KB-43BA

GPT Review — Runtime Execution Prompt for Description Policy Option 1

4 min read Revision 1
gpt-reviewdescription-policyruntime-promptoption1approve-with-patchesiu-0

GPT Review — Runtime Execution Prompt for Description Policy Option 1

Date: 2026-05-04 Reviewer: GPT-5.5 Thinking / Incomex Hội đồng AI Reviewed: Opus draft prompt "Description Policy Option 1 — Runtime Execution"

Verdict

Prompt is close, but requires 4 small hardening patches before dispatch.

The prompt correctly separates runtime DB work from KB/H11 doc patching, preserves Pack 2B hard stop, and includes preflight/DDL/seed/function/verify/report phases.

However, because this runtime prompt includes DDL and function replacement, it needs additional execution-safety controls before Claude Code is allowed to run it.

Required patches before dispatch

P1 — Wrap Phase 1–3 in explicit transaction where possible

Use a transaction for DDL + seed + function replace:

BEGIN;
-- DDL
-- seed
-- CREATE OR REPLACE FUNCTION
-- verification queries that can run inside transaction
COMMIT;

If any error before COMMIT → ROLLBACK.

If Directus auto-detect behavior requires commit before visibility, state that clearly:

  • commit only after PG + fn checks pass;
  • Directus visibility check may happen after commit;
  • if Directus absent after commit, STOP and report; do not continue to any further mutation.

P2 — Do not allow Agent to synthesize function blindly

Current prompt says "take function source and add 3 changes". That is still risky.

Add:

  • Agent must print full new CREATE OR REPLACE FUNCTION before running.
  • Agent must compare old vs new and state that only these allowed tokens/blocks changed:
    • _desc_policy TEXT;
    • SELECT governance_role, description_policy ...
    • structured_exempt early return;
    • unclassified warning.
  • If the source shape is not straightforward or the diff touches other logic → STOP and report.

P3 — Phase 0j schema tool check must inspect more broadly

Current query only checks DOT-SCHEMA-MIGRATE% and DOT-COLLECTION-ALTER%. Expand to include schema/apply/pg migration conventions:

SELECT code, name, status, operation, file_path
FROM dot_tools
WHERE code ILIKE '%SCHEMA%'
   OR code ILIKE '%MIGR%'
   OR code ILIKE '%ALTER%'
   OR name ILIKE '%schema%'
   OR name ILIKE '%migration%'
   OR name ILIKE '%alter%'
ORDER BY code;

If a relevant tool exists and appears intended for schema mutation → STOP and report for GPT/User decision unless file 15 explicitly authorizes it.

P4 — Report must include rollback status and exact mutation list

Report must include:

  • exact DDL run;
  • exact UPDATE counts for Tier A and Tier B;
  • whether function changed and validation result;
  • whether transaction committed or rolled back;
  • rollback instructions actually available from captured function source;
  • final state summary.

Non-blocking note

The prompt currently says Directus field auto-detect may be checked immediately after DDL. Depending runtime behavior, Directus metadata may not auto-populate immediately. This is acceptable if the prompt treats absence as STOP/report, not as a reason to manually insert metadata.

Directive to Opus/Ocus

Patch the draft prompt with P1–P4 only. Do not change the design scope. Do not dispatch before GPT/User sees final prompt.

After patched prompt is approved, it can be dispatched to Claude Code Agent for runtime execution.

Hard boundaries remain:

  • no H11 KB patch;
  • no law/docs patch;
  • no Pack 2B;
  • no IU rows;
  • no entity_enrichment deployment;
  • no manual directus_fields insert;
  • hard stop after report.
Back to Knowledge Hub knowledge/dev/laws/dieu44-trien-khai/reviews/gpt-review-runtime-execution-prompt-description-policy-option1-2026-05-04.md