implementation-planning

Implementation Planning Skill

Standard workflow only. If task_complexity is simple and workflow mode is Express, do not activate this skill. Simple tasks use the Express workflow, which does not activate implementation-planning. Return to the Express Workflow section.

Activate this skill during Phase 2 of Maestro orchestration, after the design document has been approved. This skill provides the methodology for generating detailed, actionable implementation plans that map directly to subagent assignments.

Codebase Grounding

Do not generate an implementation plan from guesses about the repository.

Use the built-in codebase_investigator before phase decomposition when:

• The task modifies an existing codebase
• File ownership, integration points, or validation commands are still unclear after reading the approved design
• Parallelization decisions depend on understanding current module boundaries or likely file overlap

Ask the investigator for:

• The modules and files most likely to change
• Existing architectural boundaries and conventions the plan must preserve
• Integration seams, dependencies, and shared ownership hotspots
• Validation commands and test entry points already used by the project
• Parallelization or conflict risks that should prevent batching

Skip the investigator only for greenfield tasks, documentation-only work, or plans where the current turn already established the relevant repo structure from direct reads.

Reuse investigator findings directly in the implementation plan:

• File inventories should reflect real candidate paths, not placeholders
• Validation criteria should prefer repo-native commands the investigator surfaced
• Parallel batches should account for actual ownership overlap and conflict risk

Plan Generation Methodology

Input Analysis

Before generating the plan, thoroughly analyze the approved design document for:

• Read task_complexity from the approved design document's frontmatter. Apply phase count guidance and domain analysis scaling accordingly. Record task_complexity in implementation plan frontmatter.
• Components and their responsibilities
• Interfaces and contracts between components
• Data models and their relationships
• External dependencies and integrations
• Technology stack decisions
• Quality requirements that influence implementation order

Phase Decomposition

Break the implementation into phases following these principles:

1. Foundation First: Infrastructure, configuration, and shared types/interfaces come first
2. Dependencies Flow Downward: A phase can only depend on phases with lower IDs
3. Single Responsibility: Each phase delivers a cohesive unit of functionality
4. Agent Alignment: Each phase maps to an agent specialization. The SAME agent MAY appear in multiple parallel phases when the task decomposes into independent sub-deliverables within one domain (see Same-Agent Decomposition below)
5. Agent Capability Match: Verify the assigned agent's tool tier supports the phase deliverables (see compatibility check below)
6. Testability: Each phase should be independently validatable

Phase Ordering Strategy

Layer 1: Foundation (types, interfaces, configuration)
    |
Layer 2: Core Domain (business logic, data models)
    |
Layer 3: Infrastructure (database, external services, API layer)
    |
Layer 4: Integration (connecting components, middleware)
    |
Layer 5: Quality (testing, security review, performance)
    |
Layer 6: Documentation & Polish

Agent-Deliverable Compatibility Check

Before finalizing agent assignments, verify each phase's agent can deliver its requirements:

Phase Deliverable	Required Tier	Compatible Agents
Creates/modifies files	Full Access or Read+Write	coder, data-engineer, devops-engineer, tester, refactor, design-system-engineer, i18n-specialist, analytics-engineer, technical-writer, product-manager, ux-designer, copywriter
Runs shell commands	Full Access or Read+Shell	coder, data-engineer, devops-engineer, tester, refactor, design-system-engineer, i18n-specialist, analytics-engineer, debugger, performance-engineer, security-engineer, seo-specialist, accessibility-specialist
Analysis/review only	Any tier	All agents

<HARD-GATE> Read-Only agents (architect, api-designer, code-reviewer, content-strategist, compliance-reviewer) CANNOT be assigned to phases that create or modify files. If a phase requires file creation and domain expertise from a Read-Only agent, split it: the Read-Only agent produces a spec or analysis, then a write-capable agent (typically coder) implements the files based on that output. </HARD-GATE>

Phase Count Guidance

Scale decomposition granularity to task_complexity (read from design document frontmatter):

• simple: 1-3 phases. Prefer single-phase execution when feasible. Combine foundation + implementation. Skip separate documentation/polish phases.
• medium: 3-5 phases. Use the layer model but combine Quality and Documentation into the final implementation phase where practical.
• complex: No phase count cap. Full layer decomposition strategy applies.

