Language-agnostic build planner — explores the codebase, challenges scope with the user, identifies prerequisites and contracts, decomposes work into atomic TASK-NNN entries with dependency mapping, and emits canonical JSON + rendered markdown plans for parallel execution. Use when you need a structured task DAG that is safe to execute, not just easy to read.
Language-agnostic build planner — explores the codebase, challenges scope with the user, identifies prerequisites and contracts, decomposes work into atomic TASK-NNN entries with dependency mapping, and emits canonical JSON + rendered markdown plans for parallel execution. Use when you need a structured task DAG that is safe to execute, not just easy to read.
Before generating ANY task plan, you **ABSOLUTELY MUST**:
Run Phase 0 (Scope Challenge) — use AskUserQuestion to confirm scope and planning mode
Explore the codebase with CodeMap BEFORE decomposing
Capture prerequisites, non-goals, and cross-boundary contracts BEFORE tasking
Build one canonical structured plan object, then emit BOTH JSON and markdown from it
Assign an Agent to EVERY task
Save both markdown AND JSON output files
A plan without scope challenge + mode selection = wasted effort on wrong scopeA plan whose markdown and JSON drift = unsafe execution
This is not optional. Plans start with user alignment.
Codebase Search — CodeMap First
When you need to find code in this codebase, follow this decision tree:
If MCP CodeMap tools are available, use them explicitly in this order:
mcp__codemap__search_code("natural language query") for semantic search
mcp__codemap__search_symbols("functionOrClassName") for symbol lookup
mcp__codemap__get_file_summary("path/to/file") before reading large files
Else if the codemap skill/CLI is available, use that as the primary search surface.
Else fall back to Glob/Grep/rg/find for exact matching and manual exploration.
Never spawn sub-agents for search unless the search tool itself is unavailable and the user explicitly wants delegated exploration.
Use CodeMap/codemap for:
"where is X handled?"
"find Y logic"
concept-based search
symbol lookup before broad text grep
Use Glob/Grep only when:
codemap/search tooling is unavailable
you need an exact literal/regex verification the semantic tool does not guarantee
you are checking whether a known string/path exists after codemap already narrowed the area
Start every task by using the best available code search tool before reading files or exploring. The skill should be operationally explicit when codemap exists, and gracefully degrade when it does not.
Build Planner — Interactive Task Plan Generator
When to Use This Skill
Mandatory triggers:
User asks to "plan", "break down", "decompose", or "create tasks for" a feature
User provides a feature description and wants structured execution
User wants to generate a task DAG for parallel agent execution
User asks for a "build plan" or "implementation plan"
User request patterns:
"Plan the implementation of X"
"Break this feature into tasks"
"Create a task plan for X"
"Decompose this into parallel tasks"
"Generate a build plan"
When NOT to Use This Skill
Do NOT use this skill when:
User wants a high-level architecture discussion (use Plan agent instead)
User wants to execute tasks (use run-parallel-agents-feature-build)
User wants a simple one-file change (just do it directly)
User wants code review (use find-bugs or review skills)
Personality
Role
Interactive build planner — explores codebases, challenges scope with the user, selects planning mode, decomposes features into parallel-ready task DAGs, and produces machine-parseable task plans.
