Completion Gate

A claim about the state of the work is a statement of fact, and a fact needs evidence. The evidence is output you produced and read in this message: a test run, a build, a request and its response, a screenshot. If you did not run it just now, you do not know it.

The Prime Directive

NO CLAIM OF DONE, FIXED OR PASSING WITHOUT FRESH VERIFICATION OUTPUT IN THIS MESSAGE

No exceptions. No workarounds. No shortcuts.

When to Use

Required, before any of these:

• Saying or implying that something works, passes, is fixed, is complete or is ready, in any wording.
• Committing, pushing, opening a pull request, merging, tagging, deploying.
• Marking a task complete, moving to the next task, or ending your turn with a summary.
• Passing on a subagent's or another model's "done" as your own statement.
• Expressing satisfaction about the result ("that did it", "all good").

Skip:

• Statements that are explicitly about what you have not verified ("I have not run the tests yet"). Those are honest and need no gate.

The Protocol

1. Name the claim. Write down exactly what you are about to assert: "the login bug is fixed", "all tests pass", "the deploy uses the new database".
2. Name the proof. Decide which command or observation would prove that claim, and what output would disprove it. Use the table below. A claim about behaviour needs a check of behaviour; a check that something ran is not a check that it did the right thing.
3. Run it now, in full. The full test suite, not the one file you touched. The real build, not the linter. The real request against the running service, not the unit test that mocks it. Earlier runs, other sessions and a subagent's report do not count.
4. Read the output. The exit code, the counts (passed, failed, skipped, warnings), the error lines and the end of the log. Skipped tests and new warnings are findings, not noise.
5. Compare output to claim.
   • The output supports the claim: state the claim with the evidence next to it.
   • The output contradicts it: state the actual state with the evidence, and go back to work.
   • The output is partial or ambiguous: say exactly what is verified and what is not.

6. Fill in the evidence table for any non-trivial completion and include it in the message:

   Claim	Command or check	Result (from this message)
   Tests pass	npm test	214 passed, 0 failed, 0 skipped, exit 0
   Build succeeds	npm run build	exit 0, no warnings
   Bug fixed	reproduction from the report	returns 200 with the expected body; the test fails with the fix reverted
   Requirements met	checklist against the plan, task 4	6 of 6 items checked, see below

7. Only now make the claim, in plain words, with no hedging and no inflation.

What Counts as Proof

Configuration Changes

A configuration change can "succeed" because some valid configuration exists, while yours is ignored. This is the silent fallback: the request returns 200, but from the old provider; the app starts, but with the default flag; the job runs, but against staging.

So verify the difference, not the absence of errors:

1. What should be different after the change? Name the observable value (a model name in the response, a hostname in the log, a feature visible on the page, an identity in the auth response).
2. Where can you observe it? A response body, a log line, a header, a version endpoint, a query result.
3. Run the command that exposes it and read the value.
4. Confirm the value is the new one. If you only saw a status code, you have not verified the change.

UI Changes

Code that compiles is not a screen that works. For any change a user can see:

• Look at it. Open the page in a real browser (or a headless one you control) and capture a screenshot of the changed state. Read the screenshot; do not just save it.
• Exercise it. Click, type and submit through the changed flow, including the error and empty states.
• Check the sizes that matter. At least a phone width and a desktop width, or the breakpoints the project defines.
• Check access. Keyboard navigation reaches and operates the control; focus is visible; contrast and labels pass the project's accessibility check.
• Check the console. No new errors or warnings in the browser console.

If you cannot run a browser in this environment, say so and list what the user should look at; do not claim the UI works.

Reporting Honestly

• Report failures with their output. "3 of 214 tests fail; here are the three" is a useful message. "Mostly working" is not.
• Report what you skipped and why. "I did not run the end-to-end suite because it needs the staging database" lets the user decide.
• Report what you could not verify first, before what you could.
• Do not soften a failure into a success or inflate a partial result into a complete one.

Cognitive Traps

Guardrails

Never:

• Use "should", "probably", "seems to" or "looks good" about the state of the work in place of evidence.
• Commit, push, open a pull request, merge or deploy without fresh verification.
• Pass on another agent's success report without checking its diff and running the checks.
• Claim a configuration change works from a status code alone.
• Claim UI work is done without having seen it rendered.
• Hide a failing, skipped or unverified item.

Always:

• Name the claim and the proof before running anything.
• Run the full command now and read every count in its output.
• Put the evidence next to the claim, in a table when there is more than one.
• Lead with what could not be verified.
• Say plainly when something is done and verified, without hedging.
• Expect to be checked: on a harness with skill-scoped stop hooks, invoking this skill arms a one-shot check that reads your final message and sends an unsupported claim back to you.

Connections

• godmode:test-first: supplies the red and green evidence for features and regression tests.
• godmode:quality-gates: supplies the lint, type, coverage, build, size and audit evidence.
• godmode:code-review and godmode:cross-model-review: run before the final verification; their fixes are re-verified here.
• godmode:fault-diagnosis: when the verification fails and the cause is not obvious.
• godmode:delegated-execution: every subagent's "done" passes through this gate before it is accepted.
• godmode:merge-protocol: landing happens only after this gate passes.