wave-planner

name	wave-planner
description	Plan and execute large multi-step, multi-stack changes with contract-first waves, parallel worker ownership, user approval gates, ordered integration, and final verification/docs updates. Use only when the task spans multiple areas (for example frontend, backend, tests, errors, docs) or has dependency-heavy sequencing.

Wave Planner

Create and run a Wave-Driven Development plan for large, cross-stack tasks.

Objective

Produce:

• Vision and scope boundaries
• Decision Log
• Wave 0 contracts
• Wave 1..N parallel implementation waves
• Integration wave
• Final verification wave
• Documentation update step
• Ready-to-send worker prompts and execution checklist

Workflow

1. Decide whether to use this skill:

• Use only if work is large, multi-step, and touches multiple parts/stacks.
• If not large enough, do not use this skill.

2. Ask for Gate 1 user approval before planning:

• Send a short reason: "I recommend Wave method because . Do you agree?"
• Stop until user accepts or declines.

3. If accepted, draft the plan:

• Clarify done criteria and out-of-scope.
• Build Decision Log (naming, API shape, error model, logging, tests, style constraints).
• Define Wave 0 contracts first (types/interfaces/schemas, boundaries, API shapes, constants, flags/migrations).
• Because worker execution uses gpt-5.4-mini, make the plan unusually clear and concrete so smaller agents can execute reliably without guessing.
• Save the plan by default inside .planning/; create the folder if it does not exist.
• Use a clear feature-based filename for the plan.

4. Choose wave and agent count with balancing rules:

• Prefer 1-2 waves unless more waves are clearly needed to avoid conflicts or sequencing problems.
• Use the minimum number of waves that still avoids conflicts.
• Group independent tasks into the same wave.
• Move dependency-bound tasks to later waves.
• Keep each worker task medium-sized (not tiny and not overloaded).
• Ensure each worker has explicit file/module ownership.

5. Produce a one-line execution preview before full detail:

• One line per wave.
• One line per worker with task summary.

6. Produce the full plan with complete shared context:

• Include full vision and all waves so each worker sees global intent.
• Include worker-specific goals, files, constraints, and verification.
• Include explicit wave order for execution (Wave 0, then Wave 1..N, then Integration, then Final Verification, then Docs).
• Make each worker prompt highly specific: exact ownership, exact files/modules, exact handoff format, exact verification steps, and any important non-goals.
• By default, save the full plan to the plan file and do not dump the full plan in chat unless the user explicitly asks.

7. Ask for Gate 2 user approval before execution:

• "Plan is ready. Execute now?"
• Stop until user accepts or declines.

8. If accepted, execute with Codex swarm/subagents:

• Spawn one named Codex subagent per worker in the current wave.
• Use gpt-5.4-mini for those worker subagents.
• The main Codex agent controls the subagents, assigns ownership, monitors progress, validates handoffs, and decides when a wave is complete.
• Run subagents in parallel inside a wave, but execute waves sequentially.
• Wait for all subagents in the wave to finish and validate handoffs before starting the next wave.
• Close finished subagents after completion.

9. After all waves:

• Main agent reviews merged result and runs verification tests.
• Main agent updates existing docs if needed, or adds new docs for new features.
• Follow existing documentation format and conventions.
• Main agent updates the original plan file status to Completed or Blocked.
• Main agent adds final verification notes to the original plan file.
• If the project has a status-tracking document or progress board, update it during closeout so completed work and remaining open work stay accurate.

Global Rules

• Gating rule: planning and execution require explicit user approvals (Gate 1 and Gate 2).
• Ownership rule: workers edit only assigned files.
• Dependency rule: if outside scope is needed, return a Dependency Note instead of editing.
• Consistency rule: follow existing repo patterns; do not invent new architecture unless requested.
• Verifiability rule: each worker includes at least one concrete verification method.
• Ordering rule: wave order is strict; do not start next wave early.
• Subagent rule: one named Codex subagent per worker; close finished subagents.
• Planning quality rule: because workers default to gpt-5.4-mini, the main plan must be detailed, concrete, and unambiguous enough that workers do not need to invent missing structure.
• Plan-location rule: save plans in .planning/ by default and create the folder if needed.
• Worker-context rule: every worker must read the original plan file first for the full picture before editing.
• Handoff rule: every worker must append its handoff into the original plan file, not only reply in-session.
• Deduplication rule: if a worker reports more than once, keep one clean final handoff entry in the plan file.
• Commit-scope rule: before committing, inspect git status and commit only the files that belong to the planned task.
• Verification-scope rule: run targeted checks first, then broader checks; report unrelated pre-existing failures clearly as outside the feature scope.
• Status-tracking rule: if the project uses a status-tracking doc, board, or changelog-in-progress, update it as part of closeout so project status reflects what was completed and what remains open.
• Docs rule: docs update is mandatory in final step (update existing docs or add new).

Model Routing Configuration

Use Codex built-in subagents for execution.

Rules:

• Default all worker/subagent execution to gpt-5.4-mini.
• The main Codex agent remains responsible for planning, orchestration, integration, and final verification.
• Prefer a very strong plan, strict ownership, and clear contracts over increasing model size.
• Assume smaller workers perform best when instructions are explicit, scoped, and testable; plan accordingly.
• If a task is unusually complex and gpt-5.4-mini is clearly insufficient, pause and ask the user before using a stronger model.
• Do not use per-task model routing by default; prefer gpt-5.4-mini unless the user approves an exception.

Mandatory Worker Handoff Format

1. Summary (1-3 lines)
2. Files changed (exact list)
3. Patch/diffs (or exact edits)
4. How to test (commands + expected result)
5. Risks/TODOs/Dependency Notes
6. Main plan log entry (what to append to shared execution log)
7. Completion marker: WAVE {WAVE_ID} / {AGENT_NAME} DONE

Output Contract

Always output the final plan using this section order:

• A) Vision
• B) Decision Log
• C) Contracts (Wave 0)
• D) Waves (1..N with agents)
• E) Integration + Final Verification + Docs
• F) Worker prompts
• G) Execution checklist (Codex subagents + wave order)

Use templates from templates/ when helpful.

Final Response Contract

After execution, keep the user-facing summary short by default and include:

• what was implemented
• where the plan file lives
• what tests/checks passed
• what is still failing, blocked, or outside scope