v0.12.75

feat(computer-use): support graphical VMs inside Chrome (#146)

* feat(computer-use): support graphical VMs inside Chrome

Cindro could drive ordinary web pages through DOM refs, but not a VM or
remote desktop rendered inside a tab (noVNC/Proxmox, Guacamole, SPICE,
VMware consoles, canvas/WebRTC desktops): there are no elements to snapshot,
the guest lives in nested cross-origin iframes, and synthetic JS events never
reach it.

Engine (computer-use/):
- browser_visual.py: pure logic — measured image->CSS mapping (absorbs DPR,
  browser zoom, debugger infobar, crop/downscale), iframe-chain + css-transform
  + object-fit letterbox composition of graphical surfaces, guest-framebuffer
  coordinates, US-layout CDP key/mouse event builders, per-tab inspection
  store keyed by inspect_id for concurrent agents.
- tools_browser_visual.py: browser_visual_inspect, browser_click_xy,
  browser_move_xy, browser_drag_xy, browser_scroll_xy, browser_mouse_down/up,
  browser_key_press, browser_type_trusted, browser_input_release. verify=True
  re-captures and reports whether the screen changed. browser_snapshot now
  flags graphical pages so agents switch to pixel tools automatically.
- policy: browser_visual_inspect is plan-safe; text typed by
  browser_type_trusted goes through the destructive-command scanner.

Extensions (extension/ and computer-use/extension/, shared visual.js):
- page.inspect / page.surfaces / page.focus / input.sequence handlers:
  Page.captureScreenshot + OffscreenCanvas crop/scale/grid, per-frame surface
  scan (incl. OOPIFs and shadow DOM), allowlisted Input.* dispatch with
  release-on-error, stale viewport/scroll refusal, focus emulation. Glow
  cursor follows xy input and is hidden during captures. Versions bumped.

API brains: ApiBrain forwards MCP image results (screenshots) to
OpenAI-compatible models as a follow-up user message, keeps only the newest
two, and drops them + retries once if the backend rejects images.

Prompts/docs: co-work guide + capability catalog (v6) + side-panel primer
mention the VM tools; docs/BROWSER_VM_CONTROL.md; AGENTS.md gotchas.

Tests: unit tests for the pure logic, tool tests against a scripted bridge,
and a real headless-Chromium lab that loads BOTH shipped extensions and
drives fake noVNC/Guacamole consoles (cross-origin OOPIF, nested and scaled
iframes, letterbox, 150% zoom, DPR 2, resize/scroll/fullscreen, disconnect
and recovery). linux-ci installs chromium so the lab runs in CI.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011uV59CXB81WmJAamwRL5eH

* packaging: ship a real .DirIcon/PNG for the new appimagetool

The appimage job fetches appimagetool's continuous build, and the
2026-10-29 build now fails an AppDir whose .DirIcon is missing, a symlink
or broken ("No valid .DirIcon"), and resolves the desktop Icon= only as
.png/.xpm. linuxdeploy only symlinks the SVG, so the last successful run
(2026-10-04) predates the break rather than this branch's changes.

Install a 256x256 PNG rendering of packaging/cindro.svg as both
AppDir/cindro.png and a regular-file AppDir/.DirIcon before packaging.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011uV59CXB81WmJAamwRL5eH

---------

Co-authored-by: Claude <noreply@anthropic.com>