Run any Skill in Manus with one click

ship-plan

Stars0

Forks0

UpdatedJuly 19, 2026 at 18:25

Use when writing a ship-flow plan for a shaped entity, especially implementation tasks, TDD needs, runtime commands, or scope anchoring. Layer A delegation: superpowers:writing-plans owns plan authoring; ship-flow:test-driven-development owns fallback TDD contracts.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

iamcxa

iamcxa/spacedock-workflows

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Computer Occupations, All OtherComputer and Mathematical Occupations·SOC 15-1299

SKILL.md

readonly

Ship-Plan — PLAN Stage (2.0)

You run PLAN. Output: <entity-folder>/plan.md. Dispatched by /ship to planner teammate via SendMessage (team spawned at /shape). No captain gate.

Pipeline position: reads the resolved shape artifact (shape.md, with legacy spec.md fallback via resolve-shape-artifact.sh) + parent cross-entity contracts → produces plan.md → cross-review gate → advance to execute.

Boot Self-Check

Run before any plan work. Stop and SendMessage(FO) if any check fails.

Entity status: read entity frontmatter status: — must be sharp or design (post-design stage). If draft → shape first; if plan → plan already exists, check for re-entry signal.
Hand-off present: entity body contains ### Hand-off to Plan block (from ship-shape or ship-design). If absent → SendMessage(FO): "Missing Hand-off to Plan in <entity-path> — cannot proceed without design intent."
Team context: verify planner teammate is active (this agent). Note executer teammate name for Step 6 SendMessage.
PRODUCT.md readable: repo-root PRODUCT.md exists and is readable. If absent → SendMessage(FO): "PRODUCT.md missing — constraints source unavailable."
Framework detection (if affects_ui: true): Run ship-runtime-detect Step R5 to confirm framework_detected + theme_indirection for plan's verification spec.
Density-aware skill load (T3.4): read answers_density from entity frontmatter. high → auto-load framework skills per ship-runtime-detect Step R6; skip FO ask. low|vacuum → SendMessage(FO) with proposed skill list; wait for confirmation.
Canonical context planning: read PRODUCT.md, ROADMAP.md, and relevant ARCHITECTURE.md sections when spec/design/affected files indicate product, roadmap, API, schema, domain, storage, runtime, or architecture contract impact.

Entity body contract (schema-as-prose)

Reads: resolved shape artifact (shape.md; legacy spec.md fallback alias) (## Sharp Output / ## Shape Output — problem, done criteria, size, scope, assumptions, children DAG), design.md when present, parent entity ## Cross-Entity Contracts if parent: set, PRODUCT.md constraints, ROADMAP.md status, and relevant ARCHITECTURE.md sections.
Writes: <entity-folder>/plan.md sections — ## Research Summary, ## Size Re-evaluation, ## Verification Spec, ## Plan (TDD tasks with waves), ## Plan Report (status/cost/iterations/verdict).
Full section-tag + field semantics: plugins/ship-flow/references/entity-body-schema.yaml → stages.plan.

Layer A delegation (Principle 6 Rule B)

superpowers:writing-plans owns plan authoring discipline (TDD order, wave safety, placeholder-free prose, task atomicity) when available. ship-flow:test-driven-development owns the ship-flow TDD contract and fallback artifact rules when adopter/runtime environments do not have superpowers installed. Ship-plan wraps both with Layer B augmentation:

Shape Up scope anchoring (every task maps to a Done Criterion / Scope In bullet).
Size-adaptive research team (produce + review subagents, sized S/M/L).
Runtime detection via ship-flow:ship-runtime-detect for {commands.test/build/typecheck/lint/dev}.
Assumption re-validation (Step 1.5) against current codebase state before planning.
Plan-checker multi-dimensional review + cross-review gate.

Flow

Phases (TaskCreate sub-tasks — inherit from /ship umbrella when pipeline-dispatched; create own when standalone): read-spec → assumption-revalidate → research → size-reevaluate → scope-anchor → write-plan → verification-spec → self-review → cross-review → emit-plan.md

Step 1 — Read spec + cross-entity contracts

Record stage-start ISO timestamp. For folder entities, resolve the shape artifact first: SHAPE_ARTIFACT="$(bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/resolve-shape-artifact.sh" <entity-folder>)"; this returns canonical shape.md or legacy spec.md fallback. Extract via bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/extract-section.sh" <entity-file> <section-tag> (handles folder + flat layouts). From the resolved shape artifact: problem, done criteria, size (S/M/L), scope-in bullets, musk-audit verdicts, stated assumptions.

If parent: frontmatter set: read parent ## Cross-Entity Contracts. Extract Contracts to implement, Inherited decisions (ADR overrides), Slice scope. These override any conflicting research finding.

Blocker: spec missing problem or done criteria → write ## Plan Report status: blocked, reason: missing spec and return.

Step 1.1 — Canonical context planning

Build ### Canonical Doc Actions before writing implementation tasks:

Doc	When to mark update	When to mark skip
`ROADMAP.md`	Entity is represented in Now/Next, creates/consumes deferred todos, or umbrella closeout may apply	Entity is intentionally not roadmap-listed; include `skip_rationale`
`PRODUCT.md`	Spec/design changes durable capability, user story, constraint, "Who It Serves", or "Why It Exists"	Internal-only change; include `skip_rationale`
`ARCHITECTURE.md`	Spec/design/tasks touch schema/API/domain/data-flow/storage/runtime/component boundary or architecture-impact exists	No architecture contract touched; include `skip_rationale`

Emit the schema-backed table with columns Doc | Action | Source | Rationale. For every planned task, add a canonical_doc_actions row or reference to the shared table. The row must include doc, action: update|skip, source (spec, design, plan, or touched-files), and rationale / skip_rationale when action: skip. Also mirror the table into ### Hand-off to Execute → canonical_doc_actions_summary. This does not patch canonical docs; it makes review/verify able to detect drift before ship-review performs atomic writes.

Step 1.5 — Assumption re-validation

Scan spec for file:line citations. Read each; compare current content vs spec's assumption. Stale line (shifted but claim plausible) → note (⚠ stale-evidence) inline. Contradicted (content shows opposite) → BLOCKER (status: blocked, reason: sharp assumption contradicted). Rationale: stale-premise plans = most expensive failure mode (executor BLOCKs on non-matching codebase).

Skip if no file:line citations (common for S-size).

Step 1.6 — Import Design DCs from Hand-off (G10, 2026-04-29)

Trigger (G14, 2026-04-29 disambiguation): entity body MUST contain ### Hand-off to Plan block. Two paths:

Block has design-skipped: true → valid (design ran a Phase 0 trivial-pass, per INVARIANTS Principle 11 — NOT a pipeline skip) only when the entity is non-UI (affects_ui: false), domain: unset, design_required: false, contract_decision_required: false, and no open_contract_decisions remain, OR the handoff also carries captain-approved-design-bypass: true with a rationale. For affects_ui: true without that explicit bypass, BLOCKER (status: blocked, reason: ui design handoff skipped) and bounce to design/FO. For domain: set, design_required: true, contract_decision_required: true, or non-empty open_contract_decisions with a design-skipped stub, BLOCKER (status: blocked, reason: design-bearing decision skipped) and bounce to design/FO; planner must not choose the domain contract, selector grammar, API vocabulary, tool protocol, DSL, schema, or message format. Otherwise log ## Plan Imported Design DCs: design-skipped (no UI surface, no matched domain, no design_required signal, no contract decision required, or captain-approved bypass) and proceed.
Block has design-emitted fields (design_constraints[] / render_fidelity_targets[] / whole_page_visual_targets[]) → run mechanical mapping below.
Block ABSENT entirely → BLOCKER (status: blocked, reason: hand-off-to-plan absent — neither design-skipped stub nor design output found). Either shape Phase 8 missed the stub emit OR design errored without writing hand-off. Do not silently treat as "no UI" — that's the ambiguity G14 fixes.

Read via bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/extract-section.sh" <entity-file> hand-off-to-plan (handles folder + flat layouts).

Mechanical mapping delegated to lib script:

bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/import-design-dcs.sh" <entity-folder> >> <entity-folder>/plan.md

The script reads structured ### Hand-off to Plan, emits ## Plan Imported Design DCs table directly. Falls back to MIGRATE-FIRST notice when entity is in legacy prose format. Do NOT manually transcribe — that's the LLM-drift mode this script eliminates.

C15 accounting: body-line counts exclude content inside balanced standalone <details>...</details> blocks; the plan cap and its 2× raw-total backstop remain unchanged.

Pre-import validation (run in this order):

bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/check-design-readiness-review.sh" <entity-folder>/design.md — risk-gated design readiness check. If it reports status=blocked, stop before planning. If it reports status=warn, continue only after preserving the warning in ## Plan Report.
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-handoff-schema.sh" <entity-folder> — structural check
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-d-references.sh" <entity-folder> — D{N} backref consistency
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/import-design-dcs.sh" <entity-folder> — count-preserving import; if it reports imported design_constraints count mismatch, BLOCK instead of hand-copying missing rows.

If the readiness checker blocks → BLOCKER (status: blocked, reason from checker stdout; bounce to design/FO). If either hand-off validator fails → BLOCKER (status: blocked, reason: hand-off schema invalid; details from script stderr). If importer fails or imported row count is lower than source design_constraints[], write ## Plan Report status: blocked, reason: design DC import mismatch and bounce to design/stage tooling. Do not proceed with a partial ## Plan Imported Design DCs table. If visible_surface_map[] is present and the imported visible-surface row count is lower than the source row count, BLOCK with reason: visible surface map import mismatch; row count preservation is required because a dropped surface recreates the closed-list false-positive failure mode.

Mechanical mapping (each item becomes a DC anchored to a wave):

Hand-off field	Becomes	Wave
`design_constraints[]` (UI or domain contract types)	`ui`, `e2e`, `contract`, or `api/schema` typed DC per constraint; UI constraints use `ui-verify`, domain constraints use contract/router/schema tests	task wave touching the affected component
`visible_surface_map[]` (captain-visible regions, controls, state indicators, semantic badges)	structural/mockup-parity typed DC seeds per mapped row; explicit N/A rows are imported as visible audit exclusions, not dropped	wave touching the affected route or verify helper
`render_fidelity_targets[]` (computed-style / structural assertions)	`ui`-typed DC, `verified_by: ui-verify` (calls computed-style assertion)	W0 (token-level) or task wave (component-level)
`whole_page_visual_targets[]` (route-level composition parity)	`ui`/`e2e` typed whole-page visual parity DC; verifier captures full-page screenshot and compares against `reference_artifact`	integration wave or affected route wave
`artifact_paths[]`	reference in plan `## Design Source` block; checksum on read	—
`ui_quality_contract`	import each group from `plugins/ship-flow/references/ui-quality-contract.md` into DCs, reviewer questions, or explicit N/A rows. Required groups: `copy`, `visual_hierarchy`, `color`, `typography`, `spacing`, `interaction_states`, and `source_safety`.	task wave touching the affected UI/component contract
`open_decisions[]` non-empty	BLOCKER (status: blocked, reason: design has open decisions) — bounce to design via SendMessage(designer@pitch-XX)	—

Each imported DC retains rationale_decision: D{N} cross-reference back to ship-design Phase 8 ## Captain Decisions for audit trail.

