Done is a claim

Rules for coding agents. Each one exists because breaking it cost us a real failure; the story behind each is in INCIDENTS.md.

Before you start

1. Write the acceptance command before the work. Name the observable outcome it checks. Never weaken it to make it pass; if the command is wrong, explain and record a correction that preserves the intended outcome before relying on it.
2. Check that every path you were given exists. Resolve missing inputs before relying on them. An explicitly requested new file is different: verify its intended parent and scope.

Before you say "done"

3. Re-run the acceptance yourself, on the current work, and quote its output. Name the revision and any uncommitted inputs. A check that changes those inputs does not certify the resulting tree.
4. Run the whole relevant suite, not only your test. Your test passing says nothing about the tests next to it. If a required check cannot run, name the gap instead of claiming full verification.
5. Record the command's own exit code. In cmd | tail; echo $? the $? belongs to tail. Capture the producer's status before reading its saved output, or use the shell's per-command pipeline statuses. pipefail detects a failing pipeline but is not necessarily the producer's own code.
6. "No tests ran" is a failure. An exit code of 4 or 5 from pytest, an empty collection or a skipped suite is not a pass.
7. Unknown is not pass. If a check could not run, report it as not run, with the reason.
8. Green means every required check. Read every required CI job on the revision, not the one you picked; report skipped or unavailable checks explicitly.

When you report

9. Read the artefact, not the summary. A status: ok file, a log that says "Saved" or a falling counter describes something narrower than your claim. Open the diff, the file, the frame.
10. A file that exists is not a file that works. Decode the image, load the page, run the binary.
11. Look at it yourself. A 200 response is not a working page; open it and look.
12. Verify every citation. Check repository citations with git cat-file -e <sha> and git cat-file -e <ref>:<path>; open external sources and verify that they support the claim. Identify private or historical sources and their access limits.
13. State the base, and re-check it before you publish. "Tested on abc123" stays true; "abc123 is current main" expires.

When you write tests and fixes

14. A test that repeats a constant proves nothing. If the code says MIN = 3.51 and the test checks 3.51, ask where 3.51 came from.
15. Fix the class, not the instance. Find every call site of the defect before you call it fixed, and test more than the one you changed.
16. Run a detector on a known-good case first. A check that flags 100% of its inputs is measuring itself.
17. Register cleanup before the work it cleans up. Cleanup placed after the work never runs on the path that fails.

Never

18. Never run a destructive command to answer a question. git status and git diff answer "does this match main"; git reset --hard destroys the answer.