GPT Review R4 — P9 G6 Execution Prompt v0.2

Date: 2026-04-27
Scope: Review Opus G6 execution prompt v0.2 after GPT R3 patches.

Verdict

PASS FOR PROMPT CONTENT. FINALIZATION BLOCK REQUIRED BEFORE EXECUTION GATE.

The prompt content applies the six GPT R3 patches correctly and is suitable for final execution-gate preparation. However, the KB currently contains only a summary of the full prompt and points to a session artifact/chat copy. For governed execution, the full executable prompt must be stored in Agent Data/KB as the durable source of truth before the User/GPT execution gate is opened.

Evidence checked

• knowledge/dev/laws/dieu38-trien-khai/P9-G6-execution-prompt.md rev 1 — KB summary of prompt v0.2.
• User-provided full prompt content in the current review message.
• knowledge/dev/laws/dieu38-trien-khai/P9-G6-dry-run-package.md v0.2.
• knowledge/dev/reports/gpt-review-g6-execution-prompt-r3-2026-04-27.md.
• knowledge/dev/laws/dieu38-trien-khai/index.md current S184 tracking.

R4 findings

1. R3 patch #1 PASS: PF-01 now STOPs if p9_g6_dryrun already exists; no auto-DROP.
2. R3 patch #2 PASS: PF-Pre resolves tools/env/Directus/manifest before main pre-flight.
3. R3 patch #3 PASS: V3 uses PL/pgSQL DO/EXCEPTION harness with SQLSTATE reporting.
4. R3 patch #4 PASS: Directus URL/token retrieval is env/GSM governed and fail-stop.
5. R3 patch #5 PASS: V4-05/V4-06 are schema-adaptive with column introspection and N/A evidence handling.
6. R3 patch #6 PASS: DDL bundle hash and grep tests are required on the actual generated bundle.

Remaining finalization items

These are not design rejections; they are final execution-readiness controls:

1. Persist full prompt in KB.
   Replace the current summary-only P9-G6-execution-prompt.md with the full v0.2 prompt body, or create P9-G6-execution-prompt-full-v0-2.md and make the summary link to it. Artifact/chat content is not sufficient as source-of-truth for execution.

2. Record package errata / precedence rule.
   The package v0.2 still contains older PF-01 wording that says to DROP the schema if it exists. The execution prompt v0.2 correctly supersedes that. Before execution, document precedence clearly: for execution, P9-G6-execution-prompt.md v0.2 overrides package v0.2 PF-01 behavior. Prefer adding a short errata section to the package or prompt: “PF-01 auto-DROP from package v0.2 is superseded by prompt v0.2: pre-existing schema = STOP.”

3. Generate seed manifest expected JSON.
   Produce and store seed-manifest-g6-expected.json in KB or a governed artifact path, with SHA-256 values for the eight seed SQL files and total_seed_rows = 61.

4. No execution yet.
   After these finalization items, Opus may present a single final execution authorization request. Do not run G6 until User/GPT explicitly opens the gate.

Law / constitutional check

Rule	Result	Finding
Hiến pháp / Zero Trust	PASS with finalization	Full executable prompt must be durable in KB; no reliance on transient artifact/chat.
Đ33 DB governance	PASS	Isolated schema DDL only, no production public.tac_*, no extension creation.
Đ35 DOT governance	PASS	No dot_tools/dot_action_log mutation; markdown report only.
Đ32 approval discipline	PASS	Execution still requires explicit gate.
Đ24	PASS	No taxonomy/entity label mutation.
S183 anti-patterns	PASS	Schema-qualified, env-based connection, no auto-drop, expected-error harness.

Direction

Opus should complete one bundled G6 finalization block:

1. Store the full execution prompt v0.2 in KB as durable source-of-truth.
2. Add PF-01 precedence/errata so there is no conflict between package v0.2 and prompt v0.2.
3. Generate and store the seed manifest expected JSON.
4. Update index/status to say G6 prompt content PASS, finalization complete/pending as appropriate.
5. Report back once with paths and hashes.

If these are complete, GPT can approve the final User execution-gate request without another broad review loop.