discovery

name	discovery
description	Author an orrery plan — a YAML file under the orrery plans directory that the orrery CLI executes step-by-step. Use ONLY when the user explicitly asks to create an orrery plan (e.g., "make an orrery plan", "use orrery discovery", "/discovery"). Do NOT use for general planning, architectural discussion, or decomposing ideas unless the output is an orrery plan file.
hooks	{"PostToolUse":[{"matcher":"Write","hooks":[{"type":"command","command":"orrery validate-plan"}]}]}

Discovery Skill

When to Use

Use this skill for all planning requests, regardless of size. Discovery transforms ideas into concrete, executable plans.

Triggers:

• "Build a [system/platform/module]"
• Request spans multiple features or domains
• Unclear what "done" looks like
• Requires architectural decisions before implementation

Also use Discovery when:

• Request is a single, scoped feature
• Outcomes and scope are already clear
• You can articulate the task in 1-3 sentences

---

The Decomposition Ladder

Discovery works top-down through five levels:

IDEA        "We need better analytics"
   ↓
OUTCOMES    "Users can see trends; Admins can export reports"
   ↓
CAPABILITIES "Trend visualization; Data aggregation; Export service"
   ↓
FEATURES    "Line chart component; Daily rollup job; CSV export endpoint"
   ↓
STEPS       "1.1 Create chart component; 1.2 Add API route; 1.3 Wire up data"

Each level must be concrete enough that the next level can be derived. The final output is an orchestrator-ready plan with implementation steps.

---

How to Do It

Step 1: Capture the Idea

Get the raw vision from the user. Don't judge scope yet.

• What problem are we solving?
• Who is this for?
• What sparked this idea?

Step 2: Define Outcomes (the "why")

Outcomes are user-visible results, not technical deliverables.

Ask:

• "What will users be able to do that they can't today?"
• "How will we know this succeeded?"
• "What does 'done' look like from the user's perspective?"

Output: 2-5 concrete outcome statements.

Example:

• "Implement caching" (technical, not outcome)
• "Dashboard loads in under 2 seconds" (user-visible result)

Step 3: Map Capabilities (the "what")

Capabilities are system abilities that enable outcomes.

For each outcome, ask:

• "What must the system be able to do to deliver this?"
• "What new behaviors or services are needed?"

Output: Capabilities grouped by outcome.

Example:

• Outcome: "Dashboard loads in under 2 seconds"
  • Capability: Query result caching
  • Capability: Incremental data loading
  • Capability: Pre-computed aggregations

Step 4: Decompose into Features (the "how")

Features are implementable units of work that deliver capabilities.

For each capability, ask:

• "What specific features implement this?"
• "Can this be shipped independently?"
• "What's the minimum viable version?"

Output: Feature list with clear boundaries.

Step 5: Gather Context per Feature

This is critical for "fire and forget" plans. For each feature:

1. Search the codebase - find related files, patterns, dependencies
2. Identify constraints - what must this integrate with?
3. Define acceptance criteria - specific, testable conditions
4. Note risks - what could go wrong?
5. List input files - what must an agent read to understand context?

Step 6: Decompose Features into Implementation Steps

Each feature becomes 2-5 concrete implementation steps. This is what makes the plan orchestrator-ready.

For each feature, ask:

• "What are the individual pieces of work?"
• "What order must they happen in?"
• "Can any run in parallel?"

