Skip to content

Caudra 0.2 Preview 8

View Markdown~3,535 tokens

v0.2.0-preview.8 on GitHub

  • Claude Haiku 5.5 is the Anthropic Fast model.
  • Caudra speaks the current MCP protocol, and servers can ask you questions, report progress, and share resources.
  • The experimental decision engine reranks memory notes, tool search, retained tool output, and code context, the transcript shows what each decision changed, and workflows gain rerank().
  • Favorite and 10 recent models in /model, and tasks run on the model you name.
  • Forks open in a new Herdr or Orca tab, and tabs are named after the session.
  • New sessions can open in build mode, and /sessions can delete the current session.
  • Fixes for list scrolling, thinking levels, session moves, Orca's bell, and caudra update output.

Important

Close every other Caudra session before you start Preview 8. On its first start, Preview 8 upgrades the session database in your state directory, and the upgrade needs exclusive access. This includes other terminals, editor sessions over ACP, and caudra storage commands. While any of them has the database open, Preview 8 lists the sessions to close and does not start, and nothing is changed.

The upgrade keeps your sessions and saves the old database as caudra.db.v24.bak next to caudra.db. Preview 7 and earlier cannot open the upgraded database, and caudra rollback restores only the previous binary, not the database.

Note

If you raised decisions.timeout_ms for a slow decision engine and use tool_search, set rerank_timeout_ms as well. Tool search now uses that deadline, which defaults to 2 seconds.

