Real Scenario QA

Execute release-grade QA by running an entire fictional startup project on the Compozy runtime and observing the result. The runtime drives the work; the observer never tells agents they are being evaluated. The auditor enforces real deliverables (compiled/parsed/runnable artifacts) and real collaboration (peer messages, review cycles, disagreement resolution).

The skill rejects any prompt that frames the work as QA. See references/forbidden-prompt-phrases.md.

Required Inputs

• playbook-ref (optional): Slug of the playbook to run (e.g., northstar-pay, devtool-oss-launch, consumer-saas-growth). When omitted, rotate from the previous run's PLAYBOOK_REF recorded in bootstrap-manifest.json.

Procedures

Step 1: Select the Playbook

1. Read .agents/skills/eng/eng-real-scenario-qa/references/playbook-catalog.md in full.
2. Resolve the playbook ref:
   • If the user supplied a slug, validate it exists at references/playbooks/<slug>.md.
   • Otherwise, list references/playbooks/*.md (excluding README) and rotate from the previous PLAYBOOK_REF.
3. Record PLAYBOOK_REF.

Done when: one valid playbook is selected and differs from the previous run when rotation applies.

Step 2: Bootstrap the Lab With the Playbook

1. Activate eng-qa-bootstrap with scenario $PLAYBOOK_REF and --playbook "$PLAYBOOK_REF"; complete its full procedure instead of calling its helper directly.
2. Consume the canonical BOOTSTRAP_MANIFEST and its emitted paths. Never reconstruct provider, browser, proxy, audit, or teardown state here.
3. Confirm the selected playbook, agent registrations, open-task tree, knowledge files, required deliverables/collaboration, and populated charter all belong to the same healthy manifest. Register only RUNTIME_WORKSPACE_PATH with Compozy, and capture the returned public id as RUNTIME_WORKSPACE_ID; agents must not see the lab's qa-artifacts/ or audit contracts.

Done when: bootstrap's completion criteria pass and the charter has no placeholders.

Step 3: Activate Companion Skills

1. Use qa-report with qa-docs-path=docs/qa to plan the playbook-product validation as session charters (persona + journey + tour + time-box on the playbook deliverables — never on QA itself). QA_OUTPUT_PATH remains the lab-side scratch root (journey log, observation, kickoff evidence); the living QA state lives in the repo's docs/qa/.
2. Use qa-execution with qa-docs-path=docs/qa to run those sessions against the lab (does the TSX page render? do scripts run? does the canary control respond?), driving the lab's base URL/daemon from the bootstrap env block.
3. Use eng-worktree-isolation only when concurrency was explicitly signaled by the user.
4. Use systematic-debugging and no-workarounds for any unexpected runtime behavior the observer captures.
5. Apply provider-home, Web-proxy, config-write, PID-registration, and teardown policy directly from the bootstrap manifest.

Done when: the living QA plan and execution target the bootstrapped lab, with concurrency isolation activated only when signaled.

