--- id: SECRETS-WP-0008 type: workplan title: "Evolve the Lifecycle engine to the accepted security layer model" domain: infotech repo: secrets-engine status: active flavor: implementation owner: grok topic_slug: custodian created: "2026-08-29" updated: "2026-09-21" state_hub_workstream_id: "9c9e5164-b2f5-5ea2-a557-5368d65e9fe0" --- # SECRETS-WP-0008 - Evolve the Lifecycle engine to the accepted security layer model ## Goal Close the gaps between the 2026-08-29 layer declaration (Engine / Lifecycle) and the obligations the accepted NetKingdom Security Layer Model v0.7 places on a PEP-shaped Lifecycle engine. Source review: `history/2026-08-29-layer-model-scope-intent-review.md`. Declaration surface already landed this session: `INTENT.md` frontmatter, `layer.yaml`, `pep-stance.yaml`, and `tests/test_layer_conformance.py`. This workplan does **not** restart CLI hardening owned by `SECRETS-WP-0007`. It adds the layer-model remainder: stance binding, decision-record consumption, `audit-core` evidence, and the two proposed engine surfaces. ## Non-goals - Taking ops-warden's SSH access lane or becoming a credential broker. - Rendering or caching authorization decisions. - Observation in production or automatic containment (companion §10: both are at zero estate-wide). - Service mode, runtime tenancy, or a second custody backend. - Reopening live production on State Hub status or local fixtures. ## Coordination - `SECRETS-WP-0007-T04` owns the consumer validator and the fail-closed production gate until a durable ActionAuthorization object is served. - `SECRETS-WP-0007-T05` owns delivery-session self-revocation and the KeyCape scaffold; this workplan owns the layer-model "no standing credential" bar. - railiance-platform owns production OpenBao and the JWT role. - `access-engine` / State Hub own serving the decision record. - `audit-core` owns evidence custody. - ops-warden owns the SSH Staff PEP and must assent any SSH-CA API contract. - kings-guard owns secret-abuse posture and must assent the evidence surface. - gate-house owns statute §13.1 inventory of PEP stance maps. ## Tasks ## Publish stance application as shipped behaviour ```task id: SECRETS-WP-0008-T01 status: done priority: high state_hub_task_id: "945735c6-2d3d-5f72-baf0-692e6b210889" ``` Completed 2026-08-29. `_require_lane_approval` loads `pep-stance.yaml`. `SHIPPED_STANCE` is pinned equal to the published map. Fail-closed production evidence records `stance_stage`, `stance_failure_mode`, and omits `stance_decision_id` unless a decision was rendered. `pep-stance.yaml` is published and tests pin `prod: fail_closed` to `_require_lane_approval`. The gate is still hardcoded, and a refusal is recorded as a `DecisionError` rather than named stance fields. Load the published map as the source of shipped defaults. When the unreachable-engine residue is applied, record at least stage, failure mode, and decision id present only where a decision was rendered. Keep the three-factor unsafe-demo exception out of the stance table. Assert map equality in `tests/test_layer_conformance.py` the way ops-warden's stance test does. Acceptance: - `_require_lane_approval` (or its successor) reads `pep-stance.yaml`. - Changing the YAML without changing the code fails the test, and the reverse. - Fail-closed production evidence includes named stance fields, never a secret. - Build/test remain fail-open *relative to access-engine* and still require the existing lane-approval check; that residue stays documented as a gap until T02. ## Consume access-engine decision records ```task id: SECRETS-WP-0008-T02 status: wait priority: high state_hub_task_id: "3eb9cff8-1441-5437-9e92-a2b655c82d04" ``` Progress 2026-09-02. Gate House notice `632bdad9` (`GH-DEC-2026-003`): this engine is the PEP for FLEX-WP-0017-T05 / OpenBao writes. The shared consume function now lives in `src/secrets_engine/approval_consume.py` and every live privileged production handler reaches it through `_require_lane_approval` before `OpenBaoClient.resolve`. Conflict, unavailability, or a missing binding fail closed with no OpenBao call. Production still also fail-closes on the unpublished durable ActionAuthorization / consume-binding serving path, so this task remains `wait`. Correction 2026-09-06. See `SECRETS-WP-0007-T04` for the full note. Summary: the consume-before-OpenBao gate was real, but the PIP claim + validate join was a `return None` stub and `validate_action_authorization` had no production caller, so this task's "blocked on the durable serving path" framing hid an unimplemented local seam. The join is now implemented and tested; the remaining blocker is genuinely external (approval-engine claim endpoint and access-engine Check, both unreachable as of 2026-09-06). Progress 2026-09-07. **Step 2 is no longer blocked and is proved against a real decision.