Planning

Feature Planning Pipeline

Generate quality plans through systematic discovery, synthesis, verification, and decomposition.

Pipeline Overview

USER REQUEST → Worktree Setup → Discovery → Synthesis → Verification → Decomposition → Validation → Track Planning → Ready Plan

Phase 0: Worktree Setup (Mandatory)

Why: Beads are tracked in git. Without worktrees, branch switching causes conflicts when PRs merge.

Always create a worktree before creating beads for a feature:

# From main repo root
bd worktree create .worktrees/<feature-name> --branch feature/<feature-name>
cd .worktrees/<feature-name>

This creates a redirect file so all beads operations share the main repo's .beads/ database. No merge conflicts when PR lands.

After PR merges:

cd <main-repo>
git pull
bd worktree remove .worktrees/<feature-name>

Skip worktree only if: Quick fix on main that won't create new beads.

Phase 1: Discovery (Parallel Exploration)

Launch parallel sub-agents to gather codebase intelligence:

Task() → Agent A: Architecture snapshot (gkg repo_map)
Task() → Agent B: Pattern search (find similar existing code)
Task() → Agent C: Constraints (package.json, tsconfig, deps)
Librarian → External patterns ("how do similar projects do this?")
exa → Library docs (if external integration needed)

Discovery Report Template

Save to history/<feature>/discovery.md:

# Discovery Report: <Feature Name>

## Architecture Snapshot

- Relevant packages: ...
- Key modules: ...
- Entry points: ...

## Existing Patterns

- Similar implementation: <file> does X using Y pattern
- Reusable utilities: ...
- Naming conventions: ...

## Technical Constraints

- Node version: ...
- Key dependencies: ...
- Build requirements: ...

## External References

- Library docs: ...
- Similar projects: ...

Phase 2: Synthesis (Oracle)

Feed Discovery Report to Oracle for gap analysis:

oracle(
  task: "Analyze gap between current codebase and feature requirements",
  context: "Discovery report attached. User wants: <feature>",
  files: ["history/<feature>/discovery.md"]
)

Oracle produces:

1. Gap Analysis - What exists vs what's needed
2. Approach Options - 1-3 strategies with tradeoffs
3. Risk Assessment - LOW / MEDIUM / HIGH per component

Risk Classification

Risk Indicators

Pattern exists in codebase? ─── YES → LOW base
                            └── NO  → MEDIUM+ base

External dependency? ─── YES → HIGH
                     └── NO  → Check blast radius

Blast radius >5 files? ─── YES → HIGH
                       └── NO  → MEDIUM

Save to history/<feature>/approach.md:

# Approach: <Feature Name>

## Gap Analysis

| Component | Have | Need | Gap |
| --------- | ---- | ---- | --- |
| ...       | ...  | ...  | ... |

## Recommended Approach

<Description>

### Alternative Approaches

1. <Option A> - Tradeoff: ...
2. <Option B> - Tradeoff: ...

## Risk Map

| Component   | Risk | Reason           | Verification |
| ----------- | ---- | ---------------- | ------------ |
| Stripe SDK  | HIGH | New external dep | Spike        |
| User entity | LOW  | Follows existing | Proceed      |

Phase 3: Verification (Risk-Based)

For HIGH Risk Items → Create Spike Beads

Spikes are mini-plans executed via MULTI_AGENT_WORKFLOW:

bd create "Spike: <question to answer>" -t epic -p 0
bd create "Spike: Test X" -t task --blocks <spike-epic>
bd create "Spike: Verify Y" -t task --blocks <spike-epic>

Spike Bead Template

# Spike: <specific question>

**Time-box**: 30 minutes
**Output location**: .spikes/<spike-id>/

## Question

Can we <specific technical question>?

## Success Criteria

- [ ] Working throwaway code exists
- [ ] Answer documented (yes/no + details)
- [ ] Learnings captured for main plan

## On Completion

Close with: `bd close <id> --reason "YES: <approach>" or "NO: <blocker>"`

Execute Spikes

Use the MULTI_AGENT_WORKFLOW:

1. bv --robot-plan to parallelize spikes
2. Task() per spike with time-box
3. Workers write to .spikes/<feature>/<spike-id>/
4. Close with learnings: bd close <id> --reason "<result>"

Aggregate Spike Results

oracle(
  task: "Synthesize spike results and update approach",
  context: "Spikes completed. Results: ...",
  files: ["history/<feature>/approach.md"]
)

Update approach.md with validated learnings.

