User Story

As an operator running the production-readiness audit, I want an opt-in --workflow mode for /launch-check that fans the 10 dimensions out in parallel and adversarially verifies the FAIL/WARN findings, so that a milestone-boundary audit is faster and higher-confidence (fewer false positives) than the serial pass — without changing the default behaviour or the persisted trend history.

Background

/launch-check runs 10 independent, read-only dimensions, each with a dedicated expert skill. That is a textbook fan-out: no merges, no human gates, pure audit + synthesize. The Claude Code Workflow primitive (deterministic multi-agent fan-out + adversarial verify + synthesize) fits it directly. The serial path must remain the default — a workflow run spawns many agents and costs more tokens — so the workflow path is opt-in via a flag.

Acceptance Criteria

• /launch-check --workflow [project-path] runs the 10 dimensions as a fan-out (one agent per dimension), each dimension reusing its existing PASS/WARN/FAIL criteria
• FAIL and WARN findings are adversarially verified (independent skeptic pass) before they land in the verdict — reduces false positives
• The synthesized output is the SAME verdict table + verdict vocabulary (go / go-with-warnings / conditional-go / no-go) as the serial path
• Results persist via the SAME _lib-audit-history.sh (audit_run_persist + audit_render_trend) — trend history and render-trend.sh chart stay intact; no schema change
• Default (no flag) behaviour is byte-for-byte unchanged; --workflow is purely additive
• The skill documents the canonical workflow-script shape (phases, schema, dedupe, persist) so the agent authors a consistent workflow each run
• AgDR records the decision (Workflow vs serial vs /fan-out; opt-in rationale; token-cost note)

Out of Scope

• Making --workflow the default (stays opt-in)
• Changing any dimension's pass/fail criteria
• A new persisted-history schema (reuse the existing superset)

Glossary