Paw Planning

Implementation Planning

Execution Context: This skill runs directly in the PAW session (not a subagent), preserving user interactivity for phase decisions and blocker handling.

Create detailed implementation plans through interactive refinement. Plans describe WHAT to build (components, interfaces, behaviors) at the architectural level, delegating HOW to implement to the implementer.

Reference: Follow Core Implementation Principles from paw-workflow skill.

Configuration

Step 1: Read Plan Generation Mode

Read WorkflowContext.md for:

• Work ID and target branch
• Plan Generation Mode: single-model | multi-model
• Plan Generation Models: comma-separated model names (for multi-model modes)

The plan generation mode is set during paw-init. If the field is missing (legacy workflow), default to single-model (backwards compatibility).

{{#cli}} If mode is multi-model, parse the models list. Default: latest GPT, latest Gemini, latest Claude Opus. {{/cli}} {{#vscode}} Note: VS Code only supports single-model mode. If multi-model is configured, report to user: "Multi-model plan generation not available in VS Code; running single-model planning." Proceed with single-model. {{/vscode}}

Capabilities

• Create implementation plan from spec and research
• Revise plan based on learnings or feedback
• Address PR review comments (load paw-review-response for mechanics)
• Plan documentation phases when documentation updates are warranted

Be Skeptical

• Question vague requirements—ask "why" and "what about"
• Identify potential issues early
• Don't assume—verify with code research
• If user corrects a misunderstanding, verify the correction against code before accepting

Strategic Planning Guidelines

Operate at the C4 container/component abstraction level:

Do:

• Describe component responsibilities and interfaces
• Reference file paths, module names, existing patterns
• Present 2-3 options with trade-offs for significant decisions
• Define success via observable outcomes (tests, SLOs, acceptance criteria)
• Limit code snippets to 3-10 lines for critical architectural concepts only

Don't:

• Include complete function implementations
• Provide pseudo-code walkthroughs
• Specify library choices without trade-off analysis
• Write tutorial-style code examples

Example descriptions:

• ✅ "Implement UserRepository interface with CRUD methods, following pattern in models/repositories/"
• ✅ "Add auth middleware validating JWT tokens, integrating with services/auth/"
• ❌ "Create function that loops through array and filters..."

Documentation Phase Planning

Include documentation as the final implementation phase. The documentation phase is standard for all non-trivial changes.

What the Documentation Phase Produces

1. Docs.md (always): Technical reference capturing implementation details, usage, and verification approach—serves as the "as-built" record even for internal changes
2. Project documentation updates (when warranted): README, CHANGELOG, guides following project conventions
3. Documentation build verification: Command from CodeResearch.md (if framework discovered)

When to Include Project Docs

• Work creates user-facing features
• APIs are added or changed
• Existing behavior is modified in ways users should know about

When to Omit Project Docs

• Purely internal changes (Docs.md still required)
• Refactors without behavior changes (Docs.md still required)
• User explicitly indicates no project docs needed

Note: Implementer loads paw-docs-guidance utility skill for templates and conventions during documentation phases.

ImplementationPlan.md Template

Save to: .paw/work/<work-id>/ImplementationPlan.md

# [Feature/Task Name] Implementation Plan

## Overview
[What we're implementing and why]

## Current State Analysis
[Existing state, gaps, key constraints from research]

## Desired End State
[Target state specification and verification approach]

## What We're NOT Doing
[Out-of-scope items]

## Phase Status
- [ ] **Phase 1: [Name]** - [Objective]
- [ ] **Phase 2: [Name]** - [Objective]
- [ ] **Phase N: Documentation** - [If warranted]

## Phase Candidates
<!-- Use checkbox format for each candidate so transition checks can detect unresolved items.
     Example: - [ ] Moderator mode support -->

---

## Phase 1: [Name]

### Changes Required:
- **`path/to/file.ext`**: [Component changes, pattern references]
- **Tests**: [Test file, key scenarios]

### Success Criteria:

#### Automated Verification:
- [ ] Tests pass: `<command>`
- [ ] Lint/typecheck: `<command>`

#### Manual Verification:
- [ ] [User-observable behavior]
- [ ] [Edge cases requiring human judgment]

---

## Phase N: Documentation (if warranted)

### Changes Required:
- **`.paw/work/<work-id>/Docs.md`**: Technical reference (load `paw-docs-guidance`)
- **Project docs**: Per CodeResearch.md findings

### Success Criteria:
- [ ] Docs build: `<command>`
- [ ] Content accurate, style consistent

---

## References
- Issue: [link or 'none']
- Spec: `.paw/work/<work-id>/Spec.md`
- Research: `.paw/work/<work-id>/SpecResearch.md`, `.paw/work/<work-id>/CodeResearch.md`

Phase Status format: Use checkboxes - [ ] for pending, - [x] for complete. This provides at-a-glance status without heavy tables.

Execution Contexts

Initial Planning

Desired end state: Complete ImplementationPlan.md with all phases defined

If single-model plan generation mode (default):

1. Read all context: Issue, Spec.md, SpecResearch.md, CodeResearch.md
2. Analyze and verify requirements against actual code
3. Present understanding and resolve blocking questions
4. Research patterns and design options (if significant choices exist)
5. Write plan incrementally (outline, then phase by phase)
6. Handle branching per Review Strategy (see below)

{{#cli}} If multi-model plan generation mode:

1. Read all context: Issue, Spec.md, SpecResearch.md, CodeResearch.md

2. Create .paw/work/<work-id>/planning/ directory if it doesn't exist

3. Create .paw/work/<work-id>/planning/.gitignore with content * (if not already present). This is a local-only scratch ignore marker — do NOT stage or commit it.

