CCCC v0.4.59 Release Notes
CCCC v0.4.59 lets you maintain CCCC while your agents work. On Linux and native Apple Silicon Macs, you can restart CCCC or install a compatible update while a managed agent finishes its turn. CCCC reconnects to the same process with its terminal and output intact, and a retained browser member can continue in the same page. When CCCC returns, it reports which members it found, restarted or could not confirm. When something goes wrong, CCCC now says what it confirmed and what to do next.
Upgrading from v0.4.58
Let active turns finish before this first upgrade. The v0.4.58 updater stops the agents it manages when it replaces CCCC. Use your usual installation method, then start CCCC with the same home. Your Groups, settings, messages, tasks and unread Mail are still there. You can choose to keep running agents during later compatible updates.
Choose how to update
In Web, open Settings > This instance > Update to check for a release. If the website installer installed CCCC and the Web host started its own daemon, select Update now, then Keep agents running or Restart agents. The terminal interface offers Keep agents; update where supported and Update and restart agents under /update. To use a locally rebuilt executable, choose Restart CCCC in Web or /restart in the terminal. The terminal offers Keep agents and restart CCCC where supported and Restart agents with CCCC. CCCC checks for a release when you ask. Other installations show instructions for their update method.
For a daemon you manage separately, use cccc daemon stop --keep-agents followed by cccc daemon start. To update a website installation in between, run the commands together.
cccc daemon stop --keep-agents && cccc update && cccc daemon startA plain cccc daemon stop stops managed agents. Sessions you launched yourself continue independently. If a plain stop cannot confirm that an agent stopped, it says so and suggests running cccc daemon stop again: with CCCC stopped, that run reaches the process holding the agent directly. When the receipt names that process, the review command comes second. After cccc daemon stop --keep-agents, CCCC suggests only the review command, so following the advice does not stop the agents you kept. Windows users can update through their installation method and restart agents. Keeping managed agents running requires Linux or the native arm64 macOS build.
Returning to v0.4.58
Let active work finish. If you stopped CCCC with --keep-agents, start the same newer executable with the same CCCC_HOME so it can reconnect to the members. If you use the current website installer, keep that daemon running while you install v0.4.58. The installer stops it and checks that the managed agents exited before replacing the executable. For a manual replacement, use an ordinary cccc daemon stop from the newer installation and confirm that the managed agents exited before replacing the executable. A managed agent still running under the newer version can otherwise be started a second time by v0.4.58.
Messages sent while CCCC briefly cannot reach a member
On Linux and native Apple Silicon Macs, each managed member runs under a small background process, its session owner, that keeps it running while CCCC restarts. If CCCC briefly cannot reach that process, a message sent in that moment is now delivered once the process answers again, without another message or a restart. Later messages to the same member wait behind it, so they arrive in order. A delivery that may already have reached the terminal is still never sent again on its own.
Messages sent during startup
When a Kilo terminal opens, CCCC waits for an empty chat input before sending the first queued message. In Hermes, CCCC sends prompts and tools through the same member session you see in its terminal. Messages queued during initial setup are available when that session is ready.
DeepSeek setup and managed turns
DeepSeek members use the official Harness developer preview, pinned to 0.2.2-rc.5. Before starting or restarting a DeepSeek member, run cccc setup --runtime deepseek on the host. A member kept running during an update can finish with its original tools. Restart it after setup to use the new tools in a fresh Harness session. Its earlier session history is preserved.
Batch cccc setup counts a runtime as configured when it writes or confirms its CCCC MCP entry. Runtimes that receive MCP settings at launch can report ready separately.
For managed members using generic ACP, activity shows tool starts and outcomes.
When an accepted prompt on the shared ACP path returns a provider error after you request cancellation, its activity shows the failure reason. This applies to members such as Hermes, Copilot and Cursor. OpenCode continues to settle its admitted input as cancelled.
If the recipient panel shows Delivery uncertain, check the member’s session before choosing Retry anyway. Retrying can run the same input twice.
Browse message attachments
Open a previewable attachment in the full-screen viewer to browse the other attachments in that message in order. Closing the viewer returns you to the chat or resource tab you opened it from. The resource panel can stay open after its last tab closes, ready for the next attachment.
Task details beside the board
On larger screens, opening a task in Project Context > Tasks shows its details beside the board. Select another card to switch directly. Drag the divider or move it with the arrow keys. CCCC remembers your preferred width and fits it to the available space.
On a phone or in a narrow window, details fill the Tasks area. Back to tasks returns to the board at the same scroll position. CCCC asks before discarding unsaved edits and lets the current save finish before you switch tasks.
Clearer failure reasons and less idle work
Managed-member failures keep ordinary text that used to be masked because it matched an inherited environment value. Values configured for the member or its Profile and recognized credentials are still redacted. A member kept running through an update uses the diagnostic rules it started with. Restart that member when it is idle to apply the new rules.
Browser-based Web Model members skip rewriting the Group file on each idle poll when their recorded waiting state is already current. Each round of member-status publication also shares one snapshot per session owner across Groups, which reduces repeated lookups on an instance with several Groups.
When you switch back and forth between two Groups with long chat histories, CCCC can now keep both message-status indexes in memory when each fits the base cache budget. Previously, if the two indexes exceeded that budget together, switching Groups could reread an entire ledger on every visit.
When a member starts its next turn immediately, a delayed status update for the previous turn could leave the Web UI showing the member as working after it finished. The Web UI now keeps the newer observation.
A member that reads another member's agent state through MCP now gets that member's state instead of its own.
Smaller fixes
- In a managed Claude Code terminal, a local command such as a confirmed
/modelno longer makes the member appear to be working until its next reply. - CCCC no longer crashes when an environment variable holds a value that is not valid Unicode.
