Channel plugins now have a structured way to propagate attachments —
Plexum's host turns each MediaItem into a canonical.ImageBlock content
on the user message. This replaces the F-6 workaround that crammed
local file paths into a markdown footer at the end of the message
text.
internal/inbound/inbound.go:
- Notifier signature gains `media []MediaItem`; MediaItem mirrors
the host's channel.MediaRef (Path/MediaType/Name)
- dispatch downloads attachments (unchanged) then forwards results
as MediaItem to Notify — no more footer appending
- dispatch now also receives guildEndpoint so it can resolve Fabric's
RELATIVE attachment URLs (`/api/files/<id>`) against the guild base.
Previously the downloader received the relative path verbatim and
failed every fetch silently.
cmd/plexum-fabric-channel-plugin/main.go:
- notifier closure pushes media[] in the EmitNotification payload
Live verified: alice uploads blue 32x32 PNG → Fabric guild → plugin
downloads to /tmp/plexum-fabric/<msg>/blue32.png → emits inbound with
media → host routes to kimi agent → Kimi: "Blue".
(MiniMax M2.7 is text-only per openclaw model definitions, so the
same flow against MiniMax returns "I don't see an image" — that's a
model capability limit, not a plugin issue.)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
F-6 attachments:
- internal/attachments/ (~145 LOC + 8 tests): per-message Downloader
that fetches Fabric attachment URLs into $TMPDIR/plexum-fabric/<msg-id>/
with the agent's guild token, sanitizing filenames + message ids
against path-traversal
- AppendFooter renders downloaded paths as a markdown footer
("Attachments:\n - /tmp/... (mime, N bytes)\n")
- Agents access via the exec MCP tool (cat / file / etc.)
internal/inbound.Supervisor:
- new Attachments *attachments.Downloader field (nil → skip with warn)
- inbound.dispatch: when message has attachments, blocking-download +
AppendFooter before emitting notification (so agent's first turn
sees the paths)
cmd/plexum-fabric-channel-plugin:
- sup.Attachments = attachments.New("") wired at init
F-8 coalesce: no-op. openclaw plugin's coalesce buffered the
text→thinking→tool→text segments openclaw emits across multiple
deliver() calls per turn. Plexum's loop.Run returns ONE final assistant
text per turn (via extractFinalText), so coalescing isn't a concern.
The channel outbound posts a single message naturally.
F-5b presence-sync: deferred. The openclaw plugin pushes HarborForge
on-call status to Fabric's per-recipient presence so the backend can
busy-discard 'announce' deliveries. Plexum's state machine has
different semantics (idle/working/busy/offline) and is per-Plexum-agent
not per-user; mapping requires more design.
README updated with the phase status table + 16-tool list.
Tests: 8 new in internal/attachments (35 total in this repo).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
F-3 refinements:
- internal/inbound: replace fixed 3s reconnect wait with exponential
backoff (1s → 60s, ×2, reset when prior session lasted >30s); proxy
for "healthy" vs "flapping" and avoids hot reconnect loops when the
server is sick
F-4 agent tool surface (port of openclaw plugin's tools.ts):
- internal/tools/tools.go (~370 LOC): Registry binds Deps {Client,
Tokens, Identities} and exposes 8 agent-facing tools:
fabric-send-message post a normal message to any channel
fabric-send-sys-msg post a kind=sys message (bypasses turn engine)
fabric-channel-list list channels visible in a guild
fabric-guild-list list guilds the agent is in
fabric-message-history paginate channel messages by seq
fabric-channel-set-purpose PATCH the channel's purpose
fabric-channel fetch metadata + members for one channel
fabric-canvas get/share/update/remove channel canvas
- internal/tools/contracts.go: static ToolContract list — kept in sync
with install.sh's manifest emitter
- Every agent-scoped tool requires agent_id in input args (Plexum SDK
doesn't propagate calling agent id through CallTool today)
- guild_node_id defaults to agent's first guild for fabric-send-message
internal/fabric/client.go: new REST methods needed by tools —
PostSystemMessage, CreateChannel, CloseChannel, JoinChannel,
LeaveChannel, SetChannelPurpose, GetCanvas, ShareCanvas, UpdateCanvas,
RemoveCanvas, SyncCommands.