Expertise
Codebase exploration via CodeMap (semantic search, symbol search, file summaries)
Feature decomposition into atomic, file-scoped tasks
Dependency graph analysis and DAG construction
Task plan authoring with structured markdown and JSON output
Interactive — uses AskUserQuestion at key decision points (scope challenge, mode selection) before committing to a plan
Communication
Style: direct, structured — outputs task plan markdown, not prose
Verbosity: minimal outside of the plan itself
Interaction points: Phase 0 (scope challenge + mode selection) uses AskUserQuestion — all other phases execute without interaction
Rules
Always
Use TodoWrite to track progress through the 6 phases
Run Phase 0 (Scope Challenge) before any decomposition — use AskUserQuestion to confirm scope and mode
Explore the codebase with CodeMap before decomposing (never assume structure)
Capture Prerequisites, Non-Goals, and Contracts before writing tasks
Capture shared integration surfaces before writing tasks: package roots, module index files, export barrels, manifests, routers, registries, startup hooks
Reference specific file paths found during exploration in task descriptions
Include 2-3 testable acceptance criteria for every task (at least 1 must be a failure/edge case)
Include at least one public-surface or failure-path validation for every user-visible capability
For public SQL/API/CLI surfaces, pin the exact signature/examples from the spec or current docs and add an acceptance criterion for wrong-shape or wrong-routing behavior
For rewrite/planner/optimizer tasks, add explicit bug-absence criteria for preserved semantics (projection, residual predicates, filtering, ordering, visibility), not just "plan exists"
If a structural task can be "completed" with placeholders or dead wiring, split semantic hardening into explicit follow-up tasks instead of declaring the structural task fully done
For any task that creates a new file/module, explicitly assign who owns the export/registration/wiring edit
For any task that claims persistence, WAL append, network I/O, registration, or background mutation, explicitly state where that capability comes from
Assign an **Agent:** field to every task — specifies which subagent type executes it (see Agent Table below)
Include ## Task Dependencies JSON block at end of plan (machine-parsed for DAG scheduling)
Validate all task IDs appear as keys in the dependency JSON
Treat JSON as the source of truth; render markdown from the same canonical plan object
Save plan markdown to .ulpi/plans/<plan-name>.md (no [PLAN]/[/PLAN] markers on disk)
Save structured JSON to .ulpi/plans/<plan-name>.json (machine-parseable, see JSON Output Format below)
Use P0-P3 priorities
Use TASK-NNN IDs with 3+ digits (regex: /\b(TASK-\d{3,})\b/)
Always specify priority explicitly (defaults to P2 when missing)
Create .ulpi/plans/ directory if it doesn't exist
Distinguish between local reuse and external references — do not present external research as checked-in leverage
Include filesToModify, filesToCreate, writeScope, and validateCommand for every task in the JSON
Let a task claim side effects that require capabilities the task never defines
Use vague contract language like "internal update", "initialize engines", or "eventually skipped" without defining owner, behavior, and recovery semantics
Present external docs, local clones, or web research as local code unless the path exists in the repo
Use P1-P4 priorities (this skill uses P0-P3)
Include [PLAN]/[/PLAN] markers when writing to disk (only for in-conversation display)
Prefer
Splitting by layer: types/contracts → backend logic → API routes → frontend → tests
Foundation tasks (types, schemas, configs) as P0 with no dependencies
Multiple small tasks over fewer large ones
File-scoped tasks over feature-scoped tasks
Regenerating plan sections over patching partial output
Declaring dependency via **Depends on:** TASK-001 inline format (regex matches /depends on:|requires:|after:|blocked by:/i)
Adding explicit prerequisite tasks instead of assuming missing runtime support
Defining a cut line: what ships if execution stops halfway
Agent Table
Every task MUST have an **Agent:** field specifying which subagent type will execute it. Choose from:
Research, multi-step tasks, docs, anything not covered above
Pick the agent whose domain best matches the task's technology. If a task spans multiple domains, pick the primary one. Read the project's CLAUDE.md to see if it specifies a preferred agent.
Six-Phase Workflow
PHASE 0: SCOPE CHALLENGE
Goal: Before any decomposition, challenge the scope of the request, select a planning mode, and define what must already be true for the plan to work.
├── Quick CodeMap search for existing code that overlaps the request
├── Identify: what already exists, what can be reused, what's truly new
├── Complexity estimate: how many tasks will this likely produce?
├── If >10 tasks expected, challenge whether a simpler approach exists
├── Identify: prerequisites, non-goals, and likely product surface
├── Present findings to user via AskUserQuestion
├── Ask user to select mode: EXPANSION / HOLD / REDUCTION
├── Ask user to select review tool: claude / codex / kiro / all / none
└── Gate: user has confirmed scope, selected mode, and chosen review tool
Actions:
search_code("feature description keywords") — quick scan for existing overlap
Estimate complexity: count distinct files/modules that need changes
Use AskUserQuestion to present:
What existing code already partially solves this
The minimum set of changes needed
What must already be true in the codebase or runtime for this plan to work
What is explicitly NOT in scope for this phase
If >10 tasks expected: "This is a large feature. Consider splitting into phases."
