Imported from curiositech/some_claude_skills (.claude/skills/port-daddy/SKILL.md). Install upstream with npx skills add curiositech/some_claude_skills --skill port-daddy. Copyright stays with the author.

Port Daddy v3.8.2 — Agent Coordination That Actually Works

NOT for

• Production deployment orchestration, rollout policy, or CI release gating.
• Docker, Kubernetes, service-mesh, or cloud network discovery problems.
• Single-agent local work where no port coordination, salvage, or shared-state workflow exists.

The Problem You Have Right Now

You're an AI agent. You're about to start a dev server. Which port? 3000? Taken. 3001? Another agent grabbed it. You pick a random port. Now nothing can find your server.

Meanwhile, another agent is editing the same file you are. Neither of you knows. You'll both commit. One of you loses work.

A third agent crashed 20 minutes ago — halfway through a migration. Its work is orphaned. Nobody knows.

Port Daddy solves all of this in one daemon.

Decision Tree — Pick the Right Primitive

Read top-to-bottom. The first matching leaf is your answer. If you skip the daemon-up check, every subsequent decision is meaningless.

flowchart TD
  A[I want to do something coordination-shaped] --> Up{pd status<br/>== running?}
  Up -- no --> Down[examples/06-debug-daemon-down.md]
  Up -- yes --> Scope{Scope?}

  Scope -- single agent, no shared state --> Out[Out of scope: just code.<br/>NOT for this skill.]
  Scope -- this agent, this session --> Life[Lifecycle path]
  Scope -- two+ agents, same repo --> Coord[Coordination path]
  Scope -- background cadence / fleet --> Fleet[Fleet path]
  Scope -- many agents, same harbor --> Swarm[Swarm path]

  Life --> L1{What do I need?}
  L1 -- "session boundary" --> L1a[pd begin / pd done<br/>schemas/note-shape.md]
  L1 -- "deterministic port" --> L1b[pd claim {project}:{stack}:{ctx}<br/>schemas/semantic-identity.md<br/>examples/07-port-collision.md]
  L1 -- "audit trail" --> L1c[pd note --type ...<br/>assets/session-note.template.md]
  L1 -- "what happened?" --> L1d[pd sitrep / catch_me_up MCP]

  Coord --> C1{Failure mode I'm preventing?}
  C1 -- "two agents edit same file" --> C1a[pd session files claim<br/>examples/02-two-agents-same-file.md]
  C1 -- "must be exclusive" --> C1b[pd with-lock<br/>or pd lock + pd unlock]
  C1 -- "agent-to-agent signal" --> C1c[pd pub + pd watch]
  C1 -- "DM a specific agent" --> C1d[pd inbox send<br/>or talk_to_agent MCP]
  C1 -- "agent died, finish their work" --> C1e[pd salvage claim<br/>schemas/salvage-entry.md<br/>examples/03-salvage]

  Fleet --> F1{Fleet exists?}
  F1 -- "no" --> F1a[pd fleet init + edit pd-fleet.yml<br/>assets/pd-fleet.starter.yml<br/>examples/04-fleet-from-zero.md]
  F1 -- "yes, broken" --> F1b[scripts/fleet-validate.sh<br/>schemas/pd-fleet.schema.md]
  F1 -- "yes, want it daemon-managed" --> F1c[POST /fleet/register<br/>survives terminal close]
  F1 -- "yes, want to chain agents" --> F1d[on_success: publish channel<br/>+ declare in channels:]

  Swarm --> S1{Shape of the data?}
  S1 -- "typed records, multi-reader" --> S1a[pd tuple out + pd tuple rd<br/>schemas/tuple-shape.md]
  S1 -- "exactly-once work" --> S1b[pd tuple in<br/>work-stealing<br/>examples/05-tuple-swarm-handoff.md]
  S1 -- "ambient gradient / heat" --> S1c[pd pheromone spray<br/>schemas/pheromone-signal.md]
  S1 -- "broadcast, ephemeral" --> S1d[pd pub on a harbor channel]

The five questions, in order:

1. Daemon up? — pd status. If not, you have no coordination at all. Fix this first.
2. Scope? — single-session, two-agent, fleet (cadence), or swarm (many-in-harbor)? They use different primitives; mixing them is theatre.
3. What failure mode am I preventing? — silent overwrite, port conflict, dead-agent loss, missed signal? The failure mode chooses the primitive.
4. Durable or ephemeral? — notes (forever) and tuples-with-no-TTL (until-removed) are durable. Pheromones and pub/sub are ephemeral. Picking the wrong one either pollutes the audit trail or loses signal.
5. Read-many or take-once? — tuple rd and pub are fan-out; tuple in and lock are exactly-one-winner. Get this wrong and you either drop work or duplicate it.