Why this exists: ship-design's hand-off-to-plan block previously said "encode render_fidelity_targets as DCs" but no plan step did the mechanical conversion — relied on planner LLM noticing the instruction at the bottom of the entity body. Result: render fidelity targets silently dropped on ~30% of UI pitches. This step is the explicit machine path.

Output: ## Plan Imported Design DCs section in plan.md listing each imported DC with source field + wave assignment + rationale_decision link.

Step 1.7 — Architecture-lens dispatch (entity 110, 2026-04-29)

After spec read + design DC import, check whether the spec touches any known cross-cutting domains and dispatch read-only architecture lens agents before research.

Science Officer (EM) stewardship for plan-stage assignments

Before dispatching any architecture lens, research producer, research reviewer, plan cross-reviewer, or gap-fill worker, render and include the shared worker-facing stewardship section:

bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/render-science-officer-em-stewardship-contract.sh"

The resulting ### Science Officer (EM) Stewardship Contract block is part of the assignment body. It carries results, guidelines, resources, accountability, consequences. FO owns workflow clock, state, worktrees, dispatch mechanics, PR lifecycle, and stage advancement. EM owns engineering judgment, delegation quality, worker stewardship quality, risk/scope challenge, and technical recommendations. EM does not mutate entity state, own worktrees, dispatch workers, create or merge PRs, or advance stages. Verification is output-shape evidence, not worker self-attestation.

Trigger matching — read plugins/ship-flow/references/architecture-lens-triggers.yaml. For each domain entry:

File-glob match: check spec ### Artifacts likely touched paths against trigger_patterns (bash glob, case-insensitive). Match if any path matches any pattern.
Spec-keyword match: grep spec body (full text) for each spec_keywords entry (case-insensitive). Match if any keyword found.
OR-semantic: domain triggers if EITHER file-glob OR spec-keyword matches.

Trivial escape hatch: if entity frontmatter appetite: trivial AND (affects_ui: false OR spec body contains docs-only: true) → skip lens dispatch entirely. Still emit ## Context Manifest with Lens dispatched: skipped (trivial escape hatch).

Dispatch — for each matching domain, dispatch Skill(lens_skill) with prompt:

Read spec at {spec_path}. Read {domain_knowledge_refs}. For each cross_cutting_concern listed in your frontmatter, output a structured YAML verdict (FLAG/PASS/SKIP with missing_dc if FLAG). ≤300 words total.

All matching lenses dispatch in parallel (no budget cap for v1). Collect structured YAML verdicts.

If no domains match → emit Lens dispatched: none (no trigger match) in Context Manifest. Proceed normally.

Lens FLAG integration — after all verdicts collected, for each FLAG verdict:

Option A: add DC covering the flagged concern to the plan (preferred)
Option B: write entry in ## Lens Findings: deferred section with explicit rationale per skipped FLAG

Gate refuses plan stage advance (Step 6.1) if any FLAG exists without Option A or Option B entry. The ## Lens Findings: deferred section MUST enumerate each deferred FLAG with: concern, rationale, accepted_by (captain or plan worker with justification).

Lens output in plan.md — record all verdicts in ## Context Manifest → Lens dispatched field. Full structured YAML verdicts stored as ## Lens Raw Verdicts appendix section (for audit; not required reading for execute stage).

Step 2 — Research (size-adaptive, produce+review)

S — skip research. Proceed to Step 3.
M — dispatch Agent A (sonnet producer: codebase impact + lib constraints, <400 words, file:line citations; include recent debrief warnings from docs/<wf>/_debriefs/ — read ## Issues — Workflow / ## Filed (backlog) sections, schema at plugins/ship-flow/references/debrief-schema.yaml) → Agent B (sonnet reviewer: APPROVED / GAPS / CONTRADICTION, <300 words). Max 1 gap-fill round.
L — dispatch 3-5 parallel producers (upstream constraints / existing patterns / lib surface / gotchas / reference examples), each <200 words, file:line citations; one producer dedicated to recent debrief warnings (docs/<wf>/_debriefs/ last 3-5 files, extract ## Issues — Workflow / ## Filed (backlog) / ## Observations). Reviewer runs cross-domain + coverage check. Max 1 gap-fill round.

Contradiction: write both verbatim as Open Question in ## Research Summary. Never silently resolve.

Step 2.5 — Size re-evaluation

Count actual affected files from research vs spec's size estimate.

Sharp	Actual	Action
S	≤3	Confirmed S
S	4-10	Upgrade to M
S	>10	Upgrade to L (run full research)
M	≤3	Downgrade to S
M	4-15	Confirmed M
M	>15	Upgrade to L
L	any	Confirmed L

Update entity frontmatter size: if changed.

Fold/extract LOC heuristic (harness-diet restructures; n=2 calibration):

Fold (A → B, A retired): fold_net_loc ≈ source_loc × 0.44-0.55. DC-1 budget worst-case = target_initial + source × 0.55. Include procedural skeleton + example tables + guidance templates (not just the Steps-1-to-N).
Extract (1 source → N callers with refs): extraction_net_loc ≈ -(source_loc × 0.55-0.70) where source = summed preamble LOC across callers.
Picking: ≤50 LOC + one parent → fold. ≥100 LOC + N≥3 callers → extract. 50-100 LOC + N=2-3 → fold into most natural parent. Extraction adds 1 skill (Principle 2 cap check first).

Next.js standalone artifact size heuristic (DC budgets on ui-dist/ or .next/standalone/):

Measure ui/.next/standalone/ SUBTREE, NOT full ui/.next/. Full .next/server/ includes all compiled chunks + source maps (~40MB+); .next/standalone/ contains only traced runtime deps + subset of maps.
Plan estimates based on full ui/.next/ over-predict .map strip reduction by 4-10×.
Ground truth methodology: run target bundler locally, then du -sh ui/.next/standalone/ + find ui/.next/standalone -name "*.map" | wc -l before locking size targets.

Step 2.7 — Scope anchoring (M/L only)

Every task maps to a spec Scope In bullet (if shape ran) OR a Done Criteria item. Produce task×mapping table. Unmapped task → drop or flag in ## Plan Report. Do NOT silently expand scope.

Step 3 — Write plan (delegate to superpowers:writing-plans + ship-flow TDD fallback)

Invoke Skill: superpowers:writing-plans for plan authoring when available. It handles TDD task order, wave safety, placeholder scan, task atomicity, verification commands per task Done. Then invoke Skill: ship-flow:test-driven-development for the ship-flow-owned TDD contract. If superpowers are unavailable in an adopter/runtime, do not block solely on that absence; ship-flow:test-driven-development is the fallback source of truth.

Layer B wrap (ship-plan owns these on top of Layer A output):

Runtime detection — invoke ship-flow:ship-runtime-detect before tasks get written; propagate {commands.test/build/typecheck/lint/dev} into every task's Done field.

Per-task skills_needed derivation (#108.1): after task files are known, populate each task with a skills_needed: [...] string array. Derive it by intersecting:

task files_modified / **Files:** globs
framework_detected, theme_indirection, and design_canonical_dir from ship-flow:ship-runtime-detect
density-classified skill set already loaded for answers_density: high
domain registry routing from registry-resolve.sh --domain=<domain> or --classify <spec>: preserve required_skills and merge skill_hints.plan into plan-stage task skills_needed
adopter file-signal routing from .claude/ship-flow/skill-routing.yaml when present. For each implementation task, run bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/resolve-skill-routing.sh" --files=<task-files> --config=.claude/ship-flow/skill-routing.yaml and merge the emitted skills_needed= list. Also record folder_guidance_files= and folder_guidance_skills= in ## Context Manifest → Folder guidance, and merge guidance skills into task skills_needed when they are not already present. If the config is absent on a non-trivial multi-surface pitch, run bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/discover-adopter-skills.sh" --root=. and surface the draft in ## Context Manifest before finalizing skills_needed.

Use concrete file-glob mapping, then trim to the smallest relevant list:

File signal	skills_needed candidates
`.tsx`, `.jsx`, `components/`, `app/`, `ui/**`	`frontend-design`, `react-best-practices`, `test`
`.css`, `tokens.css`, `design-system.md`, `design/*`	`frontend-design`, `web-design-guidelines`, `accessibility`
`apps/supabase/migrations/`, `domains//src/schema/**`	`project-db`, `test`
`domains//src/domain//{types,decider,view,saga}.ts`, `apps/deno-api/src/middlewares/fmodel-middleware.ts`	`fmodel`, `test`
`.test.`, `.spec.`, `__tests__/**`	`test`, `tdd` or `test-driven-development`
`.sh`, `bin/`, `lib//.sh`	`test`, `best-practices`
`.md`, `SKILL.md`, `docs/*`	`write-docs` when prose is user-facing; omit for stage artifacts

Adopter .claude/ship-flow/skill-routing.yaml entries override the generic table only for matching file signals. Example: apps/refine-app/src/** may add refine-expert, refine-gotchas, antd-expert, react-patterns, and tailwind-expert; apps/expo-app/** may add expo-rnr-nativewind and expo-accessibility; packages/api-contract/src/**/*.schemas.ts may add ts-rest and api-guide.

Codex overlap boundary: do not duplicate root session instructions. resolve-skill-routing.sh emits codex_context_boundary=root AGENTS.md/CLAUDE.md intentionally excluded from folder_guidance_files; this means root AGENTS.md/CLAUDE.md remain runtime/session context, while ship-flow records only non-root folder guidance such as apps/refine-app/CLAUDE.md. This guard exists because Codex may already load root AGENTS.md, but it does not guarantee every task worker or PR-feedback re-entry re-reads adopter app-folder guidance.

Domain-derived skills are additive, not adopter-specific defaults. project-db and fmodel are generic skill names only when those skills exist in the adopter project or local Codex environment; otherwise keep the registry output visible in ## Context Manifest and surface missing skills early.

Non-trivial plan guard: if a plan has ≥2 implementation tasks touching different file classes, require ≥2 distinct skills_needed lists. If every task receives the same list, treat it as boilerplate and revise before emitting plan.md.