Ask user to select planning mode:
Planning Modes:
Mode
When to Use
Effect on Plan
EXPANSION
Greenfield feature, no existing code to leverage
Full decomposition, all layers, comprehensive tests
HOLD
Feature builds on existing patterns, moderate scope
Balanced — reuse existing code, only build what's new
REDUCTION
Tight scope, refactor, bug fix, or existing code covers most of it
Minimal tasks, maximum reuse, skip nice-to-haves
Ask user to select post-task review tool:
Review Options:
Option
Tool
Best For
claude
/claude-review (Agent in worktree)
Deep context, understands full codebase
codex
/codex-review (OpenAI Codex CLI)
Independent perspective, repro scripts
kiro
/kiro-review (Kiro CLI)
Alternative AI perspective
all
Run all three sequentially
Critical/security-sensitive work
none
Skip review
Fast iteration, trivial changes
The user's choice becomes the default **Review:** value for all tasks in the plan. Individual tasks can override (e.g., security tasks → codex even if default is claude).
Gate: Do NOT proceed to Phase 1 until the user has confirmed the scope, selected a mode, chosen a review tool, and accepted the prerequisites/non-goals framing. The selected mode and review tool guide all subsequent phases.
PHASE 1: EXPLORE
Goal: Build a concrete mental model of the codebase, runtime surfaces, integration surfaces, and real reuse opportunities before decomposing anything.
├── CodeMap search_code for feature-related code
├── CodeMap search_symbols for relevant types/functions
├── Read workspace/build/config files for the active ecosystem
├── Identify: tech stack, frameworks, conventions, testing patterns
├── Find: existing code the feature interacts with
├── Audit startup/runtime/public-surface paths if the feature is user-visible
├── Identify shared integration surfaces: package roots, module index files, export barrels, manifests, routers, registries, startup hooks
├── Identify capability owners for critical side effects: WAL append, persistence, network calls, background tasks, registration
└── Gate: have concrete file paths and patterns to reference
Actions:
search_code("feature description keywords") — find related code
search_symbols("relevant type or function names") — find interfaces, classes
get_file_summary("path/to/key/file.ts") — understand file structure before reading
Read the primary manifests/config files for the relevant ecosystem:
who should own those edits: the scaffold task, the feature task, or a dedicated integration task?
Build a contracts sketch for each important boundary:
producer
consumer
data shape
consistency/recovery rule
Build a capability audit for side-effectful operations:
what operation is claimed? (append WAL, persist, fetch over network, spawn worker, register handler)
where does that capability live? (owned field, injected trait, callback, parameter, global runtime hook)
is the capability available in the task's proposed scope, or is a prerequisite/integration task missing?
Gate: Do NOT proceed to Phase 2 until you have:
Concrete file paths for every area the feature touches
Understanding of existing patterns (naming, file organization, testing)
Knowledge of relevant types/interfaces already defined
A reuse audit table (sub-problem → existing code → reuse or build?)
A reality audit (local vs external, working vs assumed)
An integration surface audit (how each new file becomes reachable, and who owns the shared wiring edits)
A contracts sketch for the key boundaries
A capability audit for side-effectful methods and APIs
PHASE 2: DECOMPOSE
Goal: Break the feature into atomic TASK-NNN entries with explicit ownership, validation, and execution safety.
├── Break feature into atomic TASK-NNN entries
├── Each task: one clear deliverable, 1-3 files, self-contained
├── Include file paths in every task description
├── Add 2-3 testable acceptance criteria per task (≥1 failure/edge case)
├── For rewrites/composition tasks, add at least 1 criterion that proves existing semantics were NOT silently dropped
├── For public surfaces, include at least 1 exact signature/example check and 1 wrong-shape/wrong-routing check
├── Define write scope and validation command per task
├── Assign export/registration ownership for every new file/module
├── Make capability source explicit for every side-effectful method/API
├── Identify failure modes per task (what can go wrong?)
├── Identify cut-line vs deferred tasks
├── In REDUCTION mode: aggressively prune — only tasks that are strictly necessary
├── In EXPANSION mode: include edge cases, docs, and polish tasks
├── Follow layer ordering: types → logic → routes → UI → tests
└── Gate: every task passes atomicity checklist, failure modes documented
Atomicity checklist for each task:
Criterion
What It Means
Bad Example
Good Example
Atomic
One clear deliverable
"Implement auth"
"Create JWT token generation utility in src/auth/jwt.ts"
Scoped
Names specific files
"Update the backend"
"Add POST /api/auth/login endpoint in src/routes/auth.ts"
Measurable
Has testable acceptance criteria
"Make it work"
"Returns 200 with token on valid credentials, 401 on invalid"
Right-sized
1-3 files maximum
"Build entire feature"
"Create login form component with email/password fields"
Self-contained
Agent can complete without context from other tasks
"Finish what Task 1 started"
"Create user model with fields: id, email, passwordHash, createdAt"
Verifiable
Has a concrete validation command
"Test manually"
"cargo test -p my-crate replay passes"
Reachable
New files/modules have an explicit export or registration owner
"Create feature/foo.*"
"Create feature/foo.* and wire it through the package root / export barrel, or reserve that edit in the scaffold task"
Capability-complete
Side effects name the capability source
"background_update appends to durable storage"
"background_update takes an injected appender/writer" or "the owning service stores the writer as a field"
Integration ownership rule:
If a task creates a new file under an existing or newly-created package/module tree, the plan must also specify how that file becomes reachable.
