5 Commits

Author SHA1 Message Date
a060ff98a2 feat(inbound): listen for backend-pushed channel.joined/left events
Companion to nav/Fabric.Backend.Guild#<TBD> which adds the server-side
emitToUser broadcast on channel membership changes. Before, the inbound
only learned about new channels via the 60s polling resync (worst-case
60s lag). Now the backend tells us directly so sub/unsub is realtime.

socket.on('channel.joined', evt) → join the socket.io room for evt.channelId
                                    and add to the local 'joined' set.
socket.on('channel.left',   evt) → leave + remove from 'joined'.

Both events are idempotent (`if (joined.has(id))` / `if (!joined.has(id))`)
so duplicate emits from server are safe. Polling resync still runs every
60s as a safety net for transient socket drops between emit and
reconnect, partial server failures, etc.

When backend lacks this support (older deployments), nothing breaks —
the event simply never fires and polling carries the load as before.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-21 08:08:33 +01:00
b9a5456d57 Merge pull request 'fix: dynamically sync inbound channel subscriptions' (#1) from fix/inbound-dynamic-channel-sync into main 2026-05-21 06:56:49 +00:00
d1d5ad10ca fix: dynamically sync inbound channel subscriptions
The fabric inbound previously called `joinAll()` once on socket.io
`connect` — it fetched the agent's channel list via
`GET /api/channels?guildId=...` and emitted `join_channel` for each.
Any channel the agent joined *after* connect (e.g. a fresh DM created
by another user that includes this agent) was unreachable until the
gateway restarted: the socket was never subscribed to that room, so
backend `message.created` push events never arrived.

Backend doesn't emit a user-scoped `channel.joined` event we could
piggy-back on (only `message.created`), so the fix is to poll. Every
60s the agent's channel list is re-fetched and diffed against a local
`joined` set:
- new channel ids → `socket.emit('join_channel', {channelId})` + add
- ids in `joined` but absent from the fresh list → `leave_channel`
  emit + remove (best-effort; cleans subs if the agent is removed from
  a channel)

Re-uses `freshGuildToken()` so the resync fetch survives token
expiry (15-min TTL). Initial `connect` resets the local `joined`
set since the server forgets prior room subscriptions on reconnect.

Timers are tracked in `channelSyncTimers` and cleared in `stop()`
alongside socket disconnect.

Verified against prod server.t2 scenario: hzhang creates DM channel
including agent 'manager' → without this fix, manager only sees the
message after a gateway restart; with this fix, manager receives the
message within at most 60s (next resync tick).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-21 07:45:59 +01:00
92945b777d feat(fabric): dm channels deliver any non-self message (no wakeup gate)
inbound: FabricMessage gains xType; the wakeup gate is bypassed when
xType==='dm' (self messages are already filtered upstream), so a 1:1
dm always reaches the model regardless of wakeup metadata.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-18 09:18:20 +01:00
8774cfd7cc feat(fabric): coalesce a split agent turn into ONE message (deterministic)
OpenClaw delivers an agent turn whose blocks are text -> thinking/tool
-> text via multiple inbound deliver() calls (a non-text block is a
delivery boundary), so one turn became N Fabric messages.

Fix: buffer deliver() segments per channel (src/coalesce.ts) and flush
them as ONE postMessage at a deterministic boundary — the finally after
dispatchInboundReplyWithBase() resolves, which provably runs only after
every deliver() of the turn (verified: deliver,deliver -> dispatch
returned -> flush). No hooks, no timers, no idle guessing. The
agent_end hook was rejected: it fires BEFORE deliver(). gateway_stop
flushes any leftover; a long safety timeout is a leak-guard only.
channels.fabric.coalesce=false restores raw per-segment posting.

