#6300·CodeWhale

Computer Use 0.8.0 acceptance + hardening pass (muse-driven)

Author: HmbownCreated Sep 17, 2026Updated Sep 17, 2026

Muse-driven acceptance + hardening pass for Computer Use 0.8.0, source of truth codewhale-cu-plugin @ main 460235a (0.8.0 + grant-on-refusal fix). Driven through the installed computer-use MCP server in muse, background mode only.

Phase 0 — muse wiring: plugin 0.8.0 enabled+active, both skills listed. Phase 1 — surface checks (repo gate: npm test → 333 tests, 318 pass, 0 fail, 15 skipped on macOS):

  1. request_access: app.version 0.8.0, bundled_version, stale false
  2. tools/list: 36 advertised; hidden aliases (left_click etc.) callable via raw JSON-RPC with clean refusals; click {clicks:4} → bad_args
  3. invoke_menu File>New on windowless TextEdit (route accessibility, front_lease false); File>Close no-op documented; close via button/cmd+w
  4. key cmd+w receipt carries front_lease AND front_restored (honesty check)
  5. CODEWHALE_CU_GRANT=read-only narrows surface; request_access reports grant on refusal receipts too
  6. trajectories: record 3 turns incl. refusal; replay stops at refusal; dry_run first
  7. list_apps installed/all; set_window_frame verified geometry; fixed-size refusal via ax_errors
  8. browser CDP against local fixture in self-owned profile; human Chrome untouched
  9. list_sessions; preview; kill_app Calculator with verified termination
  10. Regression sweep: malformed calls → bad_args/bad_target (never TypeError); stale element → element_stale; stop_computer_control → control_stopped; real pointer never moves

Phase 2 — fix forward in small slices at source if anything fails; propagate to marketplace mirror + Core vendored copy + muse bundle rebuild. Phase 3 — numbered PASS/FAIL report with receipts.

Human-gated (ask first): pushing Core repo, website deploy, cutting v0.8.0 release tag.