Treat "all checks pass" as a claim about one run: ask for the command, directory, commit, exit status and test count, rule out empty, cached, masked and timed-out results, and accept only a pass on the commit you merge.
Treat every edit that changes what a test expects as a change to the requirement: find it in the diff, decide whether the test or the code is wrong, have the requirement's owner approve a changed expectation, and prove the…
Coverage shows which lines ran, not whether a test would notice them doing the wrong thing. Break the behaviour on purpose, read what each test asserts, add refusals and state checks, and use mutation testing on changed code.
Write the expected behaviour before any code exists, read why the first test failed, lock the tests while the agent implements, and ask for a record of both runs before you accept the change.
Sort mid-build news into clarification, changed requirement or separate feature, trace each changed requirement to the work built on it, publish a revision that names what it supersedes, and review the diff against it.
Why long agent runs degrade on large features, how to cut a feature into checkable behaviours, how small each task should be, what integration checks the seams need, and which evidence should start the next task or revise the plan.
Why agents tick off placeholders, how to write acceptance criteria a stub would fail, how to trace each criterion to a change and a check, who may waive one, and which hooks keep the agent from closing its own task.
What a coding agent can learn from your codebase, what still needs a decision from you, the minimum spec for any feature, when to write more, and how to tell whether the spec saved rework.
What Claude Code, Codex and Cursor plan modes hold back, what a plan needs before you approve it, how to record that approval against one plan version, and which host checks actually stop an early edit.
Read the evidence precisely, hold yourself to explaining every accepted change without the chat, engage at the level of design, record decisions, and test your own understanding from time to time.