4. Resolve model intents to actual model names (e.g., "latest GPT" → current GPT model)

5. Present resolved models for confirmation:

   About to run multi-model plan generation with:
   - [resolved model 1]
   - [resolved model 2]
   - [resolved model 3]
   
   Estimated LLM calls: N+1 (N = number of models)
   
   Proceed?
   

6. Independent Plans (parallel): Spawn parallel subagents using task tool with model parameter for each model. Each subagent receives the planning subagent prompt below along with the full contents of Spec.md, CodeResearch.md, and SpecResearch.md. Save per-model plans to PLAN-{MODEL}.md in the planning/ subfolder.

7. Synthesis: Read all per-model plans. Produce the final ImplementationPlan.md using the synthesis prompt below, selecting the best phase structure, architecture decisions, and catching blind spots.

Failure handling: If a subagent fails, proceed with remaining results if at least 2 models completed successfully. If fewer than 2 succeed, offer user the choice to retry or fall back to single-model plan generation. Configurations with exactly 2 models are valid — synthesis works with any count ≥ 2.

Planning Subagent Prompt

Each subagent receives this prompt (self-contained, since subagents cannot load skills):

You are creating an implementation plan. You will receive a feature specification (Spec.md), codebase research (CodeResearch.md), and optionally spec research (SpecResearch.md).

Create a complete implementation plan following this template structure: Overview, Current State Analysis, Desired End State, What We're NOT Doing, Phase Status (checkboxes), Phase Candidates, then detailed phases with Changes Required (file paths, components, tests) and Success Criteria (automated and manual verification).

Guidelines:

• Operate at the C4 container/component abstraction level — describe WHAT to build, not HOW
• Reference specific file paths, module names, and existing patterns from CodeResearch.md
• Each phase should be independently reviewable with clear success criteria
• Include a Documentation phase as the final phase
• Limit code snippets to 3-10 lines for critical architectural concepts only
• Zero TBDs — all decisions must be made
• Include "What We're NOT Doing" to prevent scope creep

Synthesis Prompt

The session's agent reads all per-model plans and applies this approach:

Synthesize the best elements from all plan drafts into a single final ImplementationPlan.md. Compare section-by-section across all plans:

• Phase structure: Choose the best decomposition — consider granularity, independence, and logical ordering
• Architecture decisions: Pick the strongest approach, noting where plans converged (high confidence) vs. diverged (needs careful selection)
• Success criteria: Merge the most comprehensive and measurable criteria from all plans
• Scope boundaries: Union of all "What We're NOT Doing" items

The output must follow the standard ImplementationPlan.md template format exactly. {{/cli}}

PR Review Response

When addressing Planning PR comments, load paw-review-response utility for mechanics.

Desired end state: All review comments addressed, artifacts consistent

1. Verify on planning branch (<target>_plan)
2. Read all unresolved PR comments
3. Create TODOs for comments (group related, separate complex)
4. For each TODO: make changes → commit → push → reply
5. Verify all comments addressed and artifacts consistent

Plan Revision

When revising based on paw-plan-review feedback:

Desired end state: ImplementationPlan.md updated to address all BLOCKING issues

1. Read paw-plan-review feedback completely
2. Address BLOCKING issues first (these prevent implementation)
3. Address IMPROVE issues if scope permits
4. Acknowledge NOTE items in commit message if applicable
5. Re-run quality checklist to verify fix didn't introduce new issues

Branching and Commits

Reference: Load paw-git-operations skill for branch naming, commit mechanics, and PR descriptions.

Quality Checklist

• All phases contain specific file paths, components, or interfaces
• Every phase has measurable automated and manual success criteria
• Phases build incrementally and can be reviewed independently
• Zero open questions or TBDs remain
• All references trace to Spec.md and CodeResearch.md
• Unit tests specified alongside code they test within each phase
• "What We're NOT Doing" section prevents scope creep
• Code blocks absent or limited to <10 lines for architectural concepts
• Documentation phase included (or explicitly omitted with reason)
• Phase Status exists with checkbox phases and one-sentence objectives
• Phase Candidates section present (may be empty initially)

Completion Response

Report to PAW agent:

• Artifact path: .paw/work/<work-id>/ImplementationPlan.md
• Planning mode used: single-model or multi-model
• Models (if multi-model): list of models used and any failures
• Plan summary for quick review:
  • Architecture approach (1-2 sentences)
  • Phase overview (numbered list with one-line objective each)
  • What's explicitly NOT being done (scope boundaries)
• Review Strategy used (PR created or committed to target)
• Any items requiring user decision or observations worth noting

Blocked on Open Questions

If planning encounters questions that cannot be answered from existing artifacts (Spec.md, CodeResearch.md):

1. Do NOT write ImplementationPlan.md — partial artifacts cause confusion on re-invocation
2. Return to PAW agent with:
   • Status: blocked
   • List of specific open questions
   • What research or clarification would resolve each question
3. PAW agent handles resolution based on Review Policy:
   • final-pr-only: PAW agent conducts additional research to resolve questions autonomously
   • every-stage/milestones: PAW agent asks user for clarification
4. Re-invocation: PAW agent calls planning again with answers provided in the delegation prompt