iris_x_hermes/docs/03-gateway-plugin.md

03 — Gateway Plugin (Python)

The plugin is a community-style hermes platform plugin named iris. It follows the "Plugin Path" in hermes-agent/gateway/platforms/ADDING_A_PLATFORM.md and the canonical example hermes-agent/plugins/platforms/irc/adapter.py. Zero hermes-core changes. Zero new Python dependencies (websockets and httpx are already core deps).

Source map for every hermes integration point: see 15-hermes-reference.md.

3.1 plugin.yaml (manifest)

name: iris-platform
label: Iris
kind: platform
version: 0.1.0
description: >
  Native Android / Desktop client gateway adapter for Hermes Agent.
  Runs a WebSocket server inside the gateway; the app connects with a
  pairing token. Supports streaming, reasoning, structured tool events,
  channels/threads, media, FTS5 search, and FCM/ntfy push.
author: <you>
requires_env:
  - name: IRIS_TOKEN
    description: "Shared pairing token the app presents on connect"
    prompt: "Iris pairing token"
    password: true
optional_env:
  - name: IRIS_HTTP_HOST
    description: "HTTP bind host (default 127.0.0.1; use 0.0.0.0 for LAN)"
    prompt: "HTTP host"
    password: false
  - name: IRIS_HTTP_PORT
    description: "HTTP port (default 8791)"
    prompt: "HTTP port"
    password: false
  - name: IRIS_HOME_CHANNEL
    description: "Default chat id for cron/notification delivery (default default)"
    prompt: "Home channel"
    password: false
  - name: IRIS_ALLOWED_USERS
    description: "Comma-separated allowed device_ids (empty = token-only auth)"
    prompt: "Allowed device ids"
    password: false
  - name: IRIS_ALLOW_ALL_USERS
    description: "Allow any paired device (dev only)"
    prompt: "Allow all devices? (true/false)"
    password: false
  - name: IRIS_PUSH_BACKEND
    description: "Push backend: ntfy (default, keeps metadata off Google) or fcm"
    prompt: "Push backend"
    password: false
  - name: IRIS_FCM_SERVICE_ACCOUNT
    description: "Path to Firebase service-account JSON (FCM HTTP v1)"
    prompt: "FCM service account path"
    password: true
  - name: IRIS_FCM_SERVER_KEY
    description: "Legacy FCM server key (fallback if no service account)"
    prompt: "FCM server key"
    password: true
  - name: NTFY_TOPIC
    description: "ntfy topic for push (when IRIS_PUSH_BACKEND=ntfy)"
    prompt: "ntfy topic"
    password: false
  - name: NTFY_SERVER_URL
    description: "ntfy server URL (default https://ntfy.sh)"
    prompt: "ntfy server URL"
    password: false
  - name: IRIS_HTTP_CERT
    description: "TLS cert path for HTTPS (optional)"
    prompt: "HTTPS cert"
    password: false
  - name: IRIS_HTTP_KEY
    description: "TLS key path for HTTPS (optional)"
    prompt: "HTTPS key"
    password: false

Behavioral (non-secret) settings live in config.yaml under gateway.platforms.iris.extra (host, port, home_channel, outbox retention, max upload bytes, tls). Secrets live in .env. (hermes policy: .env = secrets only.)

3.2 register(ctx) entry point

def register(ctx):
    ctx.register_platform(
        name="iris",
        label="Iris",
        adapter_factory=lambda cfg: IrisAdapter(cfg),
        check_fn=check_requirements,          # passive: websockets importable + token set
        validate_config=validate_config,      # host/port/token present
        is_connected=is_connected,
        required_env=["IRIS_TOKEN"],
        install_hint="No extra packages needed (websockets + httpx are core deps)",
        setup_fn=interactive_setup,           # hermes gateway setup flow
        env_enablement_fn=_env_enablement,    # seed extra + home_channel from env
        cron_deliver_env_var="IRIS_HOME_CHANNEL",
        standalone_sender_fn=_standalone_send,  # best-effort out-of-proc cron (stretch)
        parse_target_ref_fn=_parse_target_ref,  # "iris:<chat>[:<thread>]"
        allowed_users_env="IRIS_ALLOWED_USERS",
        allow_all_env="IRIS_ALLOW_ALL_USERS",
        max_message_length=0,                 # 0 = no limit (WS has none)
        emoji="📱",
        pii_safe=False,
        platform_hint=(
            "You are chatting with the user through their native Iris app "
            "(Android/Desktop). It renders Markdown, inline code, images, "
            "audio and video, and shows your reasoning and tool activity. "
            "Conversations are organized into channels and optional threads. "
            "Keep formatting rich but readable."
        ),
    )

