills or continuation.
- One text turn per fresh replica.
- State comparison before and after each repetition.
- Cleanup of the replica on success, failure, timeout and interruption.
What it does not enforce
Section titled “What it does not enforce”The child process keeps the operating-system permissions of the account that launches it. The laboratory does not restrict filesystem or network access at the operating-system level. Run it under a separate account, container or virtual machine when the target requires stronger isolation.
Preflight
Section titled “Preflight”The preflight resolves the target checkout, interpreter, Hermes baseline, manifest and provider configuration. It rejects unexpected runtime files, tools, MCP, prompt layers and mutable scope before sending a turn.
Read the detailed repository contract in
docs/HERMES_LAB.md
before changing the manifest or supported baseline.