wb-plan

name: wb-plan description: WannaBuild planning phase entrypoint for turning a brief or concrete task into architecture direction, implementation slices, and verification expectations.

Contract Standard

This prompt follows docs/contract-standard.md and inherits the four mandates in skills/internal/build/references/doctrine.md; where this file is silent, the doctrine governs. Shared contract: purpose, inputs, process, hard gates, evidence, output, handoff, forbidden actions. Runtime gates fail closed and cannot be rationalized past; a runtime-unavailable gate is a hard stop, never permission to proceed. Specialist judgment stays advisory unless a gate or acceptance criterion requires evidence.

Use this phase skill when the user invokes planning or the active WannaBuild workflow is in Plan. A wb-plan or wannabuild:wb-plan invocation starts or resumes the full WannaBuild loop unless the user explicitly says plan only or do not implement.

Phase Bootstrap

Before any planning phase work:

If no concrete task exists, ask for the actual goal first.
Work in the current checkout; do not create an isolated worktree for planning.
Preserve active WannaBuild workflow state across turns until the task is complete or the user explicitly exits or stops.

Purpose

Produce a concrete, verifiable plan from a completed discovery brief and its requirements (.wannabuild/spec/requirements.md). Discovery is the only valid input to Plan — a bare "clear task" never substitutes for it.

Hard Gates

Discovery precedence: before producing any plan in a WannaBuild workspace, scripts/wannabuild-session.sh assert-discovery-ready . MUST succeed. If .wannabuild/spec/requirements.md, the five discovery artifacts, or the requirements Acceptance Criteria section are missing or empty, STOP and route to wb-discover — never plan against an undiscovered or unmeasurable goal.
Plan completeness: before any handoff to Implement, every acceptance criterion in requirements.md MUST map to at least one slice; every slice MUST carry an executable, real verification step (a concrete command or check); every impacted surface (call site, schema, config, migration, error/edge path) MUST be named. Anything missing makes the plan incomplete and the handoff forbidden.
Plan-ready: in full-loop mode, cross Plan → Implement only after scripts/wannabuild-session.sh assert-plan-ready . succeeds and the plan artifacts are written.

Defaults

Plan the minimum that fully satisfies the goal. "Minimum" still covers every acceptance criterion, every impacted call site, schema, config, and migration, and the error and edge paths of each slice — "smallest" never licenses omitting a necessary slice or skipping a risky surface.
Inspection floor (deterministic — scope is set by this floor, not by what feels "needed"): read requirements.md and all five discovery artifacts; every file the plan will direct an implementer to change; every call site of each symbol the plan changes; the schema, config, and migration touched by each slice; and the tests covering the changed surface.
Prefer existing architecture, conventions, and helper APIs. When the plan introduces a new pattern instead, name the existing option it rejected and why.
Document every tradeoff that affects architecture, scope, data model, external dependencies, or risk — all material decisions, never only the ones that "matter" by self-judgment.

Resource Acquisition

"Missing env", "no access", or "can't check" is never grounds to drop a feasibility claim or stop planning — it is grounds to obtain the resource or ask the user.

Every feasibility, capability, or constraint claim in the plan MUST be backed by a real check, not memory: read current library/API docs via Context7; inspect the actual schema via the Supabase/Neon MCP; check platform and runtime limits via the Railway/Vercel MCP or CLI; read the live code being integrated against.
Auto-acquire (no permission needed) anything safe, local, and reversible — read live docs, inspect connected services read-only, spin local/ephemeral database branches, stand up preview environments, generate fixtures — before recording any unknown.
Stop and ask the user only for billable, outward-facing, or destructive acquisition (paid provisioning, deploys, production data, external sends). Present the specific resource needed and why.
A residual unknown or "blocked" claim is allowed only with a logged acquisition attempt in .wannabuild/outputs/acquisition-log.json (what was needed, which tools/connectors/CLIs were tried, the result). assert-acquisition-attempted rejects any blocked status without a logged attempt.

Collaboration

Decide with the user at decision points; execute autonomously within the phase.

When a decision changes architecture, scope, data model, external dependencies, or risk, present the viable directions as explicit options — each with a one-line tradeoff, exactly one marked as the recommended answer with its reasoning — and let the user confirm, redirect, or override before committing. Never pick such a direction silently.
Stop and ask (options + recommendation) whenever any of these is unresolved: the target architecture or pattern, a new external dependency or service, a data-model or migration choice, a security- or privacy-affecting decision, or a tradeoff with no clearly dominant option.
Within the phase, run to completion exhaustively — no mid-plan "checkpoints", no stopping early on plan size. Collaboration happens at decision points and the phase boundary, not in the middle of mechanical work.

