Agent skills and runtime requirements#
SnagBadger skills orchestrate the host's available tools; portable modules own deterministic catalog, execution and report processing. There is no mandatory T3 runtime and no hidden browser installer/MCP client inside the product bridge.
Current skills#
The guided onboarding workflow adds these adopter-facing entry points:
| Skill | Responsibility |
|---|---|
| snagbadger-setup | Discover targets; guide approved reporting and testing integration |
| snagbadger-initialize-tests | Discuss resource-led coverage and initialize agreed scenarios/prerequisites |
| snagbadger-create-test | Create or update agreed definitions through the existing catalog contract |
| snagbadger-test | Execute ready selections through the maintained coordinator |
The underlying skills remain available:
| Skill | Responsibility |
|---|---|
| initialize-testing | Inspect tools, catalog and project settings; establish runtime prerequisites |
| run-test-plan | Coordinate selection, requests, fresh workers and retained output |
| execute-test | Execute one task packet in the supplied isolated environment |
| create-test-report | Review and render retained run results |
| create-test-bug | Promote selected observations and package reviewed bug evidence |
Invoke repository skill paths directly or install their folders in the host's skill search path. This is distinct from pnpm skills:install, which synchronizes the repository's merge-review skill for repository-agent work.
Required capabilities#
- A fresh agent per attempt, with explicit packet/context ownership.
- A distinct isolated browser context per browser attempt, explicit action targeting and cleanup.
- Durable real artifact paths/bytes for the selected evidence policy.
- An accessible declared host URL and trusted project hooks/review policy.
API/CLI scenarios do not require browser contexts but still require fresh agents and diagnostic evidence. Capabilities in runtime JSON are verified attestations, not features enabled by setting a boolean.
Runtime limitations#
Current T3 collaborative preview tabs share cookies/local storage; opening another tab does not satisfy product browser-test isolation. T3 remains the required observed review surface for repository changes.
Chrome DevTools MCP and other adapters must be checked against the installed tool schema and lifecycle behavior. Unique isolated context identities cannot be reused; closing a page may not dispose its context. Network reachability from the browser is separate from local terminal reachability.
The canonical runtime mappings explain these boundaries. Deterministic conformance tests verify module sequencing and adapter contracts, not compatibility with an unavailable live MCP version. First-run guide joins the required pieces.