CCCC v0.4.60 Release Notes
CCCC v0.4.60 adds an experimental unattended mode, off by default, in which Chet handles the messages members send you while you are away. Each browser can now choose which messages notify you, and message delivery is easier to trust. When CCCC has the evidence, the recipient list says whether a message was typed into a member's terminal or accepted by its runtime, and it can show why a message is still waiting. Deliveries missed during a short disconnection are recovered. When a restart that keeps agents running leaves two session owners and one stops responding, running members whose healthy owner CCCC can verify keep receiving messages and can be stopped on their own, as the v0.4.59 notes promised.
Upgrading from v0.4.59
Use your usual update method. On Linux and native Apple Silicon Macs, a compatible update can keep agents running, as in v0.4.59; on Windows, updates restart agents. Your Groups, settings, messages, tasks and unread Mail are kept.
- New messages to a member kept running from v0.4.59 can still show Sent to runtime. Restarting the member enables the more specific receipt when it is available.
- If you return to v0.4.59, cards closed only by Chet's answers show as waiting again; Chet's answers stay in the Group. To pause unattended mode, turn its switch off instead.
Let Chet act for you while you are away (Experimental)
- Chet's title bar has an Unattended switch for the whole instance, off by default. While it is on, each new message a member of a Group on this instance sends you with Send or Send + Reply still reaches you, with your usual notifications, and Chet also gets a copy. Chet can reply to the member, answer question cards for you, or point out a better approach or a risk; it is instructed to leave to you what only you can do, such as signing in or paying. Messages from other instances and IM are not copied. Chet's entry below the Group list says Unattended on.
- A card Chet answered is marked as answered by Chet for you, with the option it chose when the recorded answer is loaded, and counts as answered everywhere your pending replies are shown. Reply to Chet's answer to correct it: by default your reply goes to the member who asked.
- Chet's conversation shows each copy, marked as a message to you that was copied, with a link to Chet's reply when one was recorded in the Group. What Chet writes in its own conversation stays there; only the messages it sends to the Group appear in the Group.
- While the switch is on, Chet's entry also asks for attention when CCCC finds that Chet's Runtime is not ready, for example not installed or waiting for setup.
- Only messages sent after the switch is turned on are copied. When it is turned off, an already registered copy that has not yet passed CCCC's last check before handing it to Chet is recorded as not started, and card answers Chet has not written yet are refused; what Chet already wrote stays. A copy that already passed that check can still lead to an ordinary reply.
- Copies use Chet's own session, Runtime and quota, and the CCCC Web service must be running. They share Chet's session with your private conversation with Chet; keeping that conversation out of the Group relies on Chet's instructions. Chet handles copies one at a time, so when members send many messages at once its replies can be delayed. If Chet runs on OpenCode and your own conversation with it still holds a copy back ten minutes after the message was sent, that copy is not handed to Chet; the original message stays in the Group, and any unanswered card is still yours to answer.
- Chet can misjudge, and members may act on its answer before you correct it.
Choose which messages notify you
- Settings > Notifications has a new choice for each browser: All messages to me (as before) or Only when a reply is requested. With the second, ordinary updates and replies no longer notify you; requests for your reply or decision, including choice cards, still do.
- The choice is saved separately for each browser, so a phone can notify you only for requests while a computer notifies you for everything. Switching back to all messages does not notify you about messages already skipped. Browsers that already had notifications on keep All messages to me.
See where a message went
- In the recipient list, a delivered message shows Written to terminal (execution unconfirmed) or Accepted by the runtime when CCCC has that evidence. Receipts without it, including older messages, keep Sent to runtime. Neither label claims the member acted on the message.
- A message that has not been delivered yet can show why it waits: the Group is paused or stopped, the member is disabled or switching runtime, or the message has not been submitted yet. A delivery that can still be retried is no longer shown as failed; Delivery failed is kept for final failures. Mail the member has already read is never shown as waiting.
- For a plain message, Show recipient status in the message's action menu opens this list, so you can see where it went or why it waits. The list refreshes each time it opens and offers Send now for deliveries that are eligible.
cccc ledger search --with-obligation-statusand an agent reading one message by its ID withwith_obligation_statussee the same delivery facts. An agent sees its own status on messages it received and its recipients' status on messages it sent.
Recover missed deliveries
- A message whose delivery had not started, or is known not to have been submitted, during a short disconnection is delivered when the member is back. A full delivery queue now shows the message as waiting and tries again later instead of marking it failed. Long backlogs are delivered in full, and a Group that fails to reload catches up afterwards.
- When a write to a member's terminal fails partway, the delivery stays unconfirmed and CCCC does not type the message a second time.
