Caudra 0.2 Preview 9
- Code mode: the model calls your MCP servers from Python, so a large server costs a short list, results are filtered before the model reads them, and the prompt cache survives.
- Gemini 4 Argon is the Google default, with new Gemini Flash models, Grok 4.7, GLM-5.3, DeepSeek V4.1 Flash, and Mistral Large 4. Retired models are gone, and Google requests work again.
- The agent can ask questions without waiting, and you can answer later.
- Autopilot runs without a continuation limit, answers the agent's questions, and logs what it did.
- A log of every goal check, and goal checks survive an interrupted reply.
/worktree closebrings a worktree's work back to the main checkout./modelopens on your favorites, andHomeandEndreach the ends of a list.- macOS updates no longer stall, and updates use your GitHub token and show download progress.
Important
Close every other Caudra session before you start Preview 9. On its first start, Preview 9 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 still has the database open, Preview 9 does not start. It lists the open sessions to close instead, and nothing is changed.The upgrade keeps your sessions and saves a copy of the database from before it as
caudra.db.v25.baknext tocaudra.db. Preview 8 and earlier cannot open the upgraded database, andcaudra rollbackrestores only the previous binary, not the database.
Warning
Before you start Preview 9, delete
defer_toolsandalways_loadfrommcp.toml. Code mode replaces the tool search they tuned. A global or projectmcp.tomlthat still sets either fails to load: Caudra starts without that file's servers, and/mcpshows an error that names the key.
- Code mode. The model now reaches your MCP servers by writing Python in one
mcptool, instead of receiving a definition for every tool. A snippet can call several tools at once and filter or join their results before the model reads anything, so a list of 500 issues can come back as three lines. The model learns your servers from a short list in the conversation, andsearch()finds the tools the list leaves out. The tool definitions stay the same while a server connects, fails, or changes its tools, and when you switch to plan mode, so none of these makes the provider read the whole conversation again at full price. See Code mode and The prompt cache. - Approvals work as before. Each tool call in a snippet asks on its own, and permission rules,
agent.disabled_tools, and profiles still name a toolserver.tool. Plan mode and read-only agents still allow only the read-only tools of recognised servers. Snippets run in a separate worker process with no filesystem, network, subprocess, or environment access. See MCP tool calls and Where snippets run. - Snippets are checked before they run. A misspelt tool, an argument of the wrong type, or code that is not valid Python stops a snippet before any server call, naming the closest tool and the line. Slips with one meaning, such as a missing
import asyncioor a Markdown fence around the snippet, are fixed and listed on the card. See Results and errors. mcp_resourceis gone. The model reads a server's resources withresources()andread_resource()inmcp, and reads still ask first. See Resources.- Servers with a code mode of their own. Add
?codemode=falseto the URL of Cloudflare's MCP server, or?mode=passthroughto Executor's, so the model calls their plain tools. See Servers with a code mode of their own. /mcpand/tools./mcpshows the name the model calls each server by, and warns at startup when no code worker can run snippets./toolscounts each server call a snippet makes as a row namedserver.tool. See What the tools cost.- Smaller changes. Sessions from Preview 8 and earlier resume as before. File revert records an
mcprun only once it calls a tool that may change files. Python now runs on Monty 1.1.0, which supports complex numbers. See Recording.
Models
Section titled “Models”- Gemini 4 Argon.
google/gemini-4-argonis the Google Best model and the model Caudra picks when you use Google without choosing one, in place ofgemini-2.5-pro. It costs $2 per million input tokens and $10 per million output tokens, with cache reads at $0.10, until November 30, 2026, then $4 and $20 from December 1, 2026. Caudra switches to the new price on that date without an update. Argon always thinks, so thinkingoffruns atlow. It is not on Google's free tier, so it needs a key on a project with billing enabled. On the free tier, pickgoogle/gemini-3.8-flash. When a key has no quota for a model, Caudra now says so at once instead of retrying. Caudra uses an API key for Google, since Google's terms forbid third-party tools from using the sign-in of Gemini CLI or Antigravity. See Google. - Gemini 3.8 Flash, 3.6 Flash, and 3.5 Flash-Lite. 3.8 Flash and 3.6 Flash cost $0.75 per million input tokens and $3.75 per million output tokens until December 31, 2026, and $1.50 and $7.50 from January 1, 2027. 3.5 Flash-Lite is the Google Fast model, in place of
gemini-2.0-flash-lite, which Google shut down. Gemini 2.5 Pro and 2.5 Flash leave the model table, since Google no longer serves them to new keys. - Google models answer again. Every request to a Google model failed with
Invalid JSON payload received. Unknown name "$schema". Caudra now sends tool definitions in a form Google accepts. Switching to Gemini 3 or 4 in the middle of a turn no longer makes Google refuse the request. - Grok 4.7.
xai/grok-4.7is the xAI default, in place ofgrok-4.6. It costs $2 per million input tokens and $6 per million output tokens, with cache reads at $0.50. Once a prompt reaches 200,000 tokens, every rate doubles for the whole request, as xAI bills it. - GLM-5.3 on Z.AI.
zai/glm-5.3is the Z.AI Best model and the default on the pay-as-you-go plan, in place ofglm-5.1. It costs $1.40 per million input tokens and $4.40 per million output tokens, takes a 1M-token context, and always thinks. The Coding plan keepsglm-5-code.glm-5.3-flashandglm-5.3-flashxjoin as small models that read images. Caudra now billsglm-5.2andglm-5.1at Z.AI's current $1.40 and $4.40, up from $1 and $3.20. - DeepSeek V4.1 Flash.
deepseek/deepseek-flashis the DeepSeek default, in place of the retireddeepseek-v4-flash. It reads images and costs $0.15 per million input tokens and $0.60 per million output tokens off-peak, twice that at peak hours. - Mistral Large 4 and GLM 5.3 on Mistral.