Failure-Mode Triage

When something is already broken, jump straight here.

L3 Index — Where the Deep Knowledge Lives

This file is the L1/L2 entry. Don't paste deep references inline — load on demand:

Quick Start (Do This First)

# 1. Start your session — ALWAYS do this first
pd begin "Building auth module"

# 2. Claim a port — deterministic, never conflicts
PORT=$(pd claim myapp:api:main -q)

# 3. Leave breadcrumbs for other agents
pd note "JWT validation working, moving to refresh tokens"

# 4. Check who else is working here
pd salvage --project myapp    # Any dead agents to rescue?

# 5. End cleanly
pd done

Why This Matters

Without Port Daddy:

• Port conflicts every time two agents run dev servers
• No record of what agents did or decided
• Crashed agents leave orphaned work nobody finds
• No way for agents to signal each other
• File edit collisions destroy work silently

With Port Daddy:

• Deterministic ports — same identity always gets the same port
• Immutable notes — full audit trail of every decision
• Salvage queue — dead agent work is preserved and claimable
• Pub/sub + file claims — agents coordinate without stepping on each other
• Background fleet — QA, docs, testing run automatically on every commit
• Binary IPC — sub-microsecond heartbeats and pheromone sprays over Unix socket
• Pheromone trails — ambient numeric signals that decay over time for contention detection
• Tuple space — shared typed memory for swarm coordination
• Semantic trie — O(k) identity lookups replacing SQL LIKE scans

Shibboleths

• If the task only needs one temporary port and no shared session state, launching fleets and pheromone trails is theater, not coordination.
• If you need hard exclusion for a critical section, use a lock; advisory file claims are for negotiation, not safety.
• If the failure is in container ingress, DNS, or production service discovery, Port Daddy is the wrong layer.

MCP Tools Available

Start here (high-level, one call does many things):

Tuple space tools (shared swarm memory):

Discover more tools by category: Call pd_discover with a category name: magic, session-lifecycle, ports, sessions, notes, locks, messaging, agents, inbox, webhooks, integration, dns, briefing, tunnels, projects, changelog, activity, system, tuples, pheromone

Integration signals: Use integration ready and integration needs to coordinate service dependencies. When your service is ready, signal it so other agents can proceed.

Core Concepts

Semantic Identities: project:stack:context

Every service gets a semantic name. The name IS the port — deterministic hashing means the same identity always maps to the same port. Identities are indexed in an in-memory Adaptive Radix Tree for O(k) lookups (where k is key length), replacing SQL LIKE scans.

pd claim myapp:api:main           # Always gets port 3142 (or whatever hash gives)
pd claim myapp:api:feature-auth   # Different port, same project
pd find 'myapp:*'                 # Prefix search — resolves through the trie, not SQL
pd find 'myapp:*:main'            # Wildcard — all stacks with context "main"

Sessions & Notes

Sessions track what each agent is doing. Notes are immutable — once written, they can never be edited or deleted. This creates an audit trail that agents and humans can trust. Notes are encrypted at rest with AES-256-GCM (master key at ~/.port-daddy/master.key, auto-generated on first boot).

pd begin --identity myapp:api --purpose "Building auth"
pd note "Found SQL injection in token validation"
pd note "Patched. Tests green."
pd done

Salvage (Dead Agent Recovery)

When an agent crashes, its session enters the salvage queue. Another agent can claim and continue the work:

pd salvage --project myapp        # See dead agents' context
pd salvage claim dead-agent-42    # Pick up their work

IMPORTANT: Always check pd salvage at the start of a session. You might be able to continue where a crashed agent left off instead of starting from scratch.

