Caudra 0.2 Preview 12
- Caudra learns from your conversations in the background and keeps what it learned in your instruction files and project memory.
- pstack's engineering skills come built in, off until you turn them on.
- Config files upgrade themselves to a new format, with a backup first and a note on every line they change.
/settingseditscaudra.toml,websearchfollows the backends you choose, andwebfetchreaches the local servers you list and reads Reddit again.- More lookups rest as one row that says what they found, and a
codecard folds a printed dict or list to a tree. - Cards say how many tokens a result adds, with estimates that learn each model's tokenizer from what its provider billed.
- Shell cards show an ETA from a small model that learns from your own past commands, with no decision endpoint.
- Decisions show on the tool call's card, apply as soon as you save them, and content screening can withhold a flagged result.
Important
Close every other Caudra session before you start Preview 12. On its first start, Preview 12 upgrades the session database in your state directory. The upgrade needs exclusive access, including from other terminals, editor sessions over ACP, and
caudra storagecommands. While any of them still has the database open, Preview 12 does not start and changes nothing. It lists the sessions to close instead, with each one's process ID and title.The upgrade keeps your sessions and saves a copy of the database from before it as
backups/caudra.db.v32.bakin the state directory. Preview 11 and earlier cannot open the upgraded database, andcaudra rollbackrestores only the previous binary.
Warning
Before you start Preview 12, close older Caudra sessions so it can upgrade your config files.
caudra.tomlmoves to format version 3, andmcp.tomlandpermissions.tomlto version 2. Preview 12 rewrites the global files when it starts, after it saves a copy of each, but only while no other Caudra process is running. Otherwise it reads them upgraded in memory and leaves them unchanged on disk, and saving to them from inside Caudra, such as with/envor/mcp, fails until Caudra starts with no other session open.Preview 11 and earlier cannot read an upgraded file. To go back, copy the file's backup from
backups/config/global/in the state directory over it. Files in a project's.caudra/are never rewritten on their own. If you commit them, runcaudra config migrate --projectonce everyone who shares them runs Preview 12. See Config file versions.
Models
Section titled “Models”- MAI-Code-1.1-Flash on GitHub Copilot.
copilot/mai-code-1.1-flashandcopilot-enterprise/mai-code-1.1-flashnow have their price and size class before Copilot's model list arrives: $0.20 per million input tokens and $1.20 per million output tokens. Caudra counts it among the Fast models. See Copilot. - GPT-6 Astra and Claude Fable 5.1 on GitHub Copilot. Both cost $10 per million input tokens and $50 per million output tokens on Copilot. Caudra used to bill Fable 5.1's cache reads at Fable 5's $1 instead of $0.25, so a Fable 5.1 session's cost in
/usagecomes out lower from now on.
Changes
Section titled “Changes”- Continual learning: Caudra now learns from how you work with it, in the background. After at least 10 messages you typed and 2 hours since the last run, once the session has been idle for a minute, a read-only learner on the new Learn model job reads the project's recent conversations and checks each finding against the project's history. Caudra writes what will still hold to a managed block in the project's
AGENTS.md(shared conventions),AGENTS.local.md(your machine's setup, which Caudra keeps in.gitignore),~/.config/caudra/AGENTS.md(your own preferences), or project memory notes. It never edits text outside its block. The status bar says when a run changed something./learnlists the runs with what each changed, opens a run's transcript witht, runs one now withn, reverts a change withu, and pauses the project withp. It is on by default. Turn it off with[agent] continual_learning = false, or narrow where it writes withlearning_targets. Its spend shows underlearningin/usage. See Continual learning. - More lookups fold like a file read:
history,tool_output,skill,host.environment,list_sessions,read_topic, andautomationcalls, and theplan,workflow, andtask_controlcalls that only read, now rest as one row that says what they found, such as how many hits a search kept or a run's state. A click opens the whole answer. Calls that change something keep their cards open. The newplan_read,workflow_read, andtask_control_readentries of the defaultui.always_collapsedlist cover the reading calls. If you set your ownalways_collapsedlist, add the entries you want from the default. - Printed dicts and lists fold to a tree: When a
codesnippet prints or returns a dict or a list, the card shows it as a folded tree, such as▸ {…} 7 keys, and a click opens or folds any part of it. The raw view of an MCP call's card draws its JSON as the same tree. When the tree only repeats one call's answer, such asprint(await host.environment()), that call's card shows open instead. The model still reads the output as text. See In the transcript. - Clearer card titles: A skill's card reads
Skill loaded <name>.task_control,workflow, andautomationcards have their own titles with the action as a verb, such as◷ Jobs checked shell-66or⑂ Workflow started review, in place of the tool's name and raw action. Titles set the names and paths they show in italics. - Cards say what a result costs the context: A card whose result the model reads now estimates its tokens beside what it found, such as
