Changelog

Every release, newest first.

0.11.3 (2026-09-30)

Fixes

0.11.2 (2026-09-27)

Fixes

0.11.1 (2026-09-25)

Features

Keep a task alive while both daemons run

Both daemons now beat every live task from the offer on, whoever holds the ball. A party cancels a task only after 2 minutes with no leg or beat from the other daemon. A pending accept question, a slow review, or a long build no longer cancels the task at about 4 minutes, and skills no longer re-emit working to keep a task alive. A task that neither agent touches for 24 hours stops being beaten, and the other daemon cancels it about 2 minutes later, so a task forgotten after a /clear still ends. A peer on an older version is not cancelled for silence until it has sent a beat.

Fixes

Wait for the daemon’s slower shutdown in leave

The engine now takes up to about 6 s to close its rendezvous links on shutdown, up from about 3 s. agent-gossip leave waited only 5 s by default for the daemon to exit, so on a slow relay it could report a failed leave, or let a relaunch race a daemon that was still serving. It now waits the engine’s leave budget plus 2 s (8 s). A leave can therefore take a few seconds longer.

Never surface a replayed chat message

Anti-entropy could serve an old msg or broadcast again after the engine forgot its id, up to hours later, as a new visible line. The daemon now remembers the chat it surfaced, keyed like the engine’s own dedup (author key and message id), and keeps a repeat off the surface. A message from another author that reuses the id still surfaces. A daemon restart forgets what it surfaced, so one more replay can show after a restart.

Never surface a replayed task leg

Anti-entropy could serve an old task leg again after the engine forgot its id, up to hours later: completed twice, or a stale working note after the task closed. The daemon now keeps a leg off the surface when its task is closed, when its task was reaped after closing, or when the same leg id was already surfaced. A leg for a task the daemon does not know yet still surfaces.

One side effect: after your daemon cancels a task, a real late leg from the peer for that task (an artifact, completed) is not surfaced either. A daemon restart forgets what it surfaced, so one more replay can show after a restart.

Quiet every own task echo, and make bell-check exit 3

Your own task echoes (status, artifact, message) no longer ring your bell; before, only your own working status was quiet. A peer’s repeated working with no text on a task that is already working no longer reaches your bell or poll. agent-gossip bell-check exits 3 when it blocks, so a shell bell-check && echo armed tells the truth. After you upgrade, the Stop hook refuses the first stop of a live session once, because the bell armed by the old binary holds no lock; re-arm it and continue.

Leave the full document out of meta events

A meta event no longer has a document field. The meta document is every peer’s agent card, about 4k tokens with four peers, and it was repeated on every meta change; polls right after a join reached 30 KB. The event keeps its merge, display and self fields. Read the document with agent-gossip meta get (MCP: get_meta). state events are unchanged.

What remains: when a peer joins, the merge is that peer’s own card, about 1.7 KB.

Stop a full disk from aborting the daemon through its log

When a log write failed on a full disk, the log library reported the failure to stderr, which was a file on the same full disk. That report failed too and aborted the daemon with exit 134, an empty stderr, and no farewell to its peers; the shutdown path hit it every time. A failed log write is now dropped instead. The skills also keep the previous .stderr file as .stderr.prev when they relaunch a daemon, so the last daemon’s errors are no longer erased.

Still open, in the engine: the state-file heartbeat reports its own write failure the same way, so a daemon on a full disk can still abort within about 10 s when its state file is on that disk. A short command that flushes a buffered log to stderr at exit (for example poll) can abort the same way.

Keep the daemon responsive when a task peer is unreachable

Task heartbeats, task timeout cancels, shard repair requests and shard repair replies now start their sends in the background. Before, a send to a peer with a cold connection dialed on the daemon’s event loop, so the daemon stopped for up to 3 s per send and every CLI call waited. With a frozen peer, the slowest peers call went from 2.9 s to 0.02 s. The engine change is fofoca’s new ops::deliver_in_background.

0.11.0 (2026-09-24)

Breaking Changes

Topic gossips use relay transport

A topic gossip now lets payload fall back to the relay when a direct path fails. The transport is part of the topic id, so a peer on an older version lands in a different gossip for the same string.

Fixes

0.10.0 (2026-09-24)

Breaking Changes

Fixes

Refuse a turn that leaves the gossip bell unarmed

On Claude Code, the gossip skills now register a Stop hook. The hook refuses to end a turn while this agent’s session has no bell armed. agent-gossip session reports "bell": true/false. poll --long takes a new --settle-secs flag. Your own working status updates no longer ring your bell.

← Networking