Step characteristics:

• Scoped - completable in a single focused session
• Specific files - lists exactly which files to create/modify
• Testable - has clear acceptance criteria
• Self-contained - an agent can execute without asking questions

Naming convention:

• Use {feature-number}.{step-number} format: 1.1, 1.2, 2.1, etc.
• Group related steps by feature for readability

Example decomposition:

Feature: "Line chart component"
  → Step 1.1: Create TrendChart.tsx with basic Chart.js setup
  → Step 1.2: Add time range toggle (7d/30d/90d)
  → Step 1.3: Connect to trends API endpoint
  → Step 1.4: Add loading state and error handling

Dependency Installation Steps

If a plan requires installing dependencies, handle them explicitly:

• Create a dedicated install step (e.g., "0.1 Install project dependencies") as the first step in the plan (after project creation/structuring if relevant)
• All subsequent steps MUST list the install step in their deps array - otherwise parallel execution or out-of-order execution will fail because dependencies aren't available
• Include test dependencies if acceptance criteria involve testing - use npm install --include=dev, pip install -e ".[test]", or the equivalent for the project's package manager

Example:

Step 0.1: Install project dependencies (including test frameworks)
  → deps: []
Step 1.1: Create TrendChart component
  → deps: ["0.1"]  ← Must reference the install step!
Step 1.2: Add unit tests for TrendChart
  → deps: ["0.1", "1.1"]  ← Also depends on install step

Step 7: Validate with User

Before producing the plan:

• Present the step breakdown (not just features)
• Confirm priorities and ordering
• Resolve any remaining ambiguities
• Get explicit sign-off that this captures the intent

Step 8: Output the Plan

Generate an orchestrator-ready plan file with implementation steps. Each step must be self-contained - an agent should be able to execute it without asking questions.

Use the schema defined in ./schemas/plan-schema.yaml.

Output Location:

First, determine the plans directory by running:

orrery plans-dir

This prints the resolved plans directory (respects ORRERY_WORK_DIR when set).

• Directory: the path returned by orrery plans-dir
• Filename: <date>-<plan-name>.yaml
• Date format: YYYY-MM-DD (e.g., 2026-01-11)
• Plan name: kebab-case description of the task (e.g., fix-clone-agent-skills)

Document Project Context in Notes:

Use the notes field in metadata to capture project-specific context that agents need throughout execution. This is critical because when plans are broken into substeps, agents executing individual steps might not know project conventions.

Include in notes:

• Testing commands: How to run tests (e.g., uv run pytest, npm test, make test) (check the README, pyproject.toml, package.json, etc. if applicable to understand: project manager, how to run tests, project conventions)
• Build commands: How to build the project (e.g., uv run build, npm run build)
• Linting/formatting: How to check code quality (e.g., uv run ruff check .)
• Environment setup: Required environment variables, virtual environments, etc.
• Project conventions: Naming conventions, file organization, coding standards
• Tool-specific notes: e.g., "Always prefix Python commands with 'uv run'"

Validate the Plan

Plans are automatically validated via the PostToolUse hook when written. For manual validation, run:

orrery validate-plan <plans-dir>/<plan>.yaml

This catches common YAML issues like unquoted colons and normalizes formatting.

YAML Formatting Rules

• Always quote strings containing special characters (colons, brackets, etc.)
• BAD: criteria: Output shows: timestamp value
• GOOD: criteria: "Output shows: timestamp value"
• Common gotchas: colons followed by space, special character prefixes, multi-line strings
• Rule: When in doubt, use double quotes around the entire value

---

Output Format

# plan.yaml
metadata:
  created_at: "2026-01-11T10:00:00Z"
  created_by: "Discovery-Agent"
  version: "1.0"
  source_idea: "We need better analytics"
  outcomes:
    - "Users can see usage trends over time"
    - "Admins can export reports for stakeholders"
  notes: |
    This project uses uv for dependency management.
    - Run tests: uv run pytest
    - Run linting: uv run ruff check .
    Always prefix Python commands with 'uv run'.

steps:
  # ============================================================================
  # Step 0: Setup (Dependencies)
  # ============================================================================

  - id: "0.1"
    description: "Install project dependencies including test frameworks"
    status: "pending"
    deps: []
    parallel: false
    context: |
      Install all dependencies required for the project. Include dev/test
      dependencies since acceptance criteria involve running tests.
    requirements:
      - "Run npm install (or equivalent for the project)"
      - "Include dev dependencies for testing"
    criteria:
      - "All dependencies install without errors"
      - "Test framework is available (e.g., jest, pytest)"
    files: []
    commands:
      - "npm install"