** The 2026-09-06 note above says "approval-engine claim endpoint and access-engine Check, both unreachable"; the second half of that is now out of date and is corrected here rather than left to read as current. `access-engine` Check is reachable through the owner-documented access path (loopback `kubectl port-forward` to the named pod plus a bounded `TokenRequest` token; `docs/pdp-access-path.md`). A real CheckRequest for `glas-claude-agent-dev-anthropic` returned `decision:0f9c98f14545c42d`, `allow`, `catalog_lane_policy_matched`, under package `secrets-engine.catalog-lane.lifecycle` v2 with `binding.tenant: tenant:platform`. Two defects surfaced only because the request was real, both now fixed: - Our CheckRequest carried no `tenant`, which v2 treats as a `wrong_tenant` denial rather than an omission (`80eafaf`). - `validate_decision_envelope` required byte-equality between the binding and our unenriched request, which no real decision can satisfy because the evaluator normalizes and enriches from its registry before hashing. Corrected to the published consumer rule in `10baad9`. So the acceptance line "wrong digest, expired lifetime, superseded status, or missing record fail closed before OpenBao" is now exercised against a genuine envelope: the unrefreshed live decision is refused on lifetime, and a restated `resource.attributes.stage`, a foreign `subject.tenant`, or an unexpected enrichment field are each refused. **What remains is step 1 only.** approval-engine must serve the claim endpoint; `APPROVAL-WP-0002-T03` is still `wait` with no deployed base URL, so `resolve_consume_binding` returns `None` and production stays fail-closed. That is one external dependency rather than the two this task previously named. Also recorded, and not a blocker for this task: flex-auth's response channel is unauthenticated (`FLEX-DEC-2026-010`). Fail-closed protects against a PDP that is absent, not one that lies. Until `FLEX-WP-0024` ships detached signatures, the enforced loopback address shape is what stands in for responder authenticity. Blocked on the durable ActionAuthorization serving path owned with `SECRETS-WP-0007-T04` / State Hub / `access-engine`. Once that object is served, a protected side effect holds a decision record identifying the request it was rendered for. Replay is permitted iff the canonical request digest matches and the lifetime holds. Do not cache the verdict. Do not treat State Hub status, a local fixture, or catalog admission as the decision. Acceptance: - Every live privileged production handler passes one shared consume function. - Wrong digest, expired lifetime, superseded status, or missing record fail closed before OpenBao. - Evidence carries the decision id; a fail-closed residue still records the stance instead. - Local fixtures cannot authorize steady-state production. ### 2026-09-09 published replay-identity contract received Owner message `83db7e47-ba93-4d3e-ac15-2bcb2bb16a34` supplies FLEX-DEC-2026-012, flex-auth source `0bc624b`, and updated `docs/request-enrichment.md`, `docs/canonical-request-digest.md` and replay examples. This supersedes the earlier "only external dependency" reading. T02 is progress: the next work is an executable local consumer correction, with no new owner interpretation or approval needed. Implement `binding.submitted_request_digest` comparison against the exact submitted CheckRequest. Stop treating enriched `request_digest` or the consumer's chosen list of benign enrichment fields as the replay identity. Preserve policy pin, lifetime and deny checks. Refresh the owner's replay fixtures without changing their existing request_digest/approval_binding_digest pins. For the approval join, carry the validated approval claim to the PDP and compare its recorded pdp_digest to the evaluator's approval_binding_digest. Do not recompute either side from unenriched material. Prove changed request, wrong claim/digest, absent new binding field and registry-overridden caller attributes are handled under the published contract before any live native lane activates. The separate approval identity exchange is implemented in SECRETS-WP-0009-T03 and passes 350 tests plus a real KeyCape/Approval Engine component exercise. That exercise uses a PDP double and does not prove this newer replay contract. The current source has not yet implemented the changes in this return; live PDP rollout/adoption evidence remains explicit too. No synthetic fixture or Hub decision substitutes for a runtime allow and consume. ### 2026-09-09 replay consumer implementation and real component acceptance The returned correction is implemented. The PEP now checks the exact submitted request digest including the full carried approval, permits registry-owned attribute replacement, joins the evaluator-origin approval digests and rechecks claim freshness after Check. The consume CAS uses the validated evaluator's request_digest and decision id; an observation alone is not a consume binding. 