codagent-ai/agent-skills/verify-behavior

Verify Behavior

Writing the checklist

Read the change's proposal, specs, design, and test plan, but not the code. For each behavior the spec describes, including its error, failure, and cleanup cases, say how you will verify it: an automated test (unit, integration, or end-to-end), a scripted command whose exit code or output decides the result (such as a build, lint, or verify script), or inspection of the code or files. Never plan a check in which an agent operates the product as a user would; that testing happens once, in acceptance testing after implementation. Do not cite the test plan's acceptance (AT-) or human-only (HT-) tests as a check, and add no rows for them. Several behaviors may share one check. Choose the fastest check that would fail if the behavior broke: inspection or a unit or component test where one can show it, and an integration or end-to-end test only where the behavior exists only across processes, a real browser, or the built product. Plan no integration or end-to-end tests beyond those the test plan names; cite a named test's ID for the behaviors that need it, and add a row for any such test that no behavior covers. Write the checklist where you were asked to, or else as checklist.md beside the change's spec, in one table:

Recording evidence

After building, add the columns Code, Evidence, and Status.

• Run each check and save its output to a file in the evidence directory you were given, or else

in evidence/ beside the checklist: a test log, command output, or, for something that can only be inspected, a note of what you looked at.

• Put the evidence file names, in backticks, in the Evidence column. Status holds exactly one of

verified, failed, or not verified, and nothing else. For failed, Evidence says what fails; for not verified, it says why you could not run the check.

• Never mark a behavior verified unless you ran its check.
• Mention no acceptance (AT-) or human-only (HT-) test ID anywhere in the checklist, not even to

say a behavior is left to acceptance testing.

• Evidence can come from the one broader test run and from targeted runs of single tests; never

rerun a whole suite only to refresh evidence.

• After any change to code or files, rerun the check of every row that depends on them, each alone,

and update its Evidence, or set it to not verified. Evidence from before the change no longer counts for those rows; rows that do not depend on the change keep theirs, and a test you added or changed needs only its own run.