A much clearer way to watch your coding agent work

Watch each run as it happens: each run has one header with the agent, the model, its state, the time and Stop. Under it, Claude Code or Codex writes brief progress updates, and its reads, searches, commands, edits and checks appear in the order they happened.
Changes update while the task runs: you can see the files written so far, marked as not checked and not in your project folder yet.
A cleaner result: finished work folds into one line, such as "Worked 2m · 3 reads, 3 edits, 1 check", that opens to the same timeline. Answers show their formatting properly: bold, inline code, code blocks and lists.
Second Opinion from the result: beside Open Review, ask the other agent to review the finished work, read only, and send back only the issues you pick.
Clearer failed and refused states: a run that fails or is refused says what happened and what to do next. For example, when a run is not allowed to install a package, it tells you to add the package yourself.
Stop keeps your work: ask what was finished, then Continue builds on the saved changes, and recovery picks up from the work that was kept.
Safer file deletion: when you ask to delete a file by name, Winskel deletes exactly that file and nothing else. A vague request to clean up deletes nothing.
More reliable checks and memory: tests that were already failing stay recorded as failing instead of being "fixed" by unrelated changes, Project Memory corrections carry into the next Build, and many smaller fixes make runs more dependable.