334 regression tests pass. The disposable KeyCape + actual Approval Engine + compiled Flex Auth exercise passes 13 checks, including real dual-control denial then claim/check/consume, registry override, changed submitted request, wrong action, consumed claim and same/different CAS digest. Independent producer inputs replace response-reconstructed replay fixtures; existing digest pins remain unchanged. Receipt: docs/evidence/2026-09-09-replay-contract-exercise.json. No OpenBao side effect, live credential read or model request occurred. T02 is wait for the live return: deployed PDP contract adoption and the admitted Approval Engine claim/consume path (APPROVAL-WP-0002-T03 and RPF-WP-0035-T06 / AUDIT-WP-0009-T09). SECRETS-WP-0009-T03 owns native delivery acceptance. The old live decision lacks submitted_request_digest and is explicitly refused; its historical evidence was not rewritten. Local conformance does not close the remaining live task. This supersedes earlier additive-only and claim-free transitive correspondence descriptions above. ## Emit evidence under the §9.6 bound ```task id: SECRETS-WP-0008-T03 status: done priority: high state_hub_task_id: "0b54cedf-9335-5fa0-bdcb-1efa9a9cd98d" ``` Completed 2026-08-29. `evidence-classification.yaml` is the declaration. Load-bearing records are queued in `evidence_dir/outbox/` before JSONL. `secrets-engine evidence heartbeat` emits a positive claim. Drain never sits on a mutation path and an audit-core outage leaves the queue in place. Classify each privileged action as load-bearing or attributive. Today's trail is attributive local JSONL plus best-effort State Hub notes; completeness is already not claimed and must stay that way until the destination is `audit-core`. Load-bearing classes (at least live destroy, production revoke/suspend of a control, and dual-control provision once those are enabled) require atomic emission via a **local** outbox and a declared cadence or heartbeat. Do not place a synchronous `audit-core` dependency inside the state-change transaction. Attributive classes may keep the current non-blocking trade if that trade stays declared. Acceptance: - A machine-readable classification exists next to the declaration. - No control in this repository branches on presence or absence of a local evidence record. - Load-bearing emission is queued locally; an `audit-core` outage does not block a revocation. - Documentation uses the sound claim: the archive proves records it holds were not altered or truncated after arrival. ## Accept the SSH-CA signing engine surface ```task id: SECRETS-WP-0008-T04 status: done priority: medium state_hub_task_id: "581aeee3-3d99-5034-992b-7d42e6542a19" ``` Completed 2026-08-29. Contract published at `docs/ssh-ca-signing-contract.md`. No OpenBao SSH-CA write was added. `layer.yaml` owner_status remains proposed. ops-warden's `ssh-ca-signing-write` gap names this engine as intended owner. Accept the **engine API**, decline the Staff lane. Publish a contract ops-warden can call after an `access-engine` decision: sign a short-lived certificate for an `adm`/`agt`/`atm` actor under the caller's request, present no additional authority, reconstruct as the caller's action. Do not implement a replacement client until ops-warden assents the contract. Do not unpark the standalone `warden-sign` AppRole as break-glass. Acceptance: - A non-secret contract document in `docs/` states request, decision binding, lifetime, evidence, and non-responsibilities. - ops-warden has a recorded chance to assent or contest. - No OpenBao SSH-CA write is added to this CLI ahead of that assent. - `layer.yaml` owner_status remains `proposed` until the surface ships. ## Accept the secret-use evidence surface ```task id: SECRETS-WP-0008-T05 status: done priority: medium state_hub_task_id: "3100d28c-9494-5b21-9d5a-af583cd8aa55" ``` Completed 2026-08-29. Contract published at `docs/secret-use-evidence-contract.md`. `layer.yaml` owner_status remains proposed. `route`/`audit` are still not this surface. **2026-09-02:** kings-guard assented in its own voice (message `75ebd2cc-a166-4676-94aa-deef2791c0c9`). It