mistral/mistral-large-4, in public preview, costs $0.68 per million input tokens and $2.09 per million output tokens at its launch price, and $1.36 and $4.18 from October 20, 2026.mistral/glm-5-3replacesglm-5-2at the same price. Mistral retires GLM 5.2 on October 31, 2026. - Claude Sonnet 4.6 answers at full length.
anthropic/claude-sonnet-4-6can now write up to 128,000 output tokens, up from 64,000. - Retired models leave the model table. Caudra drops
claude-sonnet-4,claude-opus-4-0, andclaude-opus-4-1, which Anthropic retired. Useclaude-sonnet-5-5orclaude-opus-4-8instead. It also dropsgpt-5.1-codex,gpt-5.1-codex-max,gpt-5.1-codex-mini, andgpt-5.2-codex, which OpenAI shut down. Useopenai/gpt-5.6-sol, orgpt-5.6-terrain place ofgpt-5.1-codex-mini. - Copilot follows GitHub's model list. Copilot and Copilot Enterprise add
gemini-3.8-flash,gemini-4-argon, andgrok-4.7, and drop the models GitHub no longer serves, such asgemini-3.5-flash,kimi-k2.7-code, andclaude-opus-4.7.claude-sonnet-4.6stays, because GitHub still serves it on annual Copilot Pro and Pro+ plans.
Questions
Section titled “Questions”- Questions that do not stop the agent. The agent can ask a question it does not need answered for its next steps, such as a preference or a name, and keep working. Press
Escto put the form aside in a[? N]chip in the footer, andCtrl+X uor a click on the chip brings it back. Your answer reaches the agent when you give it, even after its turn has ended. On a question the agent waits on,F4puts it off and lets the agent carry on. Turn this off withdeferred_questions = falseunder[agent]. See Deferred questions.
Autopilot
Section titled “Autopilot”- No continuation limit.
max_continuations = 0under[autopilot]lets an engagement run until its goal ends, its cost or time limit runs out, or you stop it. See Budgets. - An audit log of what autopilot did. The status panel gains a Log tab, which
/autopilot logopens at any time. It lists every decision of the session by engagement, with what the delegate was asked, what autopilot said and why, the model and cost, and what happened next. Pressfto filter,yto copy an entry, and Enter to jump to its row in the transcript. When you answer a decision differently, the delegate sees your answer in its next decisions. SDK clients read the log with the newautopilot_logcontrol. See The log. - Autopilot answers the agent's questions. While engaged, autopilot now answers the main agent's
questionforms and a server's forms during its tool calls, as Preview 8 described.
- A log of every goal check. The
/goalpanel gains a Log tab, which/goal logopens. It lists every goal of the session and every check of it, with its status, time, evaluator model, cost, and reason. The same keys filter, copy, and jump to the transcript, and a goal's check rows now come back after a resume. SDK clients read the log with the newgoal_logcontrol. See The goal log. - Goal checks survive an interrupted reply. After you interrupted the agent while it was thinking, every later goal check on a Claude model failed with
thinking.signature: Field required. Checks now leave out the agent's reasoning, and a goal whose checks were failing is checked normally again.
Worktrees
Section titled “Worktrees”/worktree closebrings the work back. Closing a worktree is the reverse of creating one. It fast-forwards the main checkout's branch to the worktree's commit, moves the worktree's sessions back, carries its uncommitted changes into the main checkout, and removes it, keeping the branch. The removal confirmation offers it first, or says why it cannot, such as when the main checkout has commits of its own. Removing and closing now check for a pending revert or an active workflow before they change anything. Leaving a worktree's own Herdr workspace or Orca terminal brings every open session back, each in its own tab. See Closing a worktree and Removing a worktree.
Pickers
Section titled “Pickers”/modelopens on your favorites. With favorite models,/modelandCtrl+X mput the cursor on the current model's favorite row, or on the first favorite, instead of on the Chat job. See Model jobs.HomeandEndreach the ends of a list. In every picker,HomeandEndjump to the first and last row while the filter is empty. In the workflow and automation inspectors, they andPageUpandPageDownmove the list that has the focus. See Pickers.
- Spend by purpose for the session.
/usagenow shows a Per purpose table for the current session, counted for turns from Preview 9 on. In the lifetime view, click... and N moreto show the rest of a list, or presseto open every list at once. See Commands and Lifetime spend.
Sessions
Section titled “Sessions”- Tasks run on the favorite you name. When you ask for a model by a short name, as in "ask sonnet to review this", the task runs on your favorite model that matches it. The main agent is also told your favorites when it can choose a task's model. See Models and reasoning.
- Forks keep their permission mode. A fork now inherits the parent's Ask, Auto, or YOLO mode, also in another tab or a new worktree.
- Mode and model switches apply once the session settles. A
Tabtoggle now takes effect as soon as nothing runs, instead of waiting for your next message. See Selecting Plan while work runs. /btwanswers when the model reaches for a tool. Instead of stopping, Caudra tells the model that tools are unavailable, so it answers from the conversation. See Commands.
Updates
Section titled “Updates”- Updates on macOS no longer stall. Updating a Mac that still had the large license bundle from Preview 1 stayed on
downloading caudra …for about four minutes. It now takes under a second. - Updates use your GitHub token.
caudra updateand the update notice now sendGITHUB_TOKENorGH_TOKENto GitHub, as the installer does, so a shared IP no longer fails withserver returned HTTP 403. - Download progress. The shell installer shows a progress bar for the release archive. From Preview 9 on,
caudra updatedownloads the archive itself, shows its size, speed, and time left, and checks it against the release checksums.