WARDEN-WP-0033 T01/T02: two under-graded lanes, and ADR-0008

secrets-engine reviewed our catalog metadata while drafting their five entries
and graded issue-core-ingestion-api-key and reuse-surface-hub-write-token high.
We had both as standard, and had deliberately regraded them DOWN on 2026-08-19.

They are right. Both paths carry a second credential our grade never looked at --
GITEA_BACKEND_TOKEN (CCR-2026-0002, a deliberate field-set decision) and a
dual-consumer webhook HMAC (CCR-2026-0005). Neither is recovered by rotating the
credential the lane is named after.

The defect is structural: we graded the lane by its headline field, but a read
returns every field at the path. Worse, the evidence was already in the CCRs we
cite as authoritative -- not missing, unread -- and a test asserted the wrong
answer, so a correct first-pass grade got overruled by it.

ADR-0008 records the rule: a grade covers every field its path discloses.
ADR-0007 is unchanged and still governs; this says what the grade is of.

Six of the remaining standard lanes have no KV path. Two have paths and no field
evidence; per ADR-0008 they are stated as unknown rather than assumed, and left
for operator-sanctioned grading.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
tegwick 2026-08-21 08:36:42 +02:00
parent 9b7e2e9bd2
commit c357ce5908
8 changed files with 155 additions and 10 deletions

View file

@ -15,6 +15,7 @@ distinction that matters here.
| `ADR-0005` | Implement one lane narrowly, route everything else |
| `ADR-0006` | Enforcement is zone-scoped, never a global flag |
| `ADR-0007` | Build-stage permissiveness stops at credential disclosure; every lane carries an explicit `risk` grade |
| `ADR-0008` | A lane's risk grade covers every field its path discloses, not just the field it is named after |
### Owned versus inherited — check `owner:` before changing anything

View file

@ -155,6 +155,7 @@ superseding ADR, never an in-place edit.
| `ADR-0005` | Implement one lane narrowly, route everything else |
| `ADR-0006` | Enforcement is zone-scoped, never a global flag (`policy.enabled` stays off until WP-0032) |
| `ADR-0007` | Build-stage permissiveness stops at credential disclosure; every lane carries an explicit `risk` grade |
| `ADR-0008` | A lane's risk grade covers every field its path discloses, not just the field it is named after |
Rules we follow but do not own — NetKingdom canon, the IAM profile, the
credential-management standard, the-custodian's ADR-001 workplan convention — are

View file

@ -29,6 +29,7 @@
| workplan | WARDEN-WP-0030 | finished | — | workplans/WARDEN-WP-0030-delegation-register.md |
| workplan | WARDEN-WP-0031 | finished | — | workplans/WARDEN-WP-0031-policy-caller-identity.md |
| workplan | WARDEN-WP-0032 | proposed | — | workplans/WARDEN-WP-0032-security-zones.md |
| workplan | WARDEN-WP-0033 | active | — | workplans/WARDEN-WP-0033-native-lane-handoff.md |
| task | ADHOC-2026-06-27-T01 | done | — | workplans/ADHOC-2026-06-27.md |
| task | ADHOC-2026-06-29-T01 | done | — | workplans/ADHOC-2026-06-29.md |
| task | ADHOC-2026-08-11-T01 | done | — | workplans/ADHOC-2026-08-11.md |
@ -121,3 +122,7 @@
| task | WARDEN-WP-0032-T04 | wait | — | workplans/WARDEN-WP-0032-security-zones.md |
| task | WARDEN-WP-0032-T05 | done | — | workplans/WARDEN-WP-0032-security-zones.md |
| task | WARDEN-WP-0032-T06 | progress | — | workplans/WARDEN-WP-0032-security-zones.md |
| task | WARDEN-WP-0033-T01 | progress | — | workplans/WARDEN-WP-0033-native-lane-handoff.md |
| task | WARDEN-WP-0033-T02 | todo | — | workplans/WARDEN-WP-0033-native-lane-handoff.md |
| task | WARDEN-WP-0033-T03 | todo | — | workplans/WARDEN-WP-0033-native-lane-handoff.md |
| task | WARDEN-WP-0033-T04 | todo | — | workplans/WARDEN-WP-0033-native-lane-handoff.md |

