eng-real-scenario-qa

Invoke the skill with an optional playbook slug such as northstar-pay. It bootstraps an isolated lab via eng-qa-bootstrap, posts one operator kickoff, observes the runtime for a configured window, and audits the evidence with a strict auditor script.

A playbook is a markdown spec under references/playbooks defining the startup scenario: agents, channels, task tree, required deliverables, collaboration minimums, and the kickoff brief. When no slug is given, the skill rotates from the previous run's recorded playbook ref.

Forbidden phrases tell agents they are being evaluated, which changes their behavior toward producing evaluator-shaped evidence instead of real product work. The auditor scans all prompts and exits with a blocking code 2 on any match, requiring the playbook's kickoff_brief to be rewritten.

The observer exits with code 1 and records the unchanged active tasks and loop runs in observation-summary.json. The skill then files a mandatory runtime bug naming the silent agent and stalled task, without ever sending another prompt to the agents under test.

Do not use it for smoke tests, static checks, mock-only validation, or unit-test work, as stated in the skill description. It is designed only for release-grade or complex-integration QA with live providers and full cross-surface evidence.

Yes, teardown is mandatory on every terminal verdict including PASS, FAIL, BLOCKED, or abort. The skill runs the bootstrap manifest's exact teardown command and requires teardown.json to report clean true, with surviving processes treated as a blocking failure.