automation-template / Sandbox automation
Sandbox Runner
A guarded automation pattern for executing coding agents where allowed files, credentials, logs, and generated patches need stronger auditability than normal scheduled agents.
SandboxOpenAI AgentsPatches
Inputs4required context
Outputs4expected products
Secrets1credential gates
Tools3declared access
Targets1install surfaces
Contract
Runs manifest-bound coding automations with explicit file scope, logs, patches, and test artifacts.
Reusable asGuarded coding automation
Runner adaptersandbox:openai-agents
Artifact contractPatch, manifest, log, and test output artifacts must be retained per run.
Event contractEmit manifest-loaded, run-started, patch-produced, tests-passed, blocked, and failed events.
Runbook
Build manifest, run streamed automation, collect patches and tests, then attach artifacts and failure reasons.
GuardrailOnly touch manifest-staged files.
GuardrailSurface missing credentials as blockers.
GuardrailKeep sandbox runs separate from normal agents.
Installations
Project-scoped instances created from this reusable template.
No installations yetInstall this template into a project to start tracking a concrete instance.
Inputs
What an installation or run must provide.
Inputtask file
Inputmanifest
Inputallowed files
Inputtest command
Outputs
What the template should leave behind.
Outputpatch artifact
Outputstreamed log
Outputtest result
Outputblocker summary
Allowed tools
Capabilities this template can use.
Allowed toolVercel Sandbox
Allowed toolOpenAI Agents SDK
Allowed tooltest runner
Verification
Checks before treating the template as ready.
VerificatioManifest paths match task instructions.
VerificatioMissing credentials are visible.
VerificatioPatch and test artifacts link from the run page.