Caudra 0.2 Preview 17
- Images and PDFs are stored once as files beside the session database, which shrinks it, and resumed sessions keep their PDFs.
- A task whose transcript is too large to save no longer stops every background task, and a result lost that way is delivered after a restart.
/btwand/catchupanswer with read-only tools, such as file reads, searches, and read-only commands, in a larger window that shows each call.- The agent can write reminders and follow-ups for its own session, and one-shot automations can disarm themselves after acting.
- Emil Kowalski's design-engineering skills ship as an opt-in pack beside pstack and mattpocock.
- In a project with few memory notes, the agent is reminded to save one after it learns something worth keeping.
- The agent asks for an API token on a private card, and shell commands receive it as a variable without the value entering the chat.
- Caudra checks a signed model list daily, so new prices, limits, and models arrive without a Caudra update, and the bundled copy works offline.
Important
Close other Caudra sessions before starting Preview 17. Its first start moves the images saved in sessions into the media store, which can take a few minutes on a large database.
Updating from Preview 16 saves a copy first at
backups/caudra.db.v43.bakin the state directory. Older builds cannot open the upgraded database, andcaudra rollbackrestores only the previous binary.
Models
Section titled “Models”- Signed model updates. Long-running sessions check models.caudra.ai daily without delaying startup, and accept only a file signed with a key built into Caudra. New prices and limits apply at the next turn.
caudra models refresh, or/model refreshin the TUI, checks now. Set[models] refresh = falseorCAUDRA_MODELS_REFRESH=0to disable automatic requests and keep using the bundled or cached model list. A project config can only turn refresh off. Set[models] urlin the global config to use a mirror. A downloaded list cannot change the address or API key Caudra uses for a built-in provider. See Model list. - The model catalog works offline. Catalog providers such as NVIDIA resolve from the model list bundled with Caudra, without waiting for a models.dev download. Caudra no longer fetches or caches models.dev at runtime. Neon's models are no longer listed. Its catalog entry never had a usable address, so requests to them always failed.
- Groq models.
groq/<model>runs Groq's catalog models, such asgroq/openai/gpt-oss-120b, withGROQ_API_KEYorcaudra auth login groq. Caudra sends them tohttps://api.groq.com/openai/v1. Earlier versions could not route Groq, because models.dev lists no address for it. - DeepSeek V4.1 Flash on Mistral.
mistral/deepseek-v4-1-flash-1017is DeepSeek's text model hosted by Mistral, with a 1M window. It costs $0.30 per million input tokens and $1.20 per million output tokens, with cache reads at $0.007. See Mistral. - More Mistral names.
mistral/mistral-large-latestandmistral-large-2512are Mistral Large 3, at $0.50 per million input tokens and $1.50 per million output tokens.zai-glm-5andzai-glm-latestare GLM 5.3, andmistral-medium-3is Medium 3.5, at their own prices instead of none. - OpenAI pro models cost what OpenAI bills.
gpt-5.5-proandgpt-5.4-procost $30 per million input tokens and $180 per million output tokens,gpt-5.2-pro$21 and $168, ando3-pro$20 and $80. Caudra used to count them at their base models' prices. They now go through the Responses API, the only one that serves them, and take thinking frommediumtoxhigh. - Z.AI coding plan defaults to GLM-5.3.
glm-5-codeis gone from Z.AI's model list, so the coding plan now starts onzai/glm-5.3. GLM-4.6V, GLM-4.6V-FlashX, GLM-4.6V-Flash (free), GLM-4.5-X, and GLM-4.5-AirX have their own prices and vision support instead of borrowing a neighbour's. - Retiring models leave on their shutdown day. A model with an announced shutdown drops out of the model list and its lane's default on that day by itself, and sessions that used it still show what they cost. Copilot moves its Best default from Claude Opus 5 to Opus 5.5 on November 16, when GitHub retires Opus 5 and GPT-5.6 Terra, and a new Copilot setup now starts on
copilot/gpt-6.1-terrainstead ofgpt-5.6-terra. Also scheduled: Mistral'sglm-5-2on October 31, Anthropic'sclaude-sonnet-4-5on November 30, OpenAI'so3ando3-proon December 11, andgpt-5.3-codexandgpt-5.4-nanoon April 1, 2027. See Providers. - Synthetic's current models. Synthetic stopped serving Kimi K2.5 and DeepSeek V3.2, so
synthetic/hf:moonshotai/Kimi-K3is the new default. The table adds GLM-5.3, GLM-5.3-Flash, DeepSeek-V4.1-Flash, Qwen3.8-27B, Nemotron 3 Super, and gpt-oss-120b with their prices, windows, and thinking levels, and GLM-4.7-Flash now has its 196,608-token window and cache price.
