Relay-Hermes

Relay-Hermes connects Hermes Agent to Relay as an always-on messaging platform.

Relay events arrive over the Relay v1 acknowledged WebSocket. Hermes replies are sent through Relay's REST API.

Delivery model

For each Relay event, the plugin:

1. commits the complete envelope and event_id to a durable SQLite inbox;
2. sends a cumulative WebSocket ACK only after the commit;
3. deduplicates replayed event_id values;
4. starts a Hermes turn only for an inbound message.received event;
5. explicitly marks the Chat Read at intake, as soon as the message is accepted for Hermes (before Hermes decides whether it starts a new turn or folds the message into one already running);
6. sends each Message through POST /v1/chats/{chatId}/messages with an Idempotency-Key.

Transport acknowledgement and Read are separate. A WebSocket ACK never marks a Chat Read.

Relay-Hermes does not poll for events and does not expose a public HTTP server. It also does not add reactions, edits, or typing indicators.

Recovery

The SQLite inbox survives gateway restarts. If processing fails, the event returns to pending; the durable transport checkpoint does not move backward.

The state directory and SQLite database are bound to one Relay account using the normalized API origin and a one-way Agent Token fingerprint. The token is never written to state. Changing the token or API origin while reusing a state directory fails before startup requeues work or reads a FULL-sync snapshot. On supported Linux systems every directory component and the database are opened through pinned descriptors with no-follow checks before SQLite receives an already-open file descriptor path. SQLite never connects to the mutable configured pathname and cannot create through a dangling database symlink or a replaced directory. The directory is forced to mode 0700, including when it already exists. Use a separate RELAY_STATE_DIR for every staging or production Agent. Pre-binding databases are not adopted automatically; move one aside only after accounting for its pending work.

If Relay reports that a checkpoint is outside retention, the plugin follows the WebSocket FULL-sync flow. It pages through visible Chats and Messages, validates the snapshot, atomically stores it with full_sync_through, and sends full_sync_complete only after the transaction commits. Historical Messages rebuild local indexes and never become new Hermes turns.

Install

Hermes supports Git-installed directory plugins:

hermes plugins install RelayMessenger/Relay-Hermes --enable

The same repository can be installed as a Python package. Its hermes_agent.plugins entry point registers the identical platform adapter.

Create an Agent and copy its Agent Token from Relay Console, then save the token in ~/.hermes/.env:

RELAY_AGENT_TOKEN=your_agent_token

An Agent using this WebSocket must not have a saved webhook subscription. Relay rejects the WebSocket upgrade with HTTP 409 until those subscriptions are removed.

Start Hermes in the foreground:

hermes gateway run

For an always-on installation, use Hermes's service commands:

hermes gateway install
hermes gateway start

The Hermes platform id is relayapp. Hermes already reserves relay for its generic connector platform.

Configuration

RELAY_AGENT_TOKEN is the only required setting.

RELAY_BASE_URL must be an HTTPS origin, except that HTTP is accepted for loopback development. A configured invalid value fails closed and is never replaced with the production default.

Slash-command limitation

Relay Contact chat is enabled, but Relay slash commands are disabled on the pinned Hermes core. Its slash-policy resolver reads the gateway runner's primary platform config and ignores source.profile, so a primary profile operator grant could otherwise authorize a secondary-profile Contact. The adapter therefore withholds every Relay message whose first non-whitespace character is / from Hermes, installs deny-only slash policy as defense in depth, and registers relayapp as ineligible for /update. This is consistent for primary and secondary profiles; use a trusted local CLI or authenticated dashboard for operator commands.

RELAY_OPERATOR_CONTACTS and operator_contacts are not supported. Any old setting should be removed; it cannot safely grant Relay slash authority until the pinned Hermes policy becomes profile-aware.

Non-secret settings can instead be placed under gateway.platforms.relayapp.extra in ~/.hermes/config.yaml; environment variables take precedence:

gateway:
  platforms:
    relayapp:
      enabled: true
      extra:
        allowed_contacts:
          - 01993d50-ef7b-7b37-886b-23fd80c7ec12
        group_chat_policy: mentions
        reply_to_mode: auto

Display defaults

Hermes keeps a built-in table of per-platform display defaults (gateway/display_config.py); its iMessage adapters (bluebubbles, photon) sit in the quiet tier there, with tool progress, interim commentary, heartbeats, and streaming previews off, because those inboxes cannot edit a message once sent. A plugin cannot add a row to that table, so relayapp inherits the global defaults, which were written for platforms that edit in place.