Adopter gate registry: for every implementation task with non-empty owned_paths, run python3 "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/resolve-gate-registry.py" --config .claude/ship-flow/gates.yaml --files <comma-separated-owned_paths> when the adopter config exists. Merge required_gates, reviewer_questions, and evidence_required into the task's reviewer questions and domain_acceptance_checklist; record matched_routes and layers in the Context Manifest. Missing config is a WARNING for legacy/single-surface plans and a plan gap for new multi-layer adopter plans. These gates are adopter-owned; ship-flow core only resolves and carries them.
TDD contract: every non-exempt implementation task MUST include tdd_contract with red_command, expected_red_failure, green_command, and refactor_check. In prose task blocks, include the same data as RED command, Expected RED failure, GREEN command, and REFACTOR check. Execute consumes this contract for RED-before-GREEN evidence; verify audits it before PASS.
TDD ledger gate: after drafting plan.md and before ### Hand-off to Execute, run python3 "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-tdd-ledger.py" --plan <entity-folder>/plan.md > <entity-folder>/tdd-ledger.txt 2>&1 for the human-readable pass/fail summary, then run python3 "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-tdd-ledger.py" --plan <entity-folder>/plan.md --emit-jsonl > <entity-folder>/tdd-ledger.jsonl. Finally run python3 "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-tdd-ledger.py" --plan <entity-folder>/plan.md --require-ledger-jsonl <entity-folder>/tdd-ledger.jsonl to prove the persisted ledger exists and matches the current plan. Failure blocks plan advance. This makes the tdd_contract authoritative at plan time rather than post-hoc extraction. The validator rejects missing contracts, low-confidence red_command / green_command, invalid layer, stale ledger JSONL, and code-bearing tasks that declare layer: meta when the inferred layer is L1-L7.
TDD exceptions (mark inline): config / pure refactor with coverage / docs-only / migration — TDD: skip -- <reason>.
Wave rules: wave 0 = test infra; wave N+1 depends only on ≤N outputs; no files_modified overlap within a wave; no cycle.
UI quality contract import: when ### Hand-off to Plan contains ui_quality_contract, read plugins/ship-flow/references/ui-quality-contract.md and turn every group (copy, visual_hierarchy, color, typography, spacing, interaction_states, source_safety) into DCs, reviewer questions, or explicit N/A rows. Missing coverage for any group is a plan gap because verify needs concrete evidence prompts.
Parallel metadata contract: every implementation task MUST include parallel_group, depends_on, owned_paths, and integration_owner. Use parallel_group: serial when the task cannot run with peers. owned_paths is the write boundary execute uses to prove disjointness; depends_on is T{N} task-ID topology; integration_owner is the named teammate that folds the worker output into the stage artifact (normally executer@pitch-XX).
Emit plan-parallelization-manifest in ### Hand-off to Execute: one row per task with task_id (T{N}), parallel_group, depends_on, owned_paths, and integration_owner. task_id is the canonical task identifier used throughout the plan, and every depends_on entry MUST reference these same T{N} task IDs. This is contract metadata, not an execution command. Execute derives execute-dispatch-manifest from it and may only parallelize tasks whose dependencies are satisfied and whose owned_paths do not overlap.
If design emitted domain expert panel constraints, convert them into a domain-specific acceptance checklist before task writing. Each checklist item must name the domain lens, required skill/knowledge module, affected path family, and the verification command or reviewer prompt verify will reuse.
Every task with domain or framework skills_needed MUST include reviewer_questions. Convert skills_needed into reviewer questions, not just labels. For each task skill that represents a domain lens (project-db, project-seed, fmodel, refine-expert, refine-gotchas, API/router skills, Expo/mobile skills, etc.), add 1-3 concrete checks to the domain-specific acceptance checklist. Example shape: "refine-gotchas: verify list/detail cache invalidates after mutation; URL/drawer state hydrates on back/forward/shared link; ProCRUD/list hooks follow existing app pattern."
Emit domain_acceptance_checklist in ### Hand-off to Execute with columns Task ID, Verify Lens, Reviewer Question, Affected Path Family, Required Skills, and Evidence Required. This is the plan-to-verify contract consumed by ship-verify when it populates verify-check-manifest.review_lenses.
When registry or domain routing is present, emit exactly one standalone context-routing-manifest section-tagged block and a ## Context Routing Receipt. The receipt maps manifest rows to tasks[].skills_needed, task reviewer_questions, and domain_acceptance_checklist; optional rows may be skipped only with an explicit rationale. Future provider hints are append-only and non-authoritative: the local registry remains the routing source of truth for this block.
T0.X auto-indirection-sweep (T6.1, #106): when theme_indirection from ship-runtime-detect Step R5 is non-empty (e.g. tailwind-v4), auto-emit a Wave 0 task in the plan: T0.X: Audit @theme inline indirection layer — verify design tokens align with CSS custom properties, no hardcoded hex values in component files. REFUSE to emit a plan without this task when theme_indirection != "". Enforcement: bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/bin/check-invariants.sh" --check indirection-sweep-emitted (fixture-based).
Lens FLAG integration (entity 110, 2026-04-29): after Step 1.7 lens verdicts are collected, plan worker MUST for each FLAG verdict: (a) add a DC covering the flagged concern, or (b) write an entry in ## Lens Findings: deferred with explicit rationale. Gate refuses advance if any FLAG exists without Option A or Option B. Silence is NOT acceptable — the deferred section exists precisely to make punts explicit and captain-visible.

Step 3.5 — Verification spec (structural parity enforcement)

Validation task design discipline — when a plan task description is "validate X", classify X before authoring:

X is output-shape (code pattern, file presence, schema match, runtime assertion) → write as DC in ## Verification Spec, consumed by verify stage. Predicate-grep-able, evidence accumulates across N naturally.
X is process-shape (did the worker load skill, did the worker read section, did the worker echo attestation token, did the canary appear in the receipt) → do not author. Per-dispatch process-effect validation has structurally bad cost-to-signal ratio (baseline contamination + scope mismatch + timing pollution; see plugins/ship-flow/references/validation-discipline.md). Either rewrite as the equivalent output-shape DC, or drop the task.

Per typed Done Criterion, fill exact verification procedure:

Type	Procedure	Fallback
`cli`	bash command + expected exit/output	always available
`api`	`curl` with method/URL/body + status + response pattern	always available
`ui`	`curl -sfN <route>` + `grep` content; complex interaction → `e2e-test <flow.yaml>`	curl+grep skip interaction
`design-system-parity`	Runtime computed-style evidence compared to a design-system token table or explicit token contract. Required plan evidence fields: `selector`, `css_property`, `token_source`, `expected_token`, `expected_resolved_value`, `actual_computed_value`, `runtime_report`, `verdict`.	`ship-flow:ui-verify` for fixed selector/value checks; inline browser `getComputedStyle()` probe only as break-glass
`mockup-parity`	Live DOM structure evidence compared to an HTML mockup or committed design artifact. Required plan evidence fields: `mockup_artifact`, `route`, `root_selector`, `expected_structure`, `actual_dom_structure`, `comparison_method`, `runtime_report`, `verdict`.	deterministic selector hierarchy, role/name sequence, landmark order, repeated-item counts, or normalized DOM digest
`skill`	`Skill("<name>")` invoke + expected output shape	session-only verification
`e2e`	`e2e-test <flow.yaml>` with step assertions	degrade to `ui` type

Every DC MUST have a runnable verify procedure. "Manual check" = plan failure.

design-system-parity and mockup-parity are first-class Done Criterion subtypes, not optional prose. A planner must turn each into an executable procedure with runtime-backed evidence. Token tables, tokens.css, screenshots, generated artifacts, source grep, mockup files, or source-only diffs are artifact-only claims unless execute and verify can inspect the required runtime fields above.

UI entities MUST have ≥1 structural-parity DC — grep DCs prove wiring, not rendering. Column/cell count parity / grid-template vs header count / prop-type assertion / class-name presence. Evidence: entity #048 cycle — all grep DCs passed while 3 CSS-grid / prop-mis-wire bugs shipped; captain cycle wasted. Budget one structural DC per UI entity.

Note: ui type uses -sfN (Next.js 16 Turbopack SSR chunks — MEMORY #073). Backward-compatible.

Declarative e2e artifact requirement (UI-type DCs): when ≥1 DC has type ui or e2e, plan's task list MUST include authoring the declarative YAML so ship-verify Step 4.4 can run automated pre-smoke:

Static CSS / design tokens / computed-style regression → .claude/e2e/ui-verify/<entity-slug>.yaml (schema: ship-flow:ui-verify). Wraps getComputedStyle(selector)[prop] equality checks against tokens.
Dynamic DOM / navigation / event / interaction → .claude/e2e/flows/<entity-slug>.yaml (schema: e2e-pipeline:e2e-flow). Step-based browser flow with assert primitives.
Both require .claude/e2e/mappings/<mapping>.yaml (auth + base_url + test_accounts). If missing → plan's first UI task is mapping bootstrap.
Plan-stage DC: YAML file exists at canonical path AND imports valid mapping. Consumer: ship-verify Step 4.4 auto-runs matching skill.
Fallback when skill wrapper unavailable / mapping missing: inline agent-browser CLI script as a verify-time DC (declarative YAML preferred; CLI only as break-glass).

Step 4 — Self-review (plan-checker-lite)

Run 9 dimensions. Any BLOCKER → fix inline + re-review. Max 3 iterations.

Requirement coverage — every DC maps to ≥1 task.
Task completeness — every task has paths / verify command / model hint / wave.
- For #108.1+, every implementation task must include skills_needed: [...]; docs-only stage-artifact tasks may use skills_needed: [] only with TDD: skip -- docs-only/stage-artifact.
Dependency correctness — wave graph safe (no cycles / no same-wave files_modified overlap / no read_first pulling same-wave outputs).
Zero-placeholder scan — grep TBD | add appropriate | similar to Task N | as needed | fill in | \.\.\. → fix inline.
Type/signature consistency — cross-task function/type signatures match.
Task minimality (M/L) — merge adjacent non-load-bearing tasks; drop nice-to-haves not in DCs.
TDD compliance — test-first order on code tasks; TDD: skip tasks have valid reason.
Stale-line-anchor — every file:line citation re-read; content-matches / line-shifted / contradicted (BLOCKER) / phantom path (BLOCKER). Catches #1 cause of execute BLOCKED returns.
Design reference compliance — skip if no ## Design Reference section; else cross-check visual attrs (fill/stroke/colors/dimensions) against cited spec files. Flag deviations.
Stub-captain-ack scan (T6.2, #106): grep every task body for keywords stub|fake|placeholder|v1.*only|wired only for. For each match: check entity frontmatter pre-acked-stubs: true OR check that the Plan Report has a ## Stub Flag entry for this task with explicit captain rationale. If neither present → BLOCK: stub task without captain ack (literal string required for test DC). Populate ## Plan Report → Stub Flags table with all stub tasks found. Cross-review PROCEED blocked until all stubs either pre-acked or cleared.
Context Manifest completeness (entity 110, 2026-04-29): ## Context Manifest section present and all 7 fields non-empty: Skills loaded, INVARIANTS sections read, Architecture docs consulted, Domains touched, Lens dispatched, Lens findings integrated, Folder guidance. Lens dispatched field must reflect actual Step 1.7 trigger matching results (not copy-pasted from a prior entity). Folder guidance must cite every non-root folder_guidance_files= entry from resolve-skill-routing.sh, list folder_guidance_skills=, and include the literal codex_context_boundary line so reviewers can see this is not duplicating Codex root instruction loading. C8 check: bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/bin/check-invariants.sh" --check context-manifest-emitted.
skills_needed non-boilerplate (#108.1): for plans with ≥2 implementation tasks touching different file classes, grep the task blocks and confirm there are ≥2 distinct skills_needed lists. Empty lists or one repeated list across heterogeneous tasks are BLOCKERs.

