Caudra 0.2 Preview 13
- Work that a crash, a kill, or quitting cut off resumes when the session opens again.
- Autopilot is no longer experimental, and engaging it tells you what it runs without.
- JavaScript and TypeScript
codesnippets run as Python when every part of them translates without changing what it does. - Memory recall follows the work in progress and reaches subagents, and decision waits show a timer.
- A task chat can replace the child's current turn, and
/forkforks the session at its latest message. - Shell ETAs read the command itself, and
/shellsshows them with each command's run time. - pstack picks its models from your favorites and
/settings, and its statechart tool runs inside Caudra. - TLA+ files are highlighted and outlined, and Mermaid state diagrams are drawn in the chat.
Important
Close every other Caudra session before you start Preview 13. On its first start, Preview 13 upgrades the session database in your state directory to rebuild the log of shell runs that shell ETAs learn from. While another terminal, editor session over ACP, or
caudra storagecommand still has the database open, Preview 13 lists those sessions and does not start. It also moves the globalcaudra.tomlto format version 4, which it writes to disk only while no other Caudra process runs.Both upgrades save a copy first: the database as
backups/caudra.db.v37.bakandcaudra.tomlunderbackups/config/global/, both in the state directory. Preview 12 and earlier cannot open the upgraded database or read the upgradedcaudra.toml, andcaudra rollbackrestores only the previous binary. See Config file versions.
Important
pstack no longer reads
~/.config/caudra/pstack/models.md, and itssetup-pstackskill is gone. If you wrote that file, choose your models again in/settings→ pstack, then delete the file.
Changes
Section titled “Changes”- Interrupted work resumes on reopen: When Caudra crashed, was killed, or was quit while a session worked, reopening it left the work stopped. Running tasks showed as interrupted or cancelled and never reported, and an open shell call said only
[Tool result not available]. Now the main turn and every agent task the main chat started resume at once, each task under its own task ID and in the background. Every resumed agent first reads a "Resumed after interruption" reminder that lists what was cut off and asks it to check its effects before carrying on. Shell commands are not run again, and a cut-off command's result says its outcome is unknown.[agent] resume_interrupted(auto,ask, oroff) chooses what happens.autoasks instead after a resume already crashed once, afterresume_max_age_hours(default24), or while automatic wakes are suppressed. ACP clients get the resume onsession/load, with what was cut off in_meta.caudra.interrupted. Quitting also no longer records a running task as cancelled withbackground task admission is closed. See Resuming after a restart. - Autopilot is no longer experimental:
/autopilot,--autopilot, its status chip and panel, the SDK controls, and theautopilotview in automations no longer needautopilot = trueunder[experimental]. Autopilot still engages only when you ask, and the[autopilot]limits are unchanged. Caudra comments out anexperimental.autopilotline with the reason. Engaging now adds a notice when no decision endpoint is active or content screening is off, since nothing then keeps a fetched page from steering an approval, and when the model bound to the Autopilot job cannot be used. Each notice says what to do, and autopilot engages either way. SDK clients receive acaveatevent, and-pprints the notice to stderr. See Before you leave it alone. - JavaScript snippets run as Python: A
codesnippet written in JavaScript or TypeScript, such asconst r = await Promise.all([...]); return r.map(x => x.name), was rejected withThis looks like JavaScript, and the model spent a turn rewriting it. Caudra now translates it to Python and runs that, when every part has a Python form that keeps its meaning. The card and the model see the Python that ran. A snippet with a part that has no such form, such asswitch,if (items)on a bare value, or a server call that is never awaited, is still rejected, and the error names that part and its line. See Results and errors. - Memory recall follows the work:
memory_recalljudged notes against your message alone, so "commit this" or "go" could bring up notes about other tasks, and it ran only at the start of a turn in the main session. It now also weighs the to-do in progress, up to 3 open to-dos, and your previous request when the new one is short. It runs again when a to-do step starts, when a goal check finds the goal unmet, and after compaction, at most once every 5 model requests for steps and goal checks. Rare words count for more than common ones, a word matches its other forms, such as "resizing" and "resize", and notes the conversation already read are left out. Subagents recall too. Recall has its ownthresholds.recall_confidence, default0.75, instead ofrouting_confidence. See Relevance from the decision engine. - Decision waits show a timer: Memory recall, skill suggestions, the goal prescreen, the question nudge, and compaction retention waited on the decision engine with nothing on screen. Each now shows a chip on the message, to-do card, goal check, or compaction notice that started it, with a spinner and the time so far, and the chips on tool cards count up the same way. A recall names why it ran, such as
memory_recall · step started · no notes. A decision that changed nothing leaves a dimcheckedchip. Click a chip for the details, including what came closest to the threshold, and click again to fold them. See Decision engine. - Replace a task's turn from its chat: A task chat could only guide the child, so a task heading the wrong way kept going until it finished or you cancelled it.
