"All tests pass": what was seen while your agent worked, beside its tick

An agent ticks a task and writes "done, all tests pass". That is a sentence. You weren't watching the terminal, so you can believe it or go and look. With four agents finishing work through an afternoon, looking means reading four transcripts.

Hamloy now puts beside the tick what was seen while the agent worked: each test or check it ran and how it ended, the files it changed, and the pull request with its checks. To review, one of the overview's pages, lists the finished tasks with the ones that need you first.

Four tasks, one page

On a scratch project, Pebble sync, Claude Code was told only "Work through this project's open tasks." It did three. A fourth was finished the way Codex reports it, on a branch with a pull request. Nobody told Claude to report anything. Afterwards To review said:

4 finished tasks to look over · 1 with no test run seen

NO TEST RUN SEEN  1
[x] Fill in the README's Usage section with one example per function
    Done by claude 4m ago
    No test or check run seen while it was worked on.
    1 file changed: README.md

CHECKS PASSED  3
[x] titleOf gives Untitled for a note that starts with a blank line
    Done by codex 2m ago
    npm test  passed after 1 failed run · 10 tests, 10 passed, 0 failed
    2 files changed: test/words.test.js, src/words.js
    Pull request #7, merged, its checks passed
[x] wordCount counts a double space or a line break as an extra word
    Done by claude 4m ago
    npm test && npm run lint  passed · 4 passed, 0 failed
    npm test  failed · 4 tests, 3 passed, 1 failed
    1 file changed: src/words.js

The README task is first because nothing was run on it. That may be fine for a README. It is the one to glance at either way.

Agents run tests their own way

Claude ran its tests the way it likes to:

npm test 2>&1 | grep -E "^# (pass|fail)"; npm run lint 2>&1 | tail -2

The board still read it as npm test && npm run lint, passed, 4 passed and 0 failed. The first run, which failed, stays under it, drawn quietly: a test that failed before the fix isn't held against the work, but you can see the fix was needed.

What is kept is the command, how it ended, and the one line that counts the results, like 12 runs, 0 failures. Never the rest of the output, and no command that isn't a test, a linter, a type check or a build. Files are kept by name, never their contents. A word in a command that looks like a token or a password is masked.

Claude Code and Codex alike

Claude Code tells the program on your computer how a command ended. Codex hands over the output instead, so there the line that counts the results decides. Both end up on the same page, in the same words. For a task finished on a branch, the program asks GitHub's gh on that computer about the branch's pull request, so Codex's task above says #7 was merged with its checks passed.

Looked over, or opened again

Looks good takes a task off To review, and its agent hears it with its next news:

- The human looked over the work of B-l94y "titleOf gives Untitled for a note that starts with a blank line": it stays done

Open again asks what is wrong first. What you write goes on the task as a note for the agent, and the task opens for the work to be done over.

A task you tick yourself never waits here, and neither do requests, questions or work brought in already done. Only an agent's tick asks for a look.

See What was seen of the work for everything it keeps and how a run is read, and The hamloy program for the program that reports it.