209 lines · ~2.1k tokenson a file read. Caudra counts tokens locally with OpenAI's tokenizer, and now learns from what each provider billed how far that count is off for each model, and keeps it across sessions. Card counts, the/contextbreakdown, the tool ledger in/usage, and automatic compaction all use what it learned. See Watching it work. codecards keep a clock: Acodecard's header counts up while the snippet runs and then shows how long it took, such as· 2.4s, as ashellcard does. A run that slept no longer shows a stray separator, and itsbatchrow keeps what it found. See In the transcript.- Time and usage move into the
⋮menu: Tool card headers no longer end in the time and the request's usage. The first line of a message's⋮menu shows them, with the date when the message is not from today, for every prompt, reply, reasoning block, and tool call, in restored sessions too. A running turn's messages and a subagent's transcript have the menu as well. See Message actions. - Choose your web search backends:
websearchused Exa's hosted search no matter what you set. It now followsWORKCELL_WEBSEARCH_BACKENDandSEARXNG_URL, and the new[websearch]table in the globalcaudra.toml. List several backends inbackends, such as["searxng", "exa-mcp"]. In the defaultfallbackmode, the next one answers when one fails.singleuses only the first, andalllets the model pick one per call. Edits apply at once in every open session of the terminal UI. With nothing set,websearchuses Exa's hosted search as before. See Web search. - Edit
caudra.tomlfrom/settings:/settingshas a Web search tab that shows whether each backend is ready and lets you set up, reorder, and switch them, and tabs for the top-level keys,[ui],[agent], the builtin skills,[provider],[storage],[worktrees],[autopilot], and[automations]. Switches, value chips, and list editors change a key, and/filters every tab. The pane under the list explains the selected key, its default, and when an edit applies, and?opens its section of the configuration reference. It writes the globalcaudra.tomland keeps your comments. A key your project'scaudra.tomlalso sets is marked, since the project value wins. See Configuration. - Decision settings apply as soon as you save them: On the Features tab of
/decisions, click a feature's mode or pressSpaceorEnterto move to its next, such asoff,shadow, thenenforce. Caudra saves it under[decisions.features]in the globalcaudra.toml, keeping your comments. An edit to[decisions]orui.show_decisionsincaudra.toml, to the active decision endpoint's table inproviders.toml, or todecisions/permission.jsonnow applies at once in every open session, without/reload. A projectcaudra.tomlthat turns a feature off still wins. See the decision engine and Hot reload. - Decisions about a call ride on its card: A decision about a tool call, such as a
source_code_contextrerank or Auto screening, no longer draws its own row under the card. The card's header shows a chip, such as◆ reranking…and then◆ reranked, and opening the card shows the decision and its scores. Chips that warn, such as◆ asked you, use the warning colour. Decisions about a turn, such as a memory recall, keep their rows. See Decision rows in the transcript. - Rerankers judge a wider pool in one request: With a decision endpoint, a rerank now sends up to 64 candidates in one request instead of 16 at a time.
memory_search,history_search, andsource_code_contextjudge up to 64, and inenforce,source_code_contextkeeps only the rows at or abovethresholds.relevance_floorunless none clears it. A request waits onererank_timeout_msfor every 16 candidates, andrerank_timeout_msnow defaults to 3000 instead of 2000. A workflow run paused in the middle of arerank()on an older release fails when it resumes. See Reranking. - Decision requests wait up to 10 seconds by default:
[decisions] timeout_msnow defaults to 10000 instead of 800, so a busy or local engine no longer runs out of time and leaves a feature without an answer. A slow engine can now hold a screened tool call, or the start of a turn for memory recall, for up to that long. Settimeout_mslower to fall back sooner. Seedecisions. - Content screening can withhold a result: The new
content_screening = "enforce"keeps a flaggedwebfetch,websearch,code, or MCP result from the model, which reads a notice in its place asking it to tell you. Caudra deletes the result's saved copy, sotool_outputcannot read it back. The card keeps the result for you, markedDecision engine: withheld from the model. Enforce fails closed: a result is also withheld when the engine fails or times out, or when part of it could not be screened. The feature still defaults tooff. See the decision engine. - Shell ETAs from your own runs: A shell card's ETA now comes from a small model that runs on your machine and learns from the commands you run, in place of the
shell_durationdecision feature. It needs no decision endpoint and sends nothing anywhere. A command like one you ran before shows how long that took, such asETA ~40s, and others showETA secondsorETA minuteswhen the model is confident. It only labels the card and changes no timeout. On its first start, Preview 12 trains it from the newest 20,000 shell calls your saved sessions hold, keeping only hashes of each command's words, and the first shell command after Caudra starts waits while it trains. It is on by default. Turn it off with[agent] shell_eta = false. Theshell_durationfeature and its thresholds are gone, and the config upgrade comments out those keys. Seeagent. - pstack built in: Lauren Tan's pstack skills for Cursor now ship with Caudra under their MIT license, with the