Changes
Section titled “Changes”-
Secrets for shell commands. When a command needs a token, the agent asks for it on a private card instead of in chat. The session's shell commands receive it as an environment variable, such as
$CLOUDFLARE_API_TOKEN, and the value stays out of the transcript, logs, and requests to the model. When a command prints it, the output shows${CLOUDFLARE_API_TOKEN}instead. A secret lasts 24 hours in its session, or you can remember it for all sessions./secretsopens the new Secrets section of/env, where you replace, move, and delete them, andradds 24 hours to a session secret. Secrets reach shell commands on this machine only. An agent that can run commands can still read a secret and send it elsewhere, so review commands that use one as you would any other. See Secrets. -
The agent starts saving notes in new projects. While a project has few memory notes, the main session gets a one-time reminder to save one after a failed command passes, after a long search ends in an edit, or when the context nears compaction. With the new
memory_nudgedecision feature, a message in which you correct the agent or state a rule triggers one as well. Reminders get rarer as notes accumulate, come at most 3 times a day per project, and stop once the agent saves notes without them.[agent] memory_nudges = falseturns them off. A project with no notes also explains notes in its memory view, and the background learner runs after 3 typed messages instead of 10. See Memory. -
Git-aware automations. Experimental automations can use
git_compare(base)to compare a checkout with a local branch without a model turn. Caudra's contributor checkout includessync-main, which asks for a merge at a safe point when localmainmoves and skips when the worktree is current. See Automations. -
Agent-managed automations. With experimental automations on, the agent can write ad-hoc reminders and follow-ups visible only to its session. These session scripts need no digest trust and disappear when the session is deleted. Arming asks for permission unless a rule, YOLO, or autopilot approves. Session disarm and delete need no prompt by default, though your deny or ask rules still win. Project and user disarm and delete still require permission. Approving an arm also trusts an untrusted project script's exact reviewed digest, with its triggers, args and capabilities shown before approval. Reading and validating automations need no prompt. A one-shot script can call
done(reason)to disarm itself when its firing completes, without dropping its queued message. See Automations. -
emil skills.
[plugins.skill] emil = trueadds Emil Kowalski's skills, translated for Caudra, as theemilskill and a/emil:<name>command for each of the 16. They build, review, audit and remove animations, make a web app feel native on a phone, stress-test a UI with worst-case data, prototype variants, and cover Sonner and modern Swift. Audits ask which findings to plan throughquestionand write the plans into the session plan. Visual checks use screenshot tooling your project already has, and findings Caudra could not see are labelled as inferred from the code. Switching the pack applies at once in the terminal UI. See emil. -
The account chip names its quota window. The footer's account chip now says which window its percentage is left in, as
[auto·personal 41% wk]or[personal 12% 5h], and its tooltip spells it out, such as "41% left in the weekly window". The narrow form is[41% wk]. See Accounts. -
The old Lua tools are gone. Caudra no longer ships the Lua versions of its built-in tools, which stopped loading when the tools became native. Plugins can still
requirethecaudra.*modules. A plugin that required one of those tools' own files, such asedit_helpersorlang.rust, now needs its own copy. See Plugins. -
Images and PDFs are stored once. Pasted images, images that tools and tasks return, and fetched PDFs are saved as files under
media/in the state directory, and sessions keep only a reference to each. The same picture in several sessions, forks, or task transcripts takes space once, and the upgrade moves images already saved into the store. Resumed sessions now keep their fetched PDFs. If a file goes missing, the model is told the image is unavailable instead of receiving it. The storage sweep andcaudra storage pruneremove files that no session, archive, or backup refers to after seven days. Backups keep the files they refer to, so restoring one needs themedia/directory beside it./storageandcaudra storage statsshow the store's size, andcaudra storage checkreports missing or damaged files. See Sessions. -
/btwanswers with read-only tools./btw,/catchup,F2, andAsk /btwcan read and search files, run read-only shell commands, search the web, load skills, and runcodesnippets that only read, such as history searches. Caudra refuses any call that would change something. The answer shows in a window 90% of the screen wide, with tool cards and thinking, and a follow-up keeps the earlier tool results. A call that needs approval asks above the window, and denying it fails only that call. A question stops after 25 tool calls, with each host call in acodesnippet counted as one, or after 200,000 tokens beyond the conversation it reads. An answer that finishes past the token limit still shows. A question still reads the conversation's prompt cache and bills underbtwin/usage./extractand the/learntranscripts open in the same window, where dragging over text copies it. See Commands. -
