Add first-class image attachments to the interactive chat TUI. Users can attach images by pasting or drag-dropping local image paths into the prompt, and by pasting clipboard images through terminal-delivered paste input. Clipboard images are materialized to an app-specific directory under `os.tmpdir()` and then treated like normal file attachments.

Message and persistence:

- Extend `src/utils/ollama.ts` `Message` with `images?: string[]`.
- Pass `images` through unchanged to Ollama chat requests.
- Preserve `images` in session JSONL persistence and reload them on resume.

Attachment model:

- Add staged attachment state to chat input, separate from plain text.
- On submit, send one user message with:
  - `content`: typed text only
  - `images`: absolute image paths for staged attachments
- Clear staged attachments after successful submit.

Input UX:

- Do not add an `/image` command.
- Render attachments inline as bracketed atomic tokens using filenames:
  - real files: `[diagram.png]`
  - clipboard images: `[image-1.png]`, `[image-2.png]`
- Use a dedicated attachment color distinct from normal prompt text.
- Keep the cursor outside token internals; left/right moves across each token as one unit.
- `Backspace`/`Delete` removes an adjacent token as a whole.
- Normal typed text remains editable after the attachment tokens.

Attachment ingestion:

- Detect pasted or drag-dropped local image paths in the prompt and stage them as attachments instead of leaving them as plain text.
- If pasted text is not a valid readable image path, keep it as ordinary text.
- Add clipboard-image ingestion through paste/key handling where the terminal exposes it.
- Materialize clipboard images to `path.join(os.tmpdir(), 'code-ollama', 'images')` with generated names like `image-1.png`.
- Keep user-supplied file attachments at their original paths; only clipboard images are copied to temp storage.

Rendering history:

- Show sent user attachments in message history using the same bracketed filename convention.
- Do not attempt inline bitmap previews in the terminal.

Public Interface Changes:

- `Message` gains optional `images?: string[]`.
- `TextInput`/`ChatInput` move from string-only editing to structured input with attachment tokens plus text.
- Add an internal clipboard utility that reads a clipboard image and writes it to the app temp image directory.

Test Plan:

- `utils/ollama` tests: messages with `images` are forwarded to Ollama.
- Session tests: messages with `images` persist and reload correctly.
- Chat input tests:
  - valid image paths stage attachments
  - invalid paths remain plain text
  - multiple attachments render as bracketed filename tokens
  - token navigation is atomic
  - backspace/delete removes whole tokens
  - normal text editing still works
- Clipboard tests:
  - clipboard image ingestion writes to the temp image directory and stages the resulting file
  - generated names follow `image-N.ext`
  - empty/unavailable clipboard produces a user-facing error
- Submit/render tests:
  - submitting text plus attachments sends `content` and `images`
  - staged attachments clear after submit
  - history renders bracketed attachment filenames consistently
- Regression tests:
  - slash command flow
  - `@` file suggestion flow
  - prompt history navigation

Assumptions and Defaults:

- Scope is TUI chat only; the one-off `run` command remains unchanged.
- Attachment tokens always render with brackets and a distinct color.
- Clipboard-derived images are stored under `os.tmpdir()` in an app-specific subdirectory.
- If a staged clipboard image is removed before send, its temp file should be deleted.
- Sent clipboard image temp files are retained in v1 so resumed sessions can still reference them as long as the OS has not cleaned temp storage.
- Model capability checks are not enforced in the UI; unsupported image behavior is left to Ollama/model responses.

Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>