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.

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.

What works

• Typing indicators while Hermes works.
• Reactions in both directions.
• Tool-approval answers. Reply exactly /approve, /approve session, /approve always, or /deny.
• message.failed events are logged, not rejected.

Edits and unsend are not supported.

Install

Supported Hermes versions are 0.19.0 or newer, tested through 0.21.3. A newer minor version warns and runs. Install Hermes itself from upstream source, not a wheel.

Connect with the Relay CLI, then start the gateway:

npx relaymessenger@staging connect hermes
hermes gateway run

The CLI installs and enables the plugin. It writes the connection settings to <HERMES_HOME>/.env and offers to start the gateway for you.

To install the plugin manually instead:

hermes plugins install RelayMessenger/Relay-Hermes --enable

Then set the connection settings below in <HERMES_HOME>/.env.

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.

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

HERMES_HOME selects the Hermes home directory. It defaults to ~/.hermes. The CLI writes these settings to <HERMES_HOME>/.env:

• RELAY_AGENT_TOKEN is the required Agent Token.
• RELAY_BASE_URL is the agent's API origin. The adapter default is https://api.relayapp.im.
• RELAY_STATE_DIR is the durable inbox directory, set to <HERMES_HOME>/relay.
• RELAY_ALLOWED_CONTACTS is an optional comma-separated list of Contact ids allowed to start turns. The CLI writes it only when set; unset allows all reachable Contacts.

Additional adapter settings:

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 commands work like on Telegram. To restrict who may run them, set allow_admin_from (and optionally user_allowed_commands) under gateway.platforms.relayapp.extra in Hermes's config, using Relay Contact ids.

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, use these Hermes display settings. 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="${HERMES_HOME:-${HOME}/.hermes}/relay-staging"
./scripts/run-staging.sh

The staging helper refuses every other API origin.

Locked Relay contract

Current contract validation is pinned to Relay Server developer OpenAPI commit 4fe4a71e20f4f51e5f8db5714c985687c8a63d89. The exact contracts/developer/openapi.yaml SHA-256 is 46eeedd5a5e99e879e32c45972799364021143df9f81acd60837713210639735. Those exact public bytes are checked in at contracts/relay-server/4fe4a71e20f4f51e5f8db5714c985687c8a63d89/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.

Releasing

Nothing is published by hand. There are two lanes, the same two Relay-SDK uses for npm, in PEP 440:

1. Every merge to staging publishes a development build. The Publish staging to PyPI workflow decides the next version from what PyPI holds, commits release: version the staging package automatically to staging (it rewrites pyproject.toml and plugin.yaml, nothing else), runs the full CI at that commit, and uploads the wheel and sdist. The version is the target with a .devN suffix: 1.0.0rc3.dev0, 1.0.0rc3.dev1, and so on while the target is the candidate 1.0.0rc3; 1.0.0.dev0 once the target is 1.0.0.
2. A promotion is a pull request from staging to main, merged without a merge commit, so main is always a commit staging already has. The Release to PyPI workflow on main strips .devN, publishes the target (1.0.0rc3, later 1.0.0), reads it back from PyPI, and records the tag v1.0.0rc3. A target PyPI already has is skipped, so a second promotion of the same tree publishes nothing.

pip install relay-hermes never selects a .dev release. pip install --pre relay-hermes does, and pip install relay-hermes==1.0.0rc3.dev0 names one exactly. plugin.yaml carries the same version in its manifest form: 1.0.0-rc.3.dev.0 for 1.0.0rc3.dev0, 1.0.0-rc.3 for 1.0.0rc3, 1.0.0-dev.0 for 1.0.0.dev0.

To aim at a different target, edit the version in pyproject.toml (and the manifest form in plugin.yaml) in a normal pull request to staging: 1.0.0 or 1.0.0.dev0 ends the candidate line and starts the GA line, 1.1.0 starts a minor. The staging lane keeps a hand-written unpublished version as written and carries on from there. The rules are the table at the top of scripts/release_version.py, and every pull request rehearses both lanes against live PyPI without publishing (Rehearse the staging bump and Rehearse the release in CI).

Credentials

Both lanes check the distributions through one shared step (.github/actions/check-dist) and then upload with the same two steps, one per credential. Which one runs is the repository variable PYPI_TRUSTED_PUBLISHING:

• unset (today): the project-scoped PYPI_API_TOKEN secret, attestations off (PEP 740 attestations only work through Trusted Publishing);
• true: PyPI Trusted Publishing over OIDC, no token, attestations on.

To switch, register two trusted publishers on the relay-hermes project (PyPI: project settings, Publishing, GitHub), one per lane. Every field is exact:

Then set the repository variable PYPI_TRUSTED_PUBLISHING to true (GitHub: Settings, Secrets and variables, Actions, Variables). No file changes. The two workflow names are the workflows that call the upload action directly, which is what PyPI matches; a reusable workflow cannot be a trusted publisher, and the upload action cannot run from inside a composite action. The pypi-staging and pypi-release environments are where any approval rule goes; the older pypi-rc environment belongs to the manual Publish release candidate workflow, which stays until the two lanes have each published once and is then removed.

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.