/btwshows where it forks and what it cost. Two rows at the top of the/btwwindow say where the thread forks, with that conversation's context use and the model. The line under each/btwand/extractanswer gives the context the last request sent, the number of requests, output tokens, and cost. It used to add up every request's prompt. Type/in the follow-up field for/new,/catchup, and custom and skill commands. See Commands. -
/btwkeeps its thread. Closing/btwnow keeps the thread, one per chat, until you pressCtrl+L, clickClear, or the session changes.F2,/btw, and/catchupreopen it with its earlier answers. If the conversation moved on, the next question reads it as it stands, with the thread's last 10 questions and answers quoted. Threads stay in memory and are never saved. An answer still streaming when you close the window keeps running, finishes into the thread, and the status bar says when it is ready. On the Plan complete form,F2orAsk /btwasks a side question without leaving the form, and every control on that form, the question form, and the/btwfooter shows a tooltip on hover. See Commands. -
Tabcycles through slash-command matches. In the/palette,Tabcompletes as the fish shell does. When several rows match,Tabfirst fills in the part they all share, then puts each row in the draft in turn.Shift+Tabsteps back,Enterkeeps the row without running it, andEscputs back what you typed. Command arguments and theCtrl+Pargument field cycle the same way. A single match is still taken at once. See Cycling with Tab. -
Structured task results show their contents. When a task with an
output_schemareturns its result, the card now readsReturned resultfollowed by the result's top-level fields, such asReturned result summary, findings, confidence, instead ofstructured_outputandOutput recorded.The card opens with the result drawn as a JSON tree, where clicking an object or list folds or opens it. If the result fails the schema, the errors show below the tree. -
Fewer disk syncs while tools run. Caudra now writes tool statistics, token calibrations, shell timing records, and running task checkpoints together once a second, instead of syncing each one on its own. A crash can lose up to the last second of these records but never leaves one half written, and
/toolscan be up to a second behind. Sessions, permissions, task results, tool outputs, and the spend/usageshows are still saved as they happen. Shell and task records also reuse an open database connection, and a turn's spend is saved in the same commit as its session. In a test turn with 26 tool calls, Caudra held the database write lock for 1.1 s instead of 1.9 s.
-
/envkeeps its size. The text above the list grew and shrank with each row's note, so moving through/envshifted the rows under the cursor. It now keeps the height of the longest note. -
Automation state reaches the agent. With experimental automations on, messages from a completed
done()firing tell the agent that the script disarmed itself. A# Automationsreminder lists armed scripts and explains changes made through user controls or by the runtime, without repeating the agent's own tool actions. -
Renaming an account keeps its sessions. After you renamed an account in
/accountsor withcaudra auth rename, a session using it could fail withAuthentication failed, or move to another account and rebuild its prompt cache. Sessions now stay on the account under its new name and keep the cache, including sessions in other Caudra windows. See Accounts. -
Large task transcripts. A task whose transcript grew past what the session database can hold, such as one that generated many images, stopped all background tasks with
string or blob too biguntil Caudra restarted. Its result is now saved and delivered without the transcript, and continuing the task picks up from the session's own copy of its history. After you update, restart Caudra once to deliver a result already lost to this error. -
Read-only agents start no workflows. A research task, a task started in plan mode, or another read-only agent could start, stop, pause, or resume a workflow through
code. Caudra now refuses those calls. -
/learntranscripts scroll the right way. The arrow keys, page keys, and mouse wheel moved a/learntranscript in the opposite direction. -
A click outside the paste editor closes it. If you edited the paste, the editor first asks whether to discard your edits.
-
A task's footer shows the model it switched to. After you switched a running task to another model or thinking level, its footer and its card in the main chat went back to the old model as soon as the task did anything else, and a reopened session showed the old model too. The task's result also told the main agent it ran on the old model. All three now name the new model. They also update when a task moves to another account's model because the first account ran out.
-
Task rows name their task. In the main chat, a task's card dropped its description when you opened it, and a task run in a batch never showed one. Every task row now names the task in italics, followed by its model, thinking level, tool count, and time, such as
Delegating Find auth middleware (anthropic/claude-opus-5-5 · xhigh · 32 tools · 1m 14.6s). -
One resume row. Resuming a turn that a shutdown cut off showed three rows: the shutdown marker and two that both said "Resumed after interruption". It now shows the marker and a single row naming what resumed, such as
Resumed after interruption: main turn, and both stay when you reopen the session. See Sessions.