# AG2 push URL, callback and effect check

This packet keeps the first local failure, the first Actions setup failure, and the passing PR and main originals. AG2 is pinned at `fd789e3ca8a77d14c8e3e89923665fe62c00cb9a`; A2A Python is pinned at `e649325e041e44c0b0fe57e3eef0699cad164a1e`. All 453 AG2 and 127 A2A Python files are retained unchanged, with their licenses.

Both qualified runs execute 18 actual AG2 tasks: six URL/dispatch/target variants over JSON-RPC, REST and gRPC. They retain 12 configurations, six target callbacks and three Observer writes. The legacy no-policy route stores an unsafe URL without dispatching it. The target's refusal stays separate from registration and sender admission.

The host calls the unchanged SDK sender explicitly after task completion. The callback target is a controlled ASGI app; task/config gRPC uses native loopback. The same author operates the sender, target, keys and storage (PEER). These runs do not establish an outside operator, automatic task callbacks, a network callback service or consumer adoption.

Run the data-only retained-evidence check:

```sh
python3 validate_capsule.py
```

The validator authenticates complete original archives, derives the report from their native task/config/callback/effect fields, and checks the two original installed-reader outputs and 16 semantic controls. It does not execute the SDK or replace the cryptographic Observer reader. The exact installed replay is in the [Observer profile](https://github.com/probityai/agent-evidence-observer/tree/4b5ca47854c9be2814eeb5fe6f72b79b0182d2c7/interop/ag2-push-authority-2026-10-04).

The original ZIPs omit 15 empty refused-case workspace directories. Restore those directories before an installed reader replay; no file bytes change. Full Python installations and third-party dependency archives are outside these retained provider ZIPs.

The older 16-row comparison stays separately pinned. The prospective eight-task implementation-owned study has not started. This packet qualifies a finite native source route with fixed, exposed answers; it is not a model-quality study.