### IM Channels System (`app/channels/`)

Bridges external messaging platforms (Feishu, Slack, Telegram, Discord, DingTalk, GitHub) to the DeerFlow agent via Gateway's LangGraph-compatible API.

**Architecture**: Channels communicate with Gateway through the `langgraph-sdk` HTTP client (same as the frontend), ensuring threads are created and managed server-side. The internal SDK client injects process-local internal auth plus a matching CSRF cookie/header pair so Gateway accepts state-changing thread/run requests from channel workers without relying on browser session cookies.

**Components**:
- `message_bus.py` - Async pub/sub hub (`InboundMessage` → queue → dispatcher; `OutboundMessage` → callbacks → channels)
- `store.py` - JSON-file persistence mapping `channel_name:chat_id[:topic_id]` → `thread_id` (keys are `channel:chat` for root conversations and `channel:chat:topic` for threaded conversations)
- `manager.py` - Core dispatcher: creates threads via `client.threads.create()`, routes commands including `/goal` (setting a goal persists it through Gateway and then routes the objective as a chat turn), keeps Slack/Discord on `client.runs.wait()`, uses `client.runs.stream(["messages-tuple", "values"])` for Feishu/Telegram incremental outbound updates, serializes same-thread Feishu turns in-manager when the channel's `ChannelRunPolicy.serialize_thread_runs=True` so rapid follow-ups queue instead of tripping the runtime busy reply, and switches to `client.runs.create()` (fire-and-forget, returns once the run is `pending`) for channels whose `ChannelRunPolicy.fire_and_forget=True` so long autonomous runs do not hit the SDK default 300s `httpx.ReadTimeout`
  A swallowed streaming failure publishes its final outbound before releasing the inbound dedupe key, so a provider redelivery can retry without overtaking the terminal reply.
  **What may be published from the stream is an allowlist, not a denylist** (`_accumulate_stream_text`): only assistant message types — LangChain serializes `AIMessage.type` as `"ai"` and `AIMessageChunk.type` as `"AIMessageChunk"`, plus the OpenAI-style `"assistant"` spelling for foreign runtimes — become displayable text. The previous rule rejected only payloads whose `type` contained `"tool"` and therefore published everything else, which leaked DeerFlow's hidden model context to every streaming IM channel: `DynamicContextMiddleware` injects recalled memory as a hidden `HumanMessage` (`type == "human"`) and rewrites the user's own turn into a new `HumanMessage`, `DurableContextMiddleware` injects a hidden `<durable_context_data>` `HumanMessage`, and LangGraph fans those state writes out on the `messages-tuple` stream. Proved live on a Buzz relay, which published a `<memory>` fact block and, in another run, a verbatim echo of the user's own message as the assistant's reply. Matching is by prefix (`ai` / `assistant`), never substring, because ordinary words contain `"ai"` (`chain`, `domain`). The message type is resolved by `_stream_payload_type`, which handles both the `model_dump()` shape DeerFlow's own gateway emits and LangChain's `to_json()` constructor shape (whose top-level `type` is the literal `"constructor"`, with the class name at the tail of the `id` path). A bare `str` payload is no longer accepted at all: it carries no type information, so it cannot be attributed to the assistant, and nothing in DeerFlow produces one (`runtime/serialization.py::serialize_messages_tuple` always emits `[message_dict, metadata]`).
- `base.py` - Abstract `Channel` base class (start/stop/send lifecycle). Provider callbacks that submit coroutines from SDK threads must use `_submit_threadsafe_coroutine()`: it creates and retains the real `asyncio.Task` on the owner loop instead of treating `run_coroutine_threadsafe()`'s proxy Future as a completion signal. Submission is closed atomically with shutdown, and `stop()` must call `_close_and_drain_threadsafe_futures()` before tearing down SDK resources.
- `service.py` - Manages lifecycle of all configured channels from `config.yaml`. Shutdown closes manager admission first and keeps transports alive until every manager worker/follow-up watcher has exited. A successful manager stop therefore owns no live handler; if the Gateway's outer timeout cancels shutdown, the service retains its channel objects and global singleton so unfinished resources are not detached and cleanup can be retried.