Step 4: Post the Operator Kickoff

1. After runtime agents, sessions, channels, and the deterministic task ids from .compozy/tasks/open-tasks.json exist under the shared RUNTIME_WORKSPACE_PATH, prepare task activation behind a scheduler barrier (mutating): python3 .agents/skills/eng/eng-real-scenario-qa/scripts/activate-playbook-tasks.py prepare --workspace "$WORKSPACE_PATH" --qa-output-path "$QA_OUTPUT_PATH" --manifest "$BOOTSTRAP_MANIFEST" --compozy-bin "${COMPOZY_BIN:-compozy}"
2. Render and validate the kickoff payload (mutating only the inspectable payload file): python3 .agents/skills/eng/eng-real-scenario-qa/scripts/post-operator-kickoff.py --workspace "$WORKSPACE_PATH" --playbook "$PLAYBOOK_REF" --qa-output-path "$QA_OUTPUT_PATH" --manifest "$BOOTSTRAP_MANIFEST"
3. The helper aborts with exit code 2 if the rendered kickoff contains any phrase from references/forbidden-prompt-phrases.md. Rewrite the playbook's kickoff_brief when blocked.
4. Read <WORKSPACE_PATH>/.compozy/operator-kickoff.txt. Deliver that text verbatim once and capture the provider stream: compozy session prompt <operator-session-id> "$(cat $WORKSPACE_PATH/.compozy/operator-kickoff.txt)" -o jsonl > $QA_OUTPUT_PATH/qa/operator-kickoff.jsonl
5. Confirm the successful post from its non-empty evidence (mutating), then release the queued task runs (mutating): python3 .agents/skills/eng/eng-real-scenario-qa/scripts/post-operator-kickoff.py --workspace "$WORKSPACE_PATH" --playbook "$PLAYBOOK_REF" --qa-output-path "$QA_OUTPUT_PATH" --manifest "$BOOTSTRAP_MANIFEST" --confirm-posted "$QA_OUTPUT_PATH/qa/operator-kickoff.jsonl" python3 .agents/skills/eng/eng-real-scenario-qa/scripts/activate-playbook-tasks.py release --workspace "$WORKSPACE_PATH" --qa-output-path "$QA_OUTPUT_PATH" --manifest "$BOOTSTRAP_MANIFEST" --kickoff-evidence "$QA_OUTPUT_PATH/qa/operator-kickoff.jsonl" --compozy-bin "${COMPOZY_BIN:-compozy}"
6. Confirm the manifest reports KICKOFF_POSTED=true, KICKOFF_TIMESTAMP is set, task activation is released, and the scheduler is unpaused. Send no further prompt to any agent under test; a stall becomes a bug.

Done when: every declared task has one queued run behind the barrier, exactly one evidenced kickoff is confirmed, dispatch is released, and the observer has no path for a second agent prompt.

Step 5: Observe the Runtime

1. Run the observer (read-only) for the configured window: python3 .agents/skills/eng/eng-real-scenario-qa/scripts/observe-runtime.py --scenario-workspace "$WORKSPACE_PATH" --runtime-workspace "$RUNTIME_WORKSPACE_PATH" --workspace-id "$RUNTIME_WORKSPACE_ID" --api-base-url "$COMPOZY_WEB_API_PROXY_TARGET" --compozy-home "$COMPOZY_HOME" --compozy-bin "${COMPOZY_BIN:-compozy}" --qa-output-path "$QA_OUTPUT_PATH" --duration-sec 1800 --stall-threshold-sec 300
2. Before polling, the observer requires workspace info "$RUNTIME_WORKSPACE_ID" to resolve to RUNTIME_WORKSPACE_PATH. It then derives progress only from public Task catalog/detail, Loop runs, loop why, and loop events reads. It records only durable state transitions in observation-summary.json; journey-log.jsonl remains supporting evidence and never controls the stall clock.
3. While the observer polls, capture cross-surface evidence without directing agents:
   • CLI: independently capture compozy task list --workspace "$RUNTIME_WORKSPACE_ID" -o json, plus agent, channel, and session lists against the same isolated COMPOZY_HOME.
   • API: read endpoints that intersect the playbook's primary domain.
   • Web: open the Compozy web app via browser-use:browser (or the agent-browser fallback) against $COMPOZY_WEB_API_PROXY_TARGET. Capture DOM snapshot, URL, screenshot.
   • Runtime: compare the independent Task catalog capture with the observer's Task account for the same window.
4. Record observer-only or out-of-band supporting evidence with the mutating helper .agents/skills/eng/eng-real-scenario-qa/scripts/record-scenario-action.py; those rows never count as runtime progress.
5. On exit 1, open <QA_OUTPUT_PATH>/qa/observation-summary.json, identify the unchanged active Tasks or Loop runs, and proceed to Step 6 without prompting an agent. On exit 2, record the exact public-read error; a malformed or failed read is not a stall or a pass.
6. On exit 0, require the observer account and independent Task catalog capture to agree before proceeding to Step 6.

Done when: the observation window, terminal state, explicit stall, or honest read error completes with indexed CLI, API, Web, runtime, and provider evidence; the independent catalog comparison is recorded; and no observer prompt follows kickoff.