Step 5 — Cross-review gate (Principle 6 Rule C)

Dispatch cross-review to executer teammate (pipeline path) or fresh sonnet (no team). Upgrade to fresh opus when appetite: big-batch.

7-factor rubric adapted for plan stage from INVARIANTS Principle 6 Rule C #106 T1.3 plus the ship-plan skill-coverage extension:

Feasibility — tasks achievable by single agent in one dispatch each?
Executable scope — tasks are atomic commits aligned 1:1 with waves?
Quality — Verification Spec covers every DC with runnable procedure (≥1 structural-parity DC for UI)?
DC adequacy — observable DCs per task; no "works correctly" / "handles all cases"?
Canonical sync — ARCHITECTURE.md touches planned? architecture-impact draft per affected child?
Reverse-audit previous stage — does the plan's scope expose a gap in the preceding design stage's ### Hand-off to Plan block? Specifically: are all render_fidelity_targets from design encoded as DCs? Are design_constraints honored in plan tasks?
Skill Coverage (skill-coverage, #108.2) — every implementation task has non-empty skills_needed and the skills match its files_modified / **Files:** file globs. Emit exactly one grep-testable summary line when clean: skill-coverage: PASS. Emit one line per failing task when not clean: skill-coverage: FAIL — task <id>: <reason>. Failure reasons include empty skills_needed, boilerplate repeated lists across heterogeneous implementation tasks, or file-glob/skill mismatch. Use these required matches:
- *.tsx, *.jsx, components/**, app/**, ui/** → react, frontend-design, or react-best-practices
- *.css, tokens.css, design-system.md, design/** → frontend-design, web-design-guidelines, or accessibility
- apps/supabase/migrations/**, domains/**/src/schema/** → project-db
- domains/**/src/domain/**/{types,decider,view,saga}.ts, apps/deno-api/src/middlewares/fmodel-middleware.ts → fmodel
- *.test.*, *.spec.*, __tests__/** → test, tdd, or test-driven-development
- *.sh, bin/**, lib/**/*.sh → test or best-practices
- Docs-only stage-artifact tasks may use skills_needed: [] only with TDD: skip -- docs-only/stage-artifact; user-facing prose docs should use write-docs.

UI extension: Render Fidelity + captain-ack audit trail (T6.4, #106) — for UI entities: does plan have ≥1 structural-parity DC per component in design canonical? AND are all stub tasks captain-acked (either pre-acked-stubs: true in frontmatter or explicit ## Plan Report → Stub Flags entries)? This remains the render-fidelity named extension documented in INVARIANTS and is not the ship-plan skill-coverage line.

Reverse-audit prompt template (T3.2 — paste verbatim into reviewer dispatch):

Reverse-audit: Read the entity's `### Hand-off to Plan` block.
(a) List every `render_fidelity_target` — does plan.md have a DC for each? (BLOCKING if any missing)
(b) List every `design_constraint` — does each appear in at least one plan task or Verification Spec row? (WARNING if any absent)
(c) Are `open_decisions` empty? If not empty, is there a plan task that resolves each? (PROMPT_CAPTAIN if unresolved)
Coaching note: render_fidelity gaps here become silent UI regressions at verify — enforces FM#4 (fidelity gap) prevention.

Verdict: PROCEED / VETO (max 2 loops back to Step 3 with reviewer feedback; round 3 → PROMPT_CAPTAIN) / PROMPT_CAPTAIN. Each verdict MUST include a one-sentence coaching note per INVARIANTS Rule C ABC clause.

Circuit breaker: if SendMessage(executer) is unresponsive (phantom team / timeout / fresh-Agent stall), fall back per INVARIANTS Rule A Fallback — fresh sonnet by default, fresh opus on big-batch. Do not block on an unresponsive reviewer.

Step 6 — Emit plan.md

Write to <entity-folder>/plan.md via bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/write-stage-artifact.sh" --stage=plan --entity=<id>-<slug> --content=<draft-path> (Wave 5 primitive landed at commit acd73545). Primitive handles atomic commit with explicit pathspec.

Plan.md sections: ## Research Summary (findings + open questions if contradictions + reviewer verdict), ## Size Re-evaluation, ## Verification Spec (table from Step 3.5), ## Plan (TDD tasks from Step 3), ## Plan Report (status, stage_cost: dispatches×model, iterations, dimensions pass/fail, reviewer verdict, scope anchoring, task count, model split, started/completed/duration), ## Context Manifest (mandatory — see Step 1.7 and dimension 11).

Verbosity budget (INVARIANTS Principle 8 — plan.md ≤200 body lines; C15 BLOCKER): the plan body is tabular (waves + tasks + DCs). Keep it lean:

## Verification Spec — key every row by DC-N (### Done Criteria in shape.md is canonical for assertion + type; cite by DC-N, do NOT restate the assertion/type columns — 129.1 CD-2). Only Verify Procedure + Expected are the genuinely-new columns this stage produces.
Do NOT restate the shape.md Problem / User Journey / DC prose — cite via extract-section.sh, per 129.1 reference_not_restate.
Verbose research → ## Research Summary stays a findings summary; a bounded excerpt of producer/reviewer transcripts may go in <details> (excluded from the ≤200 BODY cap, but the C15 2× raw-total backstop — raw ≤ 400 — still applies, so excerpt-not-dump). Large transcripts → reference a DURABLE artifact (committed-in-repo via explicit pathspec, OR a durable URL — a bare local path is not audit evidence), don't inline.
Do NOT enumerate commit SHAs or git-reconstructable file lists in the body — git log / git diff has them.

Require a ### Metrics subsection in ## Plan Report. Use grep-friendly key: value lines:

status: passed | gaps-noted | blocked | partial
duration_minutes: wall-clock minutes for planning
iteration_count: self-review + reviewer loop count
task_count: implementation task count
verification_spec_count: DC rows in Verification Spec
model_split: dispatch model breakdown

## Context Manifest section format (mandatory output, entity 110, 2026-04-29):

## Context Manifest

- **Skills loaded**: [comma-separated skills invoked]
- **INVARIANTS sections read**: [section names with file:line citations]
- **Architecture docs consulted**: [docs read, with paths]
- **Domains touched**: [domain names from trigger registry that matched, or "none"]
- **Lens dispatched**: [list: domain (verdict summary) or "none (no trigger match)" or "skipped (trivial escape hatch)"]
- **Lens findings integrated**: [N integrated, M deferred-with-rationale, K ignored — must be 0 ignored unless all flags are deferred]
- **Folder guidance**: [for each task group: `files=<paths>` → `folder_guidance_files=<non-root AGENTS.md/CLAUDE.md>`; `folder_guidance_skills=<skills>`; `codex_context_boundary=root AGENTS.md/CLAUDE.md intentionally excluded from folder_guidance_files`]

C8 enforcement: bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/bin/check-invariants.sh" --check context-manifest-emitted — fails if section absent in any non-blocked plan.md created after 2026-04-29.

Mark TaskCreate sub-task emit-plan.md completed; return to /ship for advance to execute.

Step 6.1 — Register plan completion

After stage artifact lands, advance sibling index.md frontmatter atomically:

INDEX_MD="<entity-folder>/index.md"
H="$(if command -v sha256sum >/dev/null 2>&1; then sha256sum "$INDEX_MD" | awk '{print $1}'; else shasum -a 256 "$INDEX_MD" | awk '{print $1}'; fi)"
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/advance-stage.sh" \
  --entity="$INDEX_MD" \
  --new-status=plan \
  --stage-name=plan \
  --stage-file=plan.md \
  --if-hash="$H" \
  --lease-file="$SHIP_FLOW_COMPLETION_LEASE_FILE" --lease-token="$SHIP_FLOW_COMPLETION_LEASE_TOKEN" --worker-id="$SHIP_FLOW_COMPLETION_WORKER_ID" \
  --commit-as="plan(<id>): advance status to plan"

This registers plan completion; it does not enter execute. Return to the First Officer for a separate execute stage-entry transition. Return the receipt verbatim. FO reclaims the lease: published runs path reconcile; already-registered runs clean/no-lag. Only reconciled|ready may precede Contract 1.

On exit 6 (stale hash): write ## Plan Report status: blocked, reason: index.md stale hash; parallel session contaminated and return.

Post-ship Retro Prompt (entity 110 Captain Bet — SC-810-class prevention)

After this feature ships, re-read the ## Captain Bet block in the entity body. At the 5-carlove-pitch mark, evaluate outcome:

Bet substance: "下 5 個 carlove pitches 不再有 SC-810-class cascade-saga gap."

Retro procedure:

Review the last 5 carlove pitches that touched tag/customer/event-saga domains
For each pitch where a cascade-saga gap still occurred, categorize per 4 bet-failure paths:
- Path 1 — Domain not in lens registry (trigger didn't match) → add lens for that domain
- Path 2 — Lens fired but plan worker ignored finding → strengthen W3 gate from gate-with-explicit-deferral to hard-block-no-deferral
- Path 3 — Manifest section missing key dimension → add new manifest field
- Path 4 — Lens fired + plan integrated, but failure still occurred → format-level revisit (meta-level re-shape)
If ≥3 consecutive bets fail → freeze lens registry + trigger re-shape of this pitch

Kill criterion: 3 consecutive Bet ≠ Outcome → freeze lens registry + re-shape. Single fail → apply specific improvement per 4 paths above.

Invariants + red flags (STOP if violated)

Every DC has a runnable verify procedure. "Manual check" = plan failure.
UI entity without ≥1 structural-parity DC = silent visual-bug risk.
Contradicted sharp assumption (Step 1.5) → BLOCKER, do not plan on stale premises.
Wave graph violates safety (cycle / overlap / cross-read) → BLOCKER.
Placeholder prose (TBD / "similar to Task N" / ...) = plan failure; fix inline.
Cross-review VETO loop capped at 2 rounds; round 3 → PROMPT_CAPTAIN.
Layer A delegation (superpowers:writing-plans) owns plan authoring — re-teaching = Principle 6 Rule B violation.
Scope expansion beyond spec Scope In / DCs = Shape Up violation; drop or flag.