View file

@ -0,0 +1,94 @@
---
id: ops-warden-adr-0008
type: adr
title: "ADR-0008 — A lane's risk grade covers every field its path discloses"
domain: infotech
repo: ops-warden
status: accepted
version: "1.0"
revision: "1"
owner: ops-warden
binds: "ops-warden"
created: "2026-08-21"
updated: "2026-08-21"
last_reviewed: "2026-08-21"
review_interval: 6m
enforced_by: "registry/routing/catalog.yaml fields + risk; tests/test_routing.py::test_high_risk_lanes_classified"
supersedes: ""
successor: ""
---
# ADR-0008 — A lane's risk grade covers every field its path discloses
## Status
Accepted 2026-08-21, after `secrets-engine` found two under-graded lanes while
reviewing ops-warden's own catalog metadata.
## Context
`ADR-0007` requires every catalog lane to carry an explicit `risk` grade. It does
not say what the grade is *of*, and the omission turned out to matter.
The catalog describes a lane by a single `fetch_command` naming a single field —
`bao kv get -field=ISSUE_CORE_API_KEY <path>`. Grading followed that description.
But the unit of disclosure is not the field, it is the **path**: `bao kv get`
without `-field` returns every key stored there, and an agent session that
discloses one field has disclosed all of them.
On 2026-08-19, grading all 27 lanes, ops-warden graded
`issue-core-ingestion-api-key` and `reuse-surface-hub-write-token` as `standard`
— "ordinary internal workload secrets". Both grades read only the headline field.
`CCR-2026-0002` records a deliberate decision to keep `GITEA_BACKEND_TOKEN` at the
first path; `CCR-2026-0005` declares a dual-consumer webhook HMAC at the second.
Neither is recovered by rotating the credential the lane is named after.
Three details make this worth a record rather than a fix:
- **The evidence was already ours.** The field sets were in the CCRs the catalog
already cites as authoritative. This was not missing data; it was unread data.
- **A test held the error still.** `test_high_risk_lanes_classified` asserted
`issue-core-ingestion-api-key` was *not* high. A first grading pass had marked
it high, the test contradicted it, and the test was believed. A test that
encodes a judgement defends that judgement from correction.
- **Another repo found it.** `secrets-engine` graded both `high` independently
while drafting catalog entries whose schema records `fields`. A schema that
names the field set makes the right grade obvious; ours did not have one.
## Decision
**A lane's `risk` grade is a property of its path, and must cover the union of
everything a read of that path would disclose.**
1. Where the field set is known, the catalog records it as `fields`, with the
authority it came from.
2. The grade is argued against the most damaging field, not the named one.
3. Where the field set is unknown, that is stated — never assumed to be one
field. An unverified field set is a reason to grade conservatively, matching
the `inter-hub-bootstrap-ssh` precedent under `ADR-0007`.
4. Establishing a field set must not be done by reading the secret. Use the
owning CCR, the owner's catalog, or `bao kv metadata`. `bao kv get` on a
high-risk path is the 2026-07-16 vector and is forbidden by
`ADR-0004` for agent sessions regardless of intent.
## Consequences
`ADR-0007` is unchanged and still governs: every lane carries an explicit grade,
and absence fails safe. This record says what that grade must account for.
Grading gets more expensive: it now requires knowing what is at a path, not just
what the lane is called. That cost is the point — the cheap version produced two
wrong answers in one pass and is the reason this exists.
A test that asserts a grade is asserting a judgement. When a grade is disputed,
re-argue it from evidence before trusting the test that encodes it.
## Related
- `ADR-0007` — every lane carries an explicit grade; build-stage permissiveness
stops at credential disclosure
- `ADR-0004` — high-risk lanes refuse raw value streaming to agent sessions
- `ADR-0001` — the catalog is a pointer layer; `fields` records the owner's
declared field set with its source, and does not restate their procedure
- `WARDEN-WP-0033-T02`; `secrets-engine` `SECRETS-WP-0006`
- `history/2026-07-16-credential-disclosure-lessons.md`

View file

@ -69,3 +69,4 @@ and the publication is a defect.
| `ADR-0005` | Implement one lane narrowly, route everything else | ops-warden |
| `ADR-0006` | Enforcement is zone-scoped, never a global flag | ops-warden |
| `ADR-0007` | Build-stage permissiveness stops at credential disclosure | ops-warden |
| `ADR-0008` | A lane's risk grade covers every field its path discloses | ops-warden |