Field reference (all from PlatformEntry, gateway/platform_registry.py:63): adapter_factory, check_fn, validate_config, is_connected, required_env, install_hint, setup_fn, env_enablement_fn, apply_yaml_config_fn, cron_deliver_env_var, parse_target_ref_fn, allowed_users_env, allow_all_env, max_message_length, pii_safe, platform_hint, emoji, ensure_deps_fn.

• check_requirements() — passive probe: import websockets succeeds and IRIS_TOKEN is set. Never installs.
• _env_enablement() — returns a dict seeding PlatformConfig.extra (host/port/home_channel/push_backend) + a home_channel key {"chat_id": "default", "name": "Default"} so hermes gateway status and cron home-channel resolution work without instantiating the adapter.
• _parse_target_ref(ref) — the core strips the platform prefix first, so ref is the direct chat id (e.g. chan_7, default) with an optional :t_<n> thread suffix; friendly names resolve via the channel directory. Returns (chat_id, thread_id) or None.
• interactive_setup() — prompts for token (or generates one), host/port, push backend + credentials, prints a QR code (pairing) and the app URL.

3.3 IrisAdapter(BasePlatformAdapter)

Constructor: super().__init__(config=config, platform=Platform("iris")). Reads config.extra (env overrides win). Initializes: WS server (not started until connect()), connection registry, outbox (SQLite under get_hermes_home()/"iris"), push backend, pairing store, channel directory.

Lifecycle

• connect(*, is_reconnect=False) -> bool
  • Acquire scoped lock (gateway.status.acquire_scoped_lock("iris", key)) so two profiles can't bind the same port/identity.
  • Start the websockets server on host:port (TLS if cert/key set).
  • _mark_connected(); return True.
• disconnect()
  • Stop server, close all device sockets, release lock, _mark_disconnected().

Inbound (app → agent)

• WS message.send {text, reply_to?, media_refs?} → build SessionSource via self.build_source(chat_id, chat_name, chat_type, user_id, user_name, thread_id) → build MessageEvent(text=…, message_type=TEXT, source=…, media_urls=[cached paths], media_types=[…], reply_to_message_id=…) → await self.handle_message(event).
• Slash commands arrive as plain text starting with /; the gateway's command pipeline resolves + dispatches them (no special handling needed).
• picker.select {picker_id, value} → route to the gateway-side resolvers (model picker / choice picker / clarify / approval / slash-confirm) using the shared callback-id conventions (cl:<id>:<idx>, appr:<id>:<choice>, sc:<choice>:<id>).
• channel.create / channel.rename / channel.set_default → mutate the channel directory (SQLite) + emit channel.* frames to all devices.
• search {query, scope, chat_id?, thread_id?} → search.py → search.results.
• media.upload (chunked) → media.py → cache_*_from_bytes → media_ref.
• media.pull {media_id} → stream cached bytes as binary frames.
• fcm.register {token} / hello → update device registry.
• read.receipt {message_id} → mark delivered/read (drives ✓✓), ack.
• sync {cursor} → outbox.py → replay frames since cursor.