  # ============================================================================
  # Feature 1: Trends API Endpoint
  # ============================================================================

  - id: "1.1"
    description: "Create trends service with data aggregation logic"
    status: "pending"
    deps: ["0.1"]
    parallel: false
    context: |
      Stats are currently computed on-demand in statsService.ts. This step
      creates a new service that aggregates historical data into time-series
      format for the trends endpoint.
    requirements:
      - "Create src/api/services/trendsService.ts"
      - "Function: getTrends(range: '7d' | '30d' | '90d')"
      - "Returns { dates: string[], values: number[] }"
      - "Query existing stats table, group by date"
    criteria:
      - "Service exports getTrends function"
      - "Returns correctly shaped data for all range values"
      - "Unit test with mocked data passes"
    files:
      - "src/api/services/trendsService.ts"
      - "src/api/services/trendsService.test.ts"
    context_files:
      - "src/api/services/statsService.ts"

  - id: "1.2"
    description: "Add trends API route with caching"
    status: "pending"
    deps: ["1.1"]
    parallel: false
    context: |
      Wire up the trends service to an HTTP endpoint. Use existing cache
      middleware pattern from other routes.
    requirements:
      - "GET /api/stats/trends?range=7d|30d|90d"
      - "Cache responses for 1 hour"
      - "Validate range parameter"
    criteria:
      - "Endpoint returns 200 with valid JSON"
      - "Invalid range returns 400"
      - "Response time < 200ms (cached)"
    files:
      - "src/api/routes/stats.ts"
    context_files:
      - "src/api/middleware/cache.ts"

  # ============================================================================
  # Feature 2: Trend Visualization Component
  # ============================================================================

  - id: "2.1"
    description: "Create base TrendChart component with Chart.js"
    status: "pending"
    deps: ["0.1", "1.2"]
    parallel: false
    context: |
      Users currently see only current-day stats. This adds a line chart
      using the existing Chart.js setup in src/components/charts/.
    requirements:
      - "Create TrendChart.tsx extending BaseChart"
      - "Line chart with responsive sizing"
      - "Accept data prop: { dates: string[], values: number[] }"
    criteria:
      - "Component renders with mock data"
      - "Chart displays correctly at mobile and desktop widths"
    files:
      - "src/components/TrendChart.tsx"
    context_files:
      - "src/components/charts/BaseChart.tsx"
      - "src/styles/charts.css"
    risk_notes: "Chart.js bundle size - verify no significant increase"

  - id: "2.2"
    description: "Add time range toggle to TrendChart"
    status: "pending"
    deps: ["2.1"]
    parallel: false
    context: |
      Add toggle buttons for 7d/30d/90d. Selecting a range should trigger
      a data refetch.
    requirements:
      - "Toggle buttons: 7 days, 30 days, 90 days"
      - "Active state styling for selected range"
      - "onChange callback when range changes"
    criteria:
      - "Toggle switches time range"
      - "Visual indication of selected range"
    files:
      - "src/components/TrendChart.tsx"

  - id: "2.3"
    description: "Connect TrendChart to API and add loading states"
    status: "pending"
    deps: ["2.2"]
    parallel: false
    context: |
      Wire up the component to fetch from the trends API. Handle loading
      and error states gracefully.
    requirements:
      - "Fetch from GET /api/stats/trends on mount and range change"
      - "Loading skeleton while fetching"
      - "Error state if API fails"
    criteria:
      - "Data loads from API on initial render"
      - "Range change triggers new API call"
      - "Loading skeleton appears during fetch"
      - "Error message displays on API failure"
    files:
      - "src/components/TrendChart.tsx"
      - "src/components/TrendChart.test.tsx"

---

When Discovery is Complete

Discovery is complete when:

• All outcomes are defined and user-validated
• Each outcome maps to concrete capabilities
• Each capability has implementable features
• Each feature is decomposed into implementation steps
• Each step has sufficient context for autonomous execution
• The plan file passes schema validation
• User has approved the plan

