15 lines
1.6 KiB
Markdown
15 lines
1.6 KiB
Markdown
# Visual assistant completion verification
|
|
|
|
Changed only `_register_assistant_review`, `poll_review_assistant` and added `_assistant_attention` in each current bot variant. Preserved unrelated concurrent changes. `_review_assistant_protected` matches both pre-change backups byte-for-byte at the method level.
|
|
|
|
Both awaiting-review and needs-human states use the shared UI classifier. Registered tasks retain account/conversation identity before reminder creation. Only explicit `completed_by_human` evidence from the coordinator and non-active send state can remove pending work. Original task/send evidence is copied into a unique resumed manual-handoff audit entry; native sending journals are untouched. Old drafts are never approved or retried. Pending removal, audit record and discarded-draft marker persist together; failure restores all in-memory records. New customer continuation includes exact event rowid and dedup identity.
|
|
|
|
Validation uses temporary synthetic fixtures and the existing external-action-blocking harness. No live database/settings/queue writes, no real messages or UI input.
|
|
|
|
- Source: 57 tests, 0 failures/errors, 2 existing platform-variant skips (production-only functionality).
|
|
- Production: 57 tests, 0 failures/errors, 0 skips.
|
|
- Included modules: test_visual_review_assistant (38), test_handoff_evidence, test_visual_handoff_visibility.
|
|
- Both runs reported blocked_external_actions=[]; original desktop actions and network are patched to raise if attempted.
|
|
|
|
Artifacts: visual-handoff-source-result.json / .log; visual-handoff-production-result.json / .log; visual-method-verification.json.
|