Verified on local openclaw + Fabric with a fake text/thinking/text
model: single trigger -> exactly one merged message.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-16 22:15:46 +01:00
12 changed files with 433 additions and 24 deletions

View File

@@ -72,6 +72,12 @@ Two ways, both write the same identity registry the transport reads:
`FABRIC_BACKEND_GUILD_COMMANDS_SYNC_KEY` (Guild C-2). Read it from the
guild with `docker exec fabric-backend-guild node dist/cli/print-commands-sync-key.js`.
Sourced from config only — never from the environment.
- `channels.fabric.coalesce` — default `true`. OpenClaw splits one agent
turn whose blocks are `text → thinking/tool → text` into multiple
`deliver()` calls; this buffers them and posts ONE Fabric message at the
deterministic turn boundary (right after the inbound reply dispatch
resolves — no hooks, no timers, no idle guessing). `false` = raw
per-segment posting.
- `channels.fabric.accounts.<agentId>` = `{ fabricApiKey, enabled }`
(**agent = account**; the account id is the OpenClaw agentId)
- plugin `identityFilePath` — default `~/.openclaw/fabric-identity.json`

View File

@@ -5,6 +5,7 @@
// the OpenClawPluginApi for runtime startup (transport + tools).
import { defineChannelPluginEntry } from 'openclaw/plugin-sdk/core';
import { fabricChannelPlugin } from './src/channel.js';
import { flushAllFabric } from './src/coalesce.js';
import { FabricInbound } from './src/inbound.js';
import { listEnabledFabricAccounts } from './src/accounts.js';
import { registerFabricTools } from './src/tools.js';
@@ -60,7 +61,12 @@ export default defineChannelPluginEntry({
api.logger.info(`fabric: inbound started for ${accounts.length} account(s)`);
void syncFabricCommands(client, cfg, accounts, api.logger);
});
// Note: the per-turn coalesce flush happens deterministically in
// inbound.ts right after dispatchInboundReplyWithBase resolves (that
// is the real "all deliveries done" boundary; the agent_end hook fires
// BEFORE deliver()). gateway_stop only flushes any leftover buffer.
api.on('gateway_stop', () => {
void flushAllFabric();
inbound?.stop();
inbound = null;
});

View File

@@ -13,6 +13,11 @@ function section(cfg) {
export function resolveCommandsSyncKey(cfg) {
return (section(cfg).commandsSyncKey ?? '').trim();
}
// Whether to coalesce a split agent turn into one Fabric message
// (channel-level). Default true.
export function resolveCoalesce(cfg) {
return (cfg.channels?.fabric ?? {}).coalesce !== false;
}
export function listFabricAccountIds(cfg) {
const accts = section(cfg).accounts ?? {};
const ids = Object.keys(accts);

View File

@@ -137,7 +137,10 @@ export const fabricChannelPlugin = createChatChannelPlugin({
attachedResults: {
channel: 'fabric',
sendText: async (ctx) => {
// openclaw passes config under cfg or config depending on path
// openclaw passes config under cfg or config depending on path.
// Note: inbound agent replies go through inbound.ts `deliver`
// (where turn coalescing happens). This path is for any direct
// outbound sends and posts immediately.
const cfg = (ctx.cfg ?? ctx.config ?? {});
try {
const r = await sendToFabric(cfg, ctx.accountId ?? null, ctx.to, ctx.text);

