Caudra 0.2 Preview 10
- One
codetool replacesmcpandpython_execution, and can ask the decision engine to judge and rerank results. - Search past conversations, from the
codetool and from the TUI with/search-history. - Two source code tools instead of five, renamed
source_code_contextandsource_code_refs, that every model sees and that list call sites. - Decision models in
/loginand/model, Auto mode out of experimental, and workflowdecide()in the OpenAI Decisions question shape. - Pictures in the transcript.
- Notes on your answers to the agent's questions, and drag to copy in
/btwand the question form. - Goals stop when they make no progress, wait for CI or a deploy, and tell automations when they are stuck.
- Subscription image generation works again after GPT-5.5 left Codex.
Important
Close every other Caudra session before you start Preview 10. On its first start, Preview 10 upgrades the session database in your state directory and needs exclusive access, including from editor sessions over ACP and
caudra storagecommands. While any of them has the database open, Preview 10 lists them and does not start, and nothing is changed.The upgrade keeps your sessions and saves the old database as
caudra.db.v28.baknext tocaudra.db. Preview 9 and earlier cannot open the upgraded database, andcaudra rollbackrestores only the previous binary, not the database.
Warning
Before you start Preview 10, rename
ui.tool_output_lines.python_executiontoui.tool_output_lines.code. Acaudra.tomlthat still sets the old key fails to load, with an error that names the new one. Permission rules,agent.disabled_toolsentries, and system prompt profiles that name themcporpython_executiontool no longer match anything, so namecodein them instead. See Code.
Warning
If you run a remote Workcell server, update it with Caudra and rename its
python_executiontool group tocode.--tool-group python_executionandpython_executioninWORKCELL_MCP_TOOL_GROUPSfail at startup, and Caudra refuses a server that does not servecode. See Remote workspaces.
Warning
If your workflows call
decide(), rewrite their questions when you update. Questions are now a list, and each names itself withname.noulquestions becomepredicatequestions, andcriteriabecomeschoicesfor a choice andlevelsfor a score. Answers come back asanswers, a list in question order, andby_name, keyed by name, and a predicate's probability is inprobability. A workflow that passes the old shape fails at its firstdecide()call with a catchable error, and so does a run started before the update when you resume it, even at a call it had already answered. See Typed decisions.
Warning
If your config, profiles, or remote Workcell use the code-graph tools, update them together with Caudra.
code_contextandcode_refsare nowsource_code_contextandsource_code_refs.code_map,code_impact, andcode_expandare gone: usesource_code_contextforcode_maporcode_expand, andsource_code_refswithdepthforcode_impact. The profile tool groupcode_graphis nowsource_code,[decisions.features] code_contextis nowsource_code_context, and an old name in a config or profile is rejected as unknown. A remote Workcell must run the same release as Caudra, with--tool-group source_codeinstead of--tool-group code_graph. The SDK's tool kinds are nowSourceCodeContextandSourceCodeRefs. Your sessions,/toolsusage, decision log, and saved permission rules move to the new names on first start. See Code Intelligence.
Warning
Before you start Preview 10, move your decision endpoint out of
caudra.toml. The[decisions]table now holds only the settings every endpoint shares.base_url,protocol,model,api_key_env,allow_remote, andallow_httpin it, or incaudra.setup({ decisions = ... }), stop Caudra with an error that shows the table to write instead. To migrate:
- Run
caudra auth login --decision, or move those keys into a[decisions.<name>]table in the globalproviders.toml, such as[decisions.typesafe]. Then delete them from[decisions].- Delete
allow_remote. Configuring an endpoint is now the consent to send decision context to it. Keepallow_http = trueonly on an endpoint that uses plain HTTP to a host other than numeric loopback.- Delete
decision_enginefrom[experimental]. It is now an unknown key.- If you set
TYPESAFE_BASE_URLorOPENAI_BASE_URLto move the decision endpoint, writebase_url = "{env:TYPESAFE_BASE_URL}"in its table instead. Neither variable changes a decision endpoint any more.- Run
caudra decisions use <name>for a table you wrote by hand. Login makes a new endpoint active, but a hand-written one stays inactive until you select it.See Decision endpoints.
Important
Intel macOS users building with Nix must switch to the native build. Current nixpkgs no longer supports Intel macOS, so the flake no longer provides
x86_64-darwinoutputs. Native Intel macOS builds and release binaries remain supported.
- One
codetool replacesmcpandpython_execution. The model runs all of its Python incode: computation, MCP server calls, history search, and the decision helpers. It is offered in every session, with or without MCP servers, and its definition never changes, so the prompt cache holds. Python always runs on your machine, also in remote Workcell sessions, with 30 seconds of computation per run wherepython_executionstopped at 5 seconds, and the clock stops while a snippet waits for a call.timeoutSecis gone. Variables persist between runs, andreset: truestarts a run empty./tools,caudra tools, and/mcpmark MCP toolsvia code, and a session with anmcpcard from Preview 9 still opens. See Code and Where snippets run. - Search your past conversations.
history.search()finds what you wrote, what the model answered and reasoned, and what its tools did, in this conversation, this project, another project, or every stored session, compacted turns and subagents included.history.read()opens a hit in full, andhistory.sessions()lists sessions. Searching this project never asks, and another project or all projects ask first. With[decisions.features] history_search = "enforce", the decision engine reranks the matches and returns the 10 most relevant. See Search past conversations and Ranking. - Decisions and reranking in snippets. With a decision endpoint active, a snippet can call
decide()for typed answers about its data, andrerank()to keep only the candidates relevant to a query, for up to 256 candidates, so the rest never reach the model. Calls need no approval and work in plan mode. Each shows as a row on thecodecard with a bar per option and a warning when the winner is uncertain, and/decisionsdraws logged answers the same way. See Decisions and reranking.