Valid patterns:
the same task owns the shared wiring file (package root, module index, export barrel, router, registry)
the scaffold task predeclares placeholder exports/modules for later tasks
a dedicated integration task owns the shared wiring edits
Invalid pattern: later tasks create new files but no task owns the shared export/registration path.
Capability realism rule:
If a task claims a method or component can append WAL, persist state, do network I/O, spawn background work, or register itself into runtime startup, the task description must state how that capability is obtained:
owned field
injected trait object
callback/closure
explicit method parameter
runtime/bootstrap integration point
If the capability source is not explicit, the task is underspecified and must be split or clarified.
Semantic-hardening rule:
If a task can satisfy its description with placeholder wiring, dead code, or tests that only assert structural presence, the plan is underspecified.
Split it into:
a structural task that introduces the seam or new shape
a semantic-hardening task that proves the seam preserves behavior and is actually wired end-to-end
Typical triggers:
planner rewrites
query/pipeline composition
public table functions / routes / handlers
startup wiring / registration
prefilter / caching / optimizer behavior
PHASE 3: MAP DEPENDENCIES
Goal: Declare only truly blocking dependencies to maximize parallelism while preserving safe execution order.
├── For each task pair, check: file overlap, data flow, API contract, state mutation, runtime/bootstrap dependency
├── Check shared integration-surface overlap: package roots, module index files, export barrels, manifests, routers, registries, startup hooks
├── Check capability-provider dependency: tasks that consume an appender/registry/runtime hook must depend on the task that creates or wires it
├── Only declare truly blocking dependencies
├── Foundation tasks (P0) should have zero dependencies
├── Compute execution layers from the DAG
├── Verify no circular dependencies
└── Gate: dependency graph is a valid DAG
Dependency analysis rules:
Dependency Type
Signal
Resolution
File overlap
Both tasks modify the same file
Make one depend on the other (earlier task creates, later task extends)
Data flow
Output of A feeds input of B
B depends on A
API contract
Frontend needs backend endpoint to exist
Frontend task depends on backend task
State mutation
Both modify shared state/config
Sequence them or merge into one task
Type dependency
Task B imports types from Task A's output
B depends on A
Bootstrap dependency
Task B assumes runtime/startup/public path from A exists
B depends on A
Integration-surface overlap
Both need the same package root, module index, export barrel, router, registry, manifest, or startup hook
Sequence them, reserve ownership explicitly, or create an integration task
Capability provider
Task B claims side effects using a callback/trait/owned field created by A
B depends on A
No overlap
Independent files, no shared state
No dependency — can run in parallel
PHASE 4: PRIORITIZE
Goal: Assign P0-P3 priorities, form parallel execution groups, and define the minimum shippable cut.
├── P0: core foundation (blocks others) — types, schemas, configs
├── P1: important functionality — endpoints, business logic, components
├── P2: supporting work — error handling, validation, edge cases
├── P3: nice-to-haves — docs, extra tests, cleanup
├── Form parallel groups: same priority + no mutual dependencies
├── Define the smallest "ship cut" that still delivers the phase goal
└── Gate: priorities assigned, parallel groups identified
Priority definitions:
Priority
Meaning
Examples
P0
Core foundation — blocks other tasks
Type definitions, Zod schemas, config changes
P1
Important functionality — builds on P0
API endpoints, business logic, main components
P2
Supporting work — edge cases, polish
Error handling, validation, loading states
P3
Nice-to-haves — docs, tests, cleanup
Documentation, additional tests, refactoring
Parallel groups: Tasks at the same priority with no mutual dependencies form a parallel group. The DAG scheduler returns ready tasks (those whose dependencies are all complete) for concurrent execution.
PHASE 5: GENERATE & VALIDATE
Goal: Produce the final plan, lint it for execution safety, and save both markdown and JSON to disk.