Circuit breakers

Research producer timeout: 2 min/agent.
Research reviewer: max 1 gap-fill round.
Self-review loop: max 3 iterations → proceed with gaps noted in Plan Report.
Cross-review reviewer: max 1 round per VETO; 2 VETOs total before PROMPT_CAPTAIN.
Total stage >20 min elapsed → write plan.md with partial content + ⚠️ INCOMPLETE markers + Plan Report status=partial, then return. Never exit without emitting plan.md.

Step 6 (Hand-off): Emit Hand-off to Execute + Read Incoming Hand-off

Read incoming: at Step 1, read ### Hand-off to Plan from entity body. Verify design_constraints honored in plan tasks; encode render_fidelity_targets as DCs.

Emit ### Hand-off to Execute after plan.md is written:

tdd-ledger: path to tdd-ledger.jsonl, validator command, pass/fail summary from validate-tdd-ledger.py --plan, and persisted-ledger proof from validate-tdd-ledger.py --plan <plan.md> --require-ledger-jsonl <tdd-ledger.jsonl>; do not hand off a new non-trivial plan whose ledger validation fails.
wave_order: exact wave dispatch order (e.g., "W0 → W1 (T1.1→T1.2→T1.3→T1.4) → W2a → W2b → W2c → W3 → W4")
critical_assumptions: assumptions to re-verify at execute boot (from Assumption Re-validation section)
architecture_context: canonical doc touches required (INVARIANTS, README, schema) — drives execute commit order
stub_flags: tasks containing stub|fake|placeholder|v1.*only|wired only for — must appear in Plan Report as captain-ack flags
skills_needed_summary: per-task skills_needed lists and note whether ≥2 distinct lists were produced for heterogeneous task sets
domain_acceptance_checklist: rows with Task ID, Verify Lens, Reviewer Question, Affected Path Family, Required Skills, and Evidence Required derived from task reviewer_questions
context-routing-manifest: exactly one standalone section-tagged manifest block plus ## Context Routing Receipt, with domain_matches, knowledge_modules, required_skills, stage_hints, consumer_obligations, and future_provider_boundary

References

Entity schema: plugins/ship-flow/references/entity-body-schema.yaml → stages.plan.
Stage writer: plugins/ship-flow/lib/write-stage-artifact.sh.
Section extraction: plugins/ship-flow/lib/extract-section.sh.
Layer A: superpowers:writing-plans (plan authoring discipline).
Utility: ship-flow:ship-runtime-detect (13-ecosystem runtime detection).
Reverse-recovery audit mod (every task creating a new file/domain/route needs a confirmed-MISSING classification line; reviewers reject greenfield tasks without it): plugins/ship-flow/_mods/reverse-recovery-audit.md (plugin-canonical); adopter override docs/ship-flow/_mods/reverse-recovery-audit.md when present.
Principle 6: plugins/ship-flow/INVARIANTS.md (context continuity + 3-layer architecture + cross-review).
MEMORY: #5 (--next-id atomicity), #14/#25 (explicit pathspec / staging), #35 (dispatch discipline amended by Principle 6), #073 (Next.js 16 -sfN), opus-4.7-naturally-does (2026-04-23 harness diet).

name	ship-plan
description	Use when writing a ship-flow plan for a shaped entity, especially implementation tasks, TDD needs, runtime commands, or scope anchoring. Layer A delegation: superpowers:writing-plans owns plan authoring; ship-flow:test-driven-development owns fallback TDD contracts.
user-invocable	false
argument-hint	[entity-id \| slug]

Ship-Plan — PLAN Stage (2.0)

You run PLAN. Output: <entity-folder>/plan.md. Dispatched by /ship to planner teammate via SendMessage (team spawned at /shape). No captain gate.

Boot Self-Check

Run before any plan work. Stop and SendMessage(FO) if any check fails.

Entity status: read entity frontmatter status: — must be sharp or design (post-design stage). If draft → shape first; if plan → plan already exists, check for re-entry signal.
Hand-off present: entity body contains ### Hand-off to Plan block (from ship-shape or ship-design). If absent → SendMessage(FO): "Missing Hand-off to Plan in <entity-path> — cannot proceed without design intent."
Team context: verify planner teammate is active (this agent). Note executer teammate name for Step 6 SendMessage.
PRODUCT.md readable: repo-root PRODUCT.md exists and is readable. If absent → SendMessage(FO): "PRODUCT.md missing — constraints source unavailable."
Framework detection (if affects_ui: true): Run ship-runtime-detect Step R5 to confirm framework_detected + theme_indirection for plan's verification spec.
Density-aware skill load (T3.4): read answers_density from entity frontmatter. high → auto-load framework skills per ship-runtime-detect Step R6; skip FO ask. low|vacuum → SendMessage(FO) with proposed skill list; wait for confirmation.
Canonical context planning: read PRODUCT.md, ROADMAP.md, and relevant ARCHITECTURE.md sections when spec/design/affected files indicate product, roadmap, API, schema, domain, storage, runtime, or architecture contract impact.

Entity body contract (schema-as-prose)

Reads: resolved shape artifact (shape.md; legacy spec.md fallback alias) (## Sharp Output / ## Shape Output — problem, done criteria, size, scope, assumptions, children DAG), design.md when present, parent entity ## Cross-Entity Contracts if parent: set, PRODUCT.md constraints, ROADMAP.md status, and relevant ARCHITECTURE.md sections.
Writes: <entity-folder>/plan.md sections — ## Research Summary, ## Size Re-evaluation, ## Verification Spec, ## Plan (TDD tasks with waves), ## Plan Report (status/cost/iterations/verdict).
Full section-tag + field semantics: plugins/ship-flow/references/entity-body-schema.yaml → stages.plan.

Layer A delegation (Principle 6 Rule B)

Shape Up scope anchoring (every task maps to a Done Criterion / Scope In bullet).
Size-adaptive research team (produce + review subagents, sized S/M/L).
Runtime detection via ship-flow:ship-runtime-detect for {commands.test/build/typecheck/lint/dev}.
Assumption re-validation (Step 1.5) against current codebase state before planning.
Plan-checker multi-dimensional review + cross-review gate.

Flow

Step 1 — Read spec + cross-entity contracts

Blocker: spec missing problem or done criteria → write ## Plan Report status: blocked, reason: missing spec and return.

Step 1.1 — Canonical context planning

Build ### Canonical Doc Actions before writing implementation tasks:

Doc	When to mark update	When to mark skip
`ROADMAP.md`	Entity is represented in Now/Next, creates/consumes deferred todos, or umbrella closeout may apply	Entity is intentionally not roadmap-listed; include `skip_rationale`
`PRODUCT.md`	Spec/design changes durable capability, user story, constraint, "Who It Serves", or "Why It Exists"	Internal-only change; include `skip_rationale`
`ARCHITECTURE.md`	Spec/design/tasks touch schema/API/domain/data-flow/storage/runtime/component boundary or architecture-impact exists	No architecture contract touched; include `skip_rationale`

Step 1.5 — Assumption re-validation

Skip if no file:line citations (common for S-size).

Step 1.6 — Import Design DCs from Hand-off (G10, 2026-04-29)

Trigger (G14, 2026-04-29 disambiguation): entity body MUST contain ### Hand-off to Plan block. Two paths:

Block has design-skipped: true → valid (design ran a Phase 0 trivial-pass, per INVARIANTS Principle 11 — NOT a pipeline skip) only when the entity is non-UI (affects_ui: false), domain: unset, design_required: false, contract_decision_required: false, and no open_contract_decisions remain, OR the handoff also carries captain-approved-design-bypass: true with a rationale. For affects_ui: true without that explicit bypass, BLOCKER (status: blocked, reason: ui design handoff skipped) and bounce to design/FO. For domain: set, design_required: true, contract_decision_required: true, or non-empty open_contract_decisions with a design-skipped stub, BLOCKER (status: blocked, reason: design-bearing decision skipped) and bounce to design/FO; planner must not choose the domain contract, selector grammar, API vocabulary, tool protocol, DSL, schema, or message format. Otherwise log ## Plan Imported Design DCs: design-skipped (no UI surface, no matched domain, no design_required signal, no contract decision required, or captain-approved bypass) and proceed.
Block has design-emitted fields (design_constraints[] / render_fidelity_targets[] / whole_page_visual_targets[]) → run mechanical mapping below.
Block ABSENT entirely → BLOCKER (status: blocked, reason: hand-off-to-plan absent — neither design-skipped stub nor design output found). Either shape Phase 8 missed the stub emit OR design errored without writing hand-off. Do not silently treat as "no UI" — that's the ambiguity G14 fixes.

Read via bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/extract-section.sh" <entity-file> hand-off-to-plan (handles folder + flat layouts).

Mechanical mapping delegated to lib script:

bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/import-design-dcs.sh" <entity-folder> >> <entity-folder>/plan.md

C15 accounting: body-line counts exclude content inside balanced standalone <details>...</details> blocks; the plan cap and its 2× raw-total backstop remain unchanged.

Pre-import validation (run in this order):

bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/check-design-readiness-review.sh" <entity-folder>/design.md — risk-gated design readiness check. If it reports status=blocked, stop before planning. If it reports status=warn, continue only after preserving the warning in ## Plan Report.
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-handoff-schema.sh" <entity-folder> — structural check
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-d-references.sh" <entity-folder> — D{N} backref consistency
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/import-design-dcs.sh" <entity-folder> — count-preserving import; if it reports imported design_constraints count mismatch, BLOCK instead of hand-copying missing rows.

Mechanical mapping (each item becomes a DC anchored to a wave):

Hand-off field	Becomes	Wave
`design_constraints[]` (UI or domain contract types)	`ui`, `e2e`, `contract`, or `api/schema` typed DC per constraint; UI constraints use `ui-verify`, domain constraints use contract/router/schema tests	task wave touching the affected component
`visible_surface_map[]` (captain-visible regions, controls, state indicators, semantic badges)	structural/mockup-parity typed DC seeds per mapped row; explicit N/A rows are imported as visible audit exclusions, not dropped	wave touching the affected route or verify helper
`render_fidelity_targets[]` (computed-style / structural assertions)	`ui`-typed DC, `verified_by: ui-verify` (calls computed-style assertion)	W0 (token-level) or task wave (component-level)
`whole_page_visual_targets[]` (route-level composition parity)	`ui`/`e2e` typed whole-page visual parity DC; verifier captures full-page screenshot and compares against `reference_artifact`	integration wave or affected route wave
`artifact_paths[]`	reference in plan `## Design Source` block; checksum on read	—
`ui_quality_contract`	import each group from `plugins/ship-flow/references/ui-quality-contract.md` into DCs, reviewer questions, or explicit N/A rows. Required groups: `copy`, `visual_hierarchy`, `color`, `typography`, `spacing`, `interaction_states`, and `source_safety`.	task wave touching the affected UI/component contract
`open_decisions[]` non-empty	BLOCKER (status: blocked, reason: design has open decisions) — bounce to design via SendMessage(designer@pitch-XX)	—

Each imported DC retains rationale_decision: D{N} cross-reference back to ship-design Phase 8 ## Captain Decisions for audit trail.

