secrets-engine/SCOPE.md

365 lines
19 KiB
Markdown
Raw Permalink Normal View History

2026-06-28 09:03:37 +00:00
# SCOPE
> Implemented capability boundary for agents and contributors. Aspirational
> direction belongs in `INTENT.md`; current work and operational gates belong in
> `workplans/`.
>
> **Layer declaration (accepted model v0.7):** Engine / Lifecycle. Machine-readable
> form: `layer.yaml`. PEP stance: `pep-stance.yaml`. This file states what the
> implementation currently does inside that layer; it does not restate doctrine.
2026-06-28 09:03:37 +00:00
## One-liner
secrets-engine is the Lifecycle engine for cataloged credential work: a
decision-gated Python CLI that validates non-secret lane metadata and
orchestrates narrowly scoped OpenBao policy, AppRole, provisioning,
verification, delivery, routing, handoff, and revocation actions.
It is a deterministic API over OpenBao, not a vault, not a policy decision
point, not an identity provider, and not a general secrets API.
## Implemented Capabilities
### Catalog and validation
- Loads YAML catalog entries for `kv` and `auth-capability` lanes.
- Models `build`, `test`, and `prod` stages, exact mounts/paths/fields,
consumers, approvals, delivery modes, verification, lifecycle ownership,
workload-delivery metadata, and `standard`/`high` risk.
- Distinguishes engine-managed and existing KV mounts, and engine-managed,
existing, or absent native delivery auth.
- Rejects malformed entries, inline secret-looking values, wildcard paths,
unsupported modes, bootstrap-only approval for high-risk lanes, and high-risk
entries without rotation/deactivation owners.
- Records existing ESO/Kubernetes/OIDC workload delivery as metadata only; it
does not operate those delivery systems.
### Decision-gated planning and OpenBao metadata apply
- Resolves a catalog lane by catalog id or `approval.decision_ref`.
- Resolves legacy lane decisions from State Hub by id, with tracked local YAML
mirrors retained only for non-production and explicit throwaway demos. This is
not an access-engine decision record.
- Fails every production live command closed while there is no durable
access-engine / ActionAuthorization record. That fail-closed row is the
published unreachable-engine stance for `prod` in `pep-stance.yaml`. A local
mirror can unlock a prod-labeled lane only when an explicit unsafe-demo
switch, disabled Hub URL, and loopback OpenBao target are all present. Plans
and `apply --dry-run` remain usable.
- Builds and validates the flex-auth `ActionAuthorization` profile, including
exact lane/stage/action/target/actor/purpose matching, bounded validity,
State Hub authority, request digest and decision binding, accepted policy
package/version, and an independently required distinct-approver threshold.
Validation is consume-only; this process does not evaluate policy.
- Renders guarded OpenBao plans for exact consumer ACL policies and AppRoles.
- Applies policy and AppRole metadata idempotently. Existing mounts render a
non-mutating check and are never created by apply.
- Rejects stage mismatches, wildcard paths, broad-admin names, forbidden
`sys/`, token, identity, root, or sudo semantics, and auth-capability paths
outside an exact allowlist.
### KV provisioning
- Imports one declared field from a mode-0600 file outside Git worktrees.
- Generates random values only for non-production lanes.
- Keeps values out of CLI output and evidence records.
- Uses a CAS-aware create/patch primitive. Secret data is passed through a
mode-0600 temporary JSON input reference that is removed in a `finally` path,
never as a raw subprocess argument. Existing-path updates preserve unmentioned
sibling fields and stale-version writes fail closed.
### Verification
- KV positive verification logs in through the lane AppRole and reports field
presence without printing values. With no `--field`, it checks every declared
field; an explicit `--field` narrows the check.
- KV negative verification requires a real unrelated token supplied through a
mode-0600 file outside Git and checks that it cannot read the path. Missing
unrelated identity material fails closed without a backend probe.
- Auth-capability verification checks the AppRole token's capabilities on exact
allowlisted and denial-probe paths.
These are bounded policy/presence probes. A throwaway OpenBao integration test
proves an overlapping unrelated policy is detected. Production still needs a
reviewed owner/source for each unrelated identity; the engine does not mint or
select that identity itself. Verification does not yet exercise a provider
operation or produce OpenBao audit-request correlation.
### Exec-time delivery
- `exec-env` fetches one declared field through the lane AppRole and injects it
into a child process environment.