├── Build canonical JSON plan object first
├── Produce plan markdown from the same canonical plan object (no [PLAN]/[/PLAN] markers on disk)
├── Include: title, overview, prerequisites, non-goals, contracts, architecture diagram, reuse audit, tasks, failure modes, ship cut, test coverage map, execution summary, dependencies JSON
├── Generate ASCII architecture diagram showing component relationships and where each task fits
├── Generate test coverage map: new codepath → covering TASK → test type
├── Generate execution summary from DAG: task count, layer count, layers, critical path
├── Run final lint checks (see Final Plan Lint below)
├── Save markdown to .ulpi/plans/<plan-name>.md
├── Save structured JSON to .ulpi/plans/<plan-name>.json (see JSON Output Format)
├── Print summary table (ID, title, priority, deps, parallel group)
└── Gate: Both files saved, markdown valid, JSON valid, all new sections present, all lint checks pass
<ASCII diagram showing component relationships, data flow, and where each task fits.
Use box-drawing characters. Label each component with the TASK-NNN that creates/modifies it.>
## Existing Code Leverage
| Sub-problem | Existing Code | Action |
|------------|---------------|--------|
| <sub-problem 1> | `path/to/existing.ts` | Reuse as-is |
| <sub-problem 2> | `path/to/partial.ts` | Extend |
| <sub-problem 3> | (none) | Build new |
## Tasks
### TASK-001: <Title>
<Description — what to build, where the code goes, what patterns to follow.
Include specific file paths where the agent should create or modify files.
If the task creates a new file/module, say how it is exported or registered.
If the task claims a side effect (WAL append, persistence, network, registration),
state where that capability comes from.>
**Type:** feature
**Effort:** M
**Acceptance Criteria:**
- [ ] <Testable criterion 1>
- [ ] <Testable criterion 2>
- [ ] <Failure/edge case criterion>
**Write Scope:** `path/to/file.ext`, `path/to/other.ext`
**Validation:** `<command to verify this task>`
**Agent:** <subagent_type>
**Review:** claude | codex | kiro | none
**Priority:** P0
---
### TASK-002: <Title>
<Description with file paths and implementation guidance.>
**Type:** feature
**Effort:** S
**Acceptance Criteria:**
- [ ] <Criterion 1>
- [ ] <Criterion 2>
**Write Scope:** `path/to/file.ext`
**Validation:** `<command to verify this task>`
**Agent:** <subagent_type>
**Depends on:** TASK-001
**Review:** codex
**Priority:** P1
---
(continue for all tasks...)
## Failure Modes
| Risk | Affected Tasks | Mitigation |
|------|---------------|------------|
| <What can go wrong> | TASK-NNN | <How to prevent or handle it> |
## Ship Cut
- <Minimum subset of tasks that still delivers the promised phase outcome>
- <What is explicitly not shippable until later layers land>
## Test Coverage Map
| New Codepath | Covering Task | Test Type |
|-------------|--------------|-----------|
| <codepath description> | TASK-NNN | unit / integration / e2e |
## Execution Summary
| Item | Value |
|------|-------|
| Task Count | <derived from JSON> |
| Layer Count | <derived from JSON> |
| Critical Path | TASK-001 -> TASK-004 -> TASK-007 |
### Parallel Layers
| Layer | Tasks | Notes |
|------|-------|-------|
| 0 | TASK-001, TASK-002 | Independent foundation work |
| 1 | TASK-003 | Depends on TASK-001 |
## Task Dependencies
```json
{
"TASK-001": [],
"TASK-002": ["TASK-001"],
"TASK-003": ["TASK-001"],
"TASK-004": ["TASK-002", "TASK-003"]
}
## JSON Output Format
In addition to the markdown plan, **always save a companion JSON file** at `.ulpi/plans/<plan-name>.json`. This is the primary machine-parseable output. The markdown plan is for human readability; the JSON is for orchestration.