Output: ## Plan Imported Design DCs section in plan.md listing each imported DC with source field + wave assignment + rationale_decision link.

Step 1.7 — Architecture-lens dispatch (entity 110, 2026-04-29)

After spec read + design DC import, check whether the spec touches any known cross-cutting domains and dispatch read-only architecture lens agents before research.

Science Officer (EM) stewardship for plan-stage assignments

Before dispatching any architecture lens, research producer, research reviewer, plan cross-reviewer, or gap-fill worker, render and include the shared worker-facing stewardship section:

bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/render-science-officer-em-stewardship-contract.sh"

Trigger matching — read plugins/ship-flow/references/architecture-lens-triggers.yaml. For each domain entry:

File-glob match: check spec ### Artifacts likely touched paths against trigger_patterns (bash glob, case-insensitive). Match if any path matches any pattern.
Spec-keyword match: grep spec body (full text) for each spec_keywords entry (case-insensitive). Match if any keyword found.
OR-semantic: domain triggers if EITHER file-glob OR spec-keyword matches.

Dispatch — for each matching domain, dispatch Skill(lens_skill) with prompt:

Read spec at {spec_path}. Read {domain_knowledge_refs}. For each cross_cutting_concern listed in your frontmatter, output a structured YAML verdict (FLAG/PASS/SKIP with missing_dc if FLAG). ≤300 words total.

All matching lenses dispatch in parallel (no budget cap for v1). Collect structured YAML verdicts.

If no domains match → emit Lens dispatched: none (no trigger match) in Context Manifest. Proceed normally.

Lens FLAG integration — after all verdicts collected, for each FLAG verdict:

Option A: add DC covering the flagged concern to the plan (preferred)
Option B: write entry in ## Lens Findings: deferred section with explicit rationale per skipped FLAG

Step 2 — Research (size-adaptive, produce+review)

S — skip research. Proceed to Step 3.
M — dispatch Agent A (sonnet producer: codebase impact + lib constraints, <400 words, file:line citations; include recent debrief warnings from docs/<wf>/_debriefs/ — read ## Issues — Workflow / ## Filed (backlog) sections, schema at plugins/ship-flow/references/debrief-schema.yaml) → Agent B (sonnet reviewer: APPROVED / GAPS / CONTRADICTION, <300 words). Max 1 gap-fill round.
L — dispatch 3-5 parallel producers (upstream constraints / existing patterns / lib surface / gotchas / reference examples), each <200 words, file:line citations; one producer dedicated to recent debrief warnings (docs/<wf>/_debriefs/ last 3-5 files, extract ## Issues — Workflow / ## Filed (backlog) / ## Observations). Reviewer runs cross-domain + coverage check. Max 1 gap-fill round.

Contradiction: write both verbatim as Open Question in ## Research Summary. Never silently resolve.

Step 2.5 — Size re-evaluation

Count actual affected files from research vs spec's size estimate.

Sharp	Actual	Action
S	≤3	Confirmed S
S	4-10	Upgrade to M
S	>10	Upgrade to L (run full research)
M	≤3	Downgrade to S
M	4-15	Confirmed M
M	>15	Upgrade to L
L	any	Confirmed L

Update entity frontmatter size: if changed.

Fold/extract LOC heuristic (harness-diet restructures; n=2 calibration):

Fold (A → B, A retired): fold_net_loc ≈ source_loc × 0.44-0.55. DC-1 budget worst-case = target_initial + source × 0.55. Include procedural skeleton + example tables + guidance templates (not just the Steps-1-to-N).
Extract (1 source → N callers with refs): extraction_net_loc ≈ -(source_loc × 0.55-0.70) where source = summed preamble LOC across callers.
Picking: ≤50 LOC + one parent → fold. ≥100 LOC + N≥3 callers → extract. 50-100 LOC + N=2-3 → fold into most natural parent. Extraction adds 1 skill (Principle 2 cap check first).

Next.js standalone artifact size heuristic (DC budgets on ui-dist/ or .next/standalone/):

Measure ui/.next/standalone/ SUBTREE, NOT full ui/.next/. Full .next/server/ includes all compiled chunks + source maps (~40MB+); .next/standalone/ contains only traced runtime deps + subset of maps.
Plan estimates based on full ui/.next/ over-predict .map strip reduction by 4-10×.
Ground truth methodology: run target bundler locally, then du -sh ui/.next/standalone/ + find ui/.next/standalone -name "*.map" | wc -l before locking size targets.

Step 2.7 — Scope anchoring (M/L only)

Every task maps to a spec Scope In bullet (if shape ran) OR a Done Criteria item. Produce task×mapping table. Unmapped task → drop or flag in ## Plan Report. Do NOT silently expand scope.

Step 3 — Write plan (delegate to superpowers:writing-plans + ship-flow TDD fallback)

Layer B wrap (ship-plan owns these on top of Layer A output):

Runtime detection — invoke ship-flow:ship-runtime-detect before tasks get written; propagate {commands.test/build/typecheck/lint/dev} into every task's Done field.

Per-task skills_needed derivation (#108.1): after task files are known, populate each task with a skills_needed: [...] string array. Derive it by intersecting:

task files_modified / **Files:** globs
framework_detected, theme_indirection, and design_canonical_dir from ship-flow:ship-runtime-detect
density-classified skill set already loaded for answers_density: high
domain registry routing from registry-resolve.sh --domain=<domain> or --classify <spec>: preserve required_skills and merge skill_hints.plan into plan-stage task skills_needed
adopter file-signal routing from .claude/ship-flow/skill-routing.yaml when present. For each implementation task, run bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/resolve-skill-routing.sh" --files=<task-files> --config=.claude/ship-flow/skill-routing.yaml and merge the emitted skills_needed= list. Also record folder_guidance_files= and folder_guidance_skills= in ## Context Manifest → Folder guidance, and merge guidance skills into task skills_needed when they are not already present. If the config is absent on a non-trivial multi-surface pitch, run bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/discover-adopter-skills.sh" --root=. and surface the draft in ## Context Manifest before finalizing skills_needed.

Use concrete file-glob mapping, then trim to the smallest relevant list:

File signal	skills_needed candidates
`.tsx`, `.jsx`, `components/`, `app/`, `ui/**`	`frontend-design`, `react-best-practices`, `test`
`.css`, `tokens.css`, `design-system.md`, `design/*`	`frontend-design`, `web-design-guidelines`, `accessibility`
`apps/supabase/migrations/`, `domains//src/schema/**`	`project-db`, `test`
`domains//src/domain//{types,decider,view,saga}.ts`, `apps/deno-api/src/middlewares/fmodel-middleware.ts`	`fmodel`, `test`
`.test.`, `.spec.`, `__tests__/**`	`test`, `tdd` or `test-driven-development`
`.sh`, `bin/`, `lib//.sh`	`test`, `best-practices`
`.md`, `SKILL.md`, `docs/*`	`write-docs` when prose is user-facing; omit for stage artifacts

Adopter gate registry: for every implementation task with non-empty owned_paths, run python3 "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/resolve-gate-registry.py" --config .claude/ship-flow/gates.yaml --files <comma-separated-owned_paths> when the adopter config exists. Merge required_gates, reviewer_questions, and evidence_required into the task's reviewer questions and domain_acceptance_checklist; record matched_routes and layers in the Context Manifest. Missing config is a WARNING for legacy/single-surface plans and a plan gap for new multi-layer adopter plans. These gates are adopter-owned; ship-flow core only resolves and carries them.
TDD contract: every non-exempt implementation task MUST include tdd_contract with red_command, expected_red_failure, green_command, and refactor_check. In prose task blocks, include the same data as RED command, Expected RED failure, GREEN command, and REFACTOR check. Execute consumes this contract for RED-before-GREEN evidence; verify audits it before PASS.
TDD ledger gate: after drafting plan.md and before ### Hand-off to Execute, run python3 "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-tdd-ledger.py" --plan <entity-folder>/plan.md > <entity-folder>/tdd-ledger.txt 2>&1 for the human-readable pass/fail summary, then run python3 "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-tdd-ledger.py" --plan <entity-folder>/plan.md --emit-jsonl > <entity-folder>/tdd-ledger.jsonl. Finally run python3 "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/validate-tdd-ledger.py" --plan <entity-folder>/plan.md --require-ledger-jsonl <entity-folder>/tdd-ledger.jsonl to prove the persisted ledger exists and matches the current plan. Failure blocks plan advance. This makes the tdd_contract authoritative at plan time rather than post-hoc extraction. The validator rejects missing contracts, low-confidence red_command / green_command, invalid layer, stale ledger JSONL, and code-bearing tasks that declare layer: meta when the inferred layer is L1-L7.
TDD exceptions (mark inline): config / pure refactor with coverage / docs-only / migration — TDD: skip -- <reason>.
Wave rules: wave 0 = test infra; wave N+1 depends only on ≤N outputs; no files_modified overlap within a wave; no cycle.
UI quality contract import: when ### Hand-off to Plan contains ui_quality_contract, read plugins/ship-flow/references/ui-quality-contract.md and turn every group (copy, visual_hierarchy, color, typography, spacing, interaction_states, source_safety) into DCs, reviewer questions, or explicit N/A rows. Missing coverage for any group is a plan gap because verify needs concrete evidence prompts.
Parallel metadata contract: every implementation task MUST include parallel_group, depends_on, owned_paths, and integration_owner. Use parallel_group: serial when the task cannot run with peers. owned_paths is the write boundary execute uses to prove disjointness; depends_on is T{N} task-ID topology; integration_owner is the named teammate that folds the worker output into the stage artifact (normally executer@pitch-XX).
Emit plan-parallelization-manifest in ### Hand-off to Execute: one row per task with task_id (T{N}), parallel_group, depends_on, owned_paths, and integration_owner. task_id is the canonical task identifier used throughout the plan, and every depends_on entry MUST reference these same T{N} task IDs. This is contract metadata, not an execution command. Execute derives execute-dispatch-manifest from it and may only parallelize tasks whose dependencies are satisfied and whose owned_paths do not overlap.
If design emitted domain expert panel constraints, convert them into a domain-specific acceptance checklist before task writing. Each checklist item must name the domain lens, required skill/knowledge module, affected path family, and the verification command or reviewer prompt verify will reuse.
Every task with domain or framework skills_needed MUST include reviewer_questions. Convert skills_needed into reviewer questions, not just labels. For each task skill that represents a domain lens (project-db, project-seed, fmodel, refine-expert, refine-gotchas, API/router skills, Expo/mobile skills, etc.), add 1-3 concrete checks to the domain-specific acceptance checklist. Example shape: "refine-gotchas: verify list/detail cache invalidates after mutation; URL/drawer state hydrates on back/forward/shared link; ProCRUD/list hooks follow existing app pattern."
Emit domain_acceptance_checklist in ### Hand-off to Execute with columns Task ID, Verify Lens, Reviewer Question, Affected Path Family, Required Skills, and Evidence Required. This is the plan-to-verify contract consumed by ship-verify when it populates verify-check-manifest.review_lenses.
When registry or domain routing is present, emit exactly one standalone context-routing-manifest section-tagged block and a ## Context Routing Receipt. The receipt maps manifest rows to tasks[].skills_needed, task reviewer_questions, and domain_acceptance_checklist; optional rows may be skipped only with an explicit rationale. Future provider hints are append-only and non-authoritative: the local registry remains the routing source of truth for this block.
T0.X auto-indirection-sweep (T6.1, #106): when theme_indirection from ship-runtime-detect Step R5 is non-empty (e.g. tailwind-v4), auto-emit a Wave 0 task in the plan: T0.X: Audit @theme inline indirection layer — verify design tokens align with CSS custom properties, no hardcoded hex values in component files. REFUSE to emit a plan without this task when theme_indirection != "". Enforcement: bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/bin/check-invariants.sh" --check indirection-sweep-emitted (fixture-based).
Lens FLAG integration (entity 110, 2026-04-29): after Step 1.7 lens verdicts are collected, plan worker MUST for each FLAG verdict: (a) add a DC covering the flagged concern, or (b) write an entry in ## Lens Findings: deferred with explicit rationale. Gate refuses advance if any FLAG exists without Option A or Option B. Silence is NOT acceptable — the deferred section exists precisely to make punts explicit and captain-visible.

