Caudra 0.2 Preview 15
View Markdown~1,317 tokens
- Enable 16 of Matt Pocock's engineering skills alongside pstack.
- The main agent can send guidance to a running task or replace its current turn.
- Patch cards stream each file's diff as the model writes it, including the file in progress.
- Content-screening failures say what went wrong instead of labeling the content as flagged.
- Sessions keep subscription spend after a restart and show it beside billed API spend.
- Python snippets support exact decimal arithmetic.
- Model lists remove retired OpenAI and Copilot models and correct prices and limits for Mistral, Z.AI, and DeepSeek.
Important
Close every other Caudra session before you start Preview 15. Its first start upgrades the session database in your state directory to record content-screening outcomes. If another terminal, editor session over ACP, or
caudra storagecommand has the database open, Caudra lists those sessions and waits for them to close.Updating from Preview 14 saves a copy first at
backups/caudra.db.v39.bakin the state directory. Preview 14 and earlier cannot open the upgraded database, andcaudra rollbackrestores only the previous binary.
Changes
Section titled “Changes”- Matt Pocock's skills. Set
mattpocock = trueunder[plugins.skill]to enable 16 of Matt Pocock's skills, adapted for Caudra. Each has a/mattpocock:<name>command. They cover plan reviews, domain modeling, test-first development, debugging, code review, specs, tickets, and handoffs. The pack can run alongside pstack, and enabling it takes effect immediately in the terminal UI. See mattpocock. - Guide running tasks. The main agent can send changed requirements or reply to a task's report through
task_controlwithaction: "guide". The task reads the message at its next step. Addinginterrupt: truereplaces its current turn. Unread guidance comes back with the task's result. See Inspect and control tasks. - Patch previews stream as they are written. A
file_apply_patchcard shows each file's diff as the model writes it, including the file in progress. Line numbers appear after execution. You still approve the whole patch before any changes are applied. - Decimal arithmetic in Python. The
codetool supportsdecimal.Decimalfor exact calculations. Results keep their decimal type and digits rather than rounding to JSON floating-point numbers. - Find sessions blocking an upgrade. The list of open sessions now includes each session's project directory, so you can find the terminal to close.
Models
Section titled “Models”- OpenAI. The model list drops the retired GPT-4.1 Nano and o4-mini entries. The
gpt-5.6alias uses GPT-5.6 Sol's metadata. - Copilot. The built-in list drops GPT-5 mini, GPT-5.4 mini, GPT-5.4, GPT-5.5, and Gemini 3.7 Flash after their October 19 retirement.
- Mistral.
zai-glm-5-3-flashhas its own vision-capable entry and prices. Cached input uses the published rates for Mistral Medium, Mistral Small, and Ministral 14B. - Z.AI.
glm-4.7-flashxuses its paid rates instead of the free Flash model's pricing. - DeepSeek. Flash and V4 Pro allow 393,216 output tokens instead of 384,000.
- Screening failures no longer look like detections. When enforced screening cannot check a result, its chip says why, such as
screening timed outorengine unreachable, instead offlagged. The/decisionsoverview separates incomplete screening from flagged content. The result stays withheld and the session's restrictions still apply. Older saved rows show the neutral labelscreening result. See the decision engine. - Batched fetches share screening capacity. Caudra screens at most four chunks at once and shows
waiting to screenfor queued work. A timed-out or unreachable request can retry once within the screening time budget, using the same text without fetching the page again. - Session spend stays visible. Reopened sessions keep their subscription spend. The status bar and
/usageshow it beside billed API spend, as in$0.01 ≈$9.05, so a few cents of decision calls no longer hide it.≈replaces~for spend covered by a subscription. New sessions no longer inherit subscription spend from the previous session. - Compaction refreshes reminders. The agent gets current environment, memory, and model notices after compaction, including after a long idle, rather than waiting for your next message.
- Cancelling clears decision spinners. Decision chips no longer keep spinning after you cancel their turn.
- macOS file-watcher hangs. Watching and unwatching folders in quick succession no longer hangs the watcher.
- Go and Rust references. Initializer calls in grouped Go package declarations belong to the correct variable or constant. Calls in local Go and Rust initializers belong to the enclosing function or method.
- Expensive SVG icons. Caudra skips source icons with excessive turbulence filters before rendering them.