Parallelization Identification

Phases can run in parallel when:

• They have no shared file dependencies (no overlapping files_created or files_modified)
• They are at the same dependency depth (same layer)
• They do not share data model ownership
• Their validation can run independently

Mark parallel-eligible phases with parallel: true and group them into execution batches.

Same-Agent Parallel Decomposition

When 70%+ of the work maps to a single agent specialization, decompose by deliverable — not by agent type. The orchestrator MUST auto-detect this condition and apply it without user configuration.

Ultrawork Override: When ultrawork: true is set in session state, the 70% threshold drops to 40%. Decompose by deliverable whenever 2+ independent file sets exist within a single agent's scope, regardless of what percentage of total work that agent handles. Allow up to 8 parallel instances of the same agent (standard mode caps at 4).

Trigger Conditions (auto-detected by planner)

• The task is primarily one domain (e.g., "build a landing page" = all frontend)
• The deliverables split into 2-4 independent units with non-overlapping files (Ultrawork: 2-8 units)
• Each unit is self-contained enough for the agent to complete without the other units

Decomposition Strategies

Strategy	Example	When to Use
By section	Header+Nav, Hero+CTA, Features, Footer+Contact	Landing pages, multi-section layouts
By feature	Auth UI, Dashboard, Settings page	Feature-rich apps
By concern	Layout+Structure, Styling+Animations, Responsive+A11y	When sections share too many files

File Ownership Contract

Each parallel instance of the same agent MUST own non-overlapping files:

• Instance A: components/Header.tsx, components/Nav.tsx
• Instance B: components/Hero.tsx, components/Features.tsx
• Instance C: components/Footer.tsx, components/Contact.tsx

Shared files (e.g., globals.css, layout.tsx, page.tsx) go to a post-batch integration phase — a single agent that stitches the parallel outputs together.

Post-Batch Integration Phase

After parallel same-agent phases complete, always add a lightweight integration phase:

• Agent: same specialist (e.g., frontend_specialist)
• Task: import all parallel outputs, wire into shared layout, resolve conflicts
• Depends on: ALL parallel phases
• Files: only shared/barrel files (layout, page, globals)

Automatic Quality Review Pipeline

For any task that produces user-facing output (UI, landing page, app), the planner MUST append a review layer AFTER the build phases complete. This is automatic — the user does not request it.

When to Append (auto-detected)

• The task involves frontend_specialist, mobile_developer, or game_developer
• The build phase(s) create or modify 3+ files
• The task complexity is medium or complex

Review Pipeline Structure

Build Phase(s) → Integration (if parallel) → Review Batch (parallel) → Fix Phase

Review Batch: 2-3 parallel instances of the SAME implementing agent, each reviewing ONE quality dimension:

Review Instance	Focus	What to Check
Responsive Review	Layout at 375px, 768px, 1024px, 1440px	Breakpoints, overflow, touch targets, safe areas
Design Compliance	Style adherence	Does output match the chosen taste style? Anti-slop check, animation quality
Accessibility + Performance	A11y + Core Web Vitals	ARIA, keyboard nav, contrast, image optimization, bundle size

Each review instance produces a structured findings report (Critical/Major/Minor). The findings feed into a single Fix Phase where the implementing agent addresses Critical and Major issues.

Review Phase Template

Agent: [same as build agent]
Phase: [N]/[total]
Task: REVIEW ONLY — do NOT modify files
Focus: [Responsive / Design Compliance / Accessibility + Performance]

Review the following files created in previous phases:
[file list from build phases]

Produce a findings report with:
- Severity: Critical / Major / Minor
- File + line number
- What's wrong
- Suggested fix

Do NOT write code. Report findings only.

Fix Phase Template

Agent: [same as build agent]
Phase: [N]/[total]
Task: Fix Critical and Major findings from review batch

[Aggregated findings from all review instances]

Fix each finding. Skip Minor unless trivial.

Implementation Detail Requirements

Per-Phase Specification

Each phase in the plan must include:

Objective

A clear, measurable statement of what this phase delivers.

Agent Assignment

Which agent(s) execute this phase, with rationale for selection.