Final Output

When the plan is complete and validated, output the plan file path and present the user with their next options:

Plan created: <plans-dir>/<date>-<plan-name>.yaml

Next steps:
- /refine-plan <plans-dir>/<plan-file> — Analyze and improve the plan before execution
- /simulate-plan <plans-dir>/<plan-file> — Explore the plan through dialogue, ask "what if" questions
- orrery exec — (Command run from the terminal) Execute the plan with the orrery orchestrator

This ensures the user knows exactly what they can do next and has the full path to reference the plan.

---

Common Pitfalls

• Skipping outcome definition: Jumping to features without knowing "why" leads to building the wrong thing
• Thin context: A description like "Add caching" isn't enough. Include what, where, why, constraints.
• Implicit dependencies: If feature B needs feature A's output, say so explicitly
• No user validation: Don't assume you understood the idea correctly. Confirm the decomposition.
• Premature detail: Don't write implementation code during Discovery. Just define what to build.
• Missing dependency step dependencies: If step "0.1" installs dependencies, all steps that use those dependencies must include "0.1" in their deps array. Otherwise, parallel execution or out-of-order execution will fail.
• Forgetting test dependencies: If acceptance criteria include running tests, ensure the dependency installation step includes dev/test dependencies (e.g., npm install --include=dev, pip install -e ".[test]").

---

Dependency and Parallelization Guidelines

Dependency Detection Rules

When determining deps for each step, apply these rules:

1. Installation/Setup First: Any step that creates files depends on the step that installs dependencies or sets up the project structure.

2. File Creation Before Use: If step B reads or modifies files created by step A, then B depends on A. Check the files field.

3. Import Dependencies: If step B's code will import from modules created in step A, then B depends on A.

4. Test Infrastructure: Steps that include tests depend on the step that sets up test infrastructure (fixtures, mocks, config).

5. API Before Consumers: Backend API endpoints must exist before frontend components that call them. Frontend steps depend on relevant backend steps.

6. Schema Before Implementation: Database schema/migrations must exist before repository code. Model definitions before API routes.

Example analysis:

• Step "Create REST endpoints" lists files: routes/items.py
• Step "Create Items Browser UI" will call these endpoints
• Therefore: UI step depends on REST endpoints step

Parallelization Rules

Mark a step as parallel: true ONLY when ALL of these are true:

1. No file overlap: The step's files do not overlap with any other parallel-eligible step at the same dependency level.

2. No logical dependency: The step does not use outputs, types, or patterns established by a peer step.

3. Independent domain: The step works in a different area of the codebase (e.g., frontend vs backend, different features).

Common parallel-safe patterns:

• Backend setup + Frontend setup (different directories)
• Independent feature implementations (no shared files)
• Documentation + Docker config (no code dependencies)

Common parallel-unsafe patterns:

• Two steps modifying the same configuration file
• API routes that share model definitions
• Components that import from each other

Default to parallel: false when uncertain. Sequential execution is always safe; incorrect parallelization causes merge conflicts and bugs.

Plan ordering matters: Even with parallel: true, steps respect plan order. A serial step earlier in the plan acts as an implicit barrier for later steps. Place foundational work early in the plan.

Project Structure Analysis

Before creating the plan, understand the existing project structure:

1. Check for existing conventions: If there's an existing src/ or lib/ directory, new code goes there. Don't create parallel structures.

2. Read configuration files: Check package.json, pyproject.toml, or similar for project name, paths, and conventions.

3. Follow existing patterns: If the project has src/components/, new components go there. Don't create components/ at the root.

4. Respect monorepo structure: If there's backend/ and frontend/, keep that separation. Don't create new top-level directories.

5. Check CLAUDE.md or similar: Project instructions often specify where new code should go.

When the files field lists paths, use paths that match existing structure. Example: If project has backend/src/, use backend/src/models/ not models/.