- `slack.py` / `feishu.py` / `telegram.py` / `discord.py` / `dingtalk.py` - Platform-specific implementations (`feishu.py` tracks the running card `message_id` in memory and patches the same card in place; `telegram.py` accepts inbound text/photos/documents, preserves media captions, hands token-free attachment bytes to the shared upload pipeline, edits the "Working on it..." stream target in place via `editMessageText`, and can optionally send final Markdown replies as Rich Messages through `channels.telegram.rich_messages`; `discord.py` registers typing loops before inbound handling yields and `_start_typing()` refuses work once `_running` is false; because `stop()` runs on the main loop while typing tasks belong to `_discord_loop`, normal cross-thread cancellation, awaiting, and map cleanup are scheduled there with a bounded wait, while `_run_client()` drains tasks in its `finally` block before an exception/disconnect can make that loop unusable; an already-stopped foreign loop must never have its tasks awaited from the main loop; `dingtalk.py` optionally uses AI Card streaming for in-place updates when `card_template_id` is configured, and overrides `receive_file` to download inbound images (`picture`/`richText`) and documents (`file`) by `downloadCode` into the thread uploads bucket, mirroring `feishu.py`)
- `buzz.py` - Buzz (Nostr relay) implementation: one NIP-42-authenticated websocket, pubkey-allowlist + mention/DM/thread-follow gating, streaming replies via in-place kind-40003 edits; requires the `buzz` dependency extra. **Subscription model** (operator-facing version in [IM_CHANNEL_CONNECTIONS.md](../../docs/IM_CHANNEL_CONNECTIONS.md#buzz-subscription-model)): the relay fans kind-9 chat events out **only** to channel-scoped subscriptions, proved against a live relay — `REQ {"kinds":[9]}` is accepted and answered with `EOSE` but never receives an event (the connector authenticated and then sat silent forever), `REQ {"kinds":[9],"#h":[uuid]}` works, and a multi-value `#h` matches nothing, so it is strictly one REQ per channel (the same shape as Buzz's own `buzz-acp` harness). Every connection therefore rebuilds three kinds of subscription after NIP-42 auth: `buzz-discovery` (`{"kinds":[39000]}`, a historical query returning exactly the channels this identity belongs to, one stored event each then `EOSE` — adding `#p` returns zero, do not "narrow" it), `buzz-membership` (`{"kinds":[44100,44101],"#p":[us],"since":<connect time − `MEMBERSHIP_LOOKBACK_SECONDS`>}`, the relay-signed member-added/member-removed notifications whose `p` tag names the affected member and `h` tag the channel), and one `buzz-chat-<uuid>` per discovered channel. Chat subscriptions open as each kind-39000 arrives; the discovery `EOSE` is the completeness barrier that retries any that failed and warns when discovery found nothing. A kind-44100 for our pubkey subscribes to the new channel **live** (then re-issues discovery so its name/type reach the DM-detection cache, but only when that channel's metadata is actually missing — an unconditional refresh is how a burst of 44100s multiplied into one discovery pass each); a kind-44101 issues `buzz_nostr.close_frame` for exactly that channel's subscription and drops its metadata. **The membership subscription is scoped to LIVE events** and this is load-bearing: buzz-relay *stores* 44100/44101 and serves history newest-first (default limit 2000), so an unscoped filter replayed the whole membership history on every connect — every stored add read as live, re-running discovery once each (M+1 discovery passes × N stored kind-39000 events per connect, observed live as two `channel discovery complete` lines and one channel logged `<unnamed>` because the 44100 path subscribed before its metadata arrived), re-subscribing channels we have since been removed from, and letting a stored 44101 transiently unsubscribe a channel we are still in. `since` is anchored at the moment the socket opened (`_session_started_at`) minus `MEMBERSHIP_LOOKBACK_SECONDS` (60) of slack, which covers both relay clock skew and a membership change published *during* the connect/auth handshake; the slack can only cost an idempotent replay of the last minute. **A relay `CLOSED` frame is recovered, not merely