trust
How it's tested
How do we know Stagecoach itself works?
In one minute
- The automated suite runs the real runner against temporary repositories, local bare remotes, and labeled provider doubles, in an isolated home.
- It covers setup, customization, red and green evidence, frozen tests, scope guards, retries, fallbacks, exact-session answers, recovery, installation, delivery, and environments.
- CI runs it on Linux and macOS with Python 3.11 and 3.14.
- A provider double is not a live model. Account-authenticated runs are checked separately with the real-host smoke procedure below.
The automated suite uses temporary repositories, bare local remotes, isolated homes/configuration, and labeled provider/connector doubles. It never installs into a contributor's real agent settings or publishes to a real account.
| Area | Automated evidence |
|---|---|
| Setup | Preview/apply, stale approval rejection, re-run preservation, project membership. |
| Customization | Per-project overrides, new skills, custom stages, provider prompts. |
| Runtime | Red/green evidence, frozen tests, scope/index guards, fresh retries/fallbacks, exact-session input, stalls, interrupted recovery. |
| Installation | Global/project modes, rollback, updates, owned removal, compatible fallback, plugin relocation. |
| Git | Selected worktrees, parked/read-only repositories, base guards, partial delivery resume. |
| Integrations | Paginated feedback, connector availability/patches, bot draft restoration, rule recurrence. |
| Fixtures | Actual Python checks and an HTTP request to a disposable local application. |
| Environments | Readiness, exported connection details, shared leases, release requests, and cleanup on failed startup. |
Real-host smoke procedure
Vendor command-help and manifest validation are separate from model execution. Local native Claude lifecycle and inventory checks have also been run in an isolated configuration. A double passing does not prove account access, native discovery, or a provider's changing event format.
For each enabled host:
- Use an isolated sample project and temporary host configuration.
- Install the package or load the native plugin.
- Confirm the setup skill is discoverable and resolves its effective file.
- Create a project override; confirm direct invocation uses it.
- Select a model available to that account and run a tiny implementation-only task.
- Ask for a deliberate input pause; answer and verify the same session resumes.
- Inspect actual evidence, usage fields, permission settings, and cleanup.
Account-authenticated generation, hosted GitHub publication, GitHub Pages deployment, and native Windows are not claimed by isolated tests. See the local handoff for the specific checks performed for this release.