Phase 4: Decomposition (file-beads skill)

Load the file-beads skill and create beads with embedded learnings:

skill("file-beads")

Bead Requirements

Each bead MUST include:

• Spike learnings embedded in description (if applicable)
• Reference to .spikes/ code for HIGH risk items
• Clear acceptance criteria
• File scope for track assignment

Example Bead with Learnings

# Implement Stripe webhook handler

## Context

Spike bd-12 validated: Stripe SDK works with our Node version.
See `.spikes/billing-spike/webhook-test/` for working example.

## Learnings from Spike

- Must use `stripe.webhooks.constructEvent()` for signature verification
- Webhook secret stored in `STRIPE_WEBHOOK_SECRET` env var
- Raw body required (not parsed JSON)

## Acceptance Criteria

- [ ] Webhook endpoint at `/api/webhooks/stripe`
- [ ] Signature verification implemented
- [ ] Events: `checkout.session.completed`, `invoice.paid`

Phase 5: Validation

Run bv Analysis

bv --robot-suggest   # Find missing dependencies
bv --robot-insights  # Detect cycles, bottlenecks
bv --robot-priority  # Validate priorities

Fix Issues

bd dep add <from> <to>      # Add missing deps
bd dep remove <from> <to>   # Break cycles
bd update <id> --priority X # Adjust priorities

Oracle Final Review

oracle(
  task: "Review plan completeness and clarity",
  context: "Plan ready. Check for gaps, unclear beads, missing deps.",
  files: [".beads/"]
)

Phase 6: Track Planning

This phase creates an execution-ready plan so the orchestrator can spawn workers immediately without re-analyzing beads.

Step 1: Get Parallel Tracks

bv --robot-plan 2>/dev/null | jq '.plan.tracks'

Step 2: Assign File Scopes

For each track, determine the file scope based on beads in that track:

# For each bead, check which files it touches
bd show <bead-id>  # Look at description for file hints

Rules:

• File scopes must NOT overlap between tracks
• Use glob patterns: packages/sdk/**, apps/server/**
• If overlap unavoidable, merge into single track

Step 3: Generate Agent Names

Assign unique adjective+noun names to each track:

• BlueLake, GreenCastle, RedStone, PurpleBear, etc.
• Names are memorable identifiers, NOT role descriptions

Step 4: Create Execution Plan

Save to history/<feature>/execution-plan.md:

# Execution Plan: <Feature Name>

Epic: <epic-id>
Generated: <date>

## Tracks

| Track | Agent       | Beads (in order)      | File Scope        |
| ----- | ----------- | --------------------- | ----------------- |
| 1     | BlueLake    | bd-10 → bd-11 → bd-12 | `packages/sdk/**` |
| 2     | GreenCastle | bd-20 → bd-21         | `packages/cli/**` |
| 3     | RedStone    | bd-30 → bd-31 → bd-32 | `apps/server/**`  |

## Track Details

### Track 1: BlueLake - <track-description>

**File scope**: `packages/sdk/**`
**Beads**:

1. `bd-10`: <title> - <brief description>
2. `bd-11`: <title> - <brief description>
3. `bd-12`: <title> - <brief description>

### Track 2: GreenCastle - <track-description>

**File scope**: `packages/cli/**`
**Beads**:

1. `bd-20`: <title> - <brief description>
2. `bd-21`: <title> - <brief description>

### Track 3: RedStone - <track-description>

**File scope**: `apps/server/**`
**Beads**:

1. `bd-30`: <title> - <brief description>
2. `bd-31`: <title> - <brief description>
3. `bd-32`: <title> - <brief description>

## Cross-Track Dependencies

- Track 2 can start after bd-11 (Track 1) completes
- Track 3 has no cross-track dependencies

## Key Learnings (from Spikes)

Embedded in beads, but summarized here for orchestrator reference:

- <learning 1>
- <learning 2>

Validation

Before finalizing, verify:

# No cycles in the graph
bv --robot-insights 2>/dev/null | jq '.Cycles'

# All beads assigned to tracks
bv --robot-plan 2>/dev/null | jq '.plan.unassigned'

Output Artifacts

Quick Reference

Tool Selection

Common Mistakes

• Skipping discovery → Plan misses existing patterns
• No risk assessment → Surprises during execution
• No spikes for HIGH risk → Blocked workers
• Missing learnings in beads → Workers re-discover same issues
• No bv validation → Broken dependency graph