forgotten** — every subscription on the socket fails silently when dropped, so `_handle_closed` re-issues it, bounded by `MAX_RESUBSCRIBE_ATTEMPTS` (3) per subscription id per connection *and per auth epoch* (`_resubscribe_attempts` is reset on session start, on re-auth, and by `stop()`, so pre-auth rejections — which the auth branch already recovers wholesale — never spend the authenticated session's budget). The first retry is immediate (the common case is a one-off hiccup); later ones back off 1s then 2s, awaited inline in the read loop rather than in a background task that could outlive its own socket. **`auth-required:` before this socket has completed its NIP-42 handshake is the expected bootstrap sequence, not a refusal**, and `_handle_closed` short-circuits it ahead of every recovery path: the connector opens its control REQs immediately in case the relay serves unauthenticated reads, a closed relay answers `auth-required:` plus an `AUTH` challenge, and the auth branch re-opens everything. That case is logged at DEBUG and consumes neither the permanent-refusal branch nor the retry budget (a chat subscription is still dropped from `_chat_subscriptions`, since it genuinely is not subscribed and discovery is what re-opens it). Treating it as permanent produced an operator-facing warning claiming discovery/membership tracking was DOWN in the same run where discovery then completed, every channel was subscribed, and a brand-new channel's kind-44100 was picked up one second later. The boundary is the per-socket `_auth_completed` flag, set once the signed AUTH event has been sent and cleared on session entry, session exit, and `stop()`; the same reason *after* that stays loud, and says the subscription is down until the relay's next AUTH challenge or the next reconnect rather than borrowing the non-auth wording. Then `_is_transient_close` decides whether to retry at all: NIP-01/NIP-42 `auth-required:`/`restricted:`/`blocked:`/`mute:`/`invalid:`/`pow:` prefixes and buzz-relay's own removal/revocation prose are permanent (do not fight the relay over a channel that is no longer ours; a post-auth `auth-required:` is recovered by the AUTH branch, not by re-issuing), `rate-limited:`/`error:`/no-reason-at-all and anything unrecognized are transient — the default resolves toward "keep listening" because going silently deaf is the failure this exists to remove, and the attempt budget bounds a wrong guess. A chat `CLOSED` is only ever recovered for a channel already in `_chat_subscriptions`: a `CLOSED` is relay-supplied, so acting on an unknown one would let a relay induce a subscription just by naming a channel. Every subscription that goes unlistened is logged at WARNING, never INFO. Subscription ids are deterministic per channel precisely so one can be replaced or closed without disturbing the others on the socket, and `_chat_subscriptions` is per-socket state cleared on session end, on re-auth (a pre-auth REQ may have been rejected), and by `stop()`. Three bounds on remote-fed state: `MAX_CACHED_CHANNELS` (512) caps the kind-39000 metadata cache, the watermark map, and the resubscribe-attempt map; `MAX_CHANNEL_SUBSCRIPTIONS` (256, well under buzz-relay's own 1024-per-connection ceiling) caps live chat subscriptions — at the cap new channels are refused and named in a warning rather than evicting a working subscription. **Known bound (documented, not fixed):** the relay caps historical delivery at 2000 events per subscription, newest-first, even with a `since`, so >2000 unread messages in a *single* channel across a disconnect loses the oldest — the relay never sends them and the watermark advances past them. That is the one remaining path that can skip; everything else fails toward replay. **Trust model** (operator-facing version in [IM_CHANNEL_CONNECTIONS.md](../../docs/IM_CHANNEL_CONNECTIONS.md#buzz-trust-model)): every inbound `EVENT` is authenticated at the single `handle_relay_frame` choke point — the NIP-01 id is recomputed from the delivered payload and the BIP-340 Schnorr signature verified against the claimed `pubkey` (`buzz_nostr.verify_event`, pure and total: malformed input returns `False`, never raises) — so `ev["pubkey"]`, the authorization principal for both the allowlist and the `/connect` bind, cannot be forged by a relay the DeerFlow operator does not run. What remains trusted is the *authorship* of kind-39000 channel metadata: any member can sign one, and because per-channel subscriptions are now driven by discovery, a forged kind-39000 has two effects rather than one — it can mark a channel `type: "dm"` (relaxing `require_mention` for that channel) **and** it can induce a chat subscription for a channel of the forger's choosing, since the channels we listen to are exactly the channels we hold metadata for. Neither makes anything be *acted on*: `allowed_users` and per-event signature verification are independent gates, so an induced subscription only means the relay reads its own traffic back to a subscriber that drops it, bounded by `MAX_CHANNEL_SUBSCRIPTIONS` (which refuses rather than evicts, so it cannot displace a real channel). Same for a forged kind-44100, except its `p` tag is re-checked locally so it must at least name us. Closing this needs a configured trusted relay pubkey, which `relay_url` is not. `allowed_users` is deny-by-default (empty = nobody, unlike siblings' empty = everyone), so `start()` logs a WARNING when it is empty and each drop logs at DEBUG. The resubscribe cursor (`since`) is **per channel**, advances only for events that were actually processed, and never past `now + MAX_FUTURE_SKEW_SECONDS`, because it is peer-supplied (`created_at`) and a single future-dated event otherwise made the connector permanently deaf. Per channel rather than global is the safety-critical half: subscriptions are per channel, so one shared cursor is the newest event seen in *any* channel and a busy channel would drag it past a quiet channel's unread messages, skipping them on the next reconnect — measured on a live relay, three channels of one identity sat ~28h apart. Per-channel cursors can only ever cost duplicate delivery (absorbed by the manager's `event_id` dedupe), and an evicted cursor degrades to "no `since`", i.e. the relay's default backlog — both fail toward replay, never toward a miss. Streaming tracks every oversize chunk index (`_stream_targets` for chunk 0, `_stream_tails` for the rest), since the manager republishes cumulative text and reposting `chunks[1:]` per update flooded the channel; all of it is per-connection state cleared by `stop()`, and the remote-fed kind-39000 cache is capped at `MAX_CACHED_CHANNELS`. **`send()` refuses outright to publish text carrying a hidden model-context wrapper** (`<memory>`, `<durable_context_data>`, `<system-reminder>` — `_HIDDEN_CONTEXT_MARKERS`), logging at ERROR and clearing the stream bookkeeping on a blocked `is_final`. This is defense in depth behind the manager's allowlist, and it lives here rather than in a sibling connector because on Buzz a leak is permanent: every streaming update is an immutable public Nostr event, so a corrective edit only changes what clients render while the original leaked event stays on the relay. Matching is on the literal opening tag, so a reply that merely talks about memory is still published
- `github.py` - Webhook-driven GitHub channel. Inbound messages come from `POST /api/webhooks/github`; outbound is log-only because GitHub agents post explicitly with `gh` from their sandbox when they choose to comment or create a PR
- `app/gateway/routers/channel_connections.py` - Browser-facing user connection and disconnect APIs
- `deerflow.persistence.channel_connections` - SQL-backed user-owned connection, optional credential, connect state, and conversation store

**Message Flow**:
1. External platform -> Channel impl -> `MessageBus.publish_inbound()`
   - For GitHub, the webhook router verifies the delivery then calls `fanout_event(bus, ...)`; matching agent bindings publish one `InboundMessage` each instead of a long-polling channel worker.
   - Telegram photo/document updates use the largest photo size or document metadata, preserve `message.caption`, enforce the hosted Bot API's 20,000,000-byte download ceiling before and after download, and never expose the token-bearing Bot API file URL. Downloaded bytes cross the adapter/manager boundary only through `message_bus.INBOUND_FILE_CONTENT_KEY`; the manager consumes that transient field before persisting safe upload metadata.
