agent-shell 0.47 updates

12 min read Original article ↗

We got quite a few agent-shell additions since the last post, so let's go through the highlights as of v0.47.1.

What's agent-shell?

Agent shell is a native Emacs mode to interact with LLM agents powered by ACP (Agent Client Protocol).

Your employer can make a difference

agent-shell has been attracting quite a few users. Many of you are working in tech where employers are happily paying for IDE subscriptions and LLM tokens to improve productivity. If you are using agent-shell for work, consider getting your employer to give back by sponsoring the project.

I also know many of you work at AI companies offering paid agents like Claude Code, Copilot, Gemini, Codex, etc. all supported by agent-shell. Please nudge your employers to help fund projects like agent-shell, which are making their services available to more users.

✨ Sponsor agent-shell ✨

So what's new?

claude-code-acp renamed to claude-agent-acp [Action Required]

Let's get this one out of the way, as it needs actioning. Both the npm package and the CLI agent have been renamed from claude-code-acp to claude-agent-acp (to align with Anthropic's branding guidelines). If you're using Claude Code, you'll need to update:

npm remove -g @zed-industries/claude-code-acp
npm install -g @zed-industries/claude-agent-acp

If you had customized agent-shell-anthropic-claude-acp-command, update it to point to claude-agent-acp.

New agents supported

Bootstrapped sessions (experimental)

This was a biggie. How sessions are loaded is now configurable via agent-shell-session-strategy. When set to 'new, starting a new shell delivers a fully bootstrapped session before presenting you with the shell prompt. This means the ACP handshake, authentication, and session creation all happen upfront.

You can enable this flow with:

(setq agent-shell-session-strategy 'new)

What's the benefit? Bootstrapped sessions enable changing models and session modes (Planning, Don't ask, Skip permissions, etc…) before submitting your first prompt.

For the time being, the existing (deferred) behaviour is still offered via 'new-deferred. Just set as follows:

(setq agent-shell-session-strategy 'new-deferred)

Session resume (experimental)

Probably the most requested feature and also facilitated by the bootstrapping changes. agent-shell-session-strategy also unlocks session resume. Set it to 'prompt and every time either M-x agent-shell-new-shell or C-u M-x agent-shell are invoked, you'll be offered to resume previous sessions or start a new one.

(setq agent-shell-session-strategy 'prompt)

Alternatively, you can set to 'latest to always resume the most recent session in current project.

Under the hood, there are two ways to pick up from previous session: session/resume (lightweight, no message replay) and session/load (full history replay). By default, agent-shell prefers resuming (controlled by agent-shell-prefer-session-resume). Please favor resuming for the time being as loading has more edge cases to sort out still.

Note: Both resuming and loading sessions are agent-dependent. Some agents may not yet support either, especially as the features aren't yet considered stable in Agent Client Protocol (see session/list spec).

This feature was a collaboration between @farra, @travisjeffery, and myself.

Clipboard images

You can now use agent-shell-send-clipboard-image (#285 by @dangom) to send images straight from your clipboard into agent-shell. Clipboard images are saved to .agent-shell/screenshots in your project root and inserted into the shell buffer as context.

Note: You'll need either pngpaste or xclip installed on your system for the feature to automatically kick in.

In addition, we now have agent-shell-yank-dwim: if the clipboard has an image, it pastes it as context. Otherwise, it yanks text as usual. In other words, copy an image anywhere to your system's clipboard and paste/yank into the buffer as usual (typically via C-y).

Status display + tool calls

Status labels and tool call titles rendering got some improvements. Status reporting is generally more compact, redundant text is dropped from tool call titles, and tool status/kind shortening has been consolidated.

Image rendering

agent-shell now renders images inline. When agents output images (charts, diagrams, screenshots, etc.), they display directly in the shell buffer. You may need to nudge the agent to output image paths in the expected format so agent-shell can pick up.

Markdown images:

![alt text](/path/to/image.png)

Any of the following in a line of their own are supported also:

/path/to/image.png

file:///path/to/image.png

./output/chart.png
~/screenshots/demo.png

Recognized image formats depend on what your Emacs was built with (typically png, jpeg, gif, svg, webp, tiff, etc. via image-file-name-extensions).

Emacs skills

While on the topic of image rendering, this works particularly well when coupled with charting agent skills. I shared some of these over at emacs-skills, demoed in episode 13 of the Bending Emacs series.

Table rendering

Tables are now rendered using overlays (#17 by @ewilderj).

Usage tracking

Tracking usage now possible (#270 by @Lenbok):

Git worktree shell

If keen to run multiple agents on the same repo without stepping on each other's work, M-x agent-shell-new-worktree-shell facilitates this via git worktrees (#255 by @nhojb).

Send context to…

You can now send context to a specific shell using agent-shell-send-file-to, agent-shell-send-region-to, agent-shell-send-clipboard-image-to. and agent-shell-send-screenshot-to. These prompt you to pick a target shell.

