| name | flowguard-plan-detailing-compiler |
| description | Use only when explicitly requested or delegated by flowguard-development-process-flow's plan_detailing mode to compile a rough non-trivial plan into structured artifacts, steps, receipts, validations, failure/rework branches, and claim boundaries. |
FlowGuard Plan Detailing Compiler
Purpose
Turn rough plans into checkable PlanDetail rows; detail is not implementation proof.
Entrypoint Scope
Delegated FlowGuard mode skill; route plan_detailing_compiler, role delegated_mode, native owner development_process_flow. Generic work enters flowguard-development-process-flow first.
Local Material Routing
Read references/plan_detailing_compiler_protocol.md for rows, findings, projection order, and confidence.
Entrypoint Acceptance Map
Accept an explicit/delegated rough plan; compile scope, artifacts, state, steps, step receipts, validation, failures/rework, UI/payload gates, and claims; return typed projections.
Use When
- Use for delegated
plan_detailing with underspecified scope, steps, artifacts, UI/payload surfaces, validation, or rework.
Do Not Use When
- Do not act as a generic planner, execute work, or replace downstream evidence; return unclear routing to
model-first-function-flow.
Required Workflow
- Capture goal, assumptions, sources, risks, state/side effects, artifacts, UI actions, and payload/work-package surfaces; keep provider/task/obligation/check ids distinct.
- Keep AI steps in
agent_operation; add typed targets, order, receipts, validations, failure/rework/continue gates, freshness, and questions. If DPF optimization is active, keep only top-level process_optimization_reasons and one current required_process_optimization_evidence_ids reference—never copies on each step/validation. Run review_plan_detail().
- Project passing/scoped rows to DevelopmentProcessFlow, AgentWorkflow, UI, Model-Test Alignment, and TestMesh owners.
Hard Gates
- Model-purpose gate: before build/change, freeze this instance's task-specific failure(s) and boundary; then bind candidate plus native good/bad-per-failure/oracle/current evidence. Reusable types are not fixed-purpose; no mode/fallback; SkillGuard only supervises FlowGuard-declared checks.
- Use the real FlowGuard check engine and AGENTS.md managed record; never create a fake mini-framework.
- Prose, checkboxes, or progress cannot satisfy structured detail or terminal evidence.
- Full claims need resolved questions, final evidence, real-surface payload cases, human-operability gates, and template harvest closure.
- AgentWorkflow owns AI order; DPF owns lifecycle freshness; provider and product owners retain authority.
Output Requirements
- Return
evidence, failures, blockers, skipped_checks, residual_risk, claim_boundary, typed_next_actions, PlanDetail rows, gaps, and projections.
SkillGuard Maintenance
- Edit contract source and regenerate; SkillGuard cannot execute or create receipts.
