Local validation report
September 20, 2026 · View on GitHub
Validation date: September 19, 2026 (America/New_York). Execution environment: Linux; Python 3.13.5; Node.js 22.16.0.
Executed successfully
| Check | Result |
|---|---|
| Python unittest suite | 73 tests passed, no failures or skips |
| Node built-in test suite | 15 tests passed, no failures or skips |
| Python source compilation | Passed for runtime, adapters, installer and examples |
| Node syntax checks | Passed for every .mjs adapter/bridge |
| Offline gate demo | Prune case returned READY with one simulated assessment and zero simulated compaction calls |
| Standalone inspector fixture | One eligible exact-repeat pair; no remote call |
| Self-contained installer smoke test | Passed in an isolated temporary HOME with spaces |
The standalone smoke run covered: a no-write preview; installation into all ten target namespaces; wrappers pointing to a persistent runtime rather than the temporary bootstrap directory; repeat installation producing unchanged actions; all managed files passing integrity checks; execution of the installed Python worker; and removal of unchanged managed files on uninstall. It did not start any actual harness executable or activate a real Hermes installation.
Coverage
The Python suite exercises exact-repeat selection, protected recent/current-turn context, failed/multimodal/ambiguous records, secret-pattern vetoes, oversized inputs, strict model/answer validation, missing API consent/key, session binding, stale/missing proofs, receipt dependency checks, all three native formats, and no-network inspection. Decision-gate tests cover direct native compaction, manual prune, sufficient/insufficient/unknown budgets, failure without fallback, pause, duplicate/stale choices, compare-and-swap failure and episode replacement.
Installer/storage tests cover ten targets, root environment overrides, explicit profiles, current Codex skill location, idempotence, foreign-file conflicts, modified files, upgrade backups, safe uninstall, preserved runtimes for modified wrappers, symlink refusal, dry-run/doctor no-write behavior and receipt CAS. Hermes tests use a stub plugin context and mocked Jev answers.
Node tests exercise the real Python stdio worker for offline inspection and projection, plus stub-host Pi/OpenCode hook contracts. They verify native compact pass-through without assessment, Pi stop/cancel behavior, manual receipt persistence, changed-branch refusal, fork isolation, OpenCode in-place array mutation, command collision protection, visible command completion stop, and profile-separated receipt stores.
Not validated
- Live TypeSafe API availability, response compatibility, accuracy or calibration.
- Actual Pi, Hermes, OpenCode, OpenClaw, Codex, Claude, Gemini, Cursor or Copilot sessions; plugin loading/version compatibility; UI behavior or native scheduler invariants in those applications.
- Windows/macOS execution, Windows ACLs or platform-specific config discovery.
- Full provider-token accounting, provider-side previous-response lineage, prompt cache behavior, task quality or measured dollar/token savings.
- End-to-end recovery after crash/power loss, every host fork/replay path, all automatic compaction paths or seamless user-choice pause/resume.
- Hermes native plugin-enable command on a real installation or multi-user gateway.
88 passing local tests are not 88 live harness or model tests. Assessment responses are deterministic fixtures. This is an experimental source release, not a claim that the full native specification is installed everywhere.
Raw logs are included at tests-python.log and tests-node.log in the source ZIP.