Launch Review Pipeline

A structured, multi-phase agent pipeline that evaluates feature readiness through adversarial debate, creates a reviewed action plan, and executes fixes. Each phase runs as an isolated team to prevent context bloat.

Invocation

/launch-review [topic] [--scope "focus areas"] [--constraints "known constraints"]

Examples:

• /launch-review "checkout experience" --scope "cart, payment, confirmation"
• /launch-review "v2 API launch" --constraints "backwards compatible, feature flag gated"

Phase 0: Constraints Gathering (No Team)

Before spawning any agents, the main agent MUST:

1. Ask the user for known constraints:

• Legal restrictions (prohibited libraries, data sources, licenses)
• Business rules (budget limits, timeline, must-have vs nice-to-have)
• Scope boundaries (what's in/out of this review)
• Any prior decisions already made

1. Check memory for relevant project constraints
2. Compile a constraints block that gets injected into every agent's prompt

Format:

CONSTRAINTS (non-negotiable):
- [constraint 1]
- [constraint 2]
DO NOT propose solutions that violate these constraints.

This phase prevents the #1 failure mode: agents recommending something that's already been ruled out.

Phase 1: Debate Team

Create team: TeamCreate with name {topic}-debate

Agents (5):

Flow:

1. Spawn all 5 agents. The 4 debaters run in parallel. Moderator is blocked by all 4.
2. Each debater sends their report to moderator via SendMessage.
3. Moderator writes the report to docs/{TOPIC}_READINESS_REPORT.md.
4. Moderator sends summary to team-lead.

Handoff artifact: docs/{TOPIC}_READINESS_REPORT.md

Team transition checklist:

• [ ] Report file exists on disk
• [ ] Moderator confirmed report is written
• [ ] Shutdown all agents
• [ ] TeamDelete

Phase 2: Planning Team

Create team: TeamCreate with name {topic}-planning

Agents (2):

Flow:

1. Spawn planner first. It reads the report from disk, creates docs/{TOPIC}_ACTION_PLAN.md.
2. Once planner delivers, shut it down and spawn reviewer.
3. Reviewer reads the plan, verifies against actual code, appends review section.
4. Reviewer sends summary to team-lead.

The plan MUST include:

• All issues from the report, prioritized
• Specific files/lines affected
• Agent team structure for execution (using TeamCreate, NOT standalone subagents)
• Phasing and dependencies
• Acceptance criteria per task

Handoff artifact: docs/{TOPIC}_ACTION_PLAN.md (with review section)

Team transition checklist:

• [ ] Plan file exists on disk with review section
• [ ] Reviewer confirmed approval (with or without changes)
• [ ] Shutdown all agents
• [ ] TeamDelete

Phase 3: Execution Team

Create team: TeamCreate with name {topic}-execution

Agents: Per the plan's team structure. Typically:

• One engineer per workstream (worktree isolation)
• Each reads the action plan from disk

Critical rules:

Agents MUST commit their work

Every code-writing agent must git add + git commit in their worktree before reporting done. Uncommitted worktree changes are lost on cleanup.

Merge before delete

Before TeamDelete, a merge agent MUST:

1. List all worktrees with git worktree list
2. For each agent worktree:

• Verify changes are committed
• Merge the worktree branch into main
• Remove the worktree

1. Run git worktree prune
2. Delete merged branches
3. Confirm git status is clean

NEVER call TeamDelete while worktrees contain unmerged work.

Failure recovery

If an agent fails or auth drops:

• Check if worktree has uncommitted changes before re-running
• If changes exist, commit them first
• Respawn the agent pointing at the same worktree branch

Handoff artifact: Clean main branch with all fixes merged.

Phase 4: Backlog (No Team)

Main agent writes docs/{TOPIC}_BACKLOG.md capturing:

• What was completed
• What remains (with priority)
• Action items requiring human intervention
• Known constraints for future sessions

Anti-Patterns (Learned the Hard Way)

1. Never TeamDelete before merging worktrees. This destroyed completed work in the first run of this pipeline.
2. Never skip Phase 0. Three layers of agents all recommended a legally prohibited library because no one told them the constraint.
3. Decision agents must wait for all inputs. The manager made a call before receiving the legal report. Use addBlockedBy on tasks to enforce sequencing.
4. Don't create teams you don't need yet. One-team-at-a-time constraint means premature creation forces premature deletion.
5. Debaters and planners don't need worktrees. They write docs, not code. Only execution agents need worktree isolation.
6. Keep the main agent's context lean. It delegates, routes, and summarizes. It does not read files, grep code, or write implementations.