---
id: hall-worker-codex-railiance-platform-resolver-gate
type: worker-entry
worker_kind: agent-session
display_name: Codex
session_id: "not exposed by the harness"
created_at: "2026-08-23T20:04:03.000Z"
recorded_at: "2026-08-23"
llm_family: "OpenAI GPT-5"
exact_model: "not exposed by the harness"
harness: "Codex API session"
token_count: "total=1,376,529 input=1,307,661 (+ 15,800,576 cached) output=68,868 (reasoning 23,012)"
status: handed-forward
repos:
  - railiance-platform
  - hall-of-helix
related:
  - RAILIANCE-WP-0029
  - KEYCAPE-EXPOSURE-20260823-01
---

# Codex — railiance-platform: The gate was the work

## What happened

This session coordinated a live KeyCape Secret-exposure recovery without
reproducing any secret value. The signing-key and downstream rotation receipts
arrived, but the privacyIDEA `lldap-coulomb` resolver remained unresolved.

The useful diagnosis was precise: the realm and `platform-root` identity were
correct, while the persisted resolver bind returned LDAP `invalidCredentials
(49)`. A resolver-only update returned `PASS`, but the combined proof then
failed at replacement LLDAP authentication. No further blind retry was allowed.

## What should be remembered

The four-prompt helper conflated repair with audit. NetKingdom corrected the
design: a minimal reconcile needs only the privacyIDEA admin credential and the
provider-approved replacement LLDAP credential; lookup, MFA, and predecessor
denial belong in a separate read-only audit flow.

The remaining blocker is custody, not cleverness. The routing lanes exist, but
they remain `resolvable: false` until Railiance/OpenBao publishes the canonical
mount/path, field, policy/auth, version, expiry/revocation, and attended-handoff
metadata. Those values must never be guessed or placed in chat.

## Durable handoff

- Railiance custody contract draft: `docs/net-kingdom-credential-custody-contract.md`
- Workplan gate: `RAILIANCE-WP-0029-T06`
- Latest platform commit: `51361fb`
- Safe next step: obtain the owner-approved OpenBao metadata receipt, then use
  the minimal reconcile flow and a separate `--check` proof.

Wind down with the system intentionally blocked. A clean stop is better than a
credential retry whose authority and source are still ambiguous.