KB-49AA

IU Core Session Prompt Hygiene Audit 2026-05-21

6 min read Revision 1
reviewiu-coreprompt-hygieneno-hardcodedotopen-goal2026-05-21

IU Core Session Prompt Hygiene Audit

Date: 2026-05-21 Scope: Recent IU Core orchestration prompts and reports in the current session.

Verdict

No immediate catastrophic violation was found in the recent executed macros. The system progressed safely through production DDL, inert runtime install, controlled dry-run route activation, and event worker substrate installation with backups, gates, reports, tests, and no forbidden delivery.

However, there are hygiene issues that should be corrected before the next production-facing macro.

Confirmed good points

  • Production mutations were gated with backup, target identity, drift checks, and verification.
  • Forbidden actions were generally respected: no dry_run=false, no downstream workflow delivery, no deploy/restart, no push/merge/tag.
  • Worker remained inert/dry-run safe.
  • H3 hardcoded event_stream/payload_classification defect was discovered and fixed.
  • Evidence-first reporting was strong: object counts, backup hashes, test counts, KB reports, commit hashes.
  • Open-goal style improved after the user's correction.

Hygiene concerns / sạn

HYG-1: Some prompts were still too operational

Several prompts still included long tactical lists. They were effective, but they partially shifted responsibility from Agent outcome to GPT-issued checklist.

Correction: future prompts should lock outcome, gates, forbidden boundary, and evidence; leave implementation path to Agent except for dangerous production gates.

HYG-2: Literal route/config/table names appear frequently

Examples include route code iu.version_applied.workflow, config key iu_core.routes_master_enabled, and specific SQL object names.

This is acceptable only when they are discovered/declared SSOT objects. It becomes a violation if treated as hardcoded behavior instead of registry/config/catalog-backed references.

Correction: next macro must explicitly require Agent to classify every literal as one of:

  • DISCOVERED_SSOT
  • CONFIG_DEFAULT
  • REGISTRY_VALUE
  • NEEDS_DOT_REGISTRATION
  • TEMP_SANDBOX_ONLY
  • VIOLATION_TO_FIX

HYG-3: DOT visibility is not yet proven complete

IU Core tables/functions/views/routes/worker objects exist, but the recent reports do not conclusively prove complete DOT scan/registry coverage across all new production-visible objects.

Correction: next macro must include DOT conformance scan/check/registration package before any activation beyond dry-run.

HYG-4: Two prompt KB docs are metadata stubs, not full prompt bodies

Recent prompt docs:

  • knowledge/dev/laws/dieu44-trien-khai/prompts/iu-core-controlled-route-mastergate-structure-observability-open-goal-2026-05-21.md
  • knowledge/dev/laws/dieu44-trien-khai/prompts/iu-core-worker-live-dryrun-structure-autocut-delivery-prep-open-goal-2026-05-21.md

They contain pointers/metadata, not full prompt text. This is not production risk, but it is KB hygiene debt.

Correction: either update them with the final prompt body or mark them explicitly as pointer-only and do not treat them as prompt SSOT.

HYG-5: Five-layer sync impact not yet consistently required

Recent prompts referenced DOT and observability but did not consistently require PG → Directus → Nuxt → AgentData/KB → Qdrant impact classification.

Correction: next macro must include a five-layer sync report for all new runtime/worker/route/structure-op objects.

HYG-6: Automation-first needs stronger closure

The worker is SQL-function based and not a persistent daemon, which is safe. But invocation/scheduling mechanism is still undecided. That is acceptable before dry_run=false, but not acceptable for production delivery.

Correction: next macro should require an automated invocation plan, disabled by default, with healthcheck and rollback, not a human-memory procedure.

Required guardrail for next macro

Add this to the next prompt:

CONSTITUTION/DOT HYGIENE REQUIREMENT
Before any further activation, audit all recent IU Core artifacts for:
- no hardcode: every literal route/config/table/event name must be classified as DISCOVERED_SSOT / CONFIG_DEFAULT / REGISTRY_VALUE / NEEDS_DOT_REGISTRATION / TEMP_SANDBOX_ONLY / VIOLATION_TO_FIX;
- DOT 100%: every new table/view/function/route/worker/config must have scan/registry/check visibility or a package to create it;
- five-layer sync: classify impact on PG, Directus, Nuxt, AgentData/KB, Qdrant;
- automation-first: no safety condition may depend on a human remembering a query or sequence;
- reversible: every mutation must have disable/rollback/cleanup;
- evidence-first: PASS requires exact object counts, healthchecks, tests, report paths, and commit hash.

Final recommendation

Do not run dry_run=false or a broader activation until a short but real DOT/no-hardcode/five-layer hygiene pass is included in the next large open-goal macro.

This hygiene pass should not be a small standalone task unless the user explicitly wants cleanup only. It should be folded into the next large macro together with worker live dry-run, structure ops, and auto-cut integration progress.

Back to Knowledge Hub knowledge/dev/laws/dieu44-trien-khai/reviews/iu-core-session-prompt-hygiene-audit-2026-05-21.md