will not consume the surface as an observation input until the surface ships and publishes its cadence declaration. No Tooling contact will be opened to fill that wait. **2026-09-02 later:** shipped `secrets-engine secret-use snapshot`. Cadence is the declared 1d heartbeat on the snapshot envelope. `owner_status` stays `proposed` until kings-guard admits the snapshot as an observation input. `route`/`audit` remain distinct operator summaries. kings-guard's secret-abuse posture is fixture-driven because no engine exposes lease, revocation, mount, rotation, and delivery-session metadata. `route` and `audit` are operator summaries over local JSONL and are not that surface. Publish a read-only engine contract of non-secret metadata, bounded by §9.6: it cannot prove omission at source. Coordinate destination with `audit-core`. Do not give kings-guard a vault client, and do not return values. Acceptance: - Contract lists fields, freshness, and the evidence bound. - kings-guard has a recorded chance to assent or contest. - No secret material, token, or accessor is in the surface. - `layer.yaml` owner_status remains `proposed` until the surface ships. ## No standing engine credential ```task id: SECRETS-WP-0008-T06 status: wait priority: medium state_hub_task_id: "d7bc8bdc-a0f8-5058-a640-374ef9859148" ``` Progress 2026-09-02. Named engine auth is shipped: `service-jwt`, `bootstrap`, and `env` are exclusive providers. `--auth service-jwt` exchanges the KeyCape service JWT and logs in through `OpenBaoClient.login_jwt` (JSON file, never argv), then self-revokes. A JWT selection refuses `--bootstrap-token-file` and does not read `BAO_TOKEN` on failure. `--bootstrap-token-file` is a named break-glass provider with `auth_break_glass` evidence. The platform-owned OpenBao JWT mount/role is still unpublished, so auto keeps bootstrap/env and this task remains `wait`. Companion §7 / statute §3.4: an agent holds no long-lived credential of its own. Authority is per task, time-bounded, and attributable to the principal it acts for. Bootstrap token files remain break-glass, heavily evidenced, never an implicit fallback from service identity. Acceptance: - Steady-state CLI/API auth is the reviewed service identity, not `--bootstrap-token-file`. - Bootstrap is a named break-glass provider with distinct evidence. - No fallback from JWT/service-auth failure into bootstrap or AppRole. - Issued delivery sessions continue to self-revoke as in WP-0007-T05. ### 2026-09-21 GH-DEC-2026-017 — section 11 layer declaration ruled Read from `gate-house/decisions/decisions.md`, not from message text. `secrets-engine` is a named affected repository and carries both permitted forms, so all four parts apply here. State Hub decision id `82dd0fea-fabb-4940-92d5-d6e2ea9985f5`. What the ruling settles for this repository: 1. **`INTENT.md` governs.** Our declaration is the `layer: Engine` key in `INTENT.md` frontmatter. `layer.yaml` is a **derived artifact**: it must be marked as derived, must name `INTENT.md` as what it derives from, and must agree with it. Our `layer.yaml` already carries `declared_by: INTENT.md` and a `Voice: INTENT.md (this repository's own declaration, §11)` header comment, so we are among the repositories already treating the sidecar as derived in substance. What is missing is the explicit derived marking the ruling now requires as a field rather than a comment. 2. **The §3 vocabulary is case-insensitive and comparison must ASCII case-fold.** Our `INTENT.md` says `Engine` and our `layer.yaml` says `engine`. Under §2 of the ruling that is **conforming**, it was never a disagreement about a layer, and **nothing here is re-spelled**. flex-auth withdrew the casing finding itself (message `962bf395`) and its validator is the thing that changes. We do not touch either file's spelling, and we do not "align" them. 3. **The vocabulary is closed with four tokens** — `Taxonomy`, `Tooling`, `Engine`, `Staff`. `Engine` is in it; no action. 4. **`standard_version` comes out of the declaration.** Our `layer.yaml` carries `standard_version: "0.7"`. A declaration asserts a standing property and must not carry a standard version; version-scoped state belongs in the derived conformance record, which must name what it derives from and the version or commit it was derived at. **Deliberately not done in this session.