net-kingdom/workplans/NK-WP-0056-bounded-executor-path.md
tegwick 6c7c37aa70
All checks were successful
CI Smoke / host-smoke (push) Successful in 0s
CI Smoke / container-smoke (push) Successful in 1s
NK-WP-0056-T04: svc-keycape bind lane (source only, CCR-2026-0054)
KeyCape takes KEYCAPE_LLDAP_BIND_PW (key-cape KEY-WP-0046). ESO creates the
dedicated Secret keycape-lldap-bind. bind-lane setup reads the current bindPW
from keycape-config and requires the svc-keycape bindDN; verify compares the
delivered Secret, and the config value while it is present.
directory-bind-rollout gains clear-password, which runs only with the service
bind and the delivered env (switch restores). deployment.yaml pins image
main-e2b5fa1 and adds the env. Nothing is applied.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 896183@bnt-lap001
Assistant-Session: 3a18c01a-3205-4d02-94db-9efff25649a2
2026-10-31 10:31:30 +01:00

14 KiB

id type title domain repo status flavor owner topic_slug created updated related state_hub_workstream_id
NK-WP-0056 workplan Bounded executor path: founder admits, executors run infotech net-kingdom proposed design claude netkingdom 2026-10-30 2026-10-30
NK-WP-0054
NK-WP-0040
NK-WP-0031
NK-WP-0055
5bc74ef8-136e-5157-9a7f-da765dbfc977

NK-WP-0056 — Bounded executor path

Intake from the founder's direction on 2026-10-30, before the live LLDAP bind split: top-level decisions and guardrails stay with the founder; execution below that is delegated to automation with limited expense budgets and risk control, and the security architecture is the means. Founder-attended runbooks are a transitional state.

SI-15 (canon/standards/service-identity-privilege_v0.1.md) already names the golden path prepare → admit → execute → verify → retain and allows "a bounded executor under the admission" at step 3. Today preparation and verification are strong, admission is prose (a hub decision) and execution falls to the founder, because no executor exists that has its own identity, its own secret delivery and an authority ceiling the founder can admit and bound. SI-14 adds that an interactive agent session holds everything its tools reach, so it cannot be that executor until a session contract exists (glas-harness GLAS-WP-0018, proposed).

Why the bind split needed the founder (2026-10-30 run, step by step)

Runbook step Needed the founder because Class Delegable once
Setup: decrypt custody copy custody is a personal SOPS/age copy delegable service credentials live in OpenBao (T04)
1 prepare custody (generate passwords) writes the personal custody copy delegable executor generates into OpenBao (T04)
2 plan, 4 verify, 10 verify sign in as the human admin delegable read-only executor directory identity (T03)
3 create svc-* accounts needs lldap_admin (LLDAP has no narrower grant that can create users) root-class admitted per run; executor runs it (T03)
5-8 switch one consumer cluster-admin kubeconfig; secret values; founder sign-in check delegable scoped service account (T03), OpenBao/ESO (T04), synthetic MFA journey (T05)
9 change LLDAP's configured super-admin changes who administers the directory root-class admitted per run; executor runs it
10 change admin's own password; seal ld-admin in the password safe the founder's own credential and crisis custody human-intrinsic never: stays with the founder
Afterwards: restore drill on the next generation admin operates drills delegable executor identity with the drill's read scope

Root-class means the founder still admits each run, but no longer types it. The run also showed why personal custody does not scale: the custody copy had drifted from live (AUTHELIA_KEYCAPE_CLIENT_SECRET) and step 6 had to be done by a one-off path.

Admission contract

id: NK-WP-0056-T01
status: todo
priority: high
assignee: claude-code
state_hub_task_id: "99df0ee2-2c17-55f2-9f3e-4277da52e01c"

Draft a NetKingdom standard for a bounded-execution admission: the founder approves, at aal2, an exact plan by digest with a validity window, a scope ceiling (MODE @ SCOPE grants, SI terms), a blast-radius rule (for example one consumer at a time), a stop-and-revert policy tied to named machine checks, a risk class, and an expense budget (wall time, attempts, compute/egress cost). The executor consumes the admission once. Map it to SI-15, the Playbook Capability Contract's trust-state requirements and the execution-attribution receipt (NK-WP-0040). Include a validator and synthetic fixtures.

Done when: the standard is proposed with a schema, a validator, and the bind-split steps expressed as admissions in its examples.