Step 3.5 — Verification spec (structural parity enforcement)

Validation task design discipline — when a plan task description is "validate X", classify X before authoring:

X is output-shape (code pattern, file presence, schema match, runtime assertion) → write as DC in ## Verification Spec, consumed by verify stage. Predicate-grep-able, evidence accumulates across N naturally.
X is process-shape (did the worker load skill, did the worker read section, did the worker echo attestation token, did the canary appear in the receipt) → do not author. Per-dispatch process-effect validation has structurally bad cost-to-signal ratio (baseline contamination + scope mismatch + timing pollution; see plugins/ship-flow/references/validation-discipline.md). Either rewrite as the equivalent output-shape DC, or drop the task.

Per typed Done Criterion, fill exact verification procedure:

Type	Procedure	Fallback
`cli`	bash command + expected exit/output	always available
`api`	`curl` with method/URL/body + status + response pattern	always available
`ui`	`curl -sfN <route>` + `grep` content; complex interaction → `e2e-test <flow.yaml>`	curl+grep skip interaction
`design-system-parity`	Runtime computed-style evidence compared to a design-system token table or explicit token contract. Required plan evidence fields: `selector`, `css_property`, `token_source`, `expected_token`, `expected_resolved_value`, `actual_computed_value`, `runtime_report`, `verdict`.	`ship-flow:ui-verify` for fixed selector/value checks; inline browser `getComputedStyle()` probe only as break-glass
`mockup-parity`	Live DOM structure evidence compared to an HTML mockup or committed design artifact. Required plan evidence fields: `mockup_artifact`, `route`, `root_selector`, `expected_structure`, `actual_dom_structure`, `comparison_method`, `runtime_report`, `verdict`.	deterministic selector hierarchy, role/name sequence, landmark order, repeated-item counts, or normalized DOM digest
`skill`	`Skill("<name>")` invoke + expected output shape	session-only verification
`e2e`	`e2e-test <flow.yaml>` with step assertions	degrade to `ui` type

Every DC MUST have a runnable verify procedure. "Manual check" = plan failure.

Note: ui type uses -sfN (Next.js 16 Turbopack SSR chunks — MEMORY #073). Backward-compatible.

Static CSS / design tokens / computed-style regression → .claude/e2e/ui-verify/<entity-slug>.yaml (schema: ship-flow:ui-verify). Wraps getComputedStyle(selector)[prop] equality checks against tokens.
Dynamic DOM / navigation / event / interaction → .claude/e2e/flows/<entity-slug>.yaml (schema: e2e-pipeline:e2e-flow). Step-based browser flow with assert primitives.
Both require .claude/e2e/mappings/<mapping>.yaml (auth + base_url + test_accounts). If missing → plan's first UI task is mapping bootstrap.
Plan-stage DC: YAML file exists at canonical path AND imports valid mapping. Consumer: ship-verify Step 4.4 auto-runs matching skill.
Fallback when skill wrapper unavailable / mapping missing: inline agent-browser CLI script as a verify-time DC (declarative YAML preferred; CLI only as break-glass).

Step 4 — Self-review (plan-checker-lite)

Run 9 dimensions. Any BLOCKER → fix inline + re-review. Max 3 iterations.

Requirement coverage — every DC maps to ≥1 task.
Task completeness — every task has paths / verify command / model hint / wave.
- For #108.1+, every implementation task must include skills_needed: [...]; docs-only stage-artifact tasks may use skills_needed: [] only with TDD: skip -- docs-only/stage-artifact.
Dependency correctness — wave graph safe (no cycles / no same-wave files_modified overlap / no read_first pulling same-wave outputs).
Zero-placeholder scan — grep TBD | add appropriate | similar to Task N | as needed | fill in | \.\.\. → fix inline.
Type/signature consistency — cross-task function/type signatures match.
Task minimality (M/L) — merge adjacent non-load-bearing tasks; drop nice-to-haves not in DCs.
TDD compliance — test-first order on code tasks; TDD: skip tasks have valid reason.
Stale-line-anchor — every file:line citation re-read; content-matches / line-shifted / contradicted (BLOCKER) / phantom path (BLOCKER). Catches #1 cause of execute BLOCKED returns.
Design reference compliance — skip if no ## Design Reference section; else cross-check visual attrs (fill/stroke/colors/dimensions) against cited spec files. Flag deviations.
Stub-captain-ack scan (T6.2, #106): grep every task body for keywords stub|fake|placeholder|v1.*only|wired only for. For each match: check entity frontmatter pre-acked-stubs: true OR check that the Plan Report has a ## Stub Flag entry for this task with explicit captain rationale. If neither present → BLOCK: stub task without captain ack (literal string required for test DC). Populate ## Plan Report → Stub Flags table with all stub tasks found. Cross-review PROCEED blocked until all stubs either pre-acked or cleared.
Context Manifest completeness (entity 110, 2026-04-29): ## Context Manifest section present and all 7 fields non-empty: Skills loaded, INVARIANTS sections read, Architecture docs consulted, Domains touched, Lens dispatched, Lens findings integrated, Folder guidance. Lens dispatched field must reflect actual Step 1.7 trigger matching results (not copy-pasted from a prior entity). Folder guidance must cite every non-root folder_guidance_files= entry from resolve-skill-routing.sh, list folder_guidance_skills=, and include the literal codex_context_boundary line so reviewers can see this is not duplicating Codex root instruction loading. C8 check: bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/bin/check-invariants.sh" --check context-manifest-emitted.
skills_needed non-boilerplate (#108.1): for plans with ≥2 implementation tasks touching different file classes, grep the task blocks and confirm there are ≥2 distinct skills_needed lists. Empty lists or one repeated list across heterogeneous tasks are BLOCKERs.

Step 5 — Cross-review gate (Principle 6 Rule C)

Dispatch cross-review to executer teammate (pipeline path) or fresh sonnet (no team). Upgrade to fresh opus when appetite: big-batch.

7-factor rubric adapted for plan stage from INVARIANTS Principle 6 Rule C #106 T1.3 plus the ship-plan skill-coverage extension:

Feasibility — tasks achievable by single agent in one dispatch each?
Executable scope — tasks are atomic commits aligned 1:1 with waves?
Quality — Verification Spec covers every DC with runnable procedure (≥1 structural-parity DC for UI)?
DC adequacy — observable DCs per task; no "works correctly" / "handles all cases"?
Canonical sync — ARCHITECTURE.md touches planned? architecture-impact draft per affected child?
Reverse-audit previous stage — does the plan's scope expose a gap in the preceding design stage's ### Hand-off to Plan block? Specifically: are all render_fidelity_targets from design encoded as DCs? Are design_constraints honored in plan tasks?
Skill Coverage (skill-coverage, #108.2) — every implementation task has non-empty skills_needed and the skills match its files_modified / **Files:** file globs. Emit exactly one grep-testable summary line when clean: skill-coverage: PASS. Emit one line per failing task when not clean: skill-coverage: FAIL — task <id>: <reason>. Failure reasons include empty skills_needed, boilerplate repeated lists across heterogeneous implementation tasks, or file-glob/skill mismatch. Use these required matches:
- *.tsx, *.jsx, components/**, app/**, ui/** → react, frontend-design, or react-best-practices
- *.css, tokens.css, design-system.md, design/** → frontend-design, web-design-guidelines, or accessibility
- apps/supabase/migrations/**, domains/**/src/schema/** → project-db
- domains/**/src/domain/**/{types,decider,view,saga}.ts, apps/deno-api/src/middlewares/fmodel-middleware.ts → fmodel
- *.test.*, *.spec.*, __tests__/** → test, tdd, or test-driven-development
- *.sh, bin/**, lib/**/*.sh → test or best-practices
- Docs-only stage-artifact tasks may use skills_needed: [] only with TDD: skip -- docs-only/stage-artifact; user-facing prose docs should use write-docs.

Reverse-audit prompt template (T3.2 — paste verbatim into reviewer dispatch):

Reverse-audit: Read the entity's `### Hand-off to Plan` block.
(a) List every `render_fidelity_target` — does plan.md have a DC for each? (BLOCKING if any missing)
(b) List every `design_constraint` — does each appear in at least one plan task or Verification Spec row? (WARNING if any absent)
(c) Are `open_decisions` empty? If not empty, is there a plan task that resolves each? (PROMPT_CAPTAIN if unresolved)
Coaching note: render_fidelity gaps here become silent UI regressions at verify — enforces FM#4 (fidelity gap) prevention.

