planning

Create an execution plan ready for a coding pipeline to implement. If no task is provided, ask first. Load agent files on demand — never pre-load all at once; use references/output-format.md headers for all agent output.

Machine-checkable behavior contract: skill.spec.yml · dependency ledger: deps.toml (both generated by the skillspec CLI and git-ignored — Claude never loads them) · full phase detail: references/phases.md.

Contract

• Input: a task or requirement set (and any existing docs/deliveries/{key}/product-brief.md / PRD.md / delivery file for the same key).
• Output: docs/deliveries/{key}/codebase-map.md, docs/deliveries/delivery-{slug}-{key}.md (carrying its Reuse Map), a PLAN REVIEW result, plus docs/deliveries/{key}/epic-manifest.md (or task-manifest.md). Paths and key derivation: ../../references/delivery-and-worktree.md.
• Boundary: planning artifacts only — this skill halts after the manifest and starts no implementation. To implement: /multi-agent (epics) or /task (single task).
• Rules: Phase 0.5 (codebase discovery), Phase 2 (grill-me stress), Phase 2.5 (plan review against the real codebase) and Phase 3 (human validation of open questions) gate all code — no architecture without the codebase map, and nothing reaches the human that Priya has not read; every manifest row is independently testable; Language comes from the Architect's tech-stack decision, never inferred. Full forbids are in skill.spec.yml.

Steps

1. Requirements input — reuse product-brief.md / PRD.md if present; otherwise run Analyst and/or PM inline (see references/phases.md Phase 0).
2. Codebase discovery — dispatch a read-only mapping subagent (haiku) → docs/deliveries/{key}/codebase-map.md: blast radius, reusable symbols, prior art, conventions, extension points, contracts in force, test conventions — every entry cited file:line. Mandatory; on a greenfield repo record the empty result rather than skipping (references/phases.md Phase 0.5).
3. Architecture — derive the delivery slug + key from the feature name, then load agents/architect.md with Brief + PRD + codebase map; produce docs/deliveries/delivery-{slug}-{key}.md with its header block, security architecture, OWASP threat table, Reuse Map, Test Case Specification, and Mermaid data-flow diagram(s).
4. Stress the plan — run /grill-me against the delivery file + manifest; decide each supported gap into the delivery file, carry the rest to step 6 as open questions. Re-run until no new resolvable gaps.
5. Plan review — load agents/plan-reviewer.md (sonnet) with the delivery file, codebase map, PRD, and the repository. PLAN CHANGES REQUIRED returns to the Architect (max 2 rounds); PLAN APPROVED — or the leftover findings carried as open questions — proceeds to step 6 (references/phases.md Phase 2.5).
6. Human validation — present diagram, stack, ADRs, the plan-review result, open questions, and the api-spec.yaml contract (if produced); update and re-confirm on any change. Step 7 begins once the delivery file and api-spec.yaml (if present) are approved and open questions resolved.
7. Manifest — produce an Epic Manifest (≥ 2 epics) or Task Manifest (single epic) at project root; table shapes in references/phases.md Phase 4.
8. Plan summary — for a standalone run, print the summary and stop here. When a pipeline invoked this skill, this step is informational: print the summary and hand back to the orchestrator.