75
dist/fabric/src/coalesce.js vendored Normal file
View File

@@ -0,0 +1,75 @@
// Deterministic turn coalescer.
//
// OpenClaw calls the Fabric `deliver` callback once per assistant text
// segment; a thinking/tool block between two text blocks is a delivery
// boundary, so one agent turn of `text → thinking/tool → text` arrives as
// multiple deliver() calls. There is no turn id on the delivery, so we
// BUFFER segments by Fabric channelId and post the merged message when the
// turn truly ends. The flush is driven by inbound.ts right after
// `dispatchInboundReplyWithBase` resolves — that only happens AFTER every
// deliver() of the turn, a deterministic boundary (NOT a timer, NOT the
// agent_end hook, which fires before deliver()). `coalesce=false` posts
// each segment immediately.
const SAFETY_FLUSH_MS = 120_000; // leak-guard only; not the flush mechanism
export function normChannelId(x) {
const s = String(x ?? '');
return s.startsWith('fabric:') ? s.slice('fabric:'.length) : s;
}
const pendingByChannel = new Map();
async function flushChannel(channelId, reason) {
const p = pendingByChannel.get(channelId);
if (!p)
return;
pendingByChannel.delete(channelId);
clearTimeout(p.safety);
const text = p.parts.join('\n\n').trim();
if (!text)
return;
try {
await p.post(text);
p.log?.(`fabric: flushed ${p.parts.length} segment(s) channel=${channelId} (${reason})`);
}
catch (e) {
p.log?.(`fabric: flush FAILED channel=${channelId} (${reason}): ${String(e)}`);
}
}
// Buffer one delivered segment (or send immediately when coalesce=false).
// `post` performs the real Fabric postMessage with the caller's already
// resolved guild/token; on flush it is called once with the merged text.
export async function enqueueDelivery(params) {
const cid = normChannelId(params.channelId);
const text = (params.text ?? '').trim();
if (!text)
return;
if (!params.coalesce) {
await params.post(text);
return;
}
const existing = pendingByChannel.get(cid);
if (existing) {
existing.parts.push(text);
existing.post = params.post; // freshest guild/token closure
existing.log = params.log;
}
else {
pendingByChannel.set(cid, {
parts: [text],
post: params.post,
log: params.log,
safety: setTimeout(() => void flushChannel(cid, 'safety-timeout'), SAFETY_FLUSH_MS),
});
}
}
// Called by the agent_end hook with the hook ctx's channelId (bare or
// fabric:-prefixed). Deterministic per-turn boundary.
export async function flushFabricForChannel(rawChannelId) {
const cid = normChannelId(rawChannelId);
if (cid)
await flushChannel(cid, 'dispatch-end');
}
// gateway_stop: flush anything still buffered.
export async function flushAllFabric() {
for (const cid of [...pendingByChannel.keys()]) {
await flushChannel(cid, 'gateway_stop');
}
}