Files to Create

For each new file:

• Full relative path from project root
• Purpose and responsibility
• Key interfaces, classes, or functions to define
• Complete type signatures for public APIs

Files to Modify

For each existing file:

• Full relative path from project root
• Specific changes required and why
• Expected before/after for critical sections

Implementation Details

Provide sufficient detail for the assigned agent to execute without ambiguity:

• Interface definitions with complete type signatures
• Base class contracts with abstract method signatures
• Dependency injection patterns and registration points
• Error handling strategy (error types, propagation, recovery)
• Configuration requirements (environment variables, config files)

Validation Criteria

Specific commands to run and expected outcomes:

• Build/compile commands
• Lint/format checks
• Unit test commands
• Integration test commands (if applicable)
• Manual verification steps (if applicable)

Dependencies

• blocked_by: Phase IDs that must complete before this phase starts
• blocks: Phase IDs that cannot start until this phase completes

Dependency Minimization

List only direct blockers in blocked_by. Do not include transitive dependencies — they inflate dependency depth and prevent parallelism.

Anti-pattern (over-specified):

• Phase 2: blocked_by: [1]
• Phase 3: blocked_by: [1, 2] — Phase 1 is redundant, already reachable via Phase 2
• Phase 4: blocked_by: [1, 2, 3] — Phases 1, 2 are redundant

Result: depths 0, 1, 2, 3 — zero parallel phases.

Correct (minimized):

• Phase 2: blocked_by: [1]
• Phase 3: blocked_by: [1] — Only needs Phase 1 output, not Phase 2
• Phase 4: blocked_by: [2, 3] — Needs both done

Result: depths 0, 1, 2 — Phases 2 and 3 run in parallel at depth 1.

Ask for each dependency: "Does this phase truly need the output of that specific phase, or is it transitively covered?"

If validate_plan is available, review its parallelization_profile and redundant_dependency warnings before presenting the plan. Revise blocked_by to eliminate redundancies when possible.

Agent Assignment Criteria

Matching Tasks to Agents

Task Domain	Primary Agent	Secondary Agent	Rationale
System design, architecture	architect	-	Read-only analysis, design expertise
API contracts, endpoints	api-designer	coder	Design then implement
Feature implementation	coder	-	Full implementation access
Code quality review	code-reviewer	-	Read-only verification
Database schema, queries	data-engineer	-	Schema + implementation
Bug investigation	debugger	-	Read + shell for investigation
CI/CD, infrastructure	devops-engineer	-	Full DevOps access
Performance analysis	performance-engineer	-	Read + shell for profiling
Code restructuring	refactor	-	Write + shell access (for validation)
Security assessment	security-engineer	-	Read + shell for scanning
Test creation	tester	-	Full test implementation
Documentation	technical-writer	-	Write access for docs
Technical SEO audit	seo-specialist	-	Read + shell + web search
Marketing copy, content	copywriter	-	Read/write
Content planning	content-strategist	-	Read + web search/fetch
UX design, user flows	ux-designer	-	Read/write + web search
WCAG compliance audit	accessibility-specialist	-	Read + shell + web search
Requirements, product	product-manager	-	Read/write + web search
Tracking, analytics	analytics-engineer	coder	Implement then instrument
Internationalization	i18n-specialist	coder	Implement then localize
Design tokens, theming	design-system-engineer	coder	Tokens then consume
Legal, regulatory	compliance-reviewer	-	Read + web search/fetch

Assignment Rules

1. Match the primary task domain to the agent specialization
2. Consider tool requirements — does the task need shell access? Write access?
3. For parallel phases, assign non-overlapping file ownership to each agent
4. Prefer focused single-deliverable phases. The same agent MAY run in multiple parallel phases with distinct file ownership (see Same-Agent Decomposition)
5. Never assign more files to an agent than it can handle within its max_turns limit
6. For UI/frontend tasks, always append the Automatic Quality Review Pipeline after build phases

Token Budget Estimation

Estimate token consumption per phase based on:

• Number of files to read (input tokens)
• Complexity of output expected (output tokens)
• Agent's max_turns limit as upper bound
• Historical averages: ~500 input tokens per file read, ~200 output tokens per file written

