Plan Feature

Plan Feature

Own one planning-only workflow from ordered Jira intake through a published canonical graph plan. Remain the main coordinator; controlled agents are bounded workers.

Required contracts

Before acting, read:

• FEATURE-WORKFLOW-ROUTING.md;
• FEATURE-WORKFLOW-STATE.md; and
• FEATURE-PLANNING-CONTRACT.md.

Also follow the runtime-local problem-definition, low-level-design, feature-planner, and agent-boundary contracts.

Entry

For new work, require one or more supplied Jira keys, allocate one collision-safe plans/<feature>-feature-workflow-state.json, and init --owner plan-feature through ../_shared/scripts/feature_workflow_state.py. For resume, require the exact state path and read back its recorded owner and stage. Never scan for a newest run, and never claim a state the helper reports as owned by build-feature. Enter only on a direct human request under the routing contract's human front door; refuse invocation by a spawned agent or another skill.

Workflow

1. Close one aggregate Jira intake of at most 50 issues. For each state-frontier key, dispatch a fresh rocprof-compute-jira-intake-writer / snapshot, record its immutable file/digest, and let state add only bounded relationship context. Missing supplied issues block.
2. Dispatch one initial questionnaire from the complete snapshot inventory, record it, and pause for human completion. Resume only by exact state/artifact; focused follow-ups stay in this same lineage.
3. Dispatch one aggregate problem formulation in formulate, present its returned candidate for explicit human approval through plan mode, then publish it.
4. Dispatch rocprof-compute-lld-writer / draft from that approved formulation, its questionnaire, the snapshots, any existing design/*.md by exact path, and the repository surfaces in scope. Resolve returned questions through grill-me as the coordinator, present the returned body for explicit human approval through plan mode, publish it, and record the exact path and digest with set-lld. An lld-current outcome adopts the declared existing document and publishes nothing.
5. Dispatch the feature planner only from that approved formulation and that recorded design. Record planner questions, apply grill-me, ask them as the coordinator, record all answers, and use a fresh draft dispatch.
6. Require one parser-valid 1–9 package DAG of thin vertical slices: stable WPn identities, each package's declared behaviors delivered end-to-end by it alone, interface contracts copied unchanged from the approved design with one provider per name and every consumer depending on its provider, global path ownership, exactly one final-integration package, an optional first kind: "interface" package that claims no behavior, package RED and verbatim blind GREEN, restack/stack data, terminal review, and stacked-PR publication data.
7. Prepare and persist its exact pending chunks/digest, display those bytes, and obtain explicit human approval. Record the approval receipt, run receipt-bound publication, verify the sole canonical plan and digest, and record it.
8. Stop at released plan_ready. Return exact state/plan paths, stage/digests, Jira closure, approver/time, and recommend a later explicit implement-feature invocation.

Never enter package execution, create branches, push, create a PR, or invoke another public feature skill.

Resources

• REFERENCE.md — intake, graph, approval, and artifact checks.
• EXAMPLES.md — aggregate intake, pauses, resume, and routing.
• writing style — normative prose rules and the pre-save self-check for every human-read artifact this skill produces.