cloud-communication-protocols git:20260825.7587472A

Immutable. This exact content is served forever at /api/v1/blob/f270eb75bf78d3d9.

---
name: cloud-communication-protocols
language: en
visibility: public
description: Umbrella skill for cloud-bridged communication protocols between agents on different machines (Ping-Pong, agent-beam, listeners and future protocols). Use when coordinating work across machines via a shared sync folder or message yard.
---

<img src="banner.png" width="100%" alt="cloud-communication-protocols banner">

# cloud-communication-protocols

Family of protocols that let two or more agents on different machines
coordinate through a shared, cloud-synchronized folder (the "yard") instead of
a direct channel. This skill is the **umbrella**: it collects the protocols,
names which one is proven, and gates the ones still in concept.

## When to use

- Work spans two or more machines (verification, mirrored setups, pilots)
  and there is no direct agent-to-agent channel.
- You need a common vocabulary for assignments, answers, cadence and
  escalation across those machines.

## Protocols in this family

### 1. Ping-Pong (proven, base protocol)

Two or more workers with offset scheduled scans bridge the cloud-sync
latency: each side scans the yard for assignments addressed to itself,
executes, leaves the answer, and may issue follow-ups for the other side.

Core rules: write only in your own slot, receipts for every action,
idempotency, no secrets in the yard (public keys/fingerprints/paths only),
cooperative error deltas, readback-over-log verification.

**Cadence: adaptive polling frequency.** Cadence control is not part of the
base protocol's payload rules, but every implementation needs it — a fixed
polling interval either floods an idle yard or misses fresh work. Two
mechanisms apply and must both be handled, not left implicit:

- **A — Self-tuning (default, no operator input; the `cooldown`/backoff
  pattern).** New assignment found → jump straight to the fastest interval
  (e.g. 15 min). Empty run → cool down on a ladder: 4 consecutive empty runs
  → 30 min, 2 more → 1 h, then double each further empty streak (2 h, 4 h,
  8 h) up to a hard cap (24 h, never higher). Persist the empty-run counter
  and current interval in a small state file at the worker (not in
  conversation memory), so it survives restarts; a cadence change means
  deleting the old scheduled job and creating a new one at the new interval,
  counter carried over. Full reference scheme, state-file schema and further
  loop types (burst, wake-assist): companion skill **`cron-tuner`**.
- **B — Peer-instructed override.** Either side may set the other side's
  cadence explicitly in the assignment text — `CADENCE: match 15m` to sync
  to a tighter window, or `CADENCE: pause 1h → 2h after 3 empty` to slow a
  side doing predictable heads-down work. Whoever changes a peer's cadence
  confirms the new interval in the next delta (readback applies to clocks
  too).

Both mechanisms only change reaction time, never the base-protocol rules
(slot discipline, receipts, no secrets). A mutual `WAKE: <reason>` channel
allows out-of-cycle scans as a hint, never as authority; delete after
reading.

Full specification and reference evidence:
`dev-bricks/system-gap-master` → `docs/communications-protocols-skill.md`.

### 2. agent-beam (concept, gated)

For urgent work: a package of prompt + starter script is placed in the
target slot, and a watcher on the target machine starts a local agent run
on it — the agent "lands" with assignment and starter and begins immediately.

Gated: requires a signed-starter trust contract, quarantine/review step and
explicit operator approval before any activation.

### 3. listeners / ear-to-ear-listening (concept, gated)

Watchers observe triggers in the yard (file arrivals, flag files, registry
changes) and start agents on the other machine. "Ear-to-ear": one host's
listener watches the other host's inbox.

Gated: debouncing rules, trigger whitelists per slot and a security contract
are required first.

## Shared invariants (all protocols)

- The yard is transport, never a workspace and never a secret store.
- Every action leaves a verifiable receipt; "done" without an artifact is
  not done.
- Failure handling is cooperative: error deltas with hypothesis and
  counter-test, second opinions welcome.
- Absence of a partner is handled by the platform's vacancy rule
  (e.g. 48 h), not by escalation.

## Adding new protocols

New protocols join this family as a subchapter in
`communications-protocols-skill.md` (concept first, with its own security
contract and evidence before "proven"). List them here with one line each:
name, state (concept/pilot/proven), purpose.