- `npm-config` creates a temporary mode-0600 npm config containing an environment
reference rather than the token, injects the value only into the child, and
deletes the config afterward.
- Child stdout and stderr are combined and streamed through token-pattern and
exact-value redaction.
- Undeclared fields and undeclared/unsupported exec modes are rejected before
value fetch.
The OpenBao KV response is parsed in the parent process, so every field stored
at the path crosses that process boundary even though only the selected field is
injected. Exec and verification use scoped AppRole sessions that explicitly
self-revoke in a `finally` path. Non-secret evidence contains only an accessor
fingerprint and establishment/revocation outcome; TTL/use limits remain cleanup
backstops.
### Auth-capability handoff
- Models narrow non-KV AppRole grants such as exact `ssh/sign/<role>` update
capability.
- Mints `role_id`/`secret_id` material only after approval and writes it to
distinct mode-0600 files outside Git worktrees without printing the
`secret_id`.
The cataloged standalone `warden-sign` AppRole is currently parked by governance
and must not be applied as a break-glass or recovery path. OpenBao being sealed
cannot be recovered through that AppRole.
### Routing and evidence
- `route` returns a non-secret pointer containing lane ownership, decision
status, metadata/value-presence booleans, missing declared field names,
readiness, and a safe next command. Every declared field must be present.
- Records scrubbed local JSONL evidence and posts a minimal State Hub progress
event on a best-effort basis. Posts carry stable idempotency/source headers.
Each requested State Hub delivery receives an append-only local `delivered`,
`queued`, `failed`, or `skipped-no-topic` companion record; edge-relay queued
receipts retain only the non-secret outbox id. This trail is attributive:
completeness is not claimed, and it is not `audit-core`.
- Every live privileged CLI handler records an attempt before lane-approval
resolution and a terminal success, verification failure, rejection,
interruption, or typed backend/input failure. Failure evidence contains the
exception class and approval state, never exception prose.
- `audit` summarizes allowlisted local lane evidence: action/result counts,
canonical decision references, session cleanup outcomes, and State Hub
delivery outcomes. It never re-emits arbitrary evidence detail.
- Keeps OpenBao audit logs as the backend source of truth.
The engine recognizes edge-relay queued receipts but does not own or initiate
outbox replay. Direct State Hub failure remains locally visible. Route and audit
do not replace exact-action authorization or OpenBao audit logs.
### Steady-state service-auth scaffold
- Implements an explicit KeyCape `client_credentials` exchange for the
accepted `secrets-engine-openbao` service identity.
- Reads the confidential-client secret only from a mode-0600 file outside Git,
sends it through HTTP Basic authentication, and rejects ID/refresh tokens.
- Preflights the exact issuer, subject, audience, principal type, tenant, role,
scope, assurance, RS256 algorithm, and 15-minute lifetime. The in-memory JWT
is excluded from object representations and is renewed at the three-minute
boundary.
- Never retries into or falls back to bootstrap, operator, or AppRole auth.
The CLI now has a named `service-jwt` provider that logs in through that
scaffold when the platform JWT mount/role contract is present, then
self-revokes the issued OpenBao token. Signature verification and token
issuance remain with the platform-owned exact-bound OpenBao JWT role, whose
mount/role contract is still outstanding. Until it is published, `--auth auto`
keeps named bootstrap/env providers and never treats them as a fallback from
service-jwt failure.
Treat their output as operational guidance, not complete attestation for
high-risk lanes.
### Revocation currently available
- For engine-managed native auth, live `revoke` deletes the AppRole and policy.
- For KV lanes, ordinary revoke explicitly preserves all KV metadata/versions.
- Externally managed delivery auth and workload delivery are reported as
preserved and are not mutated.
- `lifecycle suspend` removes only the managed AppRole and preserves its policy
for reviewed re-apply.
- `lifecycle deactivate` removes managed native AppRole/policy objects while
preserving KV custody.
- `lifecycle destroy --dry-run` renders managed-auth removal plus irreversible
KV metadata deletion. Live destroy is currently disabled even with exact
catalog-id confirmation; it will remain closed until the canonical
exact-action approval contract in `SECRETS-WP-0007-T04` is enforced.
These operations do not manage external workload delivery. `session revoke
--accessor-file` revokes an already-issued token by a non-secret accessor the
operator already holds; evidence is fingerprint-only. There is currently no
rotation command, compromised state, or persistent/reversible lane state
machine.
