Cooperative Vibe Coding

Cooperative vibe coding scales Andrej Karpathy's "vibe coding" to multiple humans and AI agents working on the same codebase simultaneously across machines. Choose sync vs async collaboration modes based on team constraints and coordinate multiple AI agents without conflicts.

DECISION POINTS

1. Sync vs Async Mode Selection

Team Size ≤ 2, Same Timezone:

• If high-complexity/unfamiliar code → Real-time collaboration (Live Share/Zed)
• If independent features → Parallel worktrees + 30min sync points
• If production-critical → True pair programming with shared screen

Team Size 3-5, Mixed Timezones:

• If <4hr overlap → Async: PR-based with detailed commit messages
• If 4-6hr overlap → Hybrid: 2hr sync blocks + async handoffs
• If >6hr overlap → Mob programming with 10min rotations

Team Size >5:

• Always async with clear ownership boundaries
• Use AI agents for parallelization, humans for coordination

2. AI Agent Coordination Strategy

Single Feature, Multiple Devs:

If feature can split cleanly:
  → Each dev gets worktree + dedicated agent
  → File-level claims prevent conflicts
  
If feature is intertwined:
  → One shared agent, multiple human reviewers
  → Driver controls agent, navigator reviews output

Multiple Features, Multiple Agents:

If features share interfaces:
  → Designate interface owner, others depend on them
  → Sync interface changes before implementing
  
If features are independent:
  → Full isolation via worktrees/branches
  → Integration branch for final merge

3. Tool Selection Matrix

Scenario	Primary Tool	Voice	Why
2 devs, Windows+Mac, <2hr session	VS Code Live Share	Discord/Zoom	Universal compatibility
2-4 devs, all Mac/Linux, >2hr	Zed	Built-in	Lowest latency, native voice
Terminal-heavy workflow	tmux/SSH	External	Zero editor lag
Cross-timezone async	Git + Loom	Async video	Time-shifted collaboration
AI-heavy session	Cursor/Claude Code	External	AI-native editors

FAILURE MODES

"Merge Conflict Hell"

Symptoms: Integration takes longer than development, constant git conflicts, agents overwriting each other Diagnosis: No file-level coordination, sync points too infrequent (>45min), overlapping worktree scopes Fix: Implement file claiming (Port Daddy/manual), reduce sync to 25min, use git worktree for true isolation

"Spectator Sport Pairing"

Symptoms: Long silences, one person passive, AI generates everything while humans watch Diagnosis: No active collaboration protocol, unclear driver/navigator roles, agents doing too much Fix: 25min rotation timer, navigator must review AI output aloud, limit agent scope to functions not features

"Tool Configuration Spiral"

Symptoms: 30+ minutes setting up Live Share, audio issues, extension conflicts, "it works on my machine" Diagnosis: Over-optimization for perfect setup vs. practical collaboration Fix: 5-minute setup rule - if tools don't work quickly, fall back to screen share + voice

"Latency Death Spiral"

Symptoms: Frustrated typing, cursor lag >500ms, constant "wait, what did you just do?", productivity drops Diagnosis: Too many concurrent editors, poor network conditions, wrong tool for bandwidth Fix: Check network (ping test), reduce active collaborators to 2-3, switch to screen share if lag persists

"Context Fragmentation"

Symptoms: Frequent "wait, what are we building again?", duplicate work, misaligned implementation approaches Diagnosis: No shared mental model, insufficient sync points, parallel work without coordination Fix: Start each session with 5min goal alignment, document decisions in shared notes, 30min status checks

WORKED EXAMPLES

Scenario 1: 4 Devs, 3 Timezones, Authentication Feature

Setup: Alice (US West), Bob (US East), Chen (EU), David (Asia). Building JWT auth system with login/logout/refresh endpoints.

Decision Process:

• Timezone overlap analysis: Alice+Bob have 8hr overlap, Chen has 4hr with both, David has 2hr with Chen only
• Feature complexity: Auth requires shared types, middleware patterns - moderately intertwined
• Tool selection: Mix of sync and async due to timezone constraints

Session Structure:

# Day 1: Alice + Bob (sync block, 2 hours)
# Set up auth architecture, define interfaces

git checkout -b feature/auth-system
# Create shared types and middleware stubs
# Alice: middleware/auth.ts, Bob: routes/auth.ts
# Sync every 30min, commit WIP to shared branch

# Handoff to Chen (async)
git commit -m "WIP: auth middleware scaffolded, routes stubbed
- AuthMiddleware interface defined in types/auth.ts  
- validateJWT() needs implementation
- TODO: Chen implement token refresh logic in refresh.ts"

Chen's Session (4 hours later):

git pull origin feature/auth-system
# Chen works alone with Claude agent
claude "implement JWT refresh token logic, follow patterns in middleware/auth.ts"
# Chen tests, commits with detailed message for David
git commit -m "Refresh token implemented
- Added refreshJWT() to middleware/auth.ts
- Created /auth/refresh endpoint  
- TODO: David add logout cleanup + session management"

Integration Issues Caught:

• Alice's middleware expected userId field, Bob's routes provided user_id (snake_case vs camelCase)
• Detected during Chen's session when TypeScript failed
• Fixed by Chen, documented in commit message

What Novice Would Miss: No coordination protocol, each dev working in isolation, merge conflicts on shared files, interface mismatches discovered late.

