Devin CLI · Continuity · Oct 19–25, 2026
1 release · updated
Devin CLI adds interruptible steering and session summaries
This week, Devin CLI shipped one release with several improvements. The main change is improved steering, which now immediately interrupts Devin when a new message is received without losing progress on long-running terminal commands. This helps users who need to quickly respond to new messages while running commands.
v3000.11.205 · October 22, 2026
This release adds improved steering, a new /rename command, and a /pickup command. It also introduces a new DEVIN_SESSION_STORE=files option and several new hooks. The tool approval prompt has been updated, and the /config editor has been redesigned. Several bugs have been fixed, including issues with interrupting turns, MCP server connections, and image data handling. The release also removes the /loop command and its agent.model_mixture setting, and replaces them with the /goal command.
Official release notes
Added
- Steering improved to immediately interrupt Devin when a new message is received, without losing progress on long-running terminal commands.
/renamewith no argument generates a new session title from the conversation so far./rename <new title>still sets the title you give it./pickup(and/handofffrom a cloud session) gives the new local session a summary of the cloud session as background context.- Set
read_config_from.cloud: set it tofalseto keep your Devin Cloud account’s MCP servers and personal plugins out of the CLI. Plugins your organization requires still load. - [Experimental] Introduced a new
DEVIN_SESSION_STORE=filesoption to keep session history as a folder of files instead of SQLite, sodevin -rworks when your home directory is on NFS, SMB or another shared filesystem. - Hooks:
fail_closed: trueblocks the action a hook guards when the hook itself fails to run, instead of going ahead without it.- A hook can end the agent’s turn from any event with
stop: "<reason>"in its output (Claude-format hooks:continue: false/stopReason). PostToolUsehooks can replace a tool result before the model reads it withhookSpecificOutput.updatedToolOutput, for example to redact content a tool returned.
- New
/confetticommand.
Changed
- Tool calls show as compact single-line summaries in the transcript; press
Ctrl+Ofor the full tool cards. Commands that exit with a nonzero status show a red cross.- Use the new
agent.verbositysetting to control how much detail is shown in the transcript.
- Use the new
- The tool approval prompt is now Allow, Bypass mode and Deny. Use ←/→ on Allow to pick how long the grant lasts (Once, Session, Project, Always), and press
eto edit the command. - The
/modelpicker lists models in the same order as Devin Desktop. - The
/configeditor looks like the model picker: a search box on top, the setting’s description under the list, and values in one aligned column. /contextdraws one bar per section (system prompt, tools, messages, free) in the same style as/session-stats.- The session picker (
devin --resume,/resume,devin ls) opens as soon as the session list is read instead of taking seconds on a large sessions database; type-to-search matches session titles only./resumewith no id opens the picker over the running session. devin modelson its own lists the available models.- A session you own is made private when your machine connects through a
devin --cloudsession or Devin Desktop local access, not only throughdevin connect <code>. Only the user whose machine is connected can approve a local tool call or turn on bypass permissions from the cloud session. /loopand itsagent.model_mixturesetting are removed. Use/goalto keep Devin working on an objective until it is done.
Fixed
- Interrupting a turn (Esc-Esc, Ctrl+C, or a steer message) no longer waits for an MCP server that is still connecting or waiting for OAuth sign-in.
- MCP servers that answer the
server/discoverprobe with an incomplete result (for example MCP Toolbox for Databases) now connect instead of failing with “expected initialized result”. - GPT-5.x / GPT-6 models can use
ask_user_questionandtodo_writeagain, in normal and plan mode. - A message with more image data than the server accepts no longer leaves the session failing every later turn with “Request payload is too large”; the images that do not fit are dropped and the turn continues.
- Reopening a session (
devin -r, or a click in the Devin Desktop sidebar) no longer moves it to the top of the session list; a session moves only when you send a message, rename it, or change its mode, model, or directory. - A change saved to your personal or org plugin in Customize is picked up by the next session, instead of only after the plugin cache’s refresh interval.
- A hook matcher written as a list of strings (
"tool_name": ["shell", "bash"]) matches any of them. - Local tool use: a
local_execcommand that names a credential path without~($HOME/.aws/...,%USERPROFILE%\..., an absolute home path) now always asks for approval, as the~form already did. - Windows:
devin acpno longer leaks memory while sessions sit idle. - The thinking-level shortcut (Opt+T / Alt+T) steps through the levels in order (low, medium, high, xhigh, max) and leaves Fast Mode unchanged.
- When the connection drops mid-response, text already streamed stays in the transcript while the request is retried, and repeated failures show as one notice instead of a line per retry.
- Rendering: the input box no longer jumps by a line while a reply streams in, wrapped list lines line up under the item’s text, and the session title, tip and placeholder no longer run into the model name or folder path in narrow terminals.
Fan coverage from Devin Central — not an official Cognition announcement. Devin is a trademark of Cognition.
Follow Devin CLI changes
Get Devin CLI changes in the Monday email, with the rest of the week's Devin news.
Prefer a feed reader? Devin CLI RSS feed