Prompt context

Every prompt that launches an agent accepts @ mentions: New agent session, New worktree (single and Fan out), and Agent board drafts. Type @ at the start of a line or after a space, and a picker opens under the cursor, grouped by source. Keep typing to narrow it; ↑/↓ move, Enter or Tab insert the highlighted item, Esc closes the picker.

The picker lists issues and pull requests first, then files, even for a bare @: the most recently updated open issues and pull requests, then the worktree's top-level folders and files (dotfiles hidden). Keep typing to search. Folders can be mentioned too (@src/). When nothing matches, the picker says so instead of closing.

Files are searched in the worktree the agent will work from: the one picked in New agent session, the parent a New worktree branches from, or, for a board draft, the worktree already on the draft's branch (the selected worktree when the branch is new). Issues and pull requests list the 100 most recently updated open ones in the worktree's origin repository; only open ones are ever offered. Type an exact number (@#12) to reach an open one older than that. GitHub sources need you to be signed in to GitHub; until then the picker shows a sign-in note where issues and pull requests would be.

What gets sent

Your prompt is sent as written. When you launch, every mention still in the prompt is resolved and appended as a <context> block the agent can tell apart from your request:

Fix the crash described in @#123

<context mention="@#123" source="GitHub issues">
GitHub issue #123: Crash on launch (open)
Author: @octo
Link: https://github.com/acme/app/issues/123
Open the link (or run `gh issue view 123 --comments`) for the discussion and further details.

Steps to reproduce…
</context>

Delete a mention from the prompt and its context is not sent.

Where the prompt is saved for later, the context is resolved when you save: a board draft stores it with the card, and a fanout sends it to every attempt. Reopening a draft (or retrying a failed New worktree) shows only your text; the mentions still in it keep their saved context, so it is not fetched again.

Run commands first

An agent prompt can also hand the agent a command's output. In New agent session or New worktree (single or Fan out), type !! anywhere — even mid-sentence — and a small $ chip opens where you are typing. Type the command, then press Enter or → to leave the chip and carry on with the sentence; Backspace straight after !! gives you the two characters back. Add as many commands as you like. @ and / inside a command are just part of the command. In the saved prompt a chip is written !!`command`.

When you start, the commands run one after another, headlessly, in the worktree the agent will work in, on the machine that owns it, so an SSH project runs them on the remote host. For New worktree that's the new worktree, once it exists. For Fan out it's every attempt's own worktree, so each agent sees its own checkout's output. Once they finish, the agent starts with your prompt, each chip read as a plain `command`, followed by each command and its output. Typing Fix what !!bun run test then → reports sends:

Fix what `bun run test` reports

Before this session started, I ran these commands in the worktree. Their output follows.

<command-output command="bun run test" exit-code="1" duration="12.4s">
✗ auth › refreshes an expired token
[stderr]
error: 1 test failed
</command-output>

A failing command does not stop the launch; its exit code is included for the agent to read. Very long output is trimmed to its last 20,000 characters per stream.

If the commands are still running after 30 seconds, Pragma shows a notification with a Skip button. Skip stops the running command, skips the ones after it, and starts the agent right away with the output captured so far, marking those commands skipped="true". A session launched after that warning opens in the background, so it doesn't pull you away from what you're doing, and a notification with Open takes you to it.

In a Fan out, Skip applies to the whole fanout: the attempt running its commands stops, and the attempts still waiting start straight away with their commands marked as never run. Retrying a failed attempt runs its commands again. Fanouts created with pragma-cli fanout create or the SDK honour !!`command` chips too.

Plugins can add their own sources — see Prompt context providers.