2. `ChannelManager._dispatch_loop()` consumes from queue
3. For user-owned channel connections, incoming messages carry `connection_id`, `owner_user_id`, and `workspace_id`; `owner_user_id` becomes the DeerFlow run `user_id`, while the raw platform user id remains `channel_user_id`. The Gateway accepts `channel_user_id` only from an internally authenticated channel caller's top-level `body.context`, clears it from both free-form `body.config` sections, and writes it into runtime context only (never `configurable`, which is checkpointed). `bash_tool` exposes it to sandbox commands as the fixed env var `DEERFLOW_CHANNEL_USER_ID` — via a shell-quoted command-string prefix, NOT the `execute_command(env=...)` channel, which is reserved for request-scoped secrets and would switch `AioSandbox` onto the `bash.exec` path (image >= 1.9.3, fresh session per call). Per-call injection keeps group-chat identity correct (one thread/sandbox, many senders) **without depending on the AIO shell's session semantics**: every IM-channel command carries an explicit `export VAR=<id>; ` (valid id) or `unset VAR; ` (empty / non-str / over the 256-char cap). The AIO no-env path reuses a persistent shell session (the reason for the class lock, #1433), so a bare command could otherwise resolve a stale id an earlier sender exported; the `unset` closes the window the length/type guard would open (a dropped id would inherit the previous sender's value). Non-IM runs (no `channel_user_id` in context) are left untouched. Not injected on the Windows local sandbox (its PowerShell/cmd.exe fallback has no `export`/`unset`). Propagates across `task` delegation: `task_tool` captures the dispatching turn's id and the subagent executor forwards it into the subagent's runtime context, same as the guardrail attribution fields. The runtime-context value is authorization-grade at the Gateway/guardrail boundary, but the exported shell variable remains informational because any bash command can overwrite its own environment; skills must not treat the shell variable itself as authenticated identity. Tests: `tests/test_gateway_services.py`, `tests/test_channel_user_id_env.py`
4. For chat: look up/create thread through Gateway's LangGraph-compatible API
5. Feishu/Telegram chat: `runs.stream()` → accumulate AI text → publish multiple outbound updates (`is_final=False`) → publish final outbound (`is_final=True`)
6. Slack/Discord chat: `runs.wait()` → extract final response → publish outbound
6b. GitHub chat (`ChannelRunPolicy.fire_and_forget=True`): `runs.create()` returns once the run is `pending`; the manager does not wait for the final state and does not publish an outbound. The agent posts its own reply mid-run via `gh` from the sandbox. `ConflictError` on a busy thread still trips the standard `THREAD_BUSY_MESSAGE` path (log-only on GitHub); when the channel's policy also sets `buffer_followups_on_busy=True` (GitHub's default — see "Follow-up buffering while busy" below), the triggering message is additionally captured into a per-thread buffer instead of only logged, so a concurrent comment is not silently dropped.
7. Feishu channel sends one running reply card up front, then patches the same card for each outbound update (card JSON sets `config.update_multi=true` for Feishu's patch API requirement). Messages already sent inside an existing Feishu topic carry a compact source-message preview in that card, and queued same-thread follow-ups patch their own source message's card from queued → running → final without falling back to the generic busy reply.
8. Telegram streaming: the "Working on it..." placeholder message is registered as the stream target; non-final updates `editMessageText` it in place (channel-side throttle: 1s in private chats, 3s in groups due to Telegram's 20 msg/min group cap; 4096-char truncation; rate-limited updates dropped); the final update performs the last edit and splits >4096 texts into follow-up messages
9. DingTalk AI Card mode (when `card_template_id` configured): `runs.stream()` → create card with initial text → stream updates via `PUT /v1.0/card/streaming` → finalize on `is_final=True`. Falls back to `sampleMarkdown` if card creation or streaming fails
10. For commands (`/new`, `/status`, `/models`, `/memory`, `/goal`, `/help`): handle locally or query Gateway API
11. Outbound → channel callbacks → platform reply
    - GitHub is the exception: the channel logs the final assistant message and does **not** auto-post it to GitHub. Agents use the sandbox `gh` CLI (`gh issue comment`, `gh pr comment`, `gh pr create`, etc.) for intentional writeback, so silence is cheap when several agents fan out on the same event.

**Owner-scoped file storage**: inbound files, uploads, and output artifacts are staged under the DeerFlow owner's bucket so they land where the agent run reads/writes (`users/{user_id}/threads/{thread_id}/user-data/{uploads,outputs}`). `ChannelManager._handle_chat` resolves the storage owner once via `_channel_storage_user_id(msg)` (sanitized owner id, falling back to `safe(msg.user_id)` for unbound auth-enabled channels — mirroring `_resolve_run_params`'s run identity; `None` only when no identity is available) and threads it as the `user_id=` kwarg through the file pipeline:
- `Channel.receive_file(msg, thread_id, user_id=...)` — owner-bound channels persist downloaded files under the owner's bucket instead of the default bucket
- `_ingest_inbound_files(...)` and the underlying `ensure_uploads_dir` / `get_uploads_dir` — owner-scoped via the same kwarg
- `_resolve_attachments` / `_prepare_artifact_delivery` — resolve output artifacts from the bound owner's bucket
The cached value is reused for both the blocking (`runs.wait`) and streaming (`_handle_streaming_chat`) paths, so uploads and artifact delivery always target the same bucket even if a channel returns a rewritten `InboundMessage` from `receive_file`. The bucket id matches the memory bucket resolved by `_resolve_memory_user_id` (both normalize through `make_safe_user_id`).