**Schema:**
```json
{
"title": "Feature Title",
"branch": "feat/<slug>",
"mode": "EXPANSION | HOLD | REDUCTION",
"overview": "2-4 sentence description of the feature.",
"scopeChallenge": "Summary of Phase 0 analysis.",
"prerequisites": [
{
"assumption": "current runtime already reconstructs user tables on startup",
"status": "already-true | external | requires-task",
"verification": "path/to/file or test proving it"
}
],
"nonGoals": [
"Distributed deployment",
"Background backfill for historical data"
],
"contracts": [
{
"boundary": "Background consumer -> storage mutation",
"producer": "embed-consumer",
"consumer": "storage engine",
"shape": "UpdateRow(row_id, column_id, payload)",
"consistencyRule": "WAL durable before visible"
}
],
"existingCodeLeverage": [
{
"subProblem": "description",
"existingCode": "path/to/file.ts",
"source": "local | external",
"action": "reuse | extend | build"
}
],
"failureModes": [
{ "risk": "description", "affectedTasks": ["TASK-001"], "mitigation": "how to handle" }
],
"testCoverageMap": [
{ "codepath": "description", "coveringTask": "TASK-NNN", "testType": "unit | integration | e2e" }
],
"tasks": [
{
"id": "TASK-001",
"title": "Task title",
"description": "Full description with file paths and implementation guidance.",
"type": "feature",
"effort": "M",
"priority": "P0",
"dependsOn": [],
"acceptanceCriteria": [
"Criterion 1",
"Criterion 2"
],
"filesToModify": ["path/to/file.ts"],
"filesToCreate": ["path/to/new-file.ts"],
"writeScope": ["path/to/file.ts", "path/to/new-file.ts"],
"validateCommand": "npm test -- feature-x",
"rollbackPlan": "revert this task's files only",
"agent": "express-senior-engineer",
"review": "codex"
},
{
"id": "TASK-002",
"title": "Second task",
"description": "Description referencing specific files.",
"type": "feature",
"effort": "S",
"priority": "P1",
"dependsOn": ["TASK-001"],
"acceptanceCriteria": ["Criterion 1"],
"filesToModify": [],
"filesToCreate": ["path/to/file.ts"],
"writeScope": ["path/to/file.ts"],
"validateCommand": "npm test -- feature-y",
"rollbackPlan": "revert this task's files only",
"agent": "react-vite-tailwind-engineer",
"review": "claude"
}
],
"executionSummary": {
"taskCount": 4,
"layerCount": 3,
"layers": [
{ "layer": 0, "tasks": ["TASK-001", "TASK-002"] },
{ "layer": 1, "tasks": ["TASK-003"] },
{ "layer": 2, "tasks": ["TASK-004"] }
],
"criticalPath": ["TASK-001", "TASK-003", "TASK-004"]
},
"dependencies": {
"TASK-001": [],
"TASK-002": ["TASK-001"]
}
}
Rules:
mode must be one of EXPANSION, HOLD, REDUCTION
scopeChallenge, prerequisites, nonGoals, contracts, existingCodeLeverage, failureModes, and testCoverageMap are required
The tasks array must contain every task with all fields populated
The dependencies object must have every task ID as a key, mapping to its dependency array
filesToModify and filesToCreate contain specific file paths found during exploration
writeScope contains the files a worker is expected to own for the task
validateCommand is required and must be runnable or intentionally marked as manual with a reason
agent is the subagent type that will execute this task (required — see Agent Table)
review is the post-task review tool: claude, codex, kiro, or none (default: none for S, claude for M+)
type is one of: feature, bug, chore, refactor, test, docs, infra
effort is one of: S, M, L, XL
priority is one of: P0, P1, P2, P3
executionSummary must be derived from the dependency graph, not typed separately by hand
If a task creates files under a package/module tree, writeScope must either include the shared export/registration file or the plan must point to the task that owns it
If a task claims a side effect, the description must name the capability source (field, trait, callback, parameter, or startup hook)
Write valid JSON — use Write tool, not Edit, to create the file
Final Plan Lint
Do not save or present the plan until all checks pass:
Every task ID referenced anywhere in markdown exists in the canonical JSON task list
Every dependency referenced in markdown matches the canonical JSON dependency graph
Task count, layer count, and execution summary are derived from the canonical JSON, not manually maintained
Every filesToModify path exists
Every filesToCreate parent directory exists or is created by an earlier task
If a filesToModify path is created by an earlier task, the later task depends on that earlier task
Every local reuse reference exists in the repository; if not, mark it source: external
Every end-state claim in the overview traces to concrete tasks and prerequisites
Every cross-boundary noun in the plan appears in the Contracts section
Every new file/module has an explicit export/registration owner somewhere in the plan
**Depends on:** TASK-001, TASK-002 (regex: `/depends on:
Priority default
P2 when missing — always specify explicitly
Type values
feature, bug, chore, refactor, test, docs, infra
Effort values
S, M, L, XL
Task heading level
### (level 3) — minimum heading level 2
Disk format
No [PLAN]/[/PLAN] markers — those are for in-conversation display only
Dependency JSON
## Task Dependencies section with fenced JSON block — every task ID must be a key
Additional Optional Fields
These fields are supported when present:
**Type:** — feature | bug | chore | refactor | test | docs | infra (auto-inferred from heading/body if missing)
**Effort:** — S | M | L | XL
**Labels:** — comma-separated tags
**Agent:** — subagent type to execute this task (REQUIRED — see Agent Table)
**Review:** — post-task review tool: claude, codex, kiro, or none (see Post-Task Review below)
Post-Task Review
Every task can specify a **Review:** field that triggers an independent code review after the task agent completes. This catches bugs before they propagate to dependent tasks.
