Inbound was hardcoding `peer: { kind: 'group' }` and `ChatType: 'group'`
for every fabric channel regardless of xType. As a result:
- sessionKey for a DM was `agent:<id>:fabric:group:<chan>` instead of
`agent:<id>:fabric:direct:<chan>`
- ctx.ChatType='group' caused user-prompt metadata to render
`is_group_chat: true` on a DM
- openclaw's `isDirectMessage()` check (ChatType==='direct') returned
false, so DM-specific prompt and turn behavior never engaged
Caught by recruiter test in session 40c51de2: the model's thinking trace
acknowledged "fabric DM channel" (from the ClawPrompts chat-injector
hook) but the surrounding user-prompt metadata contradicted it with
`is_group_chat: true`, and the model reasoned its way out of running
`workflow_start`.
Fix factors a small helper `fabricPeerRoutingForXType` (and a cache-
backed `fabricPeerRoutingForChannel` for outbound) in channel.ts that
maps:
- 'dm' → { peerKind: 'direct', chatType: 'direct' }
- rest → { peerKind: 'group', chatType: 'group' } (no change)
Inbound uses m.xType directly (live, authoritative). Outbound has no
xType in its call signature, so it consults the channel-meta cache
populated by inbound (same `getChannelType` already exposed via
__fabric). Cache miss falls back to 'group' — the pre-fix default, no
regression. The proactive-DM-without-prior-inbound edge case still
routes that one outbound as 'group'; the next round agrees on 'direct'.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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>
Unlike Discord, Fabric has no message-length cap. Single-chunk chunker
(text -> [text]), textChunkLimit=MAX_SAFE_INTEGER, capabilities
blockStreaming=false, replyOptions.disableBlockStreaming=true -> every
agent reply delivered as exactly one Fabric message.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Real channel-turn dispatch (resolveAgentRoute + finalizeInboundContext +
dispatchInboundReplyWithBase), wakeup->drop/dispatch, messaging target
grammar (fabric:<id>) + outbound.sendText, tools use execute/parameters.
Verified live: human msg in Fabric -> wakeup -> openclaw agent runs ->
reply posted back into the Fabric channel as the agent.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
