Caudra 0.2 Preview 14
- Clicking the update banner, or running
/update, installs the new release and restarts Caudra on the same sessions. - A shell command still running after 10 minutes moves to the background, and tasks and shell commands have one way to run there.
- Commands list their arguments as you type, and dashed commands such as
/goal-clearbecame subcommands such as/goal clear. - To-do lists fill in while the model writes them.
/catchupsummarizes a session you lost track of, and/btwworks in subagents chats.- Decision calls count toward your spend in
/usage, goal spend, and Autopilot'smax_cost. - Long sessions stay fast, and long subagent runs no longer block background results.
Important
Close every other Caudra session before you start Preview 14. On its first start, Preview 14 upgrades the session database in your state directory to keep each background task's transcript beside its record. While another terminal, editor session over ACP, or
caudra storagecommand still has the database open, Preview 14 lists those sessions and does not start. It also moves the globalcaudra.tomlto format version 5.Both upgrades save a copy first: the database as
backups/caudra.db.v38.bakandcaudra.tomlunderbackups/config/global/, both in the state directory. Preview 13 and earlier cannot open the upgraded database or read the upgradedcaudra.toml, andcaudra rollbackrestores only the previous binary. See Config file versions.
Changes
Section titled “Changes”- Long shell commands move to the background: A shell call waited for its command however long it ran. It now waits up to 10 minutes, then the command becomes a background job without a restart, and the agent gets the output so far. The result arrives when the command exits or reaches its
timeoutSecdeadline. Change the wait withagent.shell_foreground_limit_secs.agent.shell_async_threshold_secsis retired, and Caudra comments it out with the reason. The SDKinitmessage reportsshell_execution.foreground_limit_secsin place ofasync_threshold_secs. See Execution policies. - One background mode for tasks and shell commands:
agent.task_executionandagent.shell_executioncould make every task or shell command wait or return at once. Both are retired, and Caudra comments them out with the reason. A task or shell command runs in the background when the model passesbackground: true,/tasks backgroundstill promotes a running task, and--printalways waits. The SDKinitmessage no longer has atask_executionobject. See Background tasks. - Commands guide you through their arguments: The command list shows each command's arguments, and after the command a list offers values for the slot under the cursor, such as thinking levels, task ids, workflow runs, and docs pages.
Tabtakes the highlighted value, andEnteron a mistyped word runs it, so/thinking hghsetshigh. TheCtrl+Pmodal does the same, and Lua plugins add values with the newcompleteoption ofregister_command. Commands that were verbs of another command are now its subcommands:/goal clear,/goal model,/stash pop,/stash list,/update dismiss,/model next,/decisions model,/decisions next,/sessions move,/sessions migrate, and/searchin place of/search-history. Typing an old name finds the new form. See Arguments. - To-do lists stream as the model writes them: The
todo_writecard and the sticky to-do panel show each item as soon as the model starts writing it, instead of the whole list once the call ran. If the call fails or is cancelled, the panel returns to the last stored list. ACP clients get aplanupdate as each item is finished. See Turn updates. /catchup: Asks a/btwside question that summarizes the session: the goal and where things stand, what needs you, what the agent is doing, and what is done, decided, and still open. Add a few words, such as/catchup the migration, to narrow it. Nothing enters history. See Commands./btwin task chats: In a task's chat,/btwand/catchupnow ask about that task's conversation, andF2on a question a task asked clarifies it over that task's conversation instead of refusing. See Commands.- Decision calls are priced: Calls to a decision endpoint showed in
/usagewithout a cost. TypeSafe Jev now costs $0.042 and the OpenAI Decisions API ongpt-6-luna$0.10 per 1M input tokens, and output is free on both. Decision spend counts toward the session's cost, goal spend, and Autopilot'smax_cost, including calls that timed out or were cancelled. Other models, such as a localdecision-1, stay unpriced. See Decision prices. - Long background shells ask whether to keep waiting: Every 30 minutes a background shell runs, Caudra shows the agent waiting on it the latest output and asks whether to keep waiting or cancel, so a command stuck on a prompt or a lock no longer holds the agent until its deadline. Set
agent.shell_wait_nudge_minutesto change the interval, or to0to turn it off. See Long-running background shells. - Shell results say more: A result now says how long the command ran, such as
[shell status: exit code 0; took 3m12s], and a cancelled command reports its output so far instead of onlyShell execution cancelled. See shell. - More of the composer can be undone: Clearing the draft with
Ctrl+C, completing a command, browsing input history, starting a workflow, and applying a saved draft are now undo steps, and undo brings paste chips back with their content.Ctrl+Shift+Zredoes in every text field, andSuper+Zundoes on terminals that reportSuper. See Text Fields. - Recently finished tasks come first:
/taskslists finished subagents newest first, matching/shells.
- Tab completion keeps the rest of the draft:
Tabon a command replaced the whole draft with the command name. It now replaces only the command word, so/goFix the bugcompletes to/goal Fix the bug. See Commands. - A goal set during a turn no longer waits in the queue:
/goalor/autopilot <goal>sent during a turn became a plain queued prompt, which a reopen could send as bare text. The goal now takes effect at once and starts after the current work. See Send next. - Reopened sessions hold their queued prompts: Reopening, resuming, or rewinding a session sent its saved queue at once. Saved prompts now wait under Held until sent until you choose Send now, edit, or delete them. See Recovery.
- Long sessions stay responsive: In a session with many subagents and shell commands, opening a subagent, scrolling
/tasks, and redrawing got slower the longer it ran. They now stay about as fast as in a new session. - Long subagent runs no longer block background results: A subagent transcript past 8 MiB could not be saved, so Caudra kept showing
Background results could not be saved, retryingand delivered no more results. Transcripts are now saved in the session database at any size. See Background tasks. - A replaced subagent turn can start background work again: After you replaced a task's turn from its chat, its new background shells and tasks were refused with
background task admission is closed. They now start. See Task chats. - Database upgrades show progress: An upgrade names each step with a spinner, where a full progress bar made a long upgrade look frozen.
- Plan cards no longer flicker: A plan card's preview briefly disappeared before its result arrived, which made the transcript jump.
- Update from the banner: The update banner only named a command to run in another terminal. Clicking it, or running
/update, now asks whether to update, installs the release in the terminal, where a sudo prompt can reach you, and restarts Caudra on the same tabs with the same flags. Running work is interrupted, andagent.resume_interrupteddecides whether it resumes. On Windows and in an--ephemeralrun, Caudra exits instead of restarting. A Homebrew, Nix, or Cargo installation says to update through its package manager. See ui.update_check.