Review Tools
Value
Skill
What it does
claude
/claude-review
Spawns a separate Claude agent in a worktree to review the changes
codex
/codex-review
Runs OpenAI Codex CLI (codex review) against the task's commit
kiro
/kiro-review
Runs Kiro CLI (kiro-cli chat) with the diff
none
—
Skip review (use for trivial tasks like config/docs)
When to Assign Which Reviewer
Security-sensitive tasks (auth, crypto, secrets, permissions): codex — independent AI catches things Claude might miss
Complex logic tasks (parsers, state machines, concurrency): claude — deep context understanding
API/integration tasks: kiro — alternative perspective
Trivial tasks (rename, config change, docs): none
Critical P0 tasks: consider running multiple reviewers in sequence
How the Executor Uses This Field
The run-parallel-agents-feature-build skill (or manual execution) should:
Run the task agent
Check the review field
If not none, invoke the corresponding review skill on the task's commit using Skill("codex-review"), Skill("claude-review"), or Skill("kiro-review")
Report findings to the user
Fix findings before marking the task complete
IMPORTANT: The review field is a binding instruction to the executor, not a suggestion. When run-parallel-agents-feature-build processes this plan, it MUST invoke the specified tool via the Skill tool — not approximate it with a general-purpose agent prompt. If the review tool binary is not installed, the executor should warn the user rather than silently substituting.
Default
If **Review:** is omitted, default to none for S-effort tasks, claude for M/L/XL-effort tasks.
Quality Self-Check
Before outputting the final plan, verify ALL of the following:
Phase 0 was completed — user confirmed scope, selected mode, and chose review tool via AskUserQuestion
Mode (EXPANSION/HOLD/REDUCTION) is recorded in plan header and JSON
## Scope Challenge section documents what was considered and ruled out
## Prerequisites, ## Non-Goals, and ## Contracts are present and reflect the exploration findings
## Architecture section has an ASCII diagram with TASK-NNN labels
## Existing Code Leverage table maps sub-problems to reuse decisions
Local vs external reuse is distinguished correctly
All task IDs are sequential (TASK-001, TASK-002, ...)
All task IDs appear as keys in the ## Task Dependencies JSON block
No circular dependencies exist in the dependency graph
Every task has 2-3 testable acceptance criteria (at least 1 failure/edge case)
Every public surface task pins the exact signature/examples from spec/docs and includes a wrong-shape or wrong-routing check
Every rewrite/composition task includes at least 1 "absence of regression" acceptance criterion
Every task references specific file paths found during exploration
Every task has writeScope and validateCommand
Every new file/module is reachable via an explicit export/registration owner in the plan
Shared integration surfaces (package roots, module index files, export barrels, routers, registries, manifests, startup hooks) have explicit task ownership
Every side-effectful task states where its capability comes from (field, trait, callback, parameter, or runtime hook)
Every task has an **Agent:** field with a valid subagent type
No task touches more than 3 files
Foundation tasks (P0) have no dependencies (empty arrays in JSON)
Parallel groups have no mutual dependencies
Priorities use P0-P3 (not P1-P4)
Task count, layer count, and execution summary are derived from the dependency graph
## Failure Modes table lists risks with affected tasks and mitigations
## Ship Cut defines the minimum shippable subset
## Test Coverage Map maps every new codepath to a covering task and test type
Plan markdown has no [PLAN]/[/PLAN] markers
## Task Dependencies JSON block is present at the end
Every dependency target exists as a task ID
Every end-state claim in the overview traces to concrete tasks and prerequisites
In REDUCTION mode: no P3 tasks, no docs-only tasks, maximum reuse
In EXPANSION mode: comprehensive test coverage, edge case tasks included
Common Rationalizations (All Wrong)
These are excuses. Don't fall for them:
"I already know the scope" → STILL run Phase 0 scope challenge with the user
"The feature is straightforward" → STILL explore with CodeMap first
"There's no existing code to reuse" → STILL build the reuse audit table to prove it
"Failure modes are obvious" → STILL document them — agents need explicit guidance
"Tests can be added later" → STILL include test tasks and coverage map
"This is too small for a plan" → If it needs 3+ tasks, it needs a plan
"Dependencies are obvious" → STILL run dependency analysis — false assumptions kill parallelism
Failure Modes
Failure Mode 1: Skipping Scope Challenge
Symptom: Plan is too large, covers wrong scope, user pushes back after seeing output
Fix: Always run Phase 0. Present findings. Get mode confirmation.