View File

@@ -3,6 +3,8 @@ import { tmpdir } from 'node:os';
import { join } from 'node:path';
import { io } from 'socket.io-client';
import { dispatchInboundReplyWithBase } from 'openclaw/plugin-sdk/inbound-reply-dispatch';
import { resolveCoalesce } from './accounts.js';
import { enqueueDelivery, flushFabricForChannel } from './coalesce.js';
export class FabricInbound {
core;
cfg;
@@ -12,6 +14,14 @@ export class FabricInbound {
accounts;
sockets = [];
seen = new Set();
// Timers that periodically re-sync channel membership per (agent, guild).
// Without this, the agent's socket.io subscriptions are a snapshot taken
// at connect time — any channel the agent joins later (e.g. a fresh DM
// created by another user) is unreachable until the gateway restarts.
channelSyncTimers = [];
// Resync cadence. Backend doesn't push a `channel.joined` event, so we
// poll. 60s keeps the lag bounded without hammering the backend.
static CHANNEL_SYNC_INTERVAL_MS = 60_000;
// Guild access tokens are short-lived (~15 min). The socket survives via
// socket.io reconnect, but the token captured at connect time goes stale,
// so HTTP calls (attachment download, posting the reply) start 401ing.
@@ -73,6 +83,9 @@ export class FabricInbound {
}
}
stop() {
for (const t of this.channelSyncTimers)
clearInterval(t);
this.channelSyncTimers = [];
for (const s of this.sockets)
s.disconnect();
this.sockets = [];
@@ -88,19 +101,83 @@ export class FabricInbound {
auth: { token: tok },
autoConnect: false,
});
const joinAll = async () => {
// Tracked socket.io rooms for this (agent, guild). The initial fetch
// on `connect` seeds it; the periodic resync diffs against it so we
// only emit `join_channel` for genuinely new channels (and
// `leave_channel` for ones the agent is no longer in).
const joined = new Set();
const syncChannels = async (kind) => {
let freshTok;
try {
const res = await fetch(`${g.endpoint}/api/channels?guildId=${encodeURIComponent(g.nodeId)}`, { headers: { authorization: `Bearer ${tok}` } });
const channels = res.ok ? (await res.json()) : [];
for (const c of channels)
socket.emit('join_channel', { channelId: c.id });
this.log.info(`fabric: agent ${agentId} joined ${channels.length} channel(s) on ${g.nodeId}`);
freshTok = await this.freshGuildToken(agentId, g.nodeId, session);
}
catch {
/* best effort */
freshTok = tok;
}
const authTok = freshTok ?? tok;
try {
const res = await fetch(`${g.endpoint}/api/channels?guildId=${encodeURIComponent(g.nodeId)}`, { headers: { authorization: `Bearer ${authTok}` } });
if (!res.ok)
return;
const channels = (await res.json());
const current = new Set(channels.map((c) => c.id));
let added = 0;
let removed = 0;
for (const id of current) {
if (!joined.has(id)) {
socket.emit('join_channel', { channelId: id });
joined.add(id);
added++;
}
}
for (const id of [...joined]) {
if (!current.has(id)) {
socket.emit('leave_channel', { channelId: id });
joined.delete(id);
removed++;
}
}
if (kind === 'initial') {
this.log.info(`fabric: agent ${agentId} joined ${current.size} channel(s) on ${g.nodeId}`);
}
else if (added > 0 || removed > 0) {
this.log.info(`fabric: agent ${agentId} channel resync on ${g.nodeId}: +${added} -${removed} (now ${joined.size})`);
}
}
catch {
/* best effort — next tick will retry */
}
};
socket.on('connect', () => void joinAll());
socket.on('connect', () => {
// On every (re)connect the server forgets prior subscriptions, so
// reset our local view and seed from a fresh fetch.
joined.clear();
void syncChannels('initial');
});
// Push-based membership events from the backend (companion to
// Fabric.Backend.Guild's RealtimeGateway.emitToUser). When the
// server tells us this user was added to / removed from a
// channel, we sub/unsub the socket.io room immediately — no
// 60s wait for the polling resync. Polling remains as a safety
// net for missed events.
socket.on('channel.joined', (evt) => {
const id = evt?.channelId;
if (!id || joined.has(id))
return;
socket.emit('join_channel', { channelId: id });
joined.add(id);
this.log.info(`fabric: agent ${agentId} channel.joined push on ${g.nodeId}: ${id} (now ${joined.size})`);
});
socket.on('channel.left', (evt) => {
const id = evt?.channelId;
if (!id || !joined.has(id))
return;
socket.emit('leave_channel', { channelId: id });
joined.delete(id);
this.log.info(`fabric: agent ${agentId} channel.left push on ${g.nodeId}: ${id} (now ${joined.size})`);
});
const syncTimer = setInterval(() => void syncChannels('resync'), FabricInbound.CHANNEL_SYNC_INTERVAL_MS);
this.channelSyncTimers.push(syncTimer);
socket.on('message.created', (m) => {
const channelId = m.channelId ?? '';
if (!channelId)
@@ -199,7 +276,11 @@ export class FabricInbound {
// the woken speaker emits a normal message or /no-reply). We still
// record the message into the agent's session so it has the full
// channel conversation as context whenever it IS later woken.
if (m.wakeup !== true) {
//
// Exception: dm channels are 1:1 — there is no turn/wakeup gating;
// any message that isn't the agent's own (already filtered above) is
// always delivered to the model.
if (m.xType !== 'dm' && m.wakeup !== true) {
const ctxPayload = core.channel.reply.finalizeInboundContext(baseCtx);
await core.channel.session.recordInboundSession({
storePath,
@@ -244,8 +325,16 @@ export class FabricInbound {
this.log.info(`fabric: deliver agent=${agentId} channel=${channelId} len=${text.length}`);
if (!text || !gt)
return;
await this.client.postMessage(guild.endpoint, gt, channelId, text, session.user.id);
this.log.info(`fabric: posted reply agent=${agentId} channel=${channelId}`);
// Buffer segments; the merged message is posted right after
// dispatch returns (the deterministic turn boundary, see the
// finally below). Disable per channel: channels.fabric.coalesce.
await enqueueDelivery({
channelId,
text,
coalesce: resolveCoalesce(this.cfg),
post: (t) => this.client.postMessage(guild.endpoint, gt, channelId, t, session.user.id),
log: (m) => this.log.info(m),
});
},
onRecordError: (err) => this.log.warn(`fabric: session record failed agent=${agentId}: ${String(err)}`),
onDispatchError: (err, info) => this.log.warn(`fabric: ${info.kind} dispatch failed agent=${agentId}: ${String(err)}`),
@@ -270,5 +359,12 @@ export class FabricInbound {
catch (err) {
this.log.warn(`fabric: dispatch failed agent=${agentId} channel=${channelId}: ${String(err)}`);
}
finally {
// Deterministic per-turn boundary: dispatchInboundReplyWithBase only
// resolves AFTER every deliver() call of this turn has run, so the
// buffer now holds all segments — flush them as ONE Fabric message.
// No hooks, no timers, no idle guessing.
await flushFabricForChannel(channelId);
}
}
}