Owner agreement on the components

id: NK-WP-0056-T02
status: todo
priority: high
assignee: claude-code
needs_human: true
intervention_note: "Founder chooses the executor runtime owner; owner answers are agreements, not agent decisions."
depends_on: [NK-WP-0056-T01]
state_hub_task_id: "7263417b-b227-54f7-9edc-3e7f97cf132f"

Ask the owners which part each carries: approval-engine (the admission as its consumable approval object, plan digest and budget as bound fields, single consume), access-engine (reads the admission as an input claim), secrets-engine and railiance-platform (executor credential delivery through OpenBao), glas-harness (GLAS-WP-0018 session contract, so a harness-launched agent can be an executor), audit-core (receipt holder, as in NK-WP-0040). The founder decides who builds the executor runtime itself.

Done when: each component has a named owner and an answer, and the founder has chosen the executor runtime owner.

Executor identity and authority ceiling

id: NK-WP-0056-T03
status: todo
priority: high
assignee: claude-code
depends_on: [NK-WP-0056-T02]
state_hub_task_id: "592231dd-1981-57a6-a0e8-5d6fc474c6e7"

Declare the executor as a WORKLOAD (or harness-launched AGENT) principal in declarations/ with grants no wider than the pilot: a Kubernetes service account whose Role names the sso Secrets, ConfigMaps and Deployments it may patch or restart; a read-only directory bind for plan and verify; lldap_admin only for root-class steps under an admission. Extend tools/service-identity to check it.

Done when: the declaration validates and the Role is prepared in source with tests, not applied.

Service credentials into OpenBao

id: NK-WP-0056-T04
status: todo
priority: high
assignee: claude-code
needs_human: true
intervention_note: "Founder runs the custody reconciliation and approves moving service binds out of personal custody."
depends_on: [NK-WP-0056-T02]
state_hub_task_id: "fd439c0c-9b5e-562c-9bfb-48dd0fb2c570"

With railiance-platform, move the LLDAP service binds (LLDAP_SVC_*_PASS) and the Authelia secrets from the personal custody copy into OpenBao paths delivered by ESO, so that neither a person nor an agent handles the values in a rotation. First reconcile the stale custody AUTHELIA_KEYCAPE_CLIENT_SECRET with the live value (founder-run, value-free comparison as in the 2026-10-30 evidence). Root-class material (ld-admin, unseal shares) stays in founder custody.

Done when: the four service binds are ESO-delivered from OpenBao, and the custody copy no longer needs to be decrypted for a bind rotation.

2026-10-30: reconciliation tool prepared: sso-mfa/bootstrap/reconcile-keycape-client-secret.py (value-free; accepts the live KeyCape value only if it matches Authelia's live hash, writes custody through rotation-env.py, no cluster writes; tests/test_reconcile_keycape_client_secret.py). Founder ran it 2026-10-30: live KeyCape value matches Authelia's hash; custody did not match before, was updated and matches after (value-free output); custody re-sealed and shredded. The reconciliation step is done; the ESO move remains.

2026-10-30, pilot svc-provisioner prepared (source only, nothing applied): railiance-platform CCR-2026-0052 (proposed) and policy net-kingdom-svc-provisioner-bind-eso; sso-mfa/k8s/identity-provisioner/bind-eso.yaml (ServiceAccount, SecretStore, ExternalSecret with creationPolicy: Merge into the existing Secret, password only); tools/service-binds/bind-lane.py setup|verify (approved sha256-bound CCR, attended platform-admin, custody must equal live, coding-agent fence, KV CAS=0, value-free receipts); tests/test_service_bind_lane.py (7). Next: owner approval of the CCR, then the founder-run setup, ESO apply and verify. Authelia, KeyCape and privacyIDEA follow one at a time.

2026-10-30, pilot live and verified: CCR-2026-0052 approved; founder-run setup, ESO apply and verify passed; svc-provisioner now binds with the ESO-delivered value (evidence). Finding: ESO targets need template.metadata: {} or the admission guard refuses the Secret update. Remaining: svc-authelia, svc-keycape, svc-privacyidea.

2026-10-31, svc-authelia lane prepared (source only, nothing applied): railiance-platform CCR-2026-0053 (proposed) and policy net-kingdom-svc-authelia-bind-eso; sso-mfa/k8s/authelia/bind-eso.yaml (Merge of ldap_password only into authelia-secrets, which also holds the session, storage and OIDC keys). bind-lane.py gains per-lane Secret keys: Authelia keeps the bind user in authelia-config, so setup refuses unless the live configuration binds uid=svc-authelia, and it decodes only the bind key. Authelia reads the file at start-up only, so the identical projection needs no restart; a later rotation does. 12 lane tests. Next: owner approval and binding confirmation, then the founder-run setup, ESO apply, verify and an Authelia restart with a login.

2026-10-31, svc-authelia live and verified: CCR-2026-0053 approved and binding confirmed; founder-run setup, ESO apply (Ready first time), verify and Authelia restart passed, and an admin portal sign-in with OTP worked (evidence). Two of four binds are ESO-delivered. Remaining: svc-keycape (bind inside sso/keycape-config), svc-privacyidea (resolver password stored in the privacyIDEA database, not a Secret key).