Members kept running across a restart
- Stopping one member no longer marks the whole Group stopped when CCCC briefly cannot reach another member's process.
- For a member kept running by its session process, Start, Stop, Restart, New session, Remove and enabling or disabling it first check that CCCC can reach that process. If the check fails, the request is refused and the member, its message delivery and its saved session stay as they were.
- PTY input meant for a member's terminal that was just replaced is not typed into the new terminal.
Other fixes
- Uploaded SVG attachments up to 4 MiB preview in chat, the viewer and the resource panel instead of showing "cannot preview"; larger ones can still be downloaded.
- On a phone, closing the last tab in the resource panel closes the panel.
- OpenCode members keep an
OPENCODE_CONFIG_CONTENTconfiguration inherited from the environment CCCC was started in; a member's own setting still wins. - DeepSeek setup reports the managed-session mode.
- Switches, including those in Settings, are easier to read: the knob moves right and fills in when a switch is on.
- In developer mode,
debug_snapshotalso reports ledger-index cache counters and, on Linux, the daemon's process memory.
When one of two session owners stops responding
A restart that keeps agents running can leave two session owners: the earlier process holds the members that kept running, and the newer one holds members started later. In v0.4.59, if either owner stopped responding, members held by the other owner stopped receiving new messages and could not be stopped individually. The v0.4.59 notes promised a fix in this release.
Now, while one owner is unresponsive, running members whose healthy owner CCCC can verify keep receiving new messages, including notices and messages that ask for a reply, and each of those members can be stopped on its own. A message to a member of the unresponsive owner that has not been handed off can remain pending and be sent after that owner responds. Stopping the whole Group still waits for every owner's answer.
In v0.4.59, an owner that stayed unresponsive long enough could also stall background upkeep, such as scheduled rules and catch-up deliveries, for every Group until it responded again. Background upkeep now continues. A read of the unresponsive owner can still time out, so each pass can take a few seconds longer, and if its connection queue is full, the connection fails at once.
For members kept running from v0.4.59, this covers terminal members. Managed and DeepSeek members kept running from v0.4.59 keep the earlier behavior until they are restarted.
Known limitation
If an owner is already unresponsive when CCCC starts, startup can still fail and name that owner. The daemon log gives its PID. If the owner responds again, retry cccc daemon start. As a last resort, verify the process against the reported owner ID and PID before ending only that process. Its work can outlive it. If startup reports unconfirmed work, follow When CCCC cannot confirm an agent before retrying.
Validation
- The changes were peer-reviewed, with a separate review for changes spanning several entry points.
- Linux, on this release's source: rustfmt, Clippy and the Rust workspace tests in CI's three groups, each of the 4,494 tests that are not ignored among the 4,554 listed run once; the Web check and all 5,193 Web tests in 690 files; and the repository's 128 Python tests.
- Upgrades on a Linux x86-64 trial build: from the published v0.4.59 archive, then a restart that kept agents; both times the original terminal member and its MCP process were reconnected and the task in progress kept running.
- Two session owners on a Linux x86-64 trial build. Each scenario kept one owner hung, holding its lock, from before the first request until the healthy member had stopped, while background upkeep kept running. With terminal members, either owner was the hung one; when the newer owner hung, the healthy member had been kept running from the published v0.4.59 archive. Kiro and DeepSeek members, using local ACP stand-ins, were tested with the earlier owner hung. One scenario also filled the hung owner's connection queue. In every scenario:
- sending a message and a message asking for a reply returned in 3 to 12 ms, and each reached the healthy member once, with an acceptance receipt;
- with terminal members, an earlier pending notice and the new message reached the terminal in order;
- stopping that member individually worked in 28 to 72 ms;
- neither request nor the delivery worker contacted the hung owner;
- after the hung owner responded, its own member received its waiting message once.
- Unattended mode: core, daemon and Web tests for which messages are copied and when, Chet's answers, and the copies in Chet's conversation, including after the switch is turned off or the sender leaves. On a Linux trial build, an isolated test supplied Chet's identity and exercised both answers and ordinary replies through MCP and the daemon, and an upgrade from the published v0.4.59 archive ran in an isolated home. In a separate live trial, Chet left progress reports without a reply and answered a question card in about a minute.
- In a headless browser with an isolated home: the unattended switch and Chet's entry at desktop and phone widths, the recipient list labels and waiting reasons, SVG previews, closing the last resource tab on a phone, opening a waiting plain message's recipient list from its action menu and sending it now, and the notification choice in two separate browser profiles, kept after a reload.
Windows and macOS were not tested locally; their CI results are checked before the release is published. The published packages are built by the release workflow from the tagged commit.