View File

@@ -6,6 +6,7 @@
import { defineChannelPluginEntry } from 'openclaw/plugin-sdk/core';
import type { OpenClawPluginApi } from 'openclaw/plugin-sdk/core';
import { fabricChannelPlugin } from './src/channel.js';
import { flushAllFabric } from './src/coalesce.js';
import { FabricInbound } from './src/inbound.js';
import { listEnabledFabricAccounts } from './src/accounts.js';
import { registerFabricTools } from './src/tools.js';
@@ -37,7 +38,7 @@ export default defineChannelPluginEntry({
config?: unknown;
pluginConfig?: { identityFilePath?: string };
logger: { info: (m: string) => void; warn: (m: string) => void };
on: (ev: string, fn: () => void) => void;
on: (ev: string, fn: (...args: unknown[]) => unknown) => void;
registerTool: (d: unknown) => void;
};
const cfg = (api.config ?? {}) as { channels?: { fabric?: { centerApiBase?: string } } };
@@ -86,7 +87,12 @@ export default defineChannelPluginEntry({
void syncFabricCommands(client, cfg, accounts, api.logger);
});
// Note: the per-turn coalesce flush happens deterministically in
// inbound.ts right after dispatchInboundReplyWithBase resolves (that
// is the real "all deliveries done" boundary; the agent_end hook fires
// BEFORE deliver()). gateway_stop only flushes any leftover buffer.
api.on('gateway_stop', () => {
void flushAllFabric();
inbound?.stop();
inbound = null;
});

View File

@@ -44,6 +44,10 @@
"minLength": 1,
"description": "Shared secret that must equal the guild's FABRIC_BACKEND_GUILD_COMMANDS_SYNC_KEY. Required to register the slash-command catalog (Guild C-2). Read it from the guild via: docker exec fabric-backend-guild node dist/cli/print-commands-sync-key.js"
},
"coalesce": {
"type": "boolean",
"description": "Merge a split agent turn (text → thinking/tool → text) into ONE Fabric message. Flushed deterministically on the agent_end hook. Default true; false = raw per-segment posting."
},
"dmSecurity": { "type": "string" },
"dmPolicy": { "type": "string" },
"enabled": { "type": "boolean" },

View File

@@ -18,6 +18,11 @@ export type FabricChannelConfig = {
// (Guild C-2). Required by the channel config schema; sourced from config
// only — never from the environment.
commandsSyncKey?: string;
// Coalesce an agent turn that OpenClaw split into multiple deliveries
// (text → thinking/tool → text => N sendText calls) into ONE Fabric
// message. The flush boundary is the deterministic `agent_end` hook (not
// a timer). Default true; set false for raw per-segment posting.
coalesce?: boolean;
accounts?: Record<string, FabricAccountConfig>;
defaultAccount?: string;
} & FabricAccountConfig;
@@ -46,6 +51,12 @@ export function resolveCommandsSyncKey(cfg: Cfg): string {
return (section(cfg).commandsSyncKey ?? '').trim();
}
// Whether to coalesce a split agent turn into one Fabric message
// (channel-level). Default true.
export function resolveCoalesce(cfg: Cfg): boolean {
return (cfg.channels?.fabric ?? {}).coalesce !== false;
}
export function listFabricAccountIds(cfg: Cfg): string[] {
const accts = section(cfg).accounts ?? {};
const ids = Object.keys(accts);

