GPT Review — 23-P3D2 Notification Triggers Prompt rev3

Date: 2026-05-08
Reviewer: GPT-5.5 Thinking / Incomex Hội đồng AI
Reviewed: knowledge/dev/laws/dieu44-trien-khai/prompts/23-p3d2-notification-triggers-prompt.md rev3

Verdict

Rev3 is close, but do not dispatch yet. Rev4 required.

Opus applied most fixes correctly and incorporated the User’s implicit self-read requirement. However, rev3 still leaves execution-time adaptation inside the CREATE section, and Variant B is incomplete. That is not acceptable for a production trigger deployment prompt.

Accepted rev3 improvements

• Creator/self implicit read is correctly added to board semantics:
  • own events are treated as read for the creator;
  • unread_for_actor=false for own events.
• Board function includes unread_for_actor.
• Static quoted heredoc direction is mostly adopted.
• Partial unique event index is now checked through pg_index.
• Owner assignment uses set_config + format('%I', owner).
• PUBLIC revocation is applied to all four functions.
• Source safety is broader.
• Cleanup captures event ids before deleting events.
• Final report includes board/read-state status and next context update pack.

Required rev4 fixes

P1 — BLOCKER: Variant B must be fully written, not “copy from Variant A”

Rev3 Variant B contains comments:

-- fn_iu_notif_comment and fn_iu_notif_draft: identical to Variant A
-- (Agent: copy exact bodies from Variant A above)
-- fn_iu_notification_board: identical to Variant A
-- (Agent: copy exact body from Variant A above)
-- Triggers, REVOKE, owner DO block: identical to Variant A

This violates the no-adaptation rule. A runnable execution prompt must contain complete SQL for both variants, or it must choose one variant by preflight and generate a complete block.

Patch:

• Expand Variant B fully.
• No “copy from above” instructions inside executable section.
• Agent should select exactly one complete block and run it.

P2 — BLOCKER: author_type branch still asks Agent to adapt CREATE body

Rev3 says:

If HAS_AUTHOR_TYPE=1, adapt fn_iu_notif_comment suppress condition to...

This reintroduces execution-time editing inside the CREATE section.

Patch one of two ways:

Option A — Require author_type exists and hard-code:

IF NEW.comment_kind='system' OR NEW.author_type='system' THEN RETURN NEW; END IF;

Option B — Provide complete Variant C/D for comment trigger with/without author_type.

GPT preference: Option A if runtime P3C1/P3C2 reports confirm author_type exists. P3D2 preflight can require it. If it does not exist, STOP and return for review rather than silently weakening suppression.

P3 — P3D1 read unique index check still incomplete

Rev3 checks uq_notif_event_type_ref, chk_notif_event_type_stream, and fk_notif_read_event, but it no longer verifies uq_notif_read_event_actor.

Patch preflight to verify uq_notif_read_event_actor exists and is unique via pg_index/pg_class.

P4 — Board function should include read_status_source for implicit self-read

To avoid future confusion, board JSON should distinguish explicit read rows from implicit creator-read.

Add:

read_status_source = explicit_read | implicit_self | unread | not_applicable

Rules:

• p_actor IS NULL → not_applicable;
• e.actor_ref=p_actor → implicit_self;
• read row exists → explicit_read;
• else → unread.

This directly addresses the User’s point: if Opus created it, Opus is treated as having read it even without a read row.

P5 — T16 must verify unread_for_actor=false and read_status_source=implicit_self

Rev3 T16 only verifies is_read_for_actor=true for self. Add assertions:

• self event: unread_for_actor=false;
• self event: read_status_source=implicit_self;
• GPT after mark_read: read_status_source=explicit_read for at least one event.

P6 — T15 PUBLIC revoke check can miss overloaded function signatures

proname IN (...) is acceptable if no overloads exist, but this prompt creates a board with signature (text,text,integer). Add exact signature verification for the board:

to_regprocedure('public.fn_iu_notification_board(text,text,integer)')

Also verify the three trigger functions as zero-arg regprocedures:

public.fn_iu_notif_comment()
public.fn_iu_notif_draft()
public.fn_iu_notif_version()

P7 — T14 should require insert into event table, not only forbid bad inserts

Rev3 forbids bad INSERT targets but does not assert each trigger function actually inserts into iu_notification_event.

Patch T14:

• for each of the three trigger functions, require source matches:
  • insert into public.iu_notification_event or insert into iu_notification_event;
• board function should not insert anywhere.

P8 — Cleanup with TEST_EVT_IDS can fail if array is NULL

If no test events exist due to earlier failure but cleanup path runs, ANY(NULL::uuid[]) behavior may not be intended.

Patch:

• if TEST_EVT_IDS is empty/null, set cleanup event id array to '{}'::uuid[];
• report test_event_ids_count.

P9 — Fail cleanup should also run when FN_STATUS=FAIL after partial create and verify exact objects removed

Rev3 has a partial cleanup branch, but it does not verify exact remaining trigger/function count after the partial cleanup. Add the same exact verification used in the FAIL cleanup branch. If anything remains, PHASE_STATUS=CRITICAL.

P10 — Final report should include trigger/function creation variant and author_type suppression mode

Add report fields:

version_actor_variant=A_created_by|B_draft_applied_by_fallback
comment_suppression=comment_kind_and_author_type|comment_kind_only
self_read_policy=implicit_self_read
board_read_status_source=ACTIVE

If rev4 chooses to require author_type, then comment_suppression=comment_kind_and_author_type should be expected.

Directive to Opus

Patch P3D2 prompt to rev4 with P1–P10.

Path:

knowledge/dev/laws/dieu44-trien-khai/prompts/23-p3d2-notification-triggers-prompt.md

Do not dispatch after patch. Return for GPT/User final review.

Hard boundaries remain

• No dispatch.
• No Pack 23 function changes.
• No P3D1 table changes.
• No gateway/birth trigger changes.
• No vector mutation.
• No LISTEN/NOTIFY.
• No retention/archival.
• No external queue.
• No global read flag.
• No Hermes production start.

Summary

Rev3 mostly solves the core issues but still has incomplete executable SQL and a lingering “Agent adapt” instruction. Rev4 should make all CREATE paths complete and static, encode implicit self-read explicitly in the board output, and tighten final verification. After rev4, P3D2 may be ready for final approval.