Ctrl+X x, thereplacecontrol,ron a guidance row, or Replace current run in its menu now stops the child's current turn, with its tool calls, nested tasks, and background jobs, and runs it again on the same history with your message. The agent that started the task gets the result of the new turn. See Task chats. /fork: Copies the session into a new one at its latest message, the same as Fork here in that message's⋮menu. During a turn, the fork stops before the prompt that started it and puts that prompt in the new session's composer. See Message actions.- Shell ETAs read the command: The shell ETA model now reads each command for what decides its time, such as
sleepandtimeoutvalues, loops, chains, pipes, programs, and flags, and adds up a new chain from the times of its parts. In the shell runs recorded on a test machine, it gave a right ETA for three in five commands that ran 10 seconds or more, up from under a third, with fewer wrong ones. The first shell call of each process waits a second or two while the model trains./shellsnow shows a running command's ETA and a clock, such asETA ~40s · 12.3s, and a finished command's run time. Seeagent. - pstack chooses models you can run: pstack's pages named Claude Opus and Grok for every role, so a session without Grok had its tasks rejected, and a review panel ran one model twice. The pages now name six kinds of work, such as Code, Judgment, and Panel, and end with a Model roles table of ready
taskarguments. Code work runs on your Subagent model, judgment on your newest flagship favorite, and a panel on flagship favorites from different providers. Without favorites, Caudra takes the flagship of each provider you are signed in to. The pstack tab in/settingschanges a kind's model and thinking. See Model roles. - pstack's statechart runs without Node or Java:
/pstack:statechartis nowpstack.statechart(...)in thecodetool, with the samefrom-specandfrom-chartcommands, pluscomparefor two charts, and checks specs with tla-checker inside Caudra. A spec's mapping is now a RhaiSPEC.map.rhaiinstead ofSPEC.map.mjs, and charts are Mermaid.mmdonly, so--pngis gone. See pstack scripts. - TLA+ support:
.tlafiles andtlacode blocks are highlighted, with PlusCal inside(* --algorithm ... *)comments, andtlacfgblocks highlight TLC configs.file_indexoutlines specs, andsource_code_refson an action finds theNextthat takes it. See Code Intelligence. - Mermaid state diagrams are drawn: A
stateDiagramblock, such as a chart from pstack's statechart tool, is drawn like a flowchart instead of shown as code. A transition from a state back to itself is drawn in both kinds of diagram. See Diagrams.
- Background tasks outlive a failed or cancelled turn: When the main turn failed or was cancelled, a running subagent or background shell kept running but its card stopped updating and its result never reached the agent. Its card now keeps updating, and its result arrives with your next message.
- Background work survives a brief database lock: A brief database lock no longer stops goal or autopilot sessions from continuing after background tasks finish, or cancels running subagents' shells. Caudra retries, and says so in the chat if saving keeps failing.
- A batched call's decision opens with that call: A decision on a call inside a batch was drawn open at the top of the batch card. It is now a chip on that call's row.
- A stray indent no longer makes a server call unconditional: Preview 12 lined up an over-indented line with its block. When that block names a server or a host function, the indent may be all that is left of a missing
ifor loop, so such a snippet is rejected withUnexpected indentationagain. - No stale quota rows: An OpenAI account no longer shows a
secondary 100% leftrow for a limit its plan does not have, and a window a provider stops reporting disappears. - A disabled MCP server stays disabled: A server with
enabled = falseand an invalid entry showed as failed and counted as enabled. It is now disabled like any other. - Moving a session works while another Caudra is open: Moving a session could abort with
invalid database value in checkpoint log frames: -1while another Caudra process tidied the database. Caudra now treats this as a busy database, and the move goes ahead. - JSON keys have their own colour: Keys in highlighted JSON took the string colour. They now take the theme's member colour, as YAML and TOML keys do.