coulomb/railiance-platform
4
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Consolidate platform workplans and assess intent gaps

Browse source 
Assistant: codex
Assistant-Model: gpt-6-astra
Assistant-Session: 01a06ecb-456a-71c2-b41e-0755d336e883
This commit is contained in:
codex 2026-09-05 11:14:42 +02:00
parent 9d958f8e09
commit 9f83e426c7
40 changed files with 2985 additions and 295 deletions
Download patch file Download diff file Expand all files Collapse all files

325
workplans/RPF-WP-0002-openbao-approved-automation-delegation.md
View file

@ -1,325 +0,0 @@
---
id: RPF-WP-0002
type: workplan
title: "OpenBao Approved Automation Delegation"
domain: financials
repo: railiance-platform
status: finished
owner: codex
topic_slug: railiance
planning_priority: high
planning_order: 8
created: "2026-06-28"
updated: "2026-07-01"
depends_on_workplans:
- RAIL-PL-WP-0002
- RAILIANCE-WP-0005
- RAILIANCE-WP-0006
- RAILIANCE-WP-0007
state_hub_workstream_id: "c69cb4e7-15c8-5858-aebd-73a6e46f0f21"
---
# RPF-WP-0002 - OpenBao Approved Automation Delegation
## Goal
Create a narrow OpenBao automation delegation path so approved credential and
it-sec changes can be applied without handing broad `platform-admin` power to an
agent, while still preserving different build, test, and production security
requirements.
This closes the current gap where reviewed CCRs can be approved and rendered,
but live application still fails with `403 permission denied` on
`sys/policies/acl/...` unless a human operator uses broad OpenBao admin
privileges.
## Problem
The current model has two bad modes:
- fully manual OpenBao UI work, which can unblock one credential but is slow,
hard to review, and easy to drift from the CCR;
- broad `platform-admin`, which is too powerful for routine automation and too
risky to hand to agents.
The issue appears in multiple workstreams:
- `RAILIANCE-WP-0005` token-grant apply failed on
`sys/policies/acl/credential-broker-warden-sign-issuer`;
- `RAILIANCE-WP-0007` whynot-design CCR apply failed on
`sys/policies/acl/workload-kv-read-whynot-design-npm-publish`.
Both are approved non-secret metadata changes. They should not require a human
to hand-type OpenBao policy and auth-role writes once review is complete.
## Direction
Add one or more constrained OpenBao applier identities that can apply only
reviewed, generated metadata for approved change classes.
The applier must not be a general `platform-admin` replacement. It should:
- verify a resolved State Hub decision or CCR status before live mutation;
- write only allowed policy names and auth role prefixes;
- never read or print secret values;
- record non-secret apply evidence;
- keep production secret value provisioning under approved operator custody
until a separate wrapped or dual-control flow is approved.
## Environment Model
Build and development:
- automation may apply approved sandbox metadata;
- generated test secrets may be allowed only in non-production mounts;
- approvals can be lightweight but still recorded.
Test and staging:
- automation may apply approved metadata after validation and owner review;
- verification must include positive and negative access checks;
- secret values should use wrapped delivery or operator custody depending on
risk classification.
Production:
- automation may apply only approved non-secret metadata such as narrow ACL
policies and auth roles;
- production secret values remain out-of-band operator custody unless a later
workplan creates a stronger wrapped or dual-control provisioning path;
- activation requires verification evidence before ops-warden or another front
door becomes resolvable.
## Proposed Policy Shape
Start with one production candidate policy, for example
`credential-change-prod-applier`, and keep it intentionally narrow:
- allow `create`, `update`, and `read` on `sys/policies/acl/workload-kv-read-*`;
- allow `create`, `update`, and `read` on approved credential-broker issuer
policy names such as `credential-broker-*-issuer`;
- allow `create`, `update`, and `read` on selected auth role prefixes such as
`auth/netkingdom/role/*` with local dry-run role-name constraints,
`auth/kubernetes/role/*`, and `auth/token/roles/credential-broker-*`;
- allow read/list only where needed for idempotent verification;
- deny broad `sys/*`, `auth/*`, `platform/*`, `identity/*`, `root`, and
`platform-admin` semantics.
Policy-name and role-name restrictions should be enforced in both OpenBao ACLs
and the local applier script.