File Claims (Advisory)

pd session files claim src/auth/*.ts
# Another agent tries the same file:
pd session files claim src/auth/login.ts
# → CONFLICT: claimed by agent 'myapp:api'

Claims are advisory — they warn, don't lock. Hard locks cause deadlocks. Advisory claims cause conversations.

Pub/Sub Messaging

Agents signal each other through channels:

# Agent A finishes database setup
pd pub myapp:events "database-ready"

# Agent B was watching
pd watch myapp:events --exec "npm run migrate"

Distributed Locks

For operations that truly must be exclusive:

pd with-lock deployment -- npm run deploy
# Or manually:
pd lock db-migration --ttl 300
pd unlock db-migration

Binary IPC Protocol (v3.8.2)

High-frequency agent communication over a Unix domain socket with MessagePack encoding. The IPC channel sits alongside the HTTP API — agents that need low-latency communication (heartbeats, pheromone sprays, pub/sub publish) use IPC automatically when the daemon is running.

Key properties:

• 7-byte header: [type:1][conv_id:4][payload_len:2] + MessagePack payload
• 70-80% bandwidth reduction vs HTTP JSON
• ~3us latency for fire-and-forget operations (vs ~200us HTTP)
• 13 FIPA performatives: INFORM, REQUEST, QUERY_REF, REFUSE, FAILURE, NOT_UNDERSTOOD, SUBSCRIBE, UNSUBSCRIBE, etc.
• Fire-and-forget: heartbeats, pheromone sprays, pub/sub publish (conv_id=0)
• Request-response: claims, locks, sessions (conv_id for correlation)
• Pub/sub subscriptions: with dead-man cleanup on disconnect
• Auto-reconnect: client reconnects with subscription replay on socket drop
• SDK fast paths: heartbeat(), pheromoneSpray(), publish() auto-use IPC when available

Socket location: ~/.port-daddy/daemon.ipc

Security hardening:

• Rate limiting: 500 frames/sec per connection
• Connection limit: 256 max (REFUSE for excess)
• 3-strike protocol violation budget (malformed frames disconnect)
• Backpressure via write queue + drain events
• Lock release on IPC disconnect

You don't need to use IPC directly. The SDK and CLI use it transparently for hot-path operations.

Fleet: Background Agents (v3.8.0)

Declare agents in YAML. They fire on git commits, cron schedules, or pub/sub messages. Auto-respawn on crash with circuit breaker.

pd fleet init     # Creates pd-fleet.yml + git hook
pd fleet up       # Starts the fleet
git commit -m "fix auth"  # QA, docs, cartographer fire automatically
pd fleet status   # What is the fleet doing?
pd fleet down     # Stop the fleet

The starter fleet includes: QA (bug hunting), Documentarian (docs sync), Cartographer (roadmap tracking), Spark (idea generation), Spider (cross-feature connections).

# pd-fleet.yml
fleet:
  name: myapp
  harbor: "{project}:fleet"
  agents:
    qa:
      trigger: git:committed        # React to pub/sub events
      respawn: true                  # Auto-restart on crash
      max_respawns: 3               # Circuit breaker
      backend: claude-cli
      allowedTools: "Read,Grep,Glob,Bash(npm test*)"
      prompt: "Review the last commit for bugs..."

    gardener:
      schedule: "*/10 * * * *"      # Or run on a cron schedule
      backend: custom
      prompt: "git status --porcelain"
      on_success: publish git:status  # Chain agents via channels

  channels:
    git:committed:
      description: "Fired after a successful commit"
      consumers: [qa]

Key features:

• Works with any LLM backend: claude-cli, ollama, gemini, aider, custom
• Template variables ({project}) resolve from the YAML context
• on_success: publish <channel> chains agents via pub/sub (DAG topology validated at startup)
• Fleet harbor auto-created on pd fleet up — all agents share a semantic namespace
• Each agent gets full PD coordination: registration, sessions, heartbeats, salvage on crash
• Auto-respawn with respawn: true and max_respawns circuit breaker

Tuple Space: Shared Swarm Memory (v3.8.0)

Agents write typed tuples to a shared space. Other agents query by pattern. Based on Linda (Gelernter, 1985). Harbor-scoped for fleet isolation. TTL for auto-expiry.