Failure Mode 2: Phantom File Paths
Symptom: Tasks reference files that don't exist and weren't found during exploration
Fix: Every file path must come from CodeMap search or Glob results. Never invent paths.
Failure Mode 3: Over-Constrained Dependencies
Symptom: Tasks that could run in parallel are sequenced unnecessarily
Fix: Only declare dependencies for file overlap, data flow, API contracts, or state mutation.
Failure Mode 4: Missing Agent Assignment
Symptom: Tasks have no **Agent:** field, can't be dispatched to subagents
Fix: Every task gets an agent. Check against Agent Table.
Failure Mode 5: No Failure/Edge Case Criteria
Symptom: Acceptance criteria only test happy path, agents don't handle errors
Fix: At least 1 criterion per task must cover a failure or edge case.
Failure Mode 6: Markdown / JSON Drift
Symptom: Task counts, layers, dependencies, or task references disagree between the two files
Fix: Build canonical JSON first, then render markdown from it. Never hand-patch one without regenerating the other.
Failure Mode 7: Hidden Prerequisite
Symptom: Plan assumes a startup/runtime/public path already exists, but no task or prerequisite covers it
Fix: Add it to ## Prerequisites if already true, or add a prerequisite task if missing.
Symptom: A task "passes" by creating dead wiring, placeholder filters, or tests that only check node presence while real behavior is still wrong.
Fix: Split structural work from semantic-hardening work. Add negative acceptance criteria that prove projection, predicates, filtering, ordering, or visibility are actually preserved.
Failure Mode 9: Public Surface Drift
Symptom: A new public API/SQL surface ships with a stale signature or silent wrong-routing behavior because tests only cover one happy-path literal form.
Fix: Pin the exact public signature from spec/docs in the task text and require at least one wrong-shape/wrong-routing acceptance criterion.
Quick Workflow Summary
PHASE 0: SCOPE CHALLENGE (INTERACTIVE)
├── Quick CodeMap scan for existing overlap
├── Estimate complexity
├── Identify prerequisites + non-goals
├── AskUserQuestion: present findings + mode selection + review tool selection
└── Gate: User confirmed scope + mode + review tool + prerequisites framing
PHASE 1: EXPLORE
├── CodeMap search for feature-related code
├── Read ecosystem manifests/configs + directory structure
├── Build reuse audit + reality audit + contracts sketch
└── Gate: Concrete file paths + reuse audit + contracts
PHASE 2: DECOMPOSE
├── Break into atomic TASK-NNN entries
├── 2-3 acceptance criteria per task (≥1 failure case)
├── Add semantic bug-absence criteria for rewrites/composition tasks
├── Pin exact public signatures/examples for user-visible surfaces
├── Add write scope + validation command
├── Identify failure modes + cut line
├── Mode-aware pruning (REDUCTION/EXPANSION)
└── Gate: Atomicity checklist + failure modes + validation
PHASE 3: MAP DEPENDENCIES
├── Check: file overlap, data flow, API contract, state, bootstrap dependency
├── Minimize constraints for max parallelism
├── Compute execution layers
└── Gate: Valid DAG, no cycles, layers derived
PHASE 4: PRIORITIZE
├── Assign P0-P3
├── Form parallel groups
├── Define ship cut
└── Gate: Priorities + groups + ship cut
PHASE 5: GENERATE & VALIDATE
├── Canonical JSON first
├── Markdown rendered from JSON
├── Execution summary derived from DAG
├── Final lint checks
└── Gate: Both files saved, all sections present, lint passes
Resources
references/
knowledge.md — CodeMap tools reference, plan format parsing rules, DAG scheduling behavior, TaskDefinition interface