## Tasks
## T01 - Specify delegated applier policy boundaries
```task
id: RPF-WP-0002-T01
status: done
priority: high
state_hub_task_id: "914a31ae-9fbf-5c8d-a5df-cc5a3c4633e3"
```
Define build, test, and production applier capabilities, including exact
OpenBao paths, allowed name prefixes, denied paths, and required audit evidence.
Acceptance:
- A human reviewer can see which OpenBao paths each environment can mutate.
- Production applier policy excludes secret value reads and broad admin paths.
- The proposal covers both workload KV read lanes and credential broker issuer
policies.
**2026-06-29:** Added `docs/openbao-approved-automation-delegation.md` and
`openbao/policies/credential-change-prod-applier.hcl`. The document defines
build/development, test/staging, and production boundaries, the allowed
production metadata mutation surface, denied secret/admin paths, and required
non-secret evidence. The production policy candidate allows only reviewed
metadata writes for workload KV read policies, credential-broker issuer
policies, approved auth-role prefixes, and self capability checks; it does not
grant secret value reads or writes.
## T02 - Implement a CCR-aware applier dry-run
```task
id: RPF-WP-0002-T02
status: done
priority: high
state_hub_task_id: "d02dd668-e75d-59b3-a1db-5c0e5b8ce301"
```
Extend the credential-change tooling so a proposed applier can validate a CCR,
check approval state, render the exact mutations, and refuse any out-of-policy
policy name, auth role, mount, path, or environment.
Acceptance:
- Dry-run succeeds for `CCR-2026-0001`.
- Dry-run refuses unapproved CCRs.
- Dry-run refuses attempts to create `root`, `platform-admin`, wildcard, or
unrelated policy names.
**2026-06-29:** Added `scripts/credential-change.py applier-dry-run <CCR>` and
Make target `credential-change-applier-dry-run`. The dry-run validates the CCR,
requires approved/applied/verified/active status, requires confirmed auth
bindings, verifies the OpenBao mount/path/policy/role stay inside the delegated
metadata surface, compares the policy artifact to the generated CCR policy body,
and renders only policy/auth-role mutations. It explicitly leaves secret value
writes, secret reads, and front-door activation out of scope. Unit tests cover
the active whynot-design CCR success path, unapproved CCR refusal, and rejection
of `platform-admin`/out-of-scope mount and path attempts. `make
credential-tests` passed with 28 tests.
## T03 - Add non-production applier role first
```task
id: RPF-WP-0002-T03
status: done
priority: medium
state_hub_task_id: "945a2847-ceaa-5bba-ab2b-29b4d90abac7"
```
Create a build/test applier identity and prove it can apply approved metadata in
a non-production lane without gaining unrelated OpenBao permissions.
Acceptance:
- Apply succeeds in a non-production mount or namespace.
- Negative checks prove unrelated policy/auth/secret paths are denied.
- Evidence is recorded without secret values.
**2026-06-30:** Added the non-production metadata-only policy candidate
`openbao/policies/credential-change-nonprod-applier.hcl` and documented that
generated test-secret paths require separate CCR-backed approval. Live non-prod
identity creation and positive/negative OpenBao evidence remain to close this
task.
**2026-06-30:** Added the guarded `applier-apply` execution path that reuses the
CCR dry-run guardrails, requires exact `DELEGATED APPLY <CCR-ID>` confirmation,
uses the local `bao` CLI with ambient delegated applier authority, writes only
policy/auth-role metadata, and records non-secret `delegated_metadata_apply`
evidence. Non-production task closure still needs a live build/test applier
identity plus positive and negative capability evidence.
**2026-06-30:** Added `scripts/openbao-apply-credential-change-appliers.py` and
Make target `openbao-credential-change-appliers-dry-run` to install/dry-run the
non-production applier policy plus bounded `auth/token/roles/credential-change-
nonprod-applier` role. The token role allows only the matching applier policy,
disallows `root` and `platform-admin`, disables the default policy, and does not
issue tokens by itself. Live non-production apply and denial evidence remains
the closeout gate.
**2026-07-01:** Applied the updated non-production metadata-only policy
and bounded `auth/token/roles/credential-change-nonprod-applier` role to live
OpenBao. The role attaches only `credential-change-nonprod-applier`, disables
the default policy, and disallows `root` / `platform-admin`; T03 remains open
until a non-production lane apply and denial probe are recorded.
## T04 - Add production metadata applier with human approval gate
```task
id: RPF-WP-0002-T04
status: done
priority: high
state_hub_task_id: "493744db-445d-546f-82ea-2dee4bc72214"
```
Create the production metadata applier path and require a resolved CCR/State Hub
approval before mutation.
