v0.9.0

v0.9.0 — tag v0.9.0

Tagger: JochenYang 297390763@qq.com

agent-comm-hub 0.9.0
2026-09-22

• Task ledger state machine: bridge_ack gains a non-terminal working
  status plus an optional progress (0-100), so a long-running task reports
  intermediate beats instead of one terminal jump. The four original statuses
  keep working; illegal transitions and any ack after a terminal status are
  still recorded on the timeline with applied: false and never reopen the
  task. bridge_task_status / bridge_tasks show the new working state,
  the progress of each ack, and a per-assignee row (status, first-delivery
  time, its own ack slice) for broadcast and group tasks.
• Soft task SLA (timeoutMs): bridge_task accepts an optional
  timeoutMs. When it expires without a terminal ack the hub moves a
  still-open task to the new timeout status and appends a hub-synthesized
  timeline event (system: true); terminal tasks are left alone. SLA timers
  are re-armed when the ledger is restored, so a restart does not silently
  drop a running deadline.
• Threads and hop-limited reply chains: messages now carry an optional
  threadId — settable on bridge_chat (which also accepts a ref for
  replies) and bridge_task, and filterable in bridge_wait({ threadId }) —
  so a multi-turn A2A dialog no longer has to share one mailbox slot with
  unrelated traffic. ref chains are capped at MAX_HOP (5) hops: the 6th
  send is rejected at send time with a hop limit exceeded error instead of
  letting two agents ping-pong forever.
• Ack routing survives history eviction: bridge_ack resolves its target
  from the task ledger first and only falls back to the history ring, so a
  late ack or progress beat still reaches the original sender after a busy
  hub has evicted the task message. bridge_chat may also address a group id
  directly.
• Duplicate history / relay fix: broadcast and group fan-out used to call
  the history append once per recipient, so one broadcast produced N history
  rows with the same id and fired N SSE relay pushes. A message is now
  recorded exactly once, at send time — group messages enter history and the
  relay view like any other, and the desktop relay shows a message when it is
  sent rather than when a recipient finally drains its waiter.
• Reliable state persistence: a message, an offline mailbox change, or a
  task-ledger transition (new task, ack, SLA timeout, first delivery) now
  schedules a state save through the same 500 ms debounce. Previously only
  roster and group events did, so a crash between two such events could lose
  the ledger the admin board reads back. Remote mode also persists the ledger
  to SQLite (new tasks table plus thread_id columns, migrated in place),
  restoring it — and any pending SLA — on boot.
• Admin console rewritten: /admin is now a single self-contained React
  page (Tailwind + beui components, built by npm run build:admin into
  assets/admin.html) in place of the old hand-written HTML. It shows the
  roster with client name/version, source IP and a walk-in marker for
  unclaimed anonymous identities; a members-and-tokens panel that issues,
  revokes, and toggles plaintext tokens (the same table as the CLI auth
  commands, manager token only); a copyable MCP configuration snippet
  (generic HTTP + opencode + curl probe) generated right after issuing a
  token; a task ledger board; groups; and a read-only message trace. Tables
  size themselves to their contents up to a cap and only scroll internally
  beyond it.
• Compatibility: every field and status added here is optional or additive, so
  agents running an already-installed 0.8.x SKILL.md keep working unchanged.
  Test suite grows to 282 checks (156 smoke + 37 setup + 11 ops + 35 herdr +
  23 discovery + 20 e2e).

Commits: v0.8.1...v0.9.0 · Contributors: @JochenYang · CHANGELOG