Both M-x agent-shell and agent-shell-send-dwim are now prefix-aware. C-u forces a new shell, while C-u C-u prompts you to pick a target shell.

Compose buffer improvements

Compose buffers now support file (via @) and command (via /) completions. It is now also possible to browse previous pages via C-c C-p and come back to your prompt draft.

There's also prompt history navigation/insertion when composing prompts via M-p (previous), M-n (next), and M-r (search).

Bringing context into viewport compose buffers is now more robust. For example, carrying context into a new viewport compose buffer is now supported (#383 by @liaowang11).

Viewport improvements

While viewport interaction was introduced in the previous post, it is now my preferred way of interacting with agent-shell. You can enable via (setq agent-shell-prefer-viewport-interaction t). In any case, viewport buffers got a handful of quality-of-life improvements.

Single key replies without needing to open a compose/reply buffer:

Customizable context sources

agent-shell-context-sources lets you configure which DWIM context sources are considered by M-x agent-shell, agent-shell-send-dwim, and compose buffers. You can control the order sources are checked and add custom functions. Defaults to files, region, error, and line.

I'm always on the lookout for some DWIM goodness. If you add your own context function, I'd love to hear about it.

Flycheck support

While on the topic of context sources, in addition to picking up flymake errors at point, flycheck errors are now automatically recognized (#219 by @Lenbok).

Diff improvements

Diff buffers got some love too, now with syntax highlighting (#198 by @Azkae). There's also a new agent-shell-diff-mode-map for customizing diff keybindings, which avoid inheriting unsupported features from the parent mode. You can also press f to open the modified file from a diff buffer.

Additionally, you can now press C-c C-c from an agent-shell-diff buffer to reject all changes (same binding as the shell itself).

Event subscriptions

You can now programmatically subscribe to agent-shell events like initialization steps, tool call updates, file writes, permission responses, permission requests, and turn completions. This opens the door for building integrations on top of agent-shell.

(agent-shell-subscribe-to
 :shell-buffer (current-buffer)
 :event 'file-write
 :on-event (lambda (event)
             (message "File written: %s"
                      (alist-get :path (alist-get :data event)))))

Permission UX improvements

Permission dialogs got a few improvements. Amongst them, automatic navigation to the next pending dialog and also making executed commands more prominent for some agents.

Permission responder function

You can now programmatically respond to permission requests via agent-shell-permission-responder-function. A built-in agent-shell-permission-allow-always handler is provided to auto-approve everything (use with caution):

(setq agent-shell-permission-responder-function #'agent-shell-permission-allow-always)

Claude Code OAuth token support

OAuth token now supported for Claude Code (#339 by @chemtov).

Transcript improvements

Markdown transcripts generation needed love. Thank you @Idorobots and @systemfreund for the improvements (#374, #325, and #326).

Session ID display

With (setq agent-shell-show-session-id t), session IDs are now displayed in header as well as session lists (#363 by @Cy6erBr4in).

Custom CWD function

agent-shell-cwd-function now enables customizing how agent-shell determines the working directory sent to agents.

Configurable dot-subdir location

agent-shell-dot-subdir-function now lets you customize where agent-shell keeps per-project files (#378 by @zackattackz).

Lambda MCP servers

agent-shell-mcp-servers now accept lambda functions too (#237 by @matthewbauer). Useful for setups like claude-code-ide using dynamic MCP details.

Buffer name format

agent-shell-buffer-name-format now makes buffer naming configurable. Choose between the default title case ("Claude Code Agent @ My Project"), kebab-case ("claude-code-agent @ my-project"), or provide your own function (#256 by @nhojb).

Inhibit minor modes during writes

agent-shell-write-inhibit-minor-modes lets you temporarily disable modes (like auto-formatting entire file) when agents write files (#224 by @ultronozm).

File autocompletion performance

File autocompletion is now more performant for larger repositories (#262 by @perfectayush).

Container modeline indicator

When running shells inside a container, a [C] indicator now shows up in the modeline (#250 by @ElleNajt).

Graphical header rendering is more robust now (#275 by @nhojb).

Busy throbber display options

The busy/working indicator now offers multiple visual styles, customizable via agent-shell-busy-indicator-frames (#280 by @Lenbok).

Pull requests

Thank you to all contributors for these improvements!

Bug fixes

Lots of polish

Beyond what's showcased, I've poured much love and effort into polishing the agent-shell experience. Interested in the nitty-gritty? Have a look through my regular commits.

Make the work sustainable

If agent-shell is useful to you, please consider sponsoring the project. These days, I've been working on agent-shell daily.

LLM tokens aren't free, and neither is the time dedicated to building this stuff. While I now have more time to work on agent-shell as an indie dev, I also have bills to pay ;)

Unless I can make this work sustainable, I will have to shift my focus to work on something else that is.

✨ Sponsor agent-shell ✨

powered by LMNO.lol

privacy policy · terms of service