GPT Review — 18b-P2 DOT-119 Script Rewrite Prompt rev4

Date: 2026-05-05 Reviewer: GPT-5.5 Thinking / Incomex Hội đồng AI Reviewed:

• knowledge/dev/laws/dieu44-trien-khai/design/dot119-v2-script-content.sh rev3
• knowledge/dev/laws/dieu44-trien-khai/prompts/18b-p2-dot119-script-rewrite-prompt.md rev5 / Rev 4 content

Verdict

Near PASS. Small rev5 patch required before dispatch.

Rev4 fixes the earlier hardcode and prompt completeness issues:

• prompt is now self-contained;
• script no longer hardcodes KNOWN_V2_FN_HASH;
• synthetic fallback uses capability check;
• run_pg uses ON_ERROR_STOP=1;
• --help exits 0 and no-args exits 1;
• legacy function existence is checked;
• no-clobber grep is exact;
• md5/reporting requirements are present;
• hard boundaries remain clear.

However, two small execution risks remain.

Required rev5 patches

B1 — Smoke tests expecting non-zero exits must not abort the shell

The prompt currently uses commands like:

"$DOT119" --collection=$SAMPLE_UNCLASSIFIED --dry-run 2>&1; echo "exit=$?"

If Claude Code runs the smoke block inside a shell with set -e, expected non-zero exits (10, 14, 30) can abort before echo runs.

Patch smoke tests to use a safe wrapper, for example:

run_expect_exit() {
  local expected="$1"; shift
  set +e
  "$@"
  local rc=$?
  set -e
  echo "exit=$rc expected=$expected"
  [ "$rc" -eq "$expected" ]
}

Or explicitly use:

set +e
"$DOT119" ...
RC=$?
set -e
echo "exit=$RC"

Apply this to tests expecting exit 1, 10, 14, 30.

B2 — Capability check should produce auditable PASS output

The script performs the capability check internally for synthetic_id, but the dry-run JSON does not explicitly say capability check passed.

Add one of the following:

• in script dry-run JSON/text for synthetic_id, include synthetic_capability":"pass"; or
• in prompt add a direct external capability-check query/test that proves fn_birth_registry_auto has the needed markers.

Preferred: add CAPABILITY_STATUS="pass" after successful check and include it in dry-run JSON/text.

This makes the report requirement “Capability check output” concrete.

Optional hardening, not blocker

N1 — JSON escaping for proposed_sql

The JSON output embeds SQL with single quotes only, so it is probably OK. But if future SQL includes double quotes, bash-built JSON can break. Not blocking for current known inputs.

N2 — Capability check remains heuristic

The check is capability-oriented and not a fixed hash. Good. It is still heuristic, but acceptable for P2 because P2 does not create IU trigger; 18c will test actual synthetic dry-run/execute more directly.

Directive to Opus/Ocus

Patch both files minimally:

1. knowledge/dev/laws/dieu44-trien-khai/prompts/18b-p2-dot119-script-rewrite-prompt.md

   • add safe exit-capture wrapper for smoke tests;
   • update smoke table/examples to use it for expected non-zero exits.

2. knowledge/dev/laws/dieu44-trien-khai/design/dot119-v2-script-content.sh

   • expose capability pass in dry-run output for synthetic_id, or otherwise make capability evidence explicit.

Do not expand scope.

After rev5, P2 can be approved for dispatch if no new blocker appears.

Hard boundaries remain:

• no fn_birth_registry_auto edit;
• no IU trigger install;
• no IU rows;
• no Pack 2B;
• no seed/DDL changes.