View file

@ -283,9 +283,16 @@ entries:
- id: issue-core-ingestion-api-key
title: issue-core ingestion API key (OpenBao KV + ESO)
# Ordinary internal workload secret: an ingestion key for a first-party
# service. Rotatable, no spend, no tenant data, no admin scope (WARDEN-WP-0032-T05).
risk: standard
# Regraded standard -> high 2026-08-21 (WARDEN-WP-0033-T02). The T05 grade below
# was wrong, and wrong systematically: it graded the headline field, not the path.
# was: "Ordinary internal workload secret: an ingestion key for a first-party
# service. Rotatable, no spend, no tenant data, no admin scope."
# CCR-2026-0002 records a deliberate field-set decision to keep GITEA_BACKEND_TOKEN
# at this path alongside the ingestion key, and a read discloses every field there.
# A Forgejo backend token is not recovered by rotating an ingestion key.
# Found by secrets-engine reviewing SECRETS-WP-0006 -- not by us.
risk: high
fields: [ISSUE_CORE_API_KEY, GITEA_BACKEND_TOKEN] # CCR-2026-0002
need_keywords: [issue-core, ingestion, api, key, openbao, issue_core_api_key, eso, external-secrets]
owner_repo: railiance-platform
subsystem: OpenBao + issue-core + activity-core
@ -323,9 +330,16 @@ entries:
- id: reuse-surface-hub-write-token
title: reuse-surface federation hub write bearer token
# Ordinary internal workload secret. Write access to a first-party federation
# surface; damaging to forge, but rotatable and internal (WARDEN-WP-0032-T05).
risk: standard
# Regraded standard -> high 2026-08-21 (WARDEN-WP-0033-T02), same defective T05
# pass as issue-core-ingestion-api-key.
# was: "Ordinary internal workload secret. Write access to a first-party
# federation surface; damaging to forge, but rotatable and internal."
# CCR-2026-0005 declares REUSE_SURFACE_FORGEJO_WEBHOOK_SECRET at this path: a
# dual-consumer HMAC that must stay aligned with Forgejo org webhook id=1 on
# rotation. Disclosure lets an attacker forge webhook deliveries into the
# federation hub, which rotating the write token alone does not undo.
risk: high
fields: [REUSE_SURFACE_TOKEN, REUSE_SURFACE_FORGEJO_WEBHOOK_SECRET] # CCR-2026-0005
need_keywords: [reuse-surface, reuse_surface, hub, register, federation, write, token, bearer, REUSE_SURFACE_TOKEN, REUSE_SURFACE_FORGEJO_WEBHOOK_SECRET, reuse.coulomb.social]
owner_repo: railiance-platform
subsystem: OpenBao + reuse-surface

View file

@ -543,8 +543,13 @@ def test_high_risk_lanes_classified():
assert "railiance-backup-offsite-lane" in high
assert "forgejo-admin-api-token" in high
assert "openrouter-llm-connect" in high
# Ordinary workload secrets stay standard unless reclassified.
assert catalog.get("issue-core-ingestion-api-key").is_high_risk is False
# WARDEN-WP-0033-T02: these two were asserted standard here, and the assertion
# held a defective grade still. Both paths carry a second credential the grade
# ignored -- GITEA_BACKEND_TOKEN (CCR-2026-0002) and the dual-consumer webhook
# HMAC (CCR-2026-0005). A read discloses every field at a path, so the grade
# must cover the union, not the headline field.
assert "issue-core-ingestion-api-key" in high
assert "reuse-surface-hub-write-token" in high
def test_invalid_risk_rejected(tmp_path):

View file

@ -41,7 +41,7 @@ work completed directly; this is none of those.
```task
id: WARDEN-WP-0033-T01
status: progress
status: done
priority: high
```
@ -62,7 +62,7 @@ applied to ops-warden by someone else, correctly.
```task
id: WARDEN-WP-0033-T02
status: todo
status: done
priority: high
```
@ -89,6 +89,30 @@ disclose, and it should.
Fix the grade, then fix the model that produced it: record `fields` per lane and
grade on the union. Check the other 25 lanes for the same shape.
**Done 2026-08-21 — and it was two lanes, not one.** While verifying
`reuse-surface-hub-write-token` I told secrets-engine their second field was
unevidenced and asked them to confirm it. It was evidenced: `CCR-2026-0005`'s
`kv_path` block lists both fields and its notes describe the HMAC's alignment
with Forgejo org webhook id=1. A truncated grep, and a retraction sent within
twenty minutes so they would not spend time on it. That lane was under-graded on
exactly the same reasoning, and both had been regraded *down* in the same T05
pass — which is what makes this a systematic flaw rather than one bad call.
Blast radius checked and bounded: of the 8 remaining `standard` lanes, 6 have no
KV path at all (SSH, policy check, tunnel, principals, issue-sink, key-cape
login) so nothing is disclosed by a read. Two do have paths and no CCR field
evidence — `rein-openweights-openrouter-approle` and `coulomb-social-runtime-env`
(a whole k8s env Secret, which by shape almost certainly holds several fields).
Deliberately **not** regraded: `ADR-0008` §3 says an unknown field set is stated
rather than assumed, and grading is operator-sanctioned judgement, not backfill.
They are the open item on this task.
Shipped: both regrades with their evidence and the superseded reasoning kept
inline, `fields` recorded with its CCR source, the test that had been holding the
wrong answer inverted with an explanation of why it was wrong to trust it, and
**`ADR-0008`** — a lane's grade covers every field its path discloses. That is a
new binding rule, so it is a record rather than a wiki note.
```task
id: WARDEN-WP-0033-T03
status: todo