Decisions
Section titled “Decisions”- Decision models in
/loginand/model./loginhas a Decision models section with TypeSafe Jev, OpenAI Luna, and a custom endpoint, and a new endpoint becomes the active one. In/model, the Decisions row of the Jobs section switches endpoints, turns decisions off, or adds one, and/decision-modelopens it directly./decision-nextorCtrl+X dmoves to the next endpoint. While an endpoint is active, the status bar shows[decisions: <name>]. See Decision endpoints. - Decision endpoints from the command line.
caudra auth login --decisionandcaudra auth logout --decision <name>add and remove an endpoint and its key, andcaudra decisions list,use <name|off>, andstatusshow and switch them. Without a saved key, Caudra reads the variable that the endpoint'sapi_key_envnames, such asTYPESAFE_API_KEYorOPENAI_API_KEY. See CLI. - Auto mode and the decision engine are no longer experimental.
/auto,--auto, SDK--permission-mode auto,/decisions, andcaudra decisionswork without a switch, andalways_auto = trueand sessions saved in Auto start in Auto again. With no active endpoint, Auto runs the calls that deterministic checks clear and prompts for the rest. Every decision feature stays off until you turn it on. See Auto mode. - Workflow
decide()takes OpenAI Decisions questions. Choice options and score levels can carry a description, and each answer lists the probability of every option or level. caudra auth loginkeeps your comments inproviders.toml. Login and logout replace only the table they change.
Transcript
Section titled “Transcript”- Pictures in the transcript.
view_imageandimage_generatecards, Markdown images that point at local files, and prompts with attached images show the picture as a thumbnail, at full resolution in terminals with Kitty graphics and as coloured blocks elsewhere. Click a picture to open it at full size, then pressoto open it in your system viewer,wto open it in the workbench, oryto copy its path. Markdown images are never fetched from the network. Set[ui] inline_imagesto"blocks","sextants","kitty", or"off"to choose. See Images. - Search past sessions.
/search-historysearches this project's saved sessions, andCtrl+Fstill searches the current chat.Tabswitches between This chat, Project, and All projects. Results show each match with its session and age and a preview, and words can carrykind:,tool:,since:, anduntil:filters.Enteropens the session and scrolls to the match. See Search past sessions.
Questions and /btw
Section titled “Questions and /btw”- Add a note to your answers. On a question the agent asks,
F5opens a note for anything the options leave out. It reaches the agent with your answers, or with the dismissal when you pressEsc. The form's hints now start withEnter submit. See Keybindings. - Drag to copy in
/btwand the question form. Dragging over a/btwthread copies the passage when you let go, the way the transcript copies. In the question form, a drag copies instead of picking the option it ends on, andCtrl+Ccopies a selection before it stops the agent. See Commands.
- A goal that stops getting anywhere stops its run. When three turns in a row make no progress, the run ends with the goal still active instead of spending continuations until the limit. Autopilot reviews the stall like a loop cap. Headless clients get a
stalledgoal event, and agoal_finishedautomation trigger that listsstalledinverdictsfires too. See Commands and Automations. - A goal waiting on something outside the session waits instead of stalling. When the goal waits on a change such as a CI run or a deploy, the run ends and the goal is checked again after a minute, then twice as long each time up to 30 minutes, for about two hours before it stalls. A message from you ends the wait, and a resumed session keeps it. Headless clients get a
waitinggoal event.
- Two source code tools instead of five, under new names.
source_code_context(formerlycode_context) finds the code a task touches, andsource_code_refslists who calls a symbol and from which lines. Withdepth,source_code_refsshows what a change could reach and how many of those symbols are tests, which is whatcode_impactdid. Both tools are always in the request for every model. - More precise source code results. Standard library calls such as
.clone()orVec::new()no longer link to a symbol of the same name in your tree, andsource_code_contextlifts only names written as code, so an ordinary word such as "renew" no longer pulls in everynew. /toolscounts calls made insidebatch. Each call a batch makes has its own row.
- ChatGPT subscription image generation uses a supported model. Image requests now run through GPT-6.1 Sol instead of GPT-5.5, which OpenAI retired from Codex with ChatGPT sign-in on October 14, without changing your configured image model or switching to API billing.
- Lua job cleanup. Releasing an abandoned
jobwaitno longer aborts the process. - Shell output filters. Bundled filters no longer match unrelated subcommands that start with a filtered name.
- File indexes and code graphs. PHP clone arguments and partial application no longer report parse errors, and Go type aliases are included.
- SVG source icons. An icon over the pixel budget is skipped, including through nested masks, clips, and filters.
- HTML conversion. Code blocks keep blank lines and lazy-loaded images resolve.
- Nix builds. Evaluation works without building the worker source first.