# Spider writes a connection it discovered
pd tuple out '["connection", "trie+pubsub=routing", "spider", 0.9]' --harbor myapp:fleet

# Spark reads all connections with confidence > 0.7
pd tuple rd '["connection", "*", "*", ">0.7"]' --harbor myapp:fleet

# Take (remove) a processed task from the space
pd tuple in '["task", "build-auth", "pending"]'

# Scan all tuples in a harbor
pd tuple scan --harbor myapp:fleet

# Count tuples
pd tuple count --harbor myapp:fleet

Pattern matching: exact values, * wildcard, >N/<N numeric comparisons, myapp:* semantic identity prefixes.

HTTP API:

• POST /tuples — write a tuple (body: { tuple, harbor?, writtenBy?, ttl? })
• GET /tuples — read by pattern (query: pattern, harbor, limit)
• DELETE /tuples — take (destructive read) by pattern
• GET /tuples/scan — list all tuples in a harbor
• GET /tuples/count — count tuples

Pheromone Trails: Ambient Signals (v3.8.0)

Agents spray numeric signals (0-1) onto entities. Signals decay exponentially over time at read, creating ambient awareness without polling.

# Spray a signal onto a service
pd pheromone spray --table services --id myapp:api --key urgency --strength 0.8

# Sniff pheromone values (applies read-time decay)
pd pheromone sniff --table services --id myapp:api

# View file heat map (which files are most contested)
pd pheromone files

# List all non-zero pheromone trails
pd pheromone list

Use cases: file contention detection, agent reputation scoring, hot-path identification, adaptive thresholds.

HTTP API:

• POST /pheromone/spray — set a pheromone value (body: { table, id, key, strength })
• GET /pheromone/:table/:id — read pheromone values (applies read-time decay)
• GET /pheromone — list all non-zero pheromones
• GET /pheromone/files — file heat map from session file claims (query: path, depth)

The Arbiter: Runtime Invariant Enforcement (v3.8.0)

The Arbiter monitors every state transition against 6 formally-derived invariants from the TLA+ specification:

• PID squatting — no process can claim another's port
• Capability escalation — agents can't exceed declared capabilities
• Note monotonicity — notes are append-only, never deleted
• Escrow positivity — encrypted note escrow balances stay positive
• Lock owner validity — only the owner can release a lock
• Heartbeat freshness — stale agents get reaped

In strict mode, critical violations trigger man-overboard salvage.

pd arbiter status         # Check rules and violation count
pd arbiter violations     # List recorded violations

Runtime File Locations (v3.8.2)

All runtime files live in ~/.port-daddy/ (not /tmp/). This eliminates symlink attacks, survives /tmp/ cleanup, and keeps permissions user-private (0700 directory).

Override via environment variables: PORT_DADDY_SOCK, PORT_DADDY_IPC, PORT_DADDY_PORT_FILE.

CLI Quick Reference

Decision Matrix: Which Tool When

Common Issues

Port Daddy daemon not running

Symptom: Connection refused on any pd command Fix: pd start or pd install (installs as launchd service, auto-starts on login)

Port already claimed

Symptom: You get a port but it's the "wrong" one This is correct behavior. Same identity = same port, always. If you need a different port, use a different identity context: myapp:api:feature-x instead of myapp:api:main.

Session already active

Symptom: pd begin says a session exists Fix: Call pd whoami to see the current session. Either pd done the old one or continue working in it.

File claim conflicts

Symptom: Another agent claimed files you need Fix: This is the system working. Check pd swarm_awareness to see who owns what. Coordinate via pd pub or work on different files.

IPC connection failures

Symptom: IPC-related errors in logs Fix: The SDK falls back to HTTP automatically. IPC is an optimization, not a requirement. Check that ~/.port-daddy/daemon.ipc exists and has correct permissions (should be user-only, created by the daemon).

Quality Gates

Before declaring "done" on any PD-mediated work:

• The agent registered an identity (pd whoami returns one).
• File claims match the files actually edited.
• At least one progress note per primitive milestone.
• A handoff note exists if the work is ending mid-stream (salvage-friendly).
• pd done was called — leaving sessions open is the most common dogfooding violation.