**Configuration** (`config.yaml` -> `channels`):
- `langgraph_url` - LangGraph-compatible Gateway API base URL (default: `http://localhost:8001/api`)
- `gateway_url` - Gateway API URL for auxiliary commands (default: `http://localhost:8001`)
- In Docker Compose, IM channels run inside the `gateway` container, so `localhost` points back to that container. Use `http://gateway:8001/api` for `langgraph_url` and `http://gateway:8001` for `gateway_url`, or set `DEER_FLOW_CHANNELS_LANGGRAPH_URL` / `DEER_FLOW_CHANNELS_GATEWAY_URL`.
- Per-channel configs: `feishu` (app_id, app_secret), `slack` (bot_token, app_token), `telegram` (bot_token, optional `rich_messages` for final Markdown Rich Messages), `dingtalk` (client_id, client_secret, optional `card_template_id` for AI Card streaming), `github` (operator kill-switch `enabled`, plus `default_mention_login` for mention-required GitHub triggers), `buzz` (relay_url, private_key)

**User-owned channel connections** (`config.yaml` -> `channel_connections`):
- Disabled by default. It is a user-binding layer on top of the existing `channels.*` runtime config, not a replacement for provider bot credentials.
- No public IP, OAuth callback URL, or provider webhook route is required by the current implementation.
- Telegram uses a deep-link `/start <code>` flow over the existing long-polling worker. Slack, Discord, Feishu/Lark, DingTalk, WeChat, and WeCom use `/connect <code>` over their existing outbound channel workers.
- WeChat timing settings (`polling_timeout`, `polling_retry_delay`, `qrcode_poll_interval`, `qrcode_poll_timeout`) accept only positive finite seconds; invalid values fall back to their defaults so polling cannot enter a hot loop or sleep forever.
- WeCom serializes `start()` and `stop()` for each channel instance. The SDK `connect()` task covers connection setup only; after the handshake, the SDK owns a separate receive task. Shutdown cancels an in-progress connection attempt and awaits the SDK's actual asynchronous receive-task/socket cleanup before releasing lifecycle state or allowing a restart. Cancellation of `stop()` still propagates, but only after owned cleanup finishes and lifecycle references are cleared; real connection failures remain reported by `_on_ws_task_done`.
- Frontend APIs: `GET /api/channels/providers`, `GET /api/channels/connections`, `POST /api/channels/{provider}/connect`, and `DELETE /api/channels/connections/{connection_id}`.
- Browser APIs remain protected by normal Gateway auth/CSRF. Provider messages arrive through the already-configured channel workers.
- Provider-level `connection_status` reflects the user's newest connection row. With no binding it is `not_connected`, except in auth-disabled local mode where a configured running channel reports `connected` because all channel messages already route to the default user.
- Slack replies use the configured operator bot token from `channels.slack` unless per-connection credentials are present; unreadable or corrupt stored credentials are treated as unavailable.
- Telegram, Slack, Discord, Feishu/Lark, DingTalk, WeChat, and WeCom workers resolve incoming platform identities to connection records before reaching `ChannelManager`.
- **Connect-code ordering vs `allowed_users`**: inbound workers consume a valid `/connect <code>` (or Telegram `/start <code>`) **before** applying the `allowed_users` filter, so a newly allowlisted-but-unbound user can bootstrap their first bind via the browser flow. Consequence: `allowed_users` is **not** a bind-time defense — any sender who possesses a valid code can consume it (not only allowlisted users). The bind security model rests on the code's confidentiality: `secrets.token_urlsafe(16)`, 600 s TTL, one-time `consume_oauth_state`, and codes surfaced only in the initiating browser (never echoed to chat). `allowed_users` still gates ordinary (non-bind) messages.
- **Single-active-owner transfer semantics**: an external identity is keyed by `(provider, external_account_id, workspace_id)`. The latest successful bind wins — `upsert_connection` revokes other owners' active rows for the same identity (ownership transfer). This invariant is enforced at the DB layer by the partial unique index `uq_channel_connection_active_identity` (`WHERE status != 'revoked'`), so concurrent connects from different owners cannot both end `connected`; the losing writer retries against the now-visible state. `find_connection_by_external_identity` therefore resolves deterministically.
