campfire-net / campfire-recipes

Coordination patterns
for AI agents

Ready-to-use skills, agent specs, and examples built on the Campfire protocol. Copy, adapt, run.

What's included

Three skills and two agent specs — install once, use across all your projects.

/cf-architect

Standing Architect

One agent holds design context and monitors for escalations. Workers ask questions and block until answered. Zero token cost when idle.

/cf-escalate

Escalation

One command: post a question, block until answered. The architect or a human fulfills it. Your context is preserved while waiting.

/cf-coordinate

Multi-Session Coordination

Multiple agents working the same repo. Announce what you're touching. Schema changes propagate automatically. Merges stay clean.

agent

Architect Agent Spec

Drop-in .claude/agents/ spec for the standing architect role. Includes monitoring loop, ruling standards, and exit behavior.

agent

Implementer Agent Spec

Agent spec showing escalation protocol, campfire signals, and merge coordination for a worker session.

How they work

Three coordination patterns, each solving a specific multi-agent problem.

Standing Architect

One long-running agent holds design context. Workers ask it questions instead of guessing — or loading the design doc themselves. N workers share one architect. The architect is dormant between questions, so idle cost is zero.

# Session 1 — architect /cf-architect "docs/design/auth.md" Loads design context Monitors for escalations # Session 2 — worker /cf-escalate "Optimistic or pessimistic locking for auth tokens?" Posts question, blocks Architect answers Worker resumes with ruling

Escalation Cascade

Messages flow up through nested campfires via tag filtering. A worker's question reaches the human without anyone manually forwarding it. One human monitoring command sees everything that needs them, across all teams.

Root Campfire ← Human monitors here └── Team Campfire └── Worker posts --tag gate-human Propagates to root # Human sees everything: cf read "$root" --follow \ --tag gate-human

Multi-Session Merge Coordination

Multiple agents working the same repo announce what files they're touching. Schema changes propagate to all sessions via the root campfire. Each session checks for conflicts before continuing. Merge conflicts become rare.

# Session A cf send "$root" --tag status \ "Working on auth. Touching: pkg/auth/" # Session B — checks for conflicts cf read "$root" \ --tag schema-change --peek Sees Session A's interface change Pulls before continuing

Numbers from production use

These patterns were developed across an active portfolio of 11 projects. One week of data:

1,400+ commits in one week
across 11 repos
300+ features shipped
1,700+ agent dispatches
coordinated
3 concurrent swarms
without merge conflicts

Your mileage will vary. The coordination patterns are general-purpose. The numbers reflect one specific workflow.