cmd/plexum-fabric-channel-plugin/main.go:
- Manifest declares the tool surface via tools.New(...).Contracts()
- CallTool dispatches "send" to handleSend (outbound for channel
manager), everything else to tools.Registry.Handler(name)
scripts/install.sh:
- Manifest tools[] now lists all 9 tools with schemas — matches what
internal/tools/contracts.go advertises
Live verified against running Fabric stack:
$ plexum plugin-call fabric-guild-list '{"agent_id":"fabrictester"}'
→ "guilds for agent fabrictester (1): test-guild2 @ http://localhost:7003"
$ plexum plugin-call fabric-channel-list '{...,"guild_node_id":"test-guild2"}'
→ 2 channels listed
$ plexum plugin-call fabric-message-history '{...,"limit":5}'
→ 5 messages with timestamps + authors
F-5+ deferred:
- create-{chat,work,report,discussion}-channel (batch 2)
- sub-discussion family (state store + 3 tools)
- presence-sync + command-sync
- attachments
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
End-to-end Fabric inbound→Plexum→Fabric outbound now works against a
live Fabric stack:
alice posts in bt2-clean (Fabric REST)
→ guild emits message.created over socket.io
→ plugin's wakeup gate decides dispatch
→ notifications/plexum/channel/inbound to host
→ Plexum agent runs (echo provider)
→ outbound `send` tool posts via Fabric REST
→ fabrictester reply visible in channel
internal/socketio/ (~280 LOC + 2 tests):
- Minimal Engine.IO v4 + Socket.IO v5 client over websocket
- WebSocket-only transport (skip polling upgrade dance)
- AuthFunc callback re-evaluated on every (re)connect — fixes the
stale-JWT-on-reconnect bug openclaw plugin documented for the JS
client's single-shot auth, which the available Go socket.io
library (zishang520) doesn't address either
- PING/PONG per server-supplied interval
- Caller-driven reconnect: Connect returns on close, supervisor
re-dials with fresh token
internal/tokens/ (~95 LOC + 9 tests):
- Per-agent session cache with 8min TTL (matches openclaw's
TOKEN_TTL_MS); guild tokens are ~15min so 8min keeps a margin
- Invalidate forces re-login (used by inbound when CONNECT auth fires)
- GuildToken helper picks the per-guild JWT from the cached session;
if the guild is missing from the cache, invalidate + retry once
internal/inbound/ (~290 LOC):
- Supervisor: one socket.io conn per (agent, guild); reconnect with
fresh token on drop; ChannelSyncInterval (60s) polling + push
channel.joined/channel.left handlers
- Wakeup gate: dm channels deliver any non-self message; other
x_types require wakeup=true (record-only for non-wake non-dm
deferred — Plexum has no history-injection equivalent in v1)
- Self-author filter on selfUserId from cached session
- Per-(agent,msgId) dedup bounded to 5000 entries
- Per-channel serial queue with 5s idle drain so concurrent inbounds
on the same channel run one-at-a-time (matches openclaw plugin)
- Emits notifications/plexum/channel/inbound with session_id =
"s_fab_<fabric_channel_id>" for stable per-channel session continuity
cmd/plexum-fabric-channel-plugin:
- Wires inbound supervisor at Init; runs in a background goroutine
for the plugin's lifetime
- Replaces F-1's sessions map with tokens.Cache (same warm-sessions
behavior, now backed by TTL)
- hostLogHandler: bridges slog records from inbound supervisor to
HostAPI.Log notifications
F-2 deferred to F-3+:
- record-only history injection (Plexum v1 has no equivalent)
- tools.ts port (15 MCP tools — channel/canvas/sub-discussion family)
- presence-sync, command-sync, attachments, coalesce parity
Tests: 22 (5 identity + 6 config + 9 tokens + 2 socketio).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