View File

@@ -159,7 +159,10 @@ export const fabricChannelPlugin = createChatChannelPlugin<ResolvedFabricAccount
cfg?: unknown;
config?: unknown;
}) => {
// openclaw passes config under cfg or config depending on path
// openclaw passes config under cfg or config depending on path.
// Note: inbound agent replies go through inbound.ts `deliver`
// (where turn coalescing happens). This path is for any direct
// outbound sends and posts immediately.
const cfg = (ctx.cfg ?? ctx.config ?? {}) as AnyCfg;
try {
const r = await sendToFabric(cfg, ctx.accountId ?? null, ctx.to, ctx.text);

93
src/coalesce.ts Normal file
View File

@@ -0,0 +1,93 @@
// Deterministic turn coalescer.
//
// OpenClaw calls the Fabric `deliver` callback once per assistant text
// segment; a thinking/tool block between two text blocks is a delivery
// boundary, so one agent turn of `text → thinking/tool → text` arrives as
// multiple deliver() calls. There is no turn id on the delivery, so we
// BUFFER segments by Fabric channelId and post the merged message when the
// turn truly ends. The flush is driven by inbound.ts right after
// `dispatchInboundReplyWithBase` resolves — that only happens AFTER every
// deliver() of the turn, a deterministic boundary (NOT a timer, NOT the
// agent_end hook, which fires before deliver()). `coalesce=false` posts
// each segment immediately.
const SAFETY_FLUSH_MS = 120_000; // leak-guard only; not the flush mechanism
export function normChannelId(x: string | null | undefined): string {
const s = String(x ?? '');
return s.startsWith('fabric:') ? s.slice('fabric:'.length) : s;
}
type Pending = {
parts: string[];
post: (text: string) => Promise<void>;
log?: (m: string) => void;
safety: ReturnType<typeof setTimeout>;
};
const pendingByChannel = new Map<string, Pending>();
async function flushChannel(channelId: string, reason: string): Promise<void> {
const p = pendingByChannel.get(channelId);
if (!p) return;
pendingByChannel.delete(channelId);
clearTimeout(p.safety);
const text = p.parts.join('\n\n').trim();
if (!text) return;
try {
await p.post(text);
p.log?.(`fabric: flushed ${p.parts.length} segment(s) channel=${channelId} (${reason})`);
} catch (e) {
p.log?.(`fabric: flush FAILED channel=${channelId} (${reason}): ${String(e)}`);
}
}
// Buffer one delivered segment (or send immediately when coalesce=false).
// `post` performs the real Fabric postMessage with the caller's already
// resolved guild/token; on flush it is called once with the merged text.
export async function enqueueDelivery(params: {
channelId: string;
text: string;
coalesce: boolean;
post: (text: string) => Promise<void>;
log?: (m: string) => void;
}): Promise<void> {
const cid = normChannelId(params.channelId);
const text = (params.text ?? '').trim();
if (!text) return;
if (!params.coalesce) {
await params.post(text);
return;
}
const existing = pendingByChannel.get(cid);
if (existing) {
existing.parts.push(text);
existing.post = params.post; // freshest guild/token closure
existing.log = params.log;
} else {
pendingByChannel.set(cid, {
parts: [text],
post: params.post,
log: params.log,
safety: setTimeout(
() => void flushChannel(cid, 'safety-timeout'),
SAFETY_FLUSH_MS,
),
});
}
}
// Called by the agent_end hook with the hook ctx's channelId (bare or
// fabric:-prefixed). Deterministic per-turn boundary.
export async function flushFabricForChannel(
rawChannelId: string | null | undefined,
): Promise<void> {
const cid = normChannelId(rawChannelId);
if (cid) await flushChannel(cid, 'dispatch-end');
}
// gateway_stop: flush anything still buffered.
export async function flushAllFabric(): Promise<void> {
for (const cid of [...pendingByChannel.keys()]) {
await flushChannel(cid, 'gateway_stop');
}
}