Cost Estimation

Per-Phase Cost Factors

• Model tier: Pro agents (~$0.01/1K input, ~$0.04/1K output) vs Flash agents (~$0.001/1K input, ~$0.004/1K output)
• Input complexity: Number of files read, average file size, context from previous phases
• Output complexity: Lines of code generated, number of files created/modified
• Retry budget: Add 50% buffer per phase for potential retries (max 2 retries)

Estimation Formula

Phase Cost = (input_tokens × input_rate + output_tokens × output_rate) × retry_multiplier

Where:

• input_tokens = files_to_read × 500 + context_tokens
• output_tokens = files_to_write × 200 + validation_output
• retry_multiplier = 1.5 (accounts for up to 2 retries)

Plan-Level Cost Summary

Include this table in every implementation plan:

Phase	Agent	Model	Est. Input	Est. Output	Est. Cost
1	[agent]	[model]	[tokens]	[tokens]	[$X.XX]
...	...	...	...	...	...
Total			[sum]	[sum]	[$X.XX]

Plan Document Generation

Output Location

The write path depends on whether your runtime provides a Plan Mode surface (check get_runtime_context, loaded at session start, step 0).

• Plan Mode active: Some runtimes restrict writes to a temporary staging directory during Plan Mode. Write the plan there first, then copy to the permanent location after approval. Call exit_plan_mode with the plan path to present the plan for user approval.
• Plan Mode not active or not available: Write the implementation plan directly to the project's plans directory.

Permanent location: <state_dir>/plans/YYYY-MM-DD-<topic-slug>-impl-plan.md (where <state_dir> resolves from GEM_SWARM_STATE_DIR, default docs/maestro).

If your runtime does not provide a Plan Mode transition, track planning progress using the plan-update mechanism from your runtime context, write directly to the final location, and use the user-prompt tool from runtime context for the approval gate.

Document Structure

Use the implementation-plan template loaded via get_skill_content.

Required Sections

1. Plan Overview: Summary of total phases, agents involved, estimated effort
2. Dependency Graph: Visual representation showing phase dependencies and parallel opportunities
3. Execution Strategy Table: Stage-by-stage breakdown with agent assignments and execution mode
4. Phase Details: Full specification for each phase (objective, agent, files, details, validation, dependencies)
5. File Inventory: Complete table mapping every file to its phase and purpose
6. Risk Classification: Per-phase risk assessment (LOW/MEDIUM/HIGH) with rationale
7. Execution Profile: Summary of parallel vs sequential characteristics to inform mode selection:

   Execution Profile:
   - Total phases: [N]
   - Parallelizable phases: [M] (in [B] batches)
   - Sequential-only phases: [S]
   - Estimated parallel wall time: [time estimate based on batch execution]
   - Estimated sequential wall time: [time estimate based on serial execution]
   
   Note: Native parallel execution currently runs agents in autonomous mode.
   All tool calls are auto-approved without user confirmation.
   

Completion Criteria

The implementation plan is complete when:

• Every component from the design document maps to at least one phase
• All phase dependencies are acyclic (no circular dependencies)
• Parallel opportunities are identified and marked
• Each phase has clear validation criteria
• File ownership is non-overlapping for parallel phases
• The user has given explicit approval of the complete plan

Before presenting the plan for approval, check whether validate_plan appears in your available tools. If it does, call it with the plan structure and task_complexity to verify phase count constraints, file ownership, acyclic dependencies, and agent validity. If it does not, self-check against the phase count limits above.

Post-Generation

After writing the implementation plan:

1. Confirm the file path to the user
2. Present the dependency graph and execution strategy
3. Highlight parallel execution opportunities
4. Provide token budget estimates
5. If your runtime provides Plan Mode, call exit_plan_mode with the plan path to present the plan for user approval. If Plan Mode is not available, present the completed plan for user approval using the user-prompt tool from runtime context.
6. Ensure the approved plan is at <state_dir>/plans/YYYY-MM-DD-<slug>-impl-plan.md as the permanent project reference (copy from the staging directory if Plan Mode was used)
7. Ask if the user is ready to proceed to execution (Phase 3)
8. Upon approval, create the session state file via the session-management skill