agent-completion-gate: /goal can loop; only artifacts grant complete

zhjai’s agent-completion-gate (MIT; npx skills add zhjai/agent-completion-gate; v0.4.x line) starts from a sentence factories keep learning the hard way: the goal is not the acceptance criteria. A monthly sales report page can “exist,” with green unit tests, while the CSV export is missing, the chart has one point, and the title still says Untitled — and the agent honestly believes it is finished. The kit turns done into an external acceptance check. The worker may only write candidate_complete; a protected gate reads real artifacts against a human-confirmed acceptance manifest and alone may write complete. Unknown user-visible surfaces fail closed. Self-check mode exists and is labeled SELF-CHECK-OK — deliberately not verified completion.

The timely wedge is how this composes with built-in autonomy loops. Claude Code’s /goal (and Codex’s goal mechanism) keep going until an internal, transcript-only evaluator judges the stop condition met — Claude’s docs note that evaluator does not call tools, so it can only judge what already appeared in the conversation. That is a great keep-going loop and a weak done check. The gate fills the gap: make the goal condition require a surfaced successful gate/verify_completion.sh run, so the loop terminates on transcript evidence of an artifact check, not on a narrative that sounds finished. goal-compile scaffolds the gate, drafts surfaces and machine checks for one human confirmation, then runs until the gate passes. CI can mark verify-completion required and CODEOWNERS-protect gate/ and control/ so a PR cannot rewrite the judge that judges it; --diff-base / --touched can derive touched surfaces from the real git diff instead of trusting the worker’s self-report.

Why it matters for code review

Always-on harnessed throughput compounds only when “complete” is a file-backed verdict outside the writer’s session. Peak-session mastery can eyeball the missing export; overnight volume cannot. Pair continuous agents with an acceptance gate that refuses transcript self-grading — especially under /goal-style loops — or FOMO that nothing ran overnight just means a larger pile of candidate_complete claims nobody’s artifacts supported. Review the blocked surfaces and the manifest diffs; those are the morning merge criteria.