Step 6: Audit, Diagnose, Fix, Re-Verify

1. Maintain the dated report at docs/qa/reports/<YYYY-MM-DD>-<playbook-ref>.md through qa-report/qa-execution; index lab-side evidence by path rather than copying it into the repository.
2. Diagnose and fix real runtime defects with systematic-debugging and no-workarounds; fix playbook authoring defects in the playbook source and restart from Step 2.
3. After the last code change, run make gate and record its fresh local evidence in the dated report; the enclosing workstream waits for exact-head PR CI.
4. Run the mutating strict auditor last, passing the durable report explicitly: python3 .agents/skills/eng/eng-real-scenario-qa/scripts/audit-qa-evidence.py --qa-output-path "$QA_OUTPUT_PATH" --final-report "docs/qa/reports/<YYYY-MM-DD>-<playbook-ref>.md" --strict
5. Auditor exit code 2 is a blocking failure. Read qa-audit-report.json and act per check. All durable bugs go to the repo's global registry as docs/qa/bugs/BUG-<YYYYMMDD>-<slug>.md (dedup against the registry first, per qa-report's bug-registry rules) and are linked into the affected docs/qa/scenarios/*.md files:
   • C15 forbidden phrase in a prompt → rewrite the playbook source (system_prompt or kickoff_brief), not the auditor or the regex list.
   • C16 deliverable count short → file a runtime bug (which Compozy agent failed to produce the artifact, why, what state shows the failure). Do not author the missing artifact yourself — the runtime is what's under test.
   • C17 collaboration loop short → file a runtime bug describing which channel, agent, or review cycle did not complete. Cite journey-log timestamps.
   • C18 stall → the registry bug is mandatory and must name the silent agent and stalled task.
6. Re-run the relevant scoped checks after each fix; after source freezes again, refresh the local gate evidence and rerun the strict auditor. Observer changes use the read-only verification helper: python3 .agents/skills/eng/eng-real-scenario-qa/scripts/test_observe_runtime.py.
7. Update affected scenario verdicts and append the bootstrap continuation block only when the same active loop will continue.

Done when: the dated report, fresh local gate, scenario verdicts, strict audit, and indexed evidence all describe the same execution with no blocker.

Step 7: Tear Down the Lab (MANDATORY)

1. Complete eng-qa-bootstrap Step 5 using the current manifest's exact TEARDOWN_COMMAND. This applies on every terminal verdict — PASS, FAIL, BLOCKED, or abort.
2. Cite <QA_OUTPUT_PATH>/qa/teardown.json ("clean": true) in the final summary. Survivors (exit 1) are a blocking failure.
3. Only exception: an explicitly continuing timed loop keeps the lab alive; the continuation that ends the loop inherits the teardown obligation. A stalled or aborted run tears down like any other — the stall evidence lives in files, not in live processes.

Done when: the current lab's teardown.json reports "clean": true and no owned process survives.

Error Handling

• If bootstrap fails to load the playbook, validate the playbook against .agents/skills/eng/eng-real-scenario-qa/references/playbook-schema.json; a real-scenario run never falls back to a generic charter.
• If the kickoff helper aborts on a forbidden phrase, rewrite the playbook's kickoff_brief. Do not edit references/forbidden-prompt-phrases.md to remove the rule.
• If task activation preparation fails, keep the owned scheduler barrier paused, inspect qa/task-activation.json, and retry with the same idempotency keys. Never post the kickoff with a partial task tree.
• If kickoff delivery or confirmation fails, keep dispatch paused. Retry only the same unconfirmed delivery when no provider evidence exists; once evidence exists, confirmation is the only valid next step. Release refuses an empty kickoff transcript or an unconfirmed manifest.
• If observe-runtime.py reports a stall, preserve the unchanged public snapshot and file the runtime stall without injecting a prompt. If it reports exit 2, diagnose the named public read instead of relabeling the error as a stall.
• If a required deliverable type cannot be parsed by the auditor (e.g., a TSX file with non-standard exports), fix the artifact in the workspace via the agent that authored it (re-prompting in-persona is fine; new operator prompts are not). If the agent cannot fix it, that is a runtime bug.
• If browser-use:browser is unavailable, follow the agent-browser fallback per the bootstrap browser policy. Do not silently drop the Web surface.
• If providers are unreachable, record the boundary in provider-attempt.json. The run verdict becomes BLOCKED, never PASS.
• If the auditor's playbook_compliance block reports zero counts despite agents working, confirm WORKSPACE_PATH/.compozy/playbook.json exists and journey-log.jsonl is being written. Empty counts often mean the runtime is not wired to the journey log — that is a runtime bug.