## CLI Surface
```text
secrets-engine catalog list|show
secrets-engine decision inspect
secrets-engine plan
secrets-engine apply [--dry-run]
secrets-engine provision
secrets-engine verify
secrets-engine handoff
secrets-engine exec
secrets-engine policy publication
secrets-engine route
secrets-engine revoke [--dry-run]
secrets-engine lifecycle suspend|deactivate|destroy
secrets-engine audit <catalog-id> [--json]
secrets-engine evidence heartbeat|drain|classify
```
The implemented exec adapters are `exec-env` and `npm-config`. `read-check` is
verification, `approle-login` is auth-capability handoff metadata, and
`exec-file`/`wrapped` are reserved schema names without executable adapters.
## Proven Operationally
- The whynot-design npm pilot completed a real publish through native
`secrets-engine exec` without placing the token in the parent shell.
- Throwaway OpenBao integration tests exercise plan, apply, provisioning,
positive/negative verification, exec delivery, routing, revocation, and
idempotent metadata apply.
- Five existing high-risk `platform` mount lanes have reviewed catalog metadata
and guarded dry-run plans. Their native AppRoles are not live: workplan
`SECRETS-WP-0006` waits on per-lane approval, production authority, live
positive/negative verification, and routing cutover.
## Not Implemented
- A service API, daemon, UI, queue, scheduler, or remote multi-user service.
- OpenBao JWT login and platform materialization for the implemented KeyCape
service-auth provider.
- Native `exec-file` or response-wrapped delivery.
- Provider-side rotation or coordinated multi-consumer rollout.
- First-class rotate, compromise, reactivate, lease-status, or audit report
commands; lifecycle operations currently execute plans without persistent
lane state.
- Resolution of a durable access-engine decision record / State Hub
ActionAuthorization and wiring its validated approval threshold to each
production handler. The consumer validator exists; the serving endpoint does
not, so live production remains fail-closed.
- Direct access-engine evaluation, JWT signature verification, or identity
authentication. KeyCape claims receive only a consumer preflight; OpenBao is
responsible for cryptographic JWT validation.
- Runtime tenancy isolation; `org`, `repo`, consumers, and stages are catalog
metadata plus local path/policy guards, not a tenant control plane.
- Management or health verification of ESO, Kubernetes Secrets, deployments,
provider accounts, SSH issuance, tunnels, or remote transport.
- Any backend other than the local `bao`/`vault` CLI speaking to OpenBao.
- An SSH-CA signing engine API. The contract is
`docs/ssh-ca-signing-contract.md`; ops-warden still signs through its
declared OpenBao gap.
- A secret-use evidence engine API for kings-guard. The contract is
`docs/secret-use-evidence-contract.md`. `route` and `audit` are operator
summaries over local JSONL, not that observation surface.
- Emission to `audit-core`. Load-bearing records are queued locally; drain
requires a sender binding that does not exist yet. Completeness is not
claimed.
- Security-zone membership as a request claim. PEP scope is catalog stage.
- Drain of the load-bearing outbox into a live `audit-core` sender binding.
## System Boundary
- **OpenBao / railiance-platform** (Tooling) owns custody, policy enforcement,
leases, and backend audit. secrets-engine is the Lifecycle API over it.
- **access-engine (`flex-auth`)** owns authorization decisions. secrets-engine
consumes a decision record or applies its published unreachable-engine stance;
it does not evaluate policy.
- **approval-engine** owns the durable approval object. secrets-engine may
consume an approval as an input claim and must not store or mutate one.
- **gate-house** owns security doctrine. Doctrine reaches this engine only as it
already reached the decision, never as a side channel.
- **audit-core** owns evidence custody and integrity. Local JSONL and State Hub
notes are not that archive.
- **user-engine / key-cape** own identity, OIDC, MFA, and claims.
- **ops-warden** issues SSH certificates (Staff PEP) and routes non-SSH
credential needs here; it does not vend their values. The SSH-CA write
surface is a proposed engine API, not a transferred lane.
- **ops-bridge** owns tunnels and remote execution transport and may consume a
scoped delivery path.
- **workload/platform repositories** own ESO/Kubernetes delivery, provider
rotation, and application health.
- **info-tech-canon / net-kingdom** own canonical terminology and the
cross-system security boundary. **gate-house** owns the layer model.
