tl-sprint-planning

name	tl-sprint-planning
description	Use when planning sprints, organizing parallel work, or dispatching multiple agents. Covers wave planning, collision detection, and dependency analysis.

Tech Lead Sprint Planning

Core Principles

1. Parallel work that touches the same files creates rebase hell. Always check for file collisions before dispatching parallel agents.

2. Waves are heterogeneous. A single wave can mix Claude subtrees (expensive, architectural) and Gemini leaves (cheap, focused implementation), as long as files don't overlap.

Sprint Planning Protocol

1. Identify Ready Work

gh issue list --label "ready" # Show unblocked items
gh issue list                 # Identify bottlenecks

2. Classify Tasks

Before dispatch, classify each task:

Classification	Agent	Spawn Tool	Examples
Focused implementation	Gemini leaf	spawn_gemini (worktree)	single feature, bug fix, docs
Multi-step decomposition	Claude subtree	fork_wave	architecture, refactors requiring sub-spawns
Investigation / research	Gemini worker	spawn_gemini (inline)	hypothesis testing, codebase exploration

3. Collision Check (CRITICAL)

Map files each task will touch across ALL agents:

| Task | Agent Type | Subsystem | Files | Collision Risk |
|------|------------|-----------|-------|----------------|
| auth | leaf       | handlers  | events.rs, handler.rs | Group A |
| perms| leaf       | handlers  | handler.rs, auth.rs   | Group A ⚠️ |
| proto| leaf       | proto     | agent.proto, types.rs | Group B |
| docs | leaf       | docs      | CLAUDE.md             | Group C |

Rules:

• Same files → SERIALIZE, don't parallelize
• Different subsystems, shared file → Merge first PR before second
• Completely disjoint files → Safe to parallelize

4. Wave Organization

Wave N:
├── Gemini leaves (parallel, 3-5x)
│   ├── proto-plumbing (proto/, exomonad-proto/)
│   ├── docs-update (CLAUDE.md, haskell/CLAUDE.md)
│   └── test-coverage (tests/)
│
└── Claude subtree (1x, if needed)
    └── architecture refactor (may spawn its own leaves)

5. Dispatch

# Spawn leaves for focused tasks (own branch, files PR)
spawn_gemini(name="feature-x", task="Implement X", isolation="worktree")
spawn_gemini(name="feature-y", task="Implement Y", isolation="worktree")

# Spawn workers for investigation (ephemeral panes)
spawn_gemini(name="h1", task="Investigate hypothesis 1", isolation="inline")
spawn_gemini(name="h2", task="Investigate hypothesis 2", isolation="inline")

After spawning, return immediately. Idle until [PR READY], [FIXES PUSHED], or [from: agent] messages arrive.

Anti-Patterns

Parallel agents on same subsystem

Wave 1: auth-handler + perms-handler  # Both touch handler.rs
Result: Painful rebase conflicts

Correct Pattern

Wave 1a: auth-handler (handlers/)
Wave 1b: perms-handler (handlers/) — after 1a merges
Wave 1 parallel: proto-plumbing (proto/) — different subsystem

Merge Order Strategy

When parallel PRs exist:

1. Merge smallest/most isolated first — less conflict surface
2. Rebase next PR on fresh branch (or let Copilot flag conflicts)
3. Repeat

Commands Reference

# Planning
gh issue list                # Available work

# Dispatch
spawn_gemini(isolation="worktree")  # Gemini in worktree, files PR
fork_wave                           # Claude in worktree, can sub-spawn
spawn_gemini(isolation="inline")    # Gemini panes, ephemeral

# Monitoring — idle until messages arrive
# [PR READY] — Copilot approved, merge
# [FIXES PUSHED] — leaf addressed review, merge if CI passes
# [REVIEW TIMEOUT] — no review after timeout, merge if CI passes