deslop,control-cli, andcontrol-uiskills from Cursor's team kit. Turn them on with[plugins.skill] pstack = true. The agent then sees onepstackskill whose index lists every skill and whose pages load by address, such aspstack/poteto-mode, translated to name Caudra's tools, paths, and commands. Each skill you run directly gets a/pstack:<name>command, pstack's task profiles appear, and its scripts becomepstackcalls incode, ported to Rust so they need no Bun, Node, or Bash. Ifghkeeps its login inGH_TOKENor a keyring, forward those variables with/env. Benny's Slack triage and watch-pr's drive mode ship as automations: withexperimental.automationson, they are listed in/automationsunder the newbundledscope, ready to arm. See pstack. webfetchreaches the servers you list:webfetchrefuses private, loopback, and carrier-grade NAT addresses, so it could not read a local dev server or a machine on your tailnet. The new[web]table in the globalcaudra.tomllists the ones it may reach in local sessions.allow_private_hoststakes exactHOST:PORTentries, such as"localhost:3000", andallow_private_networkstakesloopback,private,cgnat,tailscale, or a CIDR. Whenwebfetchrefuses a private target, the error names theHOST:PORTto list. Remote workspaces and sandboxes keep the public-internet rule. See[web].webfetchreads Reddit: Reddit answers page fetches with a script challenge, sowebfetchcould not read any post. It now reads threads, subreddits, user pages, and searches from Reddit's feed or JSON API, and gives the model the post and its comments as Markdown. A rate-limited fetch on any site that sendsRetry-Afternow says how many seconds to wait.- Config files upgrade themselves: Every TOML config file now has a format version in its top-level
versionkey. When a release changes a format, Caudra upgrades your files and keeps your layout and comments, with a comment above each line it changes, such as# Upgraded by Caudra 0.2.0-preview.12 on 2026-10-20: renamed .... A removed key is commented out rather than deleted. A notice says when a project file is behind.caudra config migrateupgrades the global files on demand,--projectadds the project files, and--dry-runshows a diff without writing. See Config file versions. - Old keys are upgraded instead of refused: Keys from earlier releases that stopped Caudra with an error, or silently stopped matching, now move to their new form when the file is upgraded. For example,
code_contextandcode_refsbecomesource_code_contextandsource_code_refsincaudra.tomlandpermissions.toml,[python_execution]becomes[code], and[bash]becomes[shell], so rules written under those names apply again. Removed keys, such asagent.defer_builtin_toolsand the top-leveldefer_toolsinmcp.toml, are commented out. A file that sets both an old key and its new name is left unchanged, with an error that names both. The legacy decision endpoint keys in[decisions], such asbase_url, still stop Caudra with the error that shows theproviders.tomltable to write. - Backups have their own folder: Copies Caudra saves before it changes a file now go to
backups/in the state directory, such as the session database before a schema upgrade and each config file before an upgrade. Backups are owner-only and never deleted. Database backups from earlier releases stay besidecaudra.db. See Backups.
- Content screening reads the whole result: It read only a few samples from the start of a result, so an instruction in the middle of a long page was never checked. It now reads the whole result in overlapping chunks.
- Switching decision endpoints keeps flagged content flagged: A session whose content screening had flagged web or MCP output lost that flag when you switched the decision endpoint. It now keeps it.
- Every task card names its model: Every task card and
batchrow now names the model its subagent runs on, while the task runs, after it finishes, and after a reload. - No false "statement of intent" nudge while work runs: The nudge that a turn ended on a statement of intent now waits until no task, background job, or workflow is running.
- Claude models on GitHub Copilot cache their prompts: Claude models through
copilotandcopilot-enterprisenever marked where the prompt may be cached, so every turn sent the whole conversation as fresh input. They now mark it the way theanthropicprovider does, so later turns read the conversation from the cache. - Copilot Enterprise takes a tenant name: A GitHub Enterprise hostname without a dot, such as
company, now meanscompany.ghe.com. /loginsays why it failed: A failed sign-in from the TUI now carries the reason the login gave.- Caudra starts when a provider's credentials are gone: If the model you last used belongs to a provider whose saved login was removed, Caudra now starts and opens
/logininstead of stopping. - One token reads
~1 token: Card sizes said~1 tokens. - The selected row in
/decisionsis readable: In the dark themes, the whole selected row now takes the highlight's colours. - Background commands keep their clock: A
shellcommand sent to the background now counts its time while it runs, then shows how long it took and its final state, such assucceeded, even after you reopen the session. - A stray indent no longer stops a
coderun: A snippet with a line indented deeper than the one before it no longer fails withUnexpected indentation. Caudra lines the line up with its block, and the card lists the fix. See Results and errors.