Canonical boundary:
`net-kingdom/docs/secrets-engine-security-infrastructure-boundary.md`.
Layer model: `net-kingdom/canon/standards/security-layer-model_v0.7.md`.
Working companion: `net-kingdom/SECURITY-COMPANION.md`.
## Security Rules
- Never place raw secret values in Git, State Hub, chat, prompts, workplans,
normal logs, or evidence.
- Never render or cache an authorization decision. Catalog admission, a dry-run,
an old workload CCR, or a local fixture is not an access-engine allow.
- Never treat silence from `access-engine` as permission. Production live
actions fail closed; any unreachable-engine residue must be the published
stance, recorded, never implicit.
- Never mutate an existing shared mount or replace workload delivery by
implication.
- Never add KV destruction back to ordinary `revoke`; irreversible custody
destruction requires `lifecycle destroy`, exact confirmation, and a distinct
approved action. The live path remains disabled until that approval contract
exists.
- Keep bootstrap and handoff material outside repositories with mode 0600 and
explicit expiry/revocation handling. Bootstrap is not an implicit fallback
from service identity.
- Never claim that local evidence or a missing record proves occurrence or
non-occurrence. Completeness is not claimed.
## Layer-model obligations (current vs intended)
| Obligation | Current | Intended |
| --- | --- | --- |
| Layer declaration | `INTENT.md` frontmatter + `layer.yaml` | Keep in this repository's own voice |
| One decision point | Consumer validator for ActionAuthorization; production live fail-closed from `pep-stance.yaml` | Consume an access-engine decision record before every protected side effect |
| PEP stance | Runtime loads `pep-stance.yaml`; named stance fields on privileged evidence | Unchanged map; T02 replaces fail-open residue with a decision record |
| Evidence bound | `evidence-classification.yaml`; load-bearing local outbox; heartbeat command | Drain to `audit-core` once that sender is admitted |
| SSH-CA surface | Contract at `docs/ssh-ca-signing-contract.md`; not shipped | Engine API after ops-warden assent |
| Secret-use evidence | Contract at `docs/secret-use-evidence-contract.md`; not shipped | Engine surface after kings-guard assent |
| Agent credential | Bootstrap token file still accepted; KeyCape scaffold unwired to OpenBao | Per-task, time-bounded service identity; no standing engine credential |
## Where Current Work Lives
- `workplans/` is the source of truth for open work and operational gates.
- `docs/hardening-backlog.md` tracks exit from bootstrap mode.
- `history/` contains dated capability and intent assessments.
- `INTENT.md` remains the stable aspirational direction.
- `layer.yaml`, `pep-stance.yaml`, and `evidence-classification.yaml` are the
layer-model declaration surface.
## Provided Capabilities
```capability
type: security
title: Guarded OpenBao lane planning and metadata apply
description: Validates cataloged KV and auth-capability lanes, renders exact policy/AppRole plans,
and applies approved metadata while rejecting missing decisions, stage mismatches, wildcards,
broad-admin names, and forbidden capability paths. Existing mounts are check-only.
keywords: [secrets, openbao, policy, approle, catalog, decision, stage, least-privilege]
```
```capability
type: security
title: AppRole-scoped exec delivery
description: Fetches one declared KV field through a lane AppRole and injects it into a child
process using exec-env or a temporary npm config, with output redaction, explicit scoped-token
self-revocation, and non-secret cleanup evidence. This is CLI-local delivery; exec-file, wrapping,
and service delivery are absent.
keywords: [secrets, delivery, exec, npm, injection, redaction, openbao]
```
```capability
type: security
title: Scoped OpenBao auth-capability handoff
description: Generates an exact-path policy/AppRole plan and writes role-id/secret-id material to
strict out-of-repo files after approval. Operational use remains subject to lane-specific governance.
keywords: [openbao, approle, auth-capability, handoff, least-privilege]
```
```capability
type: governance
title: Non-secret routing and evidence pointers
description: Reports decision/readiness metadata and records scrubbed local and best-effort State Hub
evidence without returning secret values. Append-only delivery receipts expose State Hub failures,
and an allowlisted audit command summarizes lane operations and cleanup. KV verification can attest a
supplied real unrelated identity, but the engine does not own identity selection. This is attributive
local evidence, not audit-core and not a secret-use observation API.
keywords: [routing, evidence, state-hub, audit, secrets]
```