Step 6 — Emit plan.md

Verbosity budget (INVARIANTS Principle 8 — plan.md ≤200 body lines; C15 BLOCKER): the plan body is tabular (waves + tasks + DCs). Keep it lean:

## Verification Spec — key every row by DC-N (### Done Criteria in shape.md is canonical for assertion + type; cite by DC-N, do NOT restate the assertion/type columns — 129.1 CD-2). Only Verify Procedure + Expected are the genuinely-new columns this stage produces.
Do NOT restate the shape.md Problem / User Journey / DC prose — cite via extract-section.sh, per 129.1 reference_not_restate.
Verbose research → ## Research Summary stays a findings summary; a bounded excerpt of producer/reviewer transcripts may go in <details> (excluded from the ≤200 BODY cap, but the C15 2× raw-total backstop — raw ≤ 400 — still applies, so excerpt-not-dump). Large transcripts → reference a DURABLE artifact (committed-in-repo via explicit pathspec, OR a durable URL — a bare local path is not audit evidence), don't inline.
Do NOT enumerate commit SHAs or git-reconstructable file lists in the body — git log / git diff has them.

Require a ### Metrics subsection in ## Plan Report. Use grep-friendly key: value lines:

status: passed | gaps-noted | blocked | partial
duration_minutes: wall-clock minutes for planning
iteration_count: self-review + reviewer loop count
task_count: implementation task count
verification_spec_count: DC rows in Verification Spec
model_split: dispatch model breakdown

## Context Manifest section format (mandatory output, entity 110, 2026-04-29):

## Context Manifest

- **Skills loaded**: [comma-separated skills invoked]
- **INVARIANTS sections read**: [section names with file:line citations]
- **Architecture docs consulted**: [docs read, with paths]
- **Domains touched**: [domain names from trigger registry that matched, or "none"]
- **Lens dispatched**: [list: domain (verdict summary) or "none (no trigger match)" or "skipped (trivial escape hatch)"]
- **Lens findings integrated**: [N integrated, M deferred-with-rationale, K ignored — must be 0 ignored unless all flags are deferred]
- **Folder guidance**: [for each task group: `files=<paths>` → `folder_guidance_files=<non-root AGENTS.md/CLAUDE.md>`; `folder_guidance_skills=<skills>`; `codex_context_boundary=root AGENTS.md/CLAUDE.md intentionally excluded from folder_guidance_files`]

Mark TaskCreate sub-task emit-plan.md completed; return to /ship for advance to execute.

Step 6.1 — Register plan completion

After stage artifact lands, advance sibling index.md frontmatter atomically:

INDEX_MD="<entity-folder>/index.md"
H="$(if command -v sha256sum >/dev/null 2>&1; then sha256sum "$INDEX_MD" | awk '{print $1}'; else shasum -a 256 "$INDEX_MD" | awk '{print $1}'; fi)"
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/advance-stage.sh" \
  --entity="$INDEX_MD" \
  --new-status=plan \
  --stage-name=plan \
  --stage-file=plan.md \
  --if-hash="$H" \
  --lease-file="$SHIP_FLOW_COMPLETION_LEASE_FILE" --lease-token="$SHIP_FLOW_COMPLETION_LEASE_TOKEN" --worker-id="$SHIP_FLOW_COMPLETION_WORKER_ID" \
  --commit-as="plan(<id>): advance status to plan"

On exit 6 (stale hash): write ## Plan Report status: blocked, reason: index.md stale hash; parallel session contaminated and return.

Post-ship Retro Prompt (entity 110 Captain Bet — SC-810-class prevention)

After this feature ships, re-read the ## Captain Bet block in the entity body. At the 5-carlove-pitch mark, evaluate outcome:

Bet substance: "下 5 個 carlove pitches 不再有 SC-810-class cascade-saga gap."

Retro procedure:

Review the last 5 carlove pitches that touched tag/customer/event-saga domains
For each pitch where a cascade-saga gap still occurred, categorize per 4 bet-failure paths:
- Path 1 — Domain not in lens registry (trigger didn't match) → add lens for that domain
- Path 2 — Lens fired but plan worker ignored finding → strengthen W3 gate from gate-with-explicit-deferral to hard-block-no-deferral
- Path 3 — Manifest section missing key dimension → add new manifest field
- Path 4 — Lens fired + plan integrated, but failure still occurred → format-level revisit (meta-level re-shape)
If ≥3 consecutive bets fail → freeze lens registry + trigger re-shape of this pitch

Kill criterion: 3 consecutive Bet ≠ Outcome → freeze lens registry + re-shape. Single fail → apply specific improvement per 4 paths above.

Invariants + red flags (STOP if violated)

Every DC has a runnable verify procedure. "Manual check" = plan failure.
UI entity without ≥1 structural-parity DC = silent visual-bug risk.
Contradicted sharp assumption (Step 1.5) → BLOCKER, do not plan on stale premises.
Wave graph violates safety (cycle / overlap / cross-read) → BLOCKER.
Placeholder prose (TBD / "similar to Task N" / ...) = plan failure; fix inline.
Cross-review VETO loop capped at 2 rounds; round 3 → PROMPT_CAPTAIN.
Layer A delegation (superpowers:writing-plans) owns plan authoring — re-teaching = Principle 6 Rule B violation.
Scope expansion beyond spec Scope In / DCs = Shape Up violation; drop or flag.

Circuit breakers

Research producer timeout: 2 min/agent.
Research reviewer: max 1 gap-fill round.
Self-review loop: max 3 iterations → proceed with gaps noted in Plan Report.
Cross-review reviewer: max 1 round per VETO; 2 VETOs total before PROMPT_CAPTAIN.
Total stage >20 min elapsed → write plan.md with partial content + ⚠️ INCOMPLETE markers + Plan Report status=partial, then return. Never exit without emitting plan.md.

Step 6 (Hand-off): Emit Hand-off to Execute + Read Incoming Hand-off

Read incoming: at Step 1, read ### Hand-off to Plan from entity body. Verify design_constraints honored in plan tasks; encode render_fidelity_targets as DCs.

Emit ### Hand-off to Execute after plan.md is written:

tdd-ledger: path to tdd-ledger.jsonl, validator command, pass/fail summary from validate-tdd-ledger.py --plan, and persisted-ledger proof from validate-tdd-ledger.py --plan <plan.md> --require-ledger-jsonl <tdd-ledger.jsonl>; do not hand off a new non-trivial plan whose ledger validation fails.
wave_order: exact wave dispatch order (e.g., "W0 → W1 (T1.1→T1.2→T1.3→T1.4) → W2a → W2b → W2c → W3 → W4")
critical_assumptions: assumptions to re-verify at execute boot (from Assumption Re-validation section)
architecture_context: canonical doc touches required (INVARIANTS, README, schema) — drives execute commit order
stub_flags: tasks containing stub|fake|placeholder|v1.*only|wired only for — must appear in Plan Report as captain-ack flags
skills_needed_summary: per-task skills_needed lists and note whether ≥2 distinct lists were produced for heterogeneous task sets
domain_acceptance_checklist: rows with Task ID, Verify Lens, Reviewer Question, Affected Path Family, Required Skills, and Evidence Required derived from task reviewer_questions
context-routing-manifest: exactly one standalone section-tagged manifest block plus ## Context Routing Receipt, with domain_matches, knowledge_modules, required_skills, stage_hints, consumer_obligations, and future_provider_boundary

References

Entity schema: plugins/ship-flow/references/entity-body-schema.yaml → stages.plan.
Stage writer: plugins/ship-flow/lib/write-stage-artifact.sh.
Section extraction: plugins/ship-flow/lib/extract-section.sh.
Layer A: superpowers:writing-plans (plan authoring discipline).
Utility: ship-flow:ship-runtime-detect (13-ecosystem runtime detection).
Reverse-recovery audit mod (every task creating a new file/domain/route needs a confirmed-MISSING classification line; reviewers reject greenfield tasks without it): plugins/ship-flow/_mods/reverse-recovery-audit.md (plugin-canonical); adopter override docs/ship-flow/_mods/reverse-recovery-audit.md when present.
Principle 6: plugins/ship-flow/INVARIANTS.md (context continuity + 3-layer architecture + cross-review).
MEMORY: #5 (--next-id atomicity), #14/#25 (explicit pathspec / staging), #35 (dispatch discipline amended by Principle 6), #073 (Next.js 16 -sfN), opus-4.7-naturally-does (2026-04-23 harness diet).

ship-plan

More from this repository

More from this repository

Ship-Plan — PLAN Stage (2.0)

Boot Self-Check

Entity body contract (schema-as-prose)

Layer A delegation (Principle 6 Rule B)

Flow

Step 1 — Read spec + cross-entity contracts

Step 1.1 — Canonical context planning

Step 1.5 — Assumption re-validation

Step 1.6 — Import Design DCs from Hand-off (G10, 2026-04-29)

Step 1.7 — Architecture-lens dispatch (entity 110, 2026-04-29)

Science Officer (EM) stewardship for plan-stage assignments

Step 2 — Research (size-adaptive, produce+review)

Step 2.5 — Size re-evaluation

Step 2.7 — Scope anchoring (M/L only)

Step 3 — Write plan (delegate to superpowers:writing-plans + ship-flow TDD fallback)

Step 3.5 — Verification spec (structural parity enforcement)

Step 4 — Self-review (plan-checker-lite)

Step 5 — Cross-review gate (Principle 6 Rule C)

Step 6 — Emit plan.md

Step 6.1 — Register plan completion

Post-ship Retro Prompt (entity 110 Captain Bet — SC-810-class prevention)

Invariants + red flags (STOP if violated)

Circuit breakers

Step 6 (Hand-off): Emit Hand-off to Execute + Read Incoming Hand-off

References

Ship-Plan — PLAN Stage (2.0)

Boot Self-Check

Entity body contract (schema-as-prose)

Layer A delegation (Principle 6 Rule B)

Flow

Step 1 — Read spec + cross-entity contracts

Step 1.1 — Canonical context planning

Step 1.5 — Assumption re-validation

Step 1.6 — Import Design DCs from Hand-off (G10, 2026-04-29)

Step 1.7 — Architecture-lens dispatch (entity 110, 2026-04-29)

Science Officer (EM) stewardship for plan-stage assignments

Step 2 — Research (size-adaptive, produce+review)

Step 2.5 — Size re-evaluation

Step 2.7 — Scope anchoring (M/L only)

Step 3 — Write plan (delegate to superpowers:writing-plans + ship-flow TDD fallback)

Step 3.5 — Verification spec (structural parity enforcement)

Step 4 — Self-review (plan-checker-lite)

Step 5 — Cross-review gate (Principle 6 Rule C)

Step 6 — Emit plan.md

Step 6.1 — Register plan completion

Post-ship Retro Prompt (entity 110 Captain Bet — SC-810-class prevention)

Invariants + red flags (STOP if violated)

Circuit breakers

Step 6 (Hand-off): Emit Hand-off to Execute + Read Incoming Hand-off

References