What Expert Catches: Clear handoff messages, interface alignment checks, timezone-aware sync points, WIP commits that enable async collaboration.

Scenario 2: Real-time Debugging Session (Agent Deadlock)

Setup: Two devs + two AI agents hit deadlock during real-time collaboration. Both agents trying to modify same function simultaneously.

Symptoms Detected:

# Terminal output shows conflict:
[Claude-A] Modifying src/utils/validation.ts lines 23-45
[Claude-B] Modifying src/utils/validation.ts lines 30-50  
[Git] error: Your local changes would be overwritten by merge
[LiveShare] Conflict in validation.ts

Expert Decision Process:

1. Immediate: Stop both agents (Ctrl+C in both terminals)
2. Diagnose: Check git status and Live Share file locks - both agents claimed overlapping ranges
3. Recover: Alice takes manual control, Bob's agent moves to different file
4. Prevent: Implement file-level agent claiming going forward

Recovery Sequence:

# Alice (taking control)
git stash  # Save Claude-A's work
# Bob 
git checkout HEAD src/utils/validation.ts  # Discard Claude-B's changes

# Coordination
Alice: "I'll finish validation.ts manually, can your agent work on auth.ts instead?"
Bob: "Yes, redirecting Claude to authentication logic"

# Resume with isolation
# Alice: manual coding on validation.ts
# Bob's Claude: working on src/middleware/auth.ts
# Sync point in 25min to review both changes

What Novice Would Miss: Panic, try to force-merge conflicts, restart entire session, blame tools.

What Expert Catches: Quick isolation, manual override of agents, clear communication, prevention protocol for future.

QUALITY GATES

Before starting any cooperative coding session:

• Voice channel established and tested - can both parties hear clearly without lag?
• Work scope defined in 1 sentence - "We're implementing X feature with Y constraints"
• File ownership mapped - who/which agent touches which directories/files?
• Sync cadence agreed - every 25min for real-time, daily for async handoffs
• AI agent boundaries set - which agent works where, who reviews output?
• Conflict resolution protocol ready - who breaks ties on technical decisions?

During the session:

• Changes committed every 30min max - no work exists only in memory/unsaved files
• Integration tested before handoff - if async, does the code build/run for next person?
• Decision rationale documented - why did we choose approach X over Y?

After the session:

• Deliverable completed or next session scheduled - no dangling incomplete work
• Shared context updated - README, docs, or team notes reflect what was built

Deterministic Session Readiness Audit

The Quality Gates above and the failure modes in this skill are encoded as a runnable check in scripts/session_readiness.mjs. It exports auditCoopSession(plan), taking a JSON session plan matching schemas/session-plan.schema.json (team size, timezone overlap, mode, voice-channel/ownership/ boundary/cadence flags) and returning { pass, findings, recommendations }. It flags exactly the failure modes documented above: Merge Conflict Hell (no file ownership mapped, or sync cadence over 45min), Spectator Sport Pairing (no agent boundaries/reviewer set), Context Fragmentation (no 1-sentence scope, or no sync points for a sync/hybrid session), and a mode mismatched to team shape (team size over 5, or under 4 hours of timezone overlap, without mode: "async"). See examples/sample-input.json for a plan that passes cleanly.

node scripts/session_readiness.mjs --input examples/sample-input.json

NOT-FOR BOUNDARIES

Do NOT use cooperative vibe coding for:

• Building collaboration infrastructure itself - For implementing CRDTs, operational transform, or real-time sync engines → use real-time-collaboration-engine
• Git workflow design decisions - For branching strategies, PR templates, merge policies → use git-best-practices
• Solo AI prompt engineering - For optimizing prompts for single-user coding, not a cooperative session
• Code review processes - For establishing review standards and approval workflows, distinct from live pairing/mobbing
• Team onboarding or training - For teaching junior developers → use dedicated mentorship approaches

When to delegate instead:

• Complex merge conflicts involving 5+ files → escalate to git expert or architectural review
• Performance issues with collaboration tools → IT support or tool vendor
• Team communication breakdown → team lead or manager intervention
• Large-scale refactoring across multiple teams → architectural planning session

<!-- BEGIN BUNDLE INDEX (auto: index_references.py) -->

Skill Bundle Index

Every file in this skill, and when to open it. Auto-generated; run scripts/index_references.py --fix.

examples/

• examples/sample-input.json — sample input (data/schema)

schemas/

• schemas/session-plan.schema.json — session plan.schema (data/schema)

scripts/

• scripts/session_readiness.mjs

<!-- END BUNDLE INDEX -->