  • Claude Haiku 5.5 is the Anthropic Fast model. anthropic/claude-haiku-5-5 costs $0.10 per million input tokens and $0.50 per million output tokens, with cache reads at $0.01. Once a prompt passes 100k tokens, cached tokens included, the whole request costs five times those rates. It runs at a 372k working window, and claude-haiku-5-5-1m opens the full 1M window. Copilot lists it as claude-haiku-5.5. It replaces claude-haiku-4-5 as the Fast model, so on an Anthropic session, session titles, completion-goal checks, and other Fast work run on it unless you bound those jobs to a model. Bind Fast to anthropic/claude-haiku-4-5 in /model to keep the older model. See Anthropic and Model jobs.
  • Sonnet 5.5 cache reads cost less. Caudra now counts Claude Sonnet 5.5 cache reads at the new price of $0.10 per million tokens, on Anthropic and Copilot.
  • Thinking levels a model rejects are no longer sent. minimal now sends the shallowest level a model has, where it used to send none and turn reasoning off. On Copilot, thinking off on a model that always reasons, such as gpt-6.1-sol, runs at its lowest level instead of failing with Unsupported value: 'none' is not supported. A none setting now means the same as off on every model. A custom openai or openai-responses model without reasoning_options takes no thinking setting. See Model defaults.
  • Favorite models. In /model, press Ctrl+F on a model to mark or unmark it as a favorite. Favorites are listed first and kept across sessions, and /favorite-next or Ctrl+X c switches to your next favorite. Recent now keeps the last 10 models instead of 4, without your favorites. Every model also keeps its row in its provider's section, marked ★ for a favorite and ↺ for a recent model. See Model jobs.
  • Run a task on a model you name. Ask for a model, as in "ask opus and sol to review this change", and the main agent starts one task per model you name. The task tool takes a model, either an exact provider/model name or a short name such as opus that resolves to the newest match, and a thinking level from off to max. Both win over a profile's subagent_model and subagent_thinking. A name that fits several models fails and lists them. A task's result, footer, and card name the model and level it actually runs on. [agent] task_model_choice = false turns this off. See Models and reasoning.
  • Reranking features. Six new experimental decision features use the engine as a reranker, and each is off by default. memory_search lists relevant memory matches first and names the ones it left out. memory_recall reminds the agent of up to three notes that may matter for your new message. memory_duplicates points out an existing note a new one may repeat. compaction_retention keeps the tool results that matter to the current goal when a session compacts, instead of only the newest three. output_pointers points a truncated tool result at the omitted lines most likely to matter. code_context moves the most relevant rows of a code_context result to the top. When the engine fails or times out, each behaves as if it were off. They use a new rerank_timeout_ms deadline (2 seconds) and thresholds.relevance_floor (0.1). With tool_search enforced, tool search now scores and reorders its first 16 keyword matches instead of picking one, and never drops a match. See Reranking.
  • Workflows can rerank. rerank(query, candidates) returns #{ index, score } entries, most relevant first, with the options model, timeout_ms, threshold, and top_k. Each batch of 16 candidates is one journaled decision, so a resumed run asks only for the rest. In /workflow, decide() and rerank() calls are labelled with their question IDs or candidate counts. A decide() or rerank() timeout may now reach the larger of timeout_ms and rerank_timeout_ms. See Reranking candidates.
  • Decision rows in the transcript. When a decision feature changes what happens, the transcript shows a row that names the feature and what it did, such as code_context · moved 3 of 12 rows up. A row also appears when the engine fails and the default applies, and a dim shadow row shows what a feature in shadow mode would have done. Click a row to see the candidates and scores behind it. Rows are saved with the session, the model never sees them, and [ui] show_decisions = false hides them. See Decision rows.
  • Shell duration estimates read plainly. A shell card now shows an estimate only for a command expected to take more than a second, in the units of its timeout, such as ETA 10s–2m.
  • The current MCP protocol. Caudra now speaks MCP 2026-07-28 and still speaks the older protocol that starts with initialize, so servers built on either work. When a server announces that its tools, prompts, or resources changed, Caudra lists them again. Caudra never offers roots, sampling, or logging, and writes server log messages to caudra.log at debug level. See Protocol versions.
  • Servers can ask you questions. During a tool call, a server can ask for input with a form of typed fields or a link to open, such as a sign-in page. The question form checks each answer and lets you review before sending. F3 declines and Esc cancels. A link opens in your browser only after you agree. Autopilot may answer a form from the main agent's call, but never a link, a form from Executor, or a form asking for a credential. ACP clients that support MCP elicitation get the question too. See Server questions.
  • Long calls report progress. A server's progress shows beside the card's label, and each report restarts the timeout wait, up to ten times timeout. Stopping the run tells the server the call was cancelled. After a dropped connection, a call that is not safe to repeat fails with connection lost; the call's outcome is unknown. See Long calls.
  • Resources and completions. Mention a server's resource as @server:uri, and Caudra reads it into your message like a mentioned file. Typing @ lists resources beside your files, and template values and prompt arguments complete from the server. The model can list and read resources with the new mcp_resource tool, which works in plan mode and asks first. A prompt's images and embedded resources now reach the model. See Resources.
  • Server instructions reach the model. The instructions each running server gives for its tools are added to the system prompt, up to 4096 bytes per server. See Server instructions.
  • Executor and Home Assistant are recognised. Executor gets a card that shows how a run ended, its logs, and what a paused run waits for. An Executor resume call that approves a paused run always asks you, even under autopilot. Add ?elicitation_mode=native to Executor's URL so it asks you directly. Home Assistant gets cards for intents and its live context, and a call that controls a lock, alarm_control_panel, cover, or valve always asks, even with an allow rule or in Auto. See Recognised servers.
  • Read-only tools work in plan mode. Tools that Executor or Home Assistant mark read-only, such as Home Assistant's GetLiveContext, can be called in plan mode and by read-only agents. They still ask, with an option to always allow the tool. See MCP tool calls.
  • MCP cards open whole. A finished MCP call's card shows its arguments and its whole result instead of the first three rows. Compact view and ui.always_collapsed still fold it. A card that ends with a click for raw row switches to the result's raw JSON and text. A result the server marks as an error draws as a failed card. See Results in the transcript.
  • /mcp shows what a server reported. Select a running server to see its name and version, the protocol version, the features it offers, and the first lines of its instructions. See Runtime toggling.
  • A sign-in the server refuses is reported. caudra mcp auth now checks that the server accepts the stored sign-in, and signs in again when it does not. When a fresh sign-in is still refused, as on Home Assistant's /api/mcp for a user who is not an administrator, /mcp shows the server as Refused instead of asking you to sign in again. Home Assistant's /api/mcp/assist works for any user. See OAuth.
  • The wheel scrolls lists. In /decisions, /automations, /workflow, /peers and /messages, the wheel over the list now scrolls it instead of moving the selection.
  • Files picked with Ctrl+S are mentioned. The file picker now inserts the path as an @ mention, so the file is attached when you send the message. See Mention a file with @.
  • The Ctrl+X panel takes the mouse. Click a chord in the panel to run it, or click anywhere else to cancel.
  • Hidden Orca tabs ring when a turn ends. A hidden Caudra tab in an Orca remote workspace could show working long after its turn ended. Caudra now rings the bell when a turn ends or a prompt waits while its terminal does not have focus, so Orca marks the tab unread, and sends its status again when you return. Turn on Terminal Bell in Orca's notification settings for a desktop notice as well. notifications = "off" stops the bell. See Orca.
  • Prompts, popups and scrollbars follow the theme. The permission prompt, the @ and # completion lists, the sandbox manager, and scrollbars now use the theme's colours, so a dark theme on a light terminal no longer shows light patches.
  • Open new sessions in build mode. Set default_mode = "build" under [ui] to start new terminal UI sessions in build mode instead of plan mode. A resumed session still reopens in the mode it was left in. See ui.
  • Delete the current session. In /sessions, Ctrl+D on the current session now deletes it after the usual confirmation and opens a new session in its place, in the same mode and directory. A session that is still working cannot be deleted until it is idle. See Sessions.
  • Tabs are named after the session. Caudra sets the terminal window title to the focused session's title, so most terminals, Orca, and Herdr label the tab with it. A tab you named yourself keeps its name. See Tab names.
  • Forks open in a new Herdr or Orca tab. Inside a Herdr pane or an Orca terminal, Fork here resumes the new session in a new tab with a Caudra of its own, and the session you forked from stays open. Remote, sandbox, and ephemeral sessions still fork into a tab of the current Caudra. See Forks in Herdr and Orca.
  • Memory writes wait out a busy database. The memory tool failed with database is locked while another Caudra process held the database for more than 5 seconds. Memory writes now retry for up to 30 seconds, like session saves do.
  • Pending reverts no longer block moves. /move-session, /migrate-sessions, and /worktree remove were refused or failed to start when a session already in the destination had a pending revert. Sessions in the destination no longer affect a move.
  • caudra update no longer prints its installer. It names the release it moves to, installs it without asking, and ends with the caudra rollback command that returns to your current version. caudra update --show-script prints the checksum-verified installer and asks before running it, as before. -y and --no-color are still accepted, so existing scripts keep working. See caudra update.

Website privacy and analytics