ci(evals): Add isolated global sandbox egress#943
Conversation
Start one Cloudflare Quick Tunnel and egress proxy for each eval invocation instead of requiring a shared named tunnel and fixed base URL. Share credential state through an isolated loopback Redis namespace and keep fixture control in the global setup process. Install a pinned cloudflared binary in CI, run deterministic lifecycle coverage, and document the supported local eval workflow. Co-Authored-By: OpenAI Codex <noreply@openai.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
Keep the invocation-generated state prefix when Vitest loads the eval config in worker processes so the global egress proxy and eval workers share one Redis namespace. Co-Authored-By: OpenAI Codex <noreply@openai.com>
Register the eval OAuth provider as an inline manifest so invocation-wide setup never changes process cwd. Scenario plugin fixtures can now create and clean their temporary apps without restoring into deleted global fixture directories. Co-Authored-By: OpenAI Codex <noreply@openai.com>
Load eval plugin manifests and skill roots directly instead of changing process cwd to temporary plugin apps. Timed-out scenarios can no longer delete a cwd that later eval cases attempt to restore. Co-Authored-By: OpenAI Codex <noreply@openai.com>
Co-Authored-By: David Cramer <david@sentry.io>
![sentry-warden[bot]](https://avatars.githubusercontent.com/in/2755838?s=60&v=4){kind=small}
Co-Authored-By: David Cramer <david@sentry.io>
Co-Authored-By: David Cramer <david@sentry.io>
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit b68db06. Configure here.
| onToolResult: (result) => { | ||
| const pendingIndex = pendingToolInvocations.findIndex( | ||
| (candidate) => candidate.params === result.params, | ||
| (candidate) => candidate.tool === result.toolName, |
There was a problem hiding this comment.
Tool results match by name
Medium Severity
The eval harness pairs onToolResult callbacks to pending invocations using only result.toolName, but the prior logic matched the shared params object reference. When more than one call of the same tool is pending, a result can attach to the wrong invocation or leave others incomplete, so transcript tool_result events and toolCalls status can be wrong.
Reviewed by Cursor Bugbot for commit b68db06. Configure here.
1 participant
PR evals now start one invocation-wide Cloudflare Quick Tunnel and local sandbox egress proxy instead of depending on a shared named tunnel, fixed
JUNIOR_BASE_URL, and repository Cloudflare credentials. Vitest global setup owns the tunnel, provider fixtures, and authenticated control channel, while an isolated loopback Redis namespace shares credential state with eval workers without touching configured remote Redis instances.The workflow installs a pinned, checksummed
cloudflaredbinary and runs deterministic egress lifecycle coverage before the model eval suite. Local eval setup and commands are documented, and the OAuth workflow assertion now accepts equivalent shell setup around the required identity-check curl command.The eval workflow now uses vitest-evals 0.15.0, emits info-level terminal diagnostics, and always publishes the native GitHub eval summary and failure annotations from
vitest-results.json. The harness output was migrated to the current ordered transcript-event contract so tool results and deterministic failures remain visible in that report.