Acceptance:
- Approved `CCR-2026-0001` metadata can be applied without `platform-admin`.
- Unapproved CCRs fail closed.
- Secret value provisioning is still not automated in production.
**2026-06-30:** Strengthened the production gate by adding source-artifact
checks to the CCR applier dry-run and documenting that unapproved CCRs fail
closed before OpenBao mutation rendering. The production policy candidate exists
and remains metadata-only; live delegated identity creation/application evidence
still needs an operator-held OpenBao step.
**2026-06-30:** Added `applier-apply` and Make targets
`credential-change-applier-apply-plan` / `credential-change-applier-apply`. The
command fails closed for unapproved CCRs, renders the dry-run payload before
mutation, requires exact confirmation, does not accept tokens in argv, leaves
secret values out of scope, and appends State Hub/file-backed non-secret apply
evidence when requested. Production closure still requires live execution using
the constrained applier identity rather than broad `platform-admin`.
**2026-06-30:** Added `scripts/openbao-apply-credential-change-appliers.py` and
Make targets `openbao-credential-change-appliers-dry-run` /
`openbao-configure-credential-change-appliers` to configure the production
`credential-change-prod-applier` policy and bounded token role. The role allows
only `credential-change-prod-applier`, disallows `root` and `platform-admin`,
uses service tokens, disables default policy attachment, and keeps token issuance
outside the setup script. Production closure still needs a live run and
capability evidence using this constrained identity.
**2026-07-01:** Updated the delegated applier ACLs to use the OpenBao-matchable
`auth/netkingdom/role/*` path while keeping role-name and bound-claim
constraints in the local CCR dry-run. Applied the live prod/nonprod applier
policies and token roles, then issued a 15-minute
`credential-change-prod-applier` child token and used it to run
`scripts/credential-change.py applier-apply CCR-2026-0001`. The delegated
run wrote the workload KV policy and OIDC role metadata without
`platform-admin`. A `sys/capabilities-self` probe on
`platform/data/workloads/coulomb/whynot-design/npm-publish` returned
`deny`, and the matching short-lived child token accessor was revoked.
## T05 - Close the whynot-design pilot
```task
id: RPF-WP-0002-T05
status: done
priority: high
state_hub_task_id: "b5959c20-7ac7-53a8-ba0d-ee569c122a4f"
```
Use the delegated production metadata applier to finish the whynot-design npm
publish token lane after the actual token is provisioned through approved
custody.
Acceptance:
- Policy and auth role are applied by the delegated applier.
- `NPM_AUTH_TOKEN` is provisioned through approved custody.
- Positive and negative verification pass without printing the token.
- `CCR-2026-0001` can move to `active`.
- ops-warden can mark `whynot-design-npm-publish` ready/resolvable.
**2026-07-01:** Closed the whynot-design pilot. `CCR-2026-0001` is
active, the front-door metadata is ready/resolvable, prior approved-custody
provisioning plus positive and negative verification are recorded without
secret values, and the delegated prod applier evidence is now recorded on the
CCR.
## Exit Criteria
- Routine approved OpenBao metadata changes no longer require broad
`platform-admin`.
- Production automation cannot read or exfiltrate managed secret values.
- Build, test, and production each have distinct, documented security
requirements.
- CCR approval, apply, verification, and front-door activation form one
reviewable chain.
## Completion 2026-07-02 — T03 live probe and workplan finish
T03 closed with live positive and negative evidence from a
`credential-change-nonprod-applier` child token (accessor
`pCznHtid1O0vy36QHqMbzu5Y`, revoked after use):
- allowed: `policy_write workload-kv-read-nonprod-probe-test` (test artifact
deleted afterwards by the operator session) and `policy_read
workload-kv-read-issue-core-runtime`;
- denied: `policy_read platform-admin`, out-of-pattern `policy_write
evil-probe-test`, KV secret read on the issue-core path, and
`auth/token/roles/credential-change-nonprod-applier` write;
- all recorded in `/openbao/audit/openbao-audit.log` (2026-07-02T10:09Z
window).
The production applier path was proven the same day: both `CCR-2026-0002`
and `CCR-2026-0003` were applied with a `credential-change-prod-applier`
child token holding only that policy — no `platform-admin` handoff. All
tasks are done; the workplan is finished.