** The ruling states that `standard_version` is not one repository's habit but **ops-warden's reference form**, which this repository copied along with seven others, and that ops-warden is asked to make the change there rather than each adopter deciding independently. ops-warden is updating that reference form in a parallel session. Inventing our own sidecar shape ahead of it would produce exactly the divergence the ruling exists to close, so we record the requirement and wait for the reference form. What `secrets-engine` will need, once the reference form lands: - Adopt the reference form's derived-artifact marking in `layer.yaml`, keeping `declared_by: INTENT.md` or whatever field the reference form names for it. - Remove `standard_version: "0.7"` from `layer.yaml`. - Move the version the conformance run was derived at into our derived conformance record, per the reference form's field for it. Our conformance checker is `scripts/check_layer_conformance.py`; it reads `layer.yaml` today and will need to read `INTENT.md` as the governing form, read the sidecar as well, and **report a disagreement between them as a finding** rather than resolve it by precedence. A casing-only difference is not such a disagreement. - Keep `framework:`, `layer:`, `role:` and the substantive §3/§5/§6/§9.6 content unchanged. `role:` is explicitly not ruled and stays as it is. - The `Framework:` header comment in `layer.yaml` cites `security-layer-model_v0.7.md` while the ruling is issued against `v0.8`. That is a citation in a comment, not a declaration version, so it is out of scope for part 4 — but it should be refreshed in the same commit so the two are not confused for each other. No task is opened here by hand. This is a note against the workplan that already owns `layer.yaml` / `pep-stance.yaml` / INTENT frontmatter staying in one voice; the work lands under that exit criterion when ops-warden's form is published. ### 2026-09-21 GH-DEC-2026-017 applied — reference form published, note closed ops-warden published the reference form (commit `a70f559`, `wiki/playbooks/netkingdom-layer-declaration.md` "Reference-form change set"). Re-read against the ruling in `gate-house/decisions/decisions.md` and amendments A9–A13 before editing; the ruling and the playbook agree. Applied in one commit, field and checker together: - `layer.yaml`: `standard_version: "0.7"` removed; `derived: true` and `derived_from: INTENT.md` added (`declared_by: INTENT.md` kept, as the note above required). Header comment marks the file derived and drops the version from the framework citation. - `INTENT.md` frontmatter: never carried `standard_version`, but its `standard:` value was a version-pinned path (`..._v0.7.md`). De-versioned to `net-kingdom/canon/standards/security-layer-model`, as the reference instance did, because a pinned path in the governing form is a standard version by another name. The field is kept; only the pin is gone. - **No layer value re-spelled.** INTENT says `Engine`, layer.yaml says `engine`; both unchanged. - `scripts/check_layer_conformance.py`: `standard_version` was a required key, so removing the field alone would have exited 2 MALFORMED. It now reads INTENT.md as the governing form, requires the derived marking, rejects a returning `standard_version` in either form, validates both layer values against the closed four-token vocabulary (Taxonomy included) after an ASCII fold, and reports a post-fold disagreement between the forms as a finding. - `tests/test_layer_conformance.py`: per-file equality on the layer spelling replaced by fold assertions; new tests for fold agreement, a real disagreement (finding, exit 1), the closed vocabulary (`surface` exit 2, `Taxonomy` admitted), and a returning `standard_version` (exit 2). Full suite: 430 passed. `role:` untouched (not ruled). `pep-stance.yaml` untouched (a stance map, not a declaration). `schema_version` left at `0.1` — ops-warden bumped its own, but the ruling does not call for it. **Still open, not ours to answer:** where the removed version lives. A12 says the derived conformance record already MUST carry it; our checker, like ops-warden's, is re-runnable and emits nothing durable. ops-warden asked gate-house (message `4220413a`); we follow that answer rather than choose. ## Exit criteria - The published PEP stance is the shipped default and stance application is recorded by name. - Production live side effects require an access-engine decision record, or remain fail-closed with a recorded stance. - Evidence classification and destination match §9.6; completeness is not claimed. - SSH-CA and secret-use surfaces are either contracted with the declaring repositories or explicitly declined in this repository's own voice. - Bootstrap is not the steady-state engine credential. - `layer.yaml` / `pep-stance.yaml` / INTENT frontmatter stay in one voice. - No raw secret values in Git, State Hub, chat, prompts, workplans, evidence, or argv.