Give the agent precise context

By the end of this walkthrough you can point the agent at exact files and sessions, attach screenshots, and write structured multi-line prompts.

Mention files and sessions

Type @ followed by part of a project path or session title. The popup lists workspace files first, then other sessions; move with the arrow keys and press Tab to insert the selected row.

A file mention inserts the path only: the agent reads the file with its normal tools, and no file contents are uploaded with your message. A session mention inserts a marker that omdsh replaces with a read-only snapshot of that session for the model; it does not resume or fork the source session. Mentions stay highlighted in the composer.

Compare @packages/tui/omdsh-tui/src/chrome/renderer.ts with @packages/tui/omdsh-tui/src/chrome/renderer.spec.ts and explain the missing edge case before editing.

Typing ./ or ~/ opens plain path completion. It inserts a path; it does not bypass tool permissions or upload contents.

Add screenshots and images

Copy an image and press Ctrl+V. The composer inserts a compact image marker; add any explanatory text and submit them as one message, or send the image with /goal or /plan. Pasting an image file path also imports that image when it can be read.

Press Alt+M to inspect draft images before sending. Use Tab and Enter for the preview actions, or P/N to switch between images and O to open the original. Esc closes the preview and keeps the draft. Stored images remain accessible through /attachments; Files and terminals describes terminal support and preview limits.

Image paste needs the platform clipboard reader. On Linux it uses wl-paste under Wayland or xclip under X11; if neither is installed, text paste keeps working but clipboard images are unavailable.

The default DeepSeek catalog includes the image-capable deepseek-flash; the other default DeepSeek models stay text-only. A deployment that wants native image requests on another model must list inputModalities: [text, image] on that model.

Pasted images are checked on admission and normalized for storage:

• Sources are accepted up to 20 MiB, 64,000,000 pixels, and 8192px per side.
• An accepted image is stored under a 2048×2048 total-pixel budget, an 8192px long-edge cap, and a 4 MiB encoded-size target. The budget governs, so an extreme aspect ratio keeps its short-edge resolution instead of being clamped to a 2048px long edge.
• A refused image shows an error notice and stays out of the prompt.
• A message with several images is admitted as one ordered batch; if any image is refused, the composer text and drafts stay put.

These checks apply at paste and storage time; the vision model applies its own image budget later, when the request is prepared. A route that rejects the request as over that budget drops the oldest images locally and retries instead of failing the turn, and the session log keeps every image, so a resumed or forked session still knows what was omitted.

Write structured prompts

Use Shift+Enter, Alt+Enter, or Ctrl+J for a newline. For a longer request, press Ctrl+X to edit the current draft in $VISUAL or $EDITOR; leaving the editor returns the text to the composer.

Goal: remove the duplicate loading row.

Constraints:
- keep the composer anchored
- preserve CJK display width
- add a regression test

Verification: run the focused TUI test and typecheck.