View File

@@ -5,6 +5,8 @@ import { io, type Socket } from 'socket.io-client';
import { dispatchInboundReplyWithBase } from 'openclaw/plugin-sdk/inbound-reply-dispatch';
import type { FabricClient, FabricSession } from './fabric-client.js';
import type { IdentityRegistry } from './identity.js';
import { resolveCoalesce } from './accounts.js';
import { enqueueDelivery, flushFabricForChannel } from './coalesce.js';
// COMPAT NOTE (openclaw v2026.5.7): the inbound path mirrors how bundled
// channels (nextcloud-talk) drive the kernel:
@@ -42,11 +44,22 @@ type FabricMessage = {
channelId?: string;
attachments?: FabricAttachment[];
wakeup?: boolean;
// x-type of the channel (sent on message.created). 'dm' bypasses the
// wakeup gate: any message that isn't the agent's own is delivered.
xType?: string;
};
export class FabricInbound {
private sockets: Socket[] = [];
private seen = new Set<string>();
// Timers that periodically re-sync channel membership per (agent, guild).
// Without this, the agent's socket.io subscriptions are a snapshot taken
// at connect time — any channel the agent joins later (e.g. a fresh DM
// created by another user) is unreachable until the gateway restarts.
private channelSyncTimers: NodeJS.Timeout[] = [];
// Resync cadence. Backend doesn't push a `channel.joined` event, so we
// poll. 60s keeps the lag bounded without hammering the backend.
private static readonly CHANNEL_SYNC_INTERVAL_MS = 60_000;
// Guild access tokens are short-lived (~15 min). The socket survives via
// socket.io reconnect, but the token captured at connect time goes stale,
// so HTTP calls (attachment download, posting the reply) start 401ing.
@@ -114,6 +127,8 @@ export class FabricInbound {
}
stop(): void {
for (const t of this.channelSyncTimers) clearInterval(t);
this.channelSyncTimers = [];
for (const s of this.sockets) s.disconnect();
this.sockets = [];
}
@@ -128,20 +143,87 @@ export class FabricInbound {
auth: { token: tok },
autoConnect: false,
});
const joinAll = async () => {
// Tracked socket.io rooms for this (agent, guild). The initial fetch
// on `connect` seeds it; the periodic resync diffs against it so we
// only emit `join_channel` for genuinely new channels (and
// `leave_channel` for ones the agent is no longer in).
const joined = new Set<string>();
const syncChannels = async (kind: 'initial' | 'resync') => {
let freshTok: string | undefined;
try {
freshTok = await this.freshGuildToken(agentId, g.nodeId, session);
} catch {
freshTok = tok;
}
const authTok = freshTok ?? tok;
try {
const res = await fetch(
`${g.endpoint}/api/channels?guildId=${encodeURIComponent(g.nodeId)}`,
{ headers: { authorization: `Bearer ${tok}` } },
{ headers: { authorization: `Bearer ${authTok}` } },
);
const channels = res.ok ? ((await res.json()) as Array<{ id: string }>) : [];
for (const c of channels) socket.emit('join_channel', { channelId: c.id });
this.log.info(`fabric: agent ${agentId} joined ${channels.length} channel(s) on ${g.nodeId}`);
if (!res.ok) return;
const channels = (await res.json()) as Array<{ id: string }>;
const current = new Set(channels.map((c) => c.id));
let added = 0;
let removed = 0;
for (const id of current) {
if (!joined.has(id)) {
socket.emit('join_channel', { channelId: id });
joined.add(id);
added++;
}
}
for (const id of [...joined]) {
if (!current.has(id)) {
socket.emit('leave_channel', { channelId: id });
joined.delete(id);
removed++;
}
}
if (kind === 'initial') {
this.log.info(
`fabric: agent ${agentId} joined ${current.size} channel(s) on ${g.nodeId}`,
);
} else if (added > 0 || removed > 0) {
this.log.info(
`fabric: agent ${agentId} channel resync on ${g.nodeId}: +${added} -${removed} (now ${joined.size})`,
