Caudra 0.2 Preview 11
- Sign in to one provider with several subscriptions and API keys, and let Caudra switch between them without breaking the prompt cache.
- Config edits keep your comments.
- Opt-in cache warming keeps the Anthropic prompt cache alive through long tool runs.
- Editors and desktop apps get more of Caudra over ACP: several sessions, session management, sign-in, steering, and background subagents.
- Image generation, the host description, workflows, and automations move into the
codetool as modules with on-demand help, andtool_searchis gone. - Snippets can sleep and read the clock, and
timeoutSecbounds a run.
Important
Close every other Caudra session before you start Preview 11. On its first start, Preview 11 upgrades the session database in your state directory, and the upgrade needs exclusive access. This includes other terminals, editor sessions over ACP, and
caudra storagecommands. While any of them has the database open, Preview 11 does not start. It lists the sessions to close, and nothing is changed.The upgrade keeps your sessions, saves the old database as
caudra.db.v30.baknext tocaudra.db, and moves each saved provider login toauth/accounts/<provider>/default.json. Preview 10 and earlier cannot open the upgraded database or read the moved logins, andcaudra rollbackrestores only the previous binary.
Warning
Before you start Preview 11, remove the settings that deferred tools.
image_generate,execution_environment,workflow, andautomationare now modules of thecodetool, andtool_searchis gone. A config or profile that still uses any of the following fails to load, with an error that names the fix:
- Delete
defer_builtin_toolsfrom[agent]incaudra.toml.- Delete
tool_searchfrom[decisions.features].- Replace
lazyin a system prompt profile'stoolsblock witheagerordisabled. An MCP tool that waslazystays callable aseager.- Remove
codefromagent.disabled_toolsand from profile overrides that disable it.codenow hosts the modules and cannot be turned off, so turn off a module's tool group instead:imagesforimage,executionforhost, anddelegationforworkflowandautomation.Permission rules,
agent.disabled_toolsentries, and profile groups that name the four tools keep working. In snippets,decide(),rerank(),search(),servers(),resources(), andread_resource()are nowdecisions.decide(),decisions.rerank(),mcp.search(),mcp.servers(),mcp.resources(), andmcp.read_resource(), and the old names raiseNameErrornaming the new call.
- Modules with on-demand help: A
codesnippet now reaches everything beyond the interpreter through modules:history,mcp,decisions,image,host,workflow, andautomation.help()lists them, andhelp("image")orhelp("workflow.start")shows the full reference with examples. The firsthistory.search,history.read,image.generate, orworkflow.startin a conversation returns its reference instead of running, unlesshelpalready showed it. Thecodedescription no longer changes with what the session connects, so the prompt cache holds. See Modules. - Tools as module calls:
await image.generate(prompt, out),await host.environment(),await workflow.start(...), andawait automation.list()run the tools they replace, with the same permission rules, plan mode checks, and cards, and return dicts to the snippet. Each call's row on thecodecard opens to the tool's own card. See Tools reached fromcode. - Sleep, clocks, and
timeoutSec:time.sleep()andawait asyncio.sleep()now wait for real, andtime.time(),time.monotonic(), anddatetime.now()read the host clock in UTC. Waiting does not count toward the 30 seconds of computation, andEscstops it.timeoutSecis back as a wall-clock deadline for the whole run, 120 seconds by default and up to 3600, so a snippet can poll a workflow. See The Python subset and Run time. - More Python: Monty 1.2 adds CSV and TOML parsing, hashing, bisection,
os.path, match statements, and dataclass field defaults. - JavaScript habits repaired: A snippet that starts an assignment with
const,let, orvar, or passes a host call one JavaScript object, now runs as the Python it meant. See Results and errors. - What a subagent did, on its task card: A running task card now lists every step of the subagent, newest last, instead of only the current one. A
coderow reads likecode 3 callsand lists the newest calls with their status. See In the transcript.
Accounts
Section titled “Accounts”- Several accounts per provider: Each
caudra auth loginadds an account, such as a second Claude subscription or an API key, and--namenames it.caudra auth statuslists every account with the quota it has left,caudra auth use <provider> <name|auto>chooses one, andcaudra auth renameandcaudra auth logout <provider> [name] [--all]manage them.--account <provider>:<name|auto>andCAUDRA_ACCOUNTchoose an account for one run. Several comma-separated keys in a provider's key variable become the accountsenv-1,env-2, and so on. See Accounts. - Automatic switching: With two or more accounts, Auto keeps each conversation on its account and moves it to another before the account runs out. It moves only where the prompt cache is cold anyway, unless the account is nearly spent or refused, and a rate limit shorter than 30 seconds waits on the same account.
orderedtries the accounts in priority order, andpaceduses the quota that resets soonest.[<provider>.switching]and[[<provider>.accounts]]inproviders.tomlset the thresholds, the order, and which API keys may take over from spent subscriptions. See How Auto switches. - Across providers: With
[switching] providers = ["anthropic", "openai"]inproviders.toml, Auto moves a conversation to another provider's subscription once every subscription of its own runs out, and back once the first provider has quota again. See Across providers. - Accounts in the TUI:
/accountsorCtrl+X ilists, pins, adds, renames, orders, and removes the accounts of every provider, with fresh quota readings. The status bar shows the account a conversation uses, each switch adds a transcript row with its cause and cache cost, and/usagesplits spend by account./loginalso offers the xAI subscription sign-in. - Sign-in details: A key saved by
caudra auth loginis now tried before the provider's key variable, and the prompt hides the key as you type it.caudra auth logoutdeletes only the credential and keeps the account entry inproviders.toml.