2026-10-31, svc-keycape design proposal (for founder decision): KeyCape's server reads lldap.bindPW only from config.yaml inside sso/keycape-config (internal/config/config.go); unlike client secrets (secretRef: env:... via secretKeyRef) it has no environment indirection. ESO Merge cannot own one field inside a YAML document, and rendering all of config.yaml from OpenBao would move every KeyCape value at once. Proposed:

  1. key-cape (owner change, new KEY-WP): config.Load takes KEYCAPE_LLDAP_BIND_PW (the name lldap-export already uses). Env set and bindPW empty: use env. Both set and equal: use it and log that bindPW is redundant. Both set and different: refuse to start. This allows a staged switch with a clean rollback. Image through the owner pipeline; check-config reports which source is used, without the value.
  2. NetKingdom: CCR-2026-0054 and a svc-keycape lane; ESO creates a new dedicated Secret keycape-lldap-bind (creationPolicy: Owner, key password), so no Secret has two writers. Setup compares custody with the live bindPW in keycape-config.
  3. Deployment: add the env from secretKeyRef keycape-lldap-bind/password with the new image; values are equal, so behaviour is unchanged. Verify, restart, admin login.
  4. Remove bindPW from keycape-config (directory-bind-rollout.py --operation clear-password, guarded on uid/resourceVersion); create-secrets.py stops writing it. Rollback at any step: re-add the value or drop

2026-10-31, founder decision: environment override, implemented by claude-code in key-cape. Done in source: key-cape e2b5fa1 (KEY-WP-0046-T01; image main-e2b5fa1 = sha256:9e082dca…6f8d, built by the Forgejo workflow; its acceptance job also fails on the previous commit a1ab8ff). NetKingdom: CCR-2026-0054 (proposed), sso-mfa/k8s/keycape/bind-eso.yaml (dedicated Secret keycape-lldap-bind, creationPolicy: Owner), bind-lane.py svc-keycape (setup reads bindPW from keycape-config and requires bindDN = svc-keycape; verify reads the dedicated Secret and requires it to equal bindPW while present), directory-bind-rollout.py --operation clear-password (only with the service bind and the delivered env; switch restores), and deployment.yaml pinned to the new image with the KEYCAPE_LLDAP_BIND_PW env. The new image also carries key-cape 67a0869..e2c5a94 (online token status, inactive while tokenStatus is unset) and identity-facts. Founder order: approve, setup, ESO apply, verify, deployment apply (image + env), login, clear-password, restart, login. the env.

Synthetic verification instead of a founder sign-in

id: NK-WP-0056-T05
status: todo
priority: medium
assignee: claude-code
depends_on: [NK-WP-0056-T04]
state_hub_task_id: "bb511d1e-6d5d-56d1-9030-2e0d94006a16"

A synthetic test user with an OTP seed held in OpenBao, and a journey that signs in through Authelia and KeyCape with MFA and resolves through privacyIDEA. Reuse the identity-provider journey acceptance. The executor runs it as the check between consumers and reverts on failure.

Done when: the journey runs unattended, and fails closed against a deliberately broken bind on a synthetic directory.

Pilot: unattended bind rotation under one admission

id: NK-WP-0056-T06
status: todo
priority: high
assignee: claude-code
needs_human: true
intervention_note: "Founder signs the admission for the single live pilot rotation."
depends_on: [NK-WP-0056-T01, NK-WP-0056-T03, NK-WP-0056-T04, NK-WP-0056-T05]
state_hub_task_id: "6c3939f8-23a9-5f70-a3b1-94f4b291706b"

Rotate svc-authelia's password with no person in the loop beyond the admission: generate into OpenBao, set in LLDAP, let ESO deliver, restart, run the synthetic journey, revert on failure, emit a receipt to audit-core. Rehearse on a synthetic directory first, then one live run under a founder-signed admission.

Done when: the live rotation completes or reverts on its own, within its budget, with a verified receipt, and the founder's only action was the admission.

Budget and risk enforcement

id: NK-WP-0056-T07
status: todo
priority: medium
assignee: claude-code
depends_on: [NK-WP-0056-T06]
state_hub_task_id: "a87c5edb-7604-5eba-9261-b1ed2ed16655"

Enforce the admission's budget and blast radius at run time (attempt and time limits, a kill switch the founder holds, refusal when a check is missing), and feed outcomes and drift into posture feedback (NK-WP-0031) so that recurring admissions can be reviewed in aggregate rather than one by one.

Done when: a run that exceeds its budget or scope stops and reverts in a synthetic test, and the pilot's receipts appear in a posture-feedback report.