omp-flow: Skill for Claude Code

OMP-Flow QbD

Gate Meaning

• QbD 1 challenges whether the selected problem, requirements, design, sources, and contracts are justified.
• QbD 2 challenges whether the authored work map, work Concepts, scope, ordering, and verification can realize the approved design.
• The independent auditor applies adversarial judgment. The human calibrates the result. Neither replaces the other.
• QbD is a material decision challenge, not an exhaustive refinement search. Practice evidence （实践检验）outranks process completeness; anti-formalism（反形式主义）never weakens mechanical authorization, data, false-authority, or irreversible-effect boundaries.

Procedure

Main/coordinator alone dispatches, correlates receipts, records human calibration, and chooses transitions. A dispatched Auditor performs only the bounded independent audit and assigned audit Concept; it does not dispatch, govern, calibrate, transition, or reinterpret coordinator clauses. It reconstructs the concrete problem without solution jargon, challenges whether the selected problem and current principal contradiction still hold before judging architecture completeness, preserves the mechanical safety analysis, and reports the required verdict, findings, evidence, and contradictions through that output boundary.

1. Identify the design or work-map entry Concept and an explicit audit output Concept. Through its useful links, bind the audit to the current decision, unacceptable consequences, current scope/change, and relevant closed findings, residual risks, and human decisions. Do not require a fixed audit-brief file or field shape; no prior findings is normal for a first audit.
2. In Main/coordinator context only, start one independent QbD operation with Bundle root, role, bounded objective, entry, output, actor ID, and optional predecessor receipt.
3. In Main/coordinator context only, dispatch one independent native qbd-auditor. Pass the paths and opaque receipt returned by the runtime by forwarding the complete returned assignment unchanged to the native qbd-auditor/omp-flow-qbd agent. Keep its strict v1 ompFlowDispatch descriptor as the first non-blank line. Do not render the Bundle, reconstruct the descriptor, or prepend audit prose.
4. Require an explicit verdict with these semantics:
   • FAIL requires evidence of a critical falsehood, authorization or data violation, irreversible harm, or an unrealizable/unverifiable core path for which safe degradation is insufficient.
   • NEEDS_EVIDENCE applies only when missing evidence prevents judging whether such a material consequence exists. Unknowns alone are not blocking.
   • PASS means no unresolved blocking finding in the current scope; it may carry advisory observations, residual risk, later verification, and deferred recommendations.
5. Require each blocking finding to link evidence through cause -> concrete consequence -> affected decision, state the violated required property separately from any non-binding example repair hint (never prescribing a repair that contradicts a locked human decision), and explain why hiding, unavailable, disabling, narrowing, refusal to write, or another safe degradation is insufficient. Findings without that chain are advisory.
6. In Main/coordinator context only, confirm the promised audit Concept exists, links to what it evaluated, and records material findings, then finish the runtime operation with the same actor ID. The already-dispatched Auditor returns that assigned audit output; it does not finish or correlate the coordinator's operation.
7. In Main/coordinator context only, present material findings and applicable governance options to the user. Only the user decides calibration; the verdict itself authorizes no repair, re-audit, or forward transition.
8. In Main/coordinator context only, record the human decision in a linked decision Concept. Do not encode approval, materiality, or risk as runtime phase or parsed state.