real dev verification, 2026-11-02, linux boat vm
baseline: github/main cc59f99e735300120a6fc74d9a0b9df7c98fa81c
original pr head: 336badcbb6a09249c4dd6133ffcc59b8a7b0afcf
verified fix: 579de19f14a17adcdcc8cae239c9547c745b3d93

setup: separate main worktree, vp i; one background dev server at a time.
browser: repository Playwright and /usr/bin/google-chrome.
project and threads created through the web client against a real git repo.
real fx custom provider and native Codex both used gpt-6.1-sol through the local proxy.
no mocks, fixtures, replay provider, or database writes.

main, fx 0.0.13: no latest-release advisory or update action.
pr, fx 0.0.13: Update available -> Update now -> Running fx upgrade.
actual /home/user/.local/bin/fx version after update: 0.0.16.
provider remained authenticated and update advisory cleared.

main Stop: persisted command_execution status=failed, output=Cancelled.
original pr with real fx 0.0.16: cancellation still failed.
captured real error envelope: {"error":{"tool":"shell","code":"Cancelled","retryable":false}}
fix accepts this envelope as well as the older details.error envelope.
after fix, running shell monitor -> Stop generation:
  persisted command_execution status=interrupted, output=<same envelope above>.
same-session next turn: printf FX_RECOVERY_OK; exit 17
  persisted command_execution status=failed, output=FX_RECOVERY_OK.
  assistant correctly reported exit status 17.
native fx subagent -> Stop parent -> parent Cancelled/Stopped;
  child view displayed readable Cancelled and Interrupted, without raw JSON.
native Codex neighbor: pwd and git rev-parse --is-inside-work-tree completed;
  assistant replied CODEX_LIVE_OK.

blacksmith: new Cancelled regression failed before fix (expected true, got false).
after fix: vp run check:changed HEAD --concurrency 2 passed:
  format, lint, both knip passes, typecheck (10 packages), tests (22 in 2 files).
earlier focused run: 122 tests passed in maintenance resolver, maintenance,
  protocol, and cancellation support.
git diff --check passed.

not exercised: native desktop/mobile, macOS/Windows hosts, actual Homebrew/mise
installations, relay/tunnel connections. Existing resolver tests cover guards;
those tests do not establish real native-install behavior.
videos are edited interaction excerpts, not performance measurements.