Add DeepSeek model integration + Fix Linux/Wayland screenshots

[papadie23]

Summary

Adds DeepSeek as a new model provider and fixes several cross-platform issues.

DeepSeek integration (operate -m deepseek-with-ocr)

• DeepSeek API uses OpenAI-compatible client (https://api.deepseek.com)
• Since DeepSeek lacks vision support, screen text is extracted via Tesseract OCR and sent as structured text
• Uses deepseek-v4-pro by default (configurable via DEEPSEEK_MODEL_NAME)
• Shows model reasoning/thinking tokens in terminal
• Custom text-only system prompt for keyboard-first interaction

Screenshot fix for Linux/Wayland

• PIL ImageGrab is broken on Wayland — replaced with multiple fallbacks:
  • flameshot (primary, works on both X11 and Wayland)
  • gnome-screenshot, mss, ImageGrab as fallbacks

OCR improvements

• Added fuzzy text matching (handles OCR errors like "Gooale" ≈ "Google")
• get_text_element() now returns None instead of crashing on miss
• All OCR-mode functions updated to gracefully skip unresolvable clicks
• Global EasyOCR reader caching (no more model re-download per loop)

Behavior fixes

• Strips premature "done" operations — model must verify success before claiming it
• Smarter delays: 4s after navigation/enter, 2s base
• Empty-screen detection gives context-aware guidance

Dependencies

• Bumped requirements.txt pins from == to >= for Python 3.13 compatibility
• Fixed numpy==1.26.1 (yanked) → >=1.26.2

Files changed (7 files, no new files)

File	Change
operate/config.py	+30 lines (deepseek_init, validation)
operate/models/apis.py	+260 lines (call_deepseek_with_ocr, fixes)
operate/models/prompts.py	+65 lines (text-only system prompt)
operate/operate.py	+27 lines (done-stripping)
operate/utils/ocr.py	+85/-100 (fuzzy matching, None return)
operate/utils/screenshot.py	+35 lines (flameshot fallback chain)
requirements.txt	dep updates (== → >=, Python 3.13)

[papadie23]

Tested on:

• Ubuntu 26.04 LTS (resolute), kernel 7.0.0-22
• GNOME on Wayland (XDG_SESSION_TYPE=wayland)
• Python 3.13 with conda
• Screenshot working via flameshot (X11 tools broken on Wayland)

[papadie23]

Testing performed

• Tested operate -m deepseek-with-ocr on Ubuntu 26.04 Wayland
• Successfully launched browsers, navigated to URLs via keyboard shortcuts
• OCR text extraction works (Tesseract finds 150-250 screen text elements)
• DeepSeek reasoning mode outputs visible chain-of-thought
• Screenshot capture works across all fallback methods

Known limitations / Future improvements

• OCR lacks visual context: Tesseract extracts raw text strings but cannot understand UI layout, window focus, or whether a page has fully loaded. The model sometimes types into the wrong window because it does not know which window is focused.
• Vision model would be ideal: A future enhancement could pair this with a local vision model (e.g. Ollama + llava) to describe the screen visually, then feed that description to DeepSeek for decision-making. This would give the model awareness of window focus, loading states, and spatial layout.
• EasyOCR click mapping: Click coordinate resolution via OCR text-to-coordinate mapping is still fragile — fuzzy matching helps but is not perfect.
• Multi-monitor: The current flameshot capture works on primary display. Multi-monitor setups may need display selection.

I plan to keep iterating on these in follow-up PRs.

[papadie23]

Note: DeepSeek currently offers file/image upload in their web chat interface (chat.deepseek.com). While their API does not yet expose multimodal/vision endpoints, this suggests vision support may be added to the API in the future. When that happens, the deepseek-with-ocr mode could be upgraded to send screenshots directly as images.

[papadie23]

The README could be updated to add DeepSeek under supported models, e.g.:

#### Try DeepSeek `-m deepseek-with-ocr`
operate -m deepseek-with-ocr

Similar to how Claude, Qwen, and LLaVA are listed. Happy to include that in this PR if maintainers want.