Outbound (agent → app)

• send(chat_id, content, reply_to=None, metadata=None) -> SendResult
  • Split reasoning prefix (see 05-streaming.md) → reasoning field.
  • If any device is connected: broadcast message frame to all.
  • Else (no live devices): append to outbox + fire push (FCM/ntfy).
  • Return SendResult(success=True, message_id=<id>).
• edit_message(chat_id, message_id, content) → message.update frame (drives streaming). If no device, no-op (outbox holds the final send).
• send_typing(chat_id, metadata=None) → typing frame.
• get_chat_info(chat_id) -> dict → {"name": <channel name>, "type": "dm"|"channel"} from the channel directory.
• Media send — send_image / send_video / send_document / send_voice / send_image_file / send_multiple_images: stage the file in the media cache, mint a media_id, emit media.offer {media_id, mime, size, filename, kind}, serve bytes on media.pull. (Base-class extract_media/extract_images already pull MEDIA:/image tags out of agent text and call these.)
• Interactive pickers — send_model_picker(...), send_choice_picker(...), send_clarify(...), send_exec_approval(...), send_slash_confirm(...): emit picker.model / picker.choice / picker.clarify / picker.approval / picker.confirm frames with options; store pending state keyed by picker_id; resolve on picker.select.
• create_handoff_thread(chat_id, name) → create a thread id, register in channel directory, return it (used by cron "continuable" threads).

Streaming hooks

The main gateway drives delivery through the legacy callback path:

• stream_delta_callback → GatewayStreamConsumer → send() (first) + edit_message() (updates) → message.start / message.update.
• tool_progress_callback → progress queue → send_progress_messages → send() → tool.* frames (structured; classified in the adapter).
• interim_assistant_callback → consumer on_commentary → send() → commentary frames.

The adapter tracks per-chat turn state (in-turn, current streaming message_id, last tool index) to classify outbound send() calls into message vs tool.* vs commentary. The exact classification markers are verified empirically in M2 (see 13-testing.md).

3.4 HTTP server (http_server.py)

• Library: stdlib http.server (ThreadingHTTPServer + BaseHTTPRequestHandler) in a daemon thread; bridges into the gateway's asyncio loop via asyncio.run_coroutine_threadsafe. Optional TLS via ssl.SSLContext (IRIS_HTTP_CERT/IRIS_HTTP_KEY). Full design: 19-http-fallback-transport.md.
• Auth: Authorization: Bearer <token> (constant-time verify_token)
  • device allowlist via X-Iris-Device; 401 on failure.
• Endpoints: GET /v1/health (unauthenticated liveness), POST /v1/frame (any JSON frame the protocol accepts), GET /v1/events?cursor=N (SSE: outbox catch-up + live frames), GET /v1/poll?cursor=N (long-poll fallback), POST /v1/media + GET /v1/media/{id} (media upload/pull).
• Routing: emit(chat_id, frame) → broadcast to all connected devices (no per-chat subscribe; single-user model). Global frames (channel.*, status) also broadcast to all.
• Limits: 64 KiB request body cap, per-device token-bucket rate limit (20/s, burst 40) → 429; media uploads bounded by the per-upload total cap. No CORS (app clients only).

3.5 State & storage (all under get_hermes_home()/"iris")

Use get_hermes_home() from hermes_constants for all paths (profile-safe). Never hardcode ~/.hermes.

• devices.db — device registry (device_id, name, caps, fcm_token, ntfy_topic, last_seen, created).
• channels.db — channel directory (chat_id, name, kind: default|channel|thread, parent_chat_id, created, is_default).
• outbox.db — undelivered frames per chat_id + monotonic cursor.
• media/ — inbound + outbound media cache (reuse hermes cache_*_from_bytes dirs where possible).

3.6 Config resolution

• Secrets (.env): IRIS_TOKEN, IRIS_FCM_SERVICE_ACCOUNT, IRIS_FCM_SERVER_KEY, IRIS_HTTP_CERT/KEY, NTFY_TOPIC (if secret).
• Behavioral (config.yaml → gateway.platforms.iris.extra): host, http_port, home_channel, allowed_users, push_backend, outbox_retention_hours, max_upload_bytes, http_cert/http_key.
• Env vars override config.yaml (hermes convention). Read secrets with the scope-aware _get_scoped_secret pattern (see plugins/platforms/irc/adapter.py:42) so multiplexed profiles don't leak each other's tokens.

3.7 Failure & lifecycle safety

• HTTP server bind failure → non-fatal: log a warning, disable the HTTP leg, show it in the inspector (the plugin keeps working for other platforms).
• All outbound sends are best-effort; a dead stream latches and the frame falls to the outbox.
• disconnect() stops the HTTP server and closes streams cleanly.
• Token/PII redaction in all logs (hermes PII policy).