);
}
} catch {
/* best effort */
/* best effort — next tick will retry */
}
};
socket.on('connect', () => void joinAll());
socket.on('connect', () => {
// On every (re)connect the server forgets prior subscriptions, so
// reset our local view and seed from a fresh fetch.
joined.clear();
void syncChannels('initial');
});
// Push-based membership events from the backend (companion to
// Fabric.Backend.Guild's RealtimeGateway.emitToUser). When the
// server tells us this user was added to / removed from a
// channel, we sub/unsub the socket.io room immediately — no
// 60s wait for the polling resync. Polling remains as a safety
// net for missed events.
socket.on('channel.joined', (evt: { channelId?: string }) => {
const id = evt?.channelId;
if (!id || joined.has(id)) return;
socket.emit('join_channel', { channelId: id });
joined.add(id);
this.log.info(`fabric: agent ${agentId} channel.joined push on ${g.nodeId}: ${id} (now ${joined.size})`);
});
socket.on('channel.left', (evt: { channelId?: string }) => {
const id = evt?.channelId;
if (!id || !joined.has(id)) return;
socket.emit('leave_channel', { channelId: id });
joined.delete(id);
this.log.info(`fabric: agent ${agentId} channel.left push on ${g.nodeId}: ${id} (now ${joined.size})`);
});
const syncTimer = setInterval(
() => void syncChannels('resync'),
FabricInbound.CHANNEL_SYNC_INTERVAL_MS,
);
this.channelSyncTimers.push(syncTimer);
socket.on('message.created', (m: FabricMessage) => {
const channelId = m.channelId ?? '';
if (!channelId) return;
@@ -250,7 +332,11 @@ export class FabricInbound {
// the woken speaker emits a normal message or /no-reply). We still
// record the message into the agent's session so it has the full
// channel conversation as context whenever it IS later woken.
if (m.wakeup !== true) {
//
// Exception: dm channels are 1:1 — there is no turn/wakeup gating;
// any message that isn't the agent's own (already filtered above) is
// always delivered to the model.
if (m.xType !== 'dm' && m.wakeup !== true) {
const ctxPayload = core.channel.reply.finalizeInboundContext(baseCtx);
await core.channel.session.recordInboundSession({
storePath,
@@ -301,8 +387,17 @@ export class FabricInbound {
const text = (payload?.text ?? '').trim();
this.log.info(`fabric: deliver agent=${agentId} channel=${channelId} len=${text.length}`);
if (!text || !gt) return;
await this.client.postMessage(guild.endpoint, gt, channelId, text, session.user.id);
this.log.info(`fabric: posted reply agent=${agentId} channel=${channelId}`);
// Buffer segments; the merged message is posted right after
// dispatch returns (the deterministic turn boundary, see the
// finally below). Disable per channel: channels.fabric.coalesce.
await enqueueDelivery({
channelId,
text,
coalesce: resolveCoalesce(this.cfg as never),
post: (t) =>
this.client.postMessage(guild.endpoint, gt, channelId, t, session.user.id) as Promise<void>,
log: (m) => this.log.info(m),
});
},
onRecordError: (err: unknown) =>
this.log.warn(`fabric: session record failed agent=${agentId}: ${String(err)}`),
@@ -327,6 +422,12 @@ export class FabricInbound {
this.log.info(`fabric: dispatch returned agent=${agentId} channel=${channelId}`);
} catch (err) {
this.log.warn(`fabric: dispatch failed agent=${agentId} channel=${channelId}: ${String(err)}`);
} finally {
// Deterministic per-turn boundary: dispatchInboundReplyWithBase only
// resolves AFTER every deliver() call of this turn has run, so the
// buffer now holds all segments — flush them as ONE Fabric message.
// No hooks, no timers, no idle guessing.
await flushFabricForChannel(channelId);
}
}
}