Cache warming
Section titled “Cache warming”- Opt-in cache warming:
cache_warming = "running"under[agent]in the globalcaudra.tomlsends the last request again withmax_tokens: 0shortly before its Anthropic prompt cache expires, while a run waits on a long tool, subagent, question, or permission prompt."idle"also warms for up to 30 minutes after a run ends. It is off by default, covers the Anthropic API only, and warms only when that costs less than the cache write it avoids./usageshows what warming cost and saved, or what it would have saved while it is off. See Cache warming.
- Several sessions: One
caudra acpprocess now keeps every session an editor opens, each with its own tools, so sessions in different projects run side by side. Editors can list, resume, close, delete, and fork Caudra sessions, and the list shows each session's title and its first and last prompt. See ACP and Sessions. - Sign-in: Editors get a
caudra-loginauth method that runscaudra auth login, andcaudra acpstarts with no signed-in provider so the editor can offer it. See Authentication. - Pickers and model catalog: Editors get pickers for reasoning effort, mode, and permission mode beside the model, and a fast mode toggle where supported. The choices are saved with the session, and editors can list models before opening one. See Config options.
- Slash commands and compaction: Editors list
/compact, your skills, and MCP prompts as slash commands, and show compaction progress and retried provider requests. See Slash commands. - Cancellation and steering: Cancelling a prompt withdraws open permission prompts and questions and answers
cancelledexactly once. A prompt sent right after a cancel runs once the cancelled turn ends, so editors such as Orca that steer by cancelling and prompting again work, and a prompt sent with_meta.caudra.delivery: "steer"joins the running turn. See Prompts, cancellation, and steering. - Turn details: Answers carry the turn's duration and usage, tool calls name their tool, shell calls report stdout, stderr, and exit code, and failures carry their text instead of a bare "Internal error". See Turn updates.
- Subagents and background tasks: Subagents and shell commands can run in the background over ACP, as in the TUI. Editors that support ACP subagent sessions get one per subagent, editors that opt in with
_meta["caudra/childWork"]can follow and stop this work, and other editors see each subagent's tool calls in the parent session. See Subagents and background tasks. - Questions as permission requests: An editor without form elicitation that sets
_meta["caudra/questionsViaPermission"]gets thequestiontool's questions as permission requests. See Permissions and questions. - Answers every request: A second prompt to a busy session is refused instead of left unanswered, and a stdin line that is not JSON or is over 64 MiB is skipped instead of breaking the connection.
Changes
Section titled “Changes”- Image viewing: The image viewer and workbench image tabs now zoom and pan. Pictures are sized to the terminal's cells and redrawn when the window or font changes, and a click outside the picture closes the viewer.
- Orca resumes Caudra: In an Orca terminal, Caudra now reports its session id, so Orca can reopen it with
caudra --resume <id>, and lists running subagents as child rows. Off Windows, clicking a row switches to that subagent's chat when Orca supports it. See Orca. - Docs from any modal: A
?on the top border of a modal such as/accountsor/usage, orF1, opens the manual section about it above the modal, soEscreturns to it. The key hints on the top border of/usageand/toolsalso answer a click. See Docs. - Tool cards fold like a file read:
source_code_context,source_code_refs, andmemoryreads now rest as one row that a click opens whole. A Home Assistant or Executor call inside acodeorbatchcard shows its whole card, amemorywrite shows the whole note, and a click on a write's title folds it in the expanded/viewtoo. If you set your ownui.always_collapsedlist, add"source_code_context","source_code_refs", and"memory_read"to it. - One proxy for every request: All of Caudra's own requests, decision endpoints and Lua's
caudra.netincluded, now useHTTPS_PROXY,HTTP_PROXY,ALL_PROXY, andNO_PROXY. A proxy set only in a project.caudra/.envno longer applies, so set it in your shell or the global.env. See Network proxies.
- Cancelling stops the running command: Cancelling a turn while a
shellcommand runs, withEscor from an ACP editor, now stops the command and every process it started. - Config edits keep your comments: Saving a provider or decision endpoint, toggling an MCP server, or saving sandbox records now changes only the keys that changed, and keeps your comments, quoting, and table order.
- Refusals end the turn: A provider refusal or content filter now ends the turn as refused instead of prompting a retry.
- Resources read as their type: An MCP resource that is not markdown now shows its text as the server wrote it.
- Purposes sorted by spend: The session's Per purpose table in
/usagenow lists the purpose that cost the most first.