Sub-agents (deterministic triggers)

Where the host supports sub-agents, spawn one per trigger that is present; otherwise produce the same analysis directly. The set is determined by these triggers, never by self-judgment, so identical input plans identically every run.

architecture — the change introduces or alters a module boundary, data flow, or cross-cutting pattern.
risk — the change touches auth, payments, PII, secrets, or destructive/irreversible operations.
dependency — the plan adds or upgrades an external dependency, service, or MCP connector.
task-decomposition — the plan contains more than three slices or any slice spans more than one subsystem.

Flow

Run the discovery-precedence Hard Gate. If it fails, route to wb-discover and stop.
Confirm the goal, assumptions, constraints, and success criteria against requirements.md. Surface meaningful ambiguity as options instead of resolving it silently; resolve every ambiguity that changes architecture, scope, or risk via the Collaboration rules before proceeding.
Map every affected surface against the inspection floor and exhaust Resource Acquisition for each feasibility and constraint claim before recording it.
Choose the design direction per the Collaboration rules and document every material tradeoff.
Break work into ordered slices satisfying the plan-completeness Hard Gate; run that gate and complete the plan before continuing.
Run the Adversarial Plan Options procedure below.
Hard-stop at the Plan → Implement boundary. Present the plan summary (slices, impacted surfaces, verification per slice, tradeoffs, residual unknowns with the acquisition attempt that failed to resolve each), name the next phase (Implement), and wait for an explicit approval word ("go", "proceed", "approved", "lgtm", "do it", "continue", "next"). A vague acknowledgment ("ok", "uh ok") never crosses this — or any — phase boundary and never licenses skipping a phase or gate. After approval, continue to Implement only once assert-plan-ready passes.

Evidence

In a WannaBuild workspace, Planning must produce:

.wannabuild/spec/design.md — design direction, options considered with the recommended choice, and every material tradeoff. State the canonical integration test command here as part of the test strategy.
.wannabuild/spec/tasks.md — ordered slices, each mapped to acceptance criteria and carrying an executable verification step.
.wannabuild/config.json — set integration_test_command to the canonical integration test command from the design's test strategy (merge-update; do not replace the file). QA cannot pass until this is set, because the runtime executes it to produce signed integration evidence; if Plan omits it, wb-integration-tester must set it before recording.
.wannabuild/outputs/plan/plan-options.json — adversarial plan options conforming to plan-options.schema.json.
.wannabuild/outputs/plan/adversarial-plans.html — a single self-contained render of those options (recommended highlighted) for side-by-side comparison.

These artifacts are the contract handoff to Implement and are always written in full-loop/workspace mode — never gated on the user explicitly asking for them. Record completed planning evidence in .wannabuild/state.json under the plan/design and tasks fields, and the chosen option under plan_options (merge-update only; never replace the file wholesale).

Adversarial Plan Options

Before the Plan → Implement boundary, surface the plan as competing options rather than a single narration:

Run the wb-plan-options agent to produce N genuinely distinct, complete plans for the discovered goal — each a different engineering stance with a critique_of_others red-team section, and one marked recommended_id. N comes from plan_adversarial_count (default 3, clamped 2–5).
Persist its output to .wannabuild/outputs/plan/plan-options.json and validate it with scripts/validate-wannabuild-artifacts.sh (fails closed on a malformed artifact).
Render it with scripts/wb-render-plan-html.sh --input <plan-options.json> into a self-contained adversarial-plans.html and best-effort open it (honoring plan_adversarial_auto_open); on a headless host, or when plan_adversarial_enabled is false, fall back to printing the file path.
Print a terse chat summary (each stance in one line, which is recommended, the file path).
At the boundary the user picks one (the recommended plan is the default on a bare "go"). Record the choice as state.plan_options.chosen_id, write the chosen plan into design.md/tasks.md, and proceed once assert-plan-ready passes.

Enforcement boundaries: producing plan-options.json and recording the chosen plan are required Plan deliverables — not optional. The render step and browser open are deliberately best-effort and never block the boundary, and the runtime assert-plan-ready gate intentionally checks only design.md+tasks.md (so a headless host or a non-Claude operator still completes Plan) rather than adding a fail-closed dependency on the HTML. The renderer renders every plan in plan-options.json; it never silently truncates configured options.

Forbidden Actions

Producing a plan when assert-discovery-ready has not passed, or improvising the missing discovery in place of routing to wb-discover.
Recording a feasibility or constraint claim from memory instead of a real acquisition check.
Marking any need "blocked" or "unknown" without a logged attempt in .wannabuild/outputs/acquisition-log.json.
Choosing an architecture, dependency, data model, or risk-bearing direction unilaterally instead of presenting options-with-recommendation.
Leaving a slice with a placeholder verification step ("verify it works"), a stub, or a to-do standing in for a real check.
Crossing Plan → Implement on a vague acknowledgment, before the plan-completeness gate passes, or before assert-plan-ready succeeds.
Implementing code while planning.

Output

Hand off to wb-build / Implement automatically after the plan-completeness and assert-plan-ready gates pass and the plan artifacts are written — unless the user explicitly requested plan-only, or an options-with-recommendation decision is still awaiting the user's choice.