Add agent PR completion protocol

AGENTS.md

@@ -83,3 +83,36 @@ This project uses `ty` for type checking. Run:
 ```bash
 uv run ty check src tests
 ```
+
+## PR completion protocol
+
+Do not consider an implementation task complete merely because code has been pushed or a pull request has been opened.
+
+After creating or updating a pull request:
+
+1. Inspect all required GitHub Actions checks for the current PR head.
+2. If checks are still running, re-check them while the session is active rather than handing the PR back to the user for manual monitoring.
+3. If a required check fails, inspect the failing job and logs and determine whether the failure is caused by the PR.
+4. If the fix is within the stated task scope, make the fix, push it, and inspect CI again.
+5. Repeat the diagnose/fix/re-check loop until all required checks pass or a genuine blocker requires user input.
+
+Escalate to the user only when resolving the failure would require one or more of the following:
+
+- changing frozen statistical semantics, contracts, or architecture;
+- broadening the agreed task scope;
+- weakening or removing a meaningful test or quality gate;
+- changing a public API or backward-compatibility promise beyond the task;
+- making a product or technical decision with multiple legitimate choices;
+- resolving an external service, permissions, infrastructure, or credential problem that the agent cannot fix safely.
+
+Routine failures such as lint errors, formatting errors, test regressions caused by the PR, snapshots/fixtures that legitimately need updating, packaging errors, documentation-build errors, and similar mechanical issues should be fixed without asking the user to manually inspect GitHub Actions.
+
+The final handoff should include:
+
+- pull request link;
+- final PR head commit;
+- tests/checks run locally when applicable;
+- final GitHub Actions status;
+- any remaining caveats or blockers.
+
+Do not ask the user to manually check whether CI passed.