- See `backend/docs/IM_CHANNEL_CONNECTIONS.md` for provider setup, operational notes, and the architecture diagrams (connect-code flow, single-active-owner transfer, sync vs streaming dispatch, owner-scoped file storage pipeline).

**GitHub event-driven agents** (webhook-driven IM channel):
- Custom agents declare a `github:` block in their `config.yaml` to bind to repos and event triggers; the webhook route is fail-closed by default (mounted only when `GITHUB_WEBHOOK_SECRET` is set) and exempt from auth/CSRF because authenticity is enforced by HMAC.
- Registry caching is keyed by the configured agent store's opaque signature: file storage uses agent config mtimes, while database storage hashes the ordered owner/name/config/soul contents so same-timestamp writes still invalidate webhook routing.
- Outbound is **log-only** by design: each agent posts its own reply mid-run via the `gh` CLI from its sandbox, so the manager uses `fire_and_forget=True` and `runs.create()` returns once pending.
- **Follow-up buffering while busy** (issue #4121): because outbound is log-only, the pre-existing `THREAD_BUSY_MESSAGE` reply on a `ConflictError` was invisible to the commenter — a comment posted while a run was already active looked like it had been silently ignored. When `ChannelRunPolicy.buffer_followups_on_busy=True` (GitHub's default), a `ConflictError` on `runs.create()` now also appends the triggering message to a per-thread, in-memory buffer (`ChannelManager._followup_buffers`) — deduped by GitHub delivery id, capped at `FOLLOWUP_BUFFER_MAX_PER_THREAD` (20, oldest dropped with a WARNING log on overflow). The first successful `runs.create()` on a thread now captures its `run_id` and spawns a background watcher that subscribes to that run's `StreamBridge` stream; once it observes `END_SENTINEL`, the watcher drains up to `FOLLOWUP_DRAIN_BATCH_SIZE` (10) buffered entries into one `<followups-while-busy>`-wrapped input and fires a follow-up `runs.create()` — itself watched the same way, so a backlog deeper than one batch chains into further drain cycles instead of growing one unbounded input. If that follow-up `runs.create()` itself hits `ConflictError` (e.g. a manual Web UI turn or a scheduled run raced onto the same thread), the batch is requeued rather than lost, and is retried whenever this manager next successfully creates and watches a run on that thread. Reactions/acknowledgment (e.g. GitHub's `eyes`/`confused` reaction API) on buffered comments are intentionally **out of scope** for this mechanism and left to a follow-up — comments are coalesced silently. **Plumbing**: the watcher needs the Gateway's `StreamBridge` singleton, which `ChannelManager` did not previously have access to; it is threaded from `app.py`'s lifespan (where `app.state.stream_bridge` is already set by `langgraph_runtime`) through `start_channel_service(get_stream_bridge=...)` → `ChannelService.__init__` → `ChannelManager.__init__`, as a zero-arg closure mirroring the existing `launch_run=lambda **kwargs: launch_scheduled_thread_run(app=app, **kwargs)` pattern used for `ScheduledTaskService` in the same lifespan function. A `ChannelManager` constructed without it (e.g. directly in a test) still buffers safely — it just has no watcher to auto-drain. **Scope limitation**: the buffer and watcher state are per-process, in-memory. Under `GATEWAY_WORKERS>1` or multi-pod, a follow-up comment routed to a different worker process than the one running the busy thread's agent will not see that buffer. This is a known, deliberately deferred limitation with the same shape as the cross-pod gap described for issue #4120 (a shared buffer store or IM-leader election would be needed to close it) — single-process/single-pod deployments, the safe default, see no correctness issue from this, only the documented per-process scope.
- See [backend/docs/GITHUB_AGENTS.md](../../docs/GITHUB_AGENTS.md) for the architecture diagrams: webhook → fan-out → `InboundMessage` dispatch, `preferred_thread_id = UUID5(repo, number, agent_name)` thread determinism, mention-handle precedence chain, GH token lifecycle via `GH_TOKEN`/`GITHUB_TOKEN` per-call `extra_env`, and the narrow `ConflictError` (HTTP 409) thread-create race recovery.