The adapter itself refuses tool-progress lines (format_tool_event returns None). Everything else is a user setting. To read like iMessage, set these under display.platforms.relayapp in ~/.hermes/config.yaml (key names from hermes_cli/config_defaults.py and gateway/display_config.py; an unset key falls through to the global display.<key>):

display:
  platforms:
    relayapp:
      tool_progress: off
      interim_assistant_messages: false
      long_running_notifications: false
      streaming: false
      busy_ack_detail: false
      tool_preview_length: 0

display.busy_input_mode (default interrupt) is global, not per platform, and is left alone: a message that arrives while Hermes is mid-turn is folded into that turn, queued behind it, or (in queue mode) buffered for 0.35 s and then queued. The adapter marks it Read at intake and leaves the reply unquoted, so it reads as it does on iMessage. Its durable inbox row is settled only when the session goes quiet: at the successful end of the turn that consumed it, unless Hermes still holds it for a later turn, in which case that later turn settles it. If the gateway dies before then the row replays on restart, which is the safe side.

Hermes core posts a busy acknowledgement bubble when a message lands mid-turn ("Redirected current run", "Interrupting current task"); its iMessage adapters show the same bubble, and this plugin does not suppress it. Hermes's own environment variable turns it off for every platform (gateway/run_busy.py: os.environ.get("HERMES_GATEWAY_BUSY_ACK_ENABLED", "true").lower() != "true" suppresses the ack):

export HERMES_GATEWAY_BUSY_ACK_ENABLED=false

Relay resolves the token, API origin, chat allowlist, state directory, and delivery settings through Hermes's shared adapter credential reader. The primary profile uses its own process environment during unscoped startup. A secondary profile's installed secret scope remains authoritative: a missing value never falls through to the primary profile's process environment. The default state path is under the active profile home, so profile inboxes are distinct even when neither profile sets RELAY_STATE_DIR.

Isolated staging

Use a staging Agent Token, the staging API origin, and a separate inbox:

export RELAY_AGENT_TOKEN='staging-agent-token'
export RELAY_BASE_URL='https://api.staging.relayapp.im'
export RELAY_STATE_DIR="$HOME/.hermes/relay-staging"
./scripts/run-staging.sh

The staging helper refuses every other API origin.

Locked Relay contract

This version was audited against Relay Server developer OpenAPI commit 99906995625ddc00348064a585ada1649313b0fc. The exact contracts/developer/openapi.yaml SHA-256 is 7094178cb01c0ddc05f9254dc91094900a0a7b6273979c0cad6257eec486f0d8. Those exact public bytes are checked in at contracts/relay-server/99906995625ddc00348064a585ada1649313b0fc/openapi.yaml; normal CI and RC publication validate that local snapshot without private repository credentials.

The runtime contract used here is:

• GET /v1/websocket for acknowledged agent events;
• GET /v1/chats and GET /v1/chats/{chatId}/messages only for a server-directed FULL sync;
• POST /v1/chats/{chatId}/read with no request body at intake;
• POST /v1/attachments, followed by the allocated raw PUT, for local files;
• POST /v1/chats/{chatId}/messages with Idempotency-Key for every Message send;
• event envelopes with api_version: v1 and webhook_version: 2026-08-30.

Relay vocabulary in the adapter is Contact, Handle, Chat, Message, and Participant.

Development

Python 3.11 through 3.13 are supported.

python -m venv .venv
.venv/bin/pip install -e '.[dev]'
HERMES_AGENT_SRC=/path/to/hermes-agent .venv/bin/python -m pytest
python -m build
hermes plugins doctor . --ci

Tests cover REST paths and bodies, Message idempotency, explicit Read timing, FULL-sync recovery, WebSocket authentication and acknowledgement ordering, replay deduplication, heartbeat and reconnect behavior, package installation, and Hermes directory-plugin registration.