- If
CCCC_HOMEis too long for the daemon's socket path,cccc daemon startnames the path and suggests a shorterCCCC_HOME, such as a symlink to the same home. - A command quoted in the top error banner no longer breaks in the middle of an option such as
--runtime. On a board with two or more columns, Planned comes first; Active comes first only when the board shows a single column. Non-image attachments such as PDFs show a file icon in the resource panel.
When a start begins a new conversation
Starting or restarting a member can begin a new conversation instead of resuming the saved one, for example after you change its model. The Web now says so at that action, with the reasons it knows, and the CLI prints a note on stderr so its JSON output stays parseable. A Group start names each member whose saved conversation was not resumed.
OpenCode members and Chet
OpenCode now uses its own settings to handle messages delivered to a Group member. When you chat privately with Chet on OpenCode, the same settings let a new message join a running turn or wait for the next one. A Group request to Chet still takes a turn of its own. It waits while messages from your own conversation with Chet have not been taken up yet. If it is still held back ten minutes after its message was sent in the Group, it is not sent: the Group message and Chet's history say it was not sent to Chet and did not run, and it can be sent again in its Group. Once a Group request is queued with Chet, messages you send afterwards go after it, unless Chet is answering another Group request at the time.
When Chet runs on Claude Code and Claude has not trusted Chet's working folder, the Chet panel names the folder and how to trust it.
If an update is interrupted
If the replacement Web host cannot listen after you chose to keep agents, CCCC keeps the members and reports how to recover. Follow the recovery command it prints. For an interrupted local restart, start a working CCCC executable with the same home. Check the reported member state before sending a prompt again.
A CCCC tool call can fail while the daemon is offline. For ChatGPT Dots and Grok Bot Web Model, CCCC retains the saved cloud connection and reports unfinished remote work as unconfirmed. Check the provider’s result before retrying.
Known limitation
After a restart that keeps agents running, CCCC can use two session owners for a while: the earlier one holds the kept members, and the new one holds members started afterwards. If one of them stops responding while CCCC runs, members held by the other may keep working, but CCCC cannot send them new messages or stop them individually. A fix is planned for the next release. Until then:
- CCCC is still running. Leave it running. When the owner responds again, CCCC delivers the pending messages it had not handed off, and new messages and individual stops work again.
- The owner has exited. Run
cccc daemon stop --keep-agents, thencccc daemon start. Members on the healthy owner can keep working; once CCCC starts successfully, it reconnects them and delivers the messages it had not handed off. In our test, the failed owner's custom terminal member stopped, and CCCC restarted it because the member was enabled in an active Group. For other runtimes, follow the startup report; for a Claude Code member, see below. - CCCC was stopped while the owner still did not respond.
cccc daemon startfails and names that owner's ID; the daemon log it points to also gives the process ID. If the owner responds again, retrycccc daemon start. As a last resort, end that owner's process only (its command line containsdaemon session-owner --owner-idfollowed by that ID), then runcccc daemon startagain. Once it starts, CCCC reconnects the members of the other owner and delivers their messages. For a Claude Code member of the owner you ended, see below.
A Claude Code member's work can outlive its owner. CCCC refuses to start a second copy while the earlier work is unconfirmed: cccc daemon start fails and prints a review command for that owner. Run claude agents --json with the Claude Code configuration the member was started with. A job still listed is still running, even when its reply looks finished; once you no longer need its work, end it with claude stop <id>. After a successful listing confirms the matching job is absent, review the records for that owner and release only those you have checked, as described in When CCCC cannot confirm an agent, then run cccc daemon start. In an active Group, CCCC then starts an enabled member again in the same Claude Code session.
Avoid a plain cccc daemon stop in this situation: it also stops the members that are still working, and CCCC cannot start again while the owner remains unresponsive and holds its lock.
Validation
- Each change was reviewed by two maintainers, and the whole range since v0.4.58 was reviewed again independently by both before this release.
- Linux: the Rust workspace tests in CI's three groups, rustfmt, Clippy, the Web check and tests, and the repository's Python tests. On the final code, each of the 4,452 Rust tests that are not ignored among the 4,512 listed ran once and passed, and the Web check and all 5,155 Web tests passed. The 128 Python tests passed; the Python code has not changed since.
- Upgrades on Linux x86-64, using a trial archive built from this release's final code: from the published v0.4.58 archive, with active turns left to finish first; from an earlier compatible build through the Web update with an agent kept running mid-task, together with its CCCC tools; and back to v0.4.58, which refused while agents were held by the newer version and then completed after they were reconnected and stopped.
- On an earlier Linux trial build, with the network disabled and local test models, members on Claude Code, Codex, Kilo, OpenCode and DeepSeek kept running across a restart that kept agents, and each replied once to the next message. In a real terminal, the terminal interface's
/restartcompleted in both modes. - Earlier builds of this release exercised command-line retention, Web restart into a local build and installation from a release archive on Linux and Apple Silicon. On Apple Silicon, six command-line agents (Claude Code, Codex, Grok, Antigravity, Cline and Hermes) kept running across an update, and in four Web cases the original terminal member and its tool process continued, then exited on an ordinary stop.
- Messages during a brief owner outage: on Linux, controlled replays through the normal daemon and send commands delivered the missed message once and in order, including a 16-second outage; repeated runs under CPU contention delivered every message exactly once.
- On Linux with the network disabled and a local test model, the latest Claude Code 2.1.330 passed the
/modelreplay, and Claude Code 2.1.330, Codex 0.171.0 and Kilo 7.8.32 passed the native session checks that CI runs with the versions it pins.
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.
Thanks
Thanks to chriscoveries for isolating the scope detection test from the host Git configuration. Pull request #230