The playbook detector is the estate reference (021 §3). Add its one
addition to the reference and the checker: any v?N.N in a standard: or
companion: value is a pin. The prose-citation note moves from pending to
not reached (021 §1, A12 r3), and intent_version is noted as a key that
must not be flagged. VALIDATED_AGAINST keeps accepted v0.7 and adds
GH-DEC-2026-021 at gate-house@39d9287.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 63291@bnt-lap001
Assistant-Session: 8bd77868-ca68-4f49-bb1e-d539ecc0d703
The layer conformance checker now prints VALIDATED_AGAINST and SCOPE on every
run, including the PASS line (kings-guard pattern), and enforces A12 r2 over
every key and value of INTENT.md frontmatter and layer.yaml: a versioned
standard: path and a companion_version are caught, schema_version and comments
are not reached, pep-stance.yaml is outside the run. Tests guard both returns.
The playbook carries the adopter change set and confirms the section 5
citation is canonical. WARDEN-WP-0034's open 4220413a note is closed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 63291@bnt-lap001
Assistant-Session: 8bd77868-ca68-4f49-bb1e-d539ecc0d703
Verified against gate-house's own committed files before editing, not the inbox
message: GH-DEC-2026-017 in decisions/decisions.md at gate-house@def0af2,
amendments A9-A13 in docs/amendments/v0.8-section-11-declaration-amendments.md,
and sections 3, 4 and 11 of net-kingdom's security-layer-model_v0.8.md. The
ruling and docs/layer-declaration-precedence.md's secondary account agreed.
INTENT.md's frontmatter is the declaration; layer.yaml is a derived artifact,
now marked derived: true / derived_from: INTENT.md, and it does not govern.
standard_version is removed from BOTH forms. The ruling's general form is that a
layer declaration must not carry a standard version, and INTENT.md is the
declaration, so removing it from the sidecar alone would have left the field in
the only file that actually declares. INTENT.md's version-pinned `standard:`
path is de-versioned for the same reason: a pinned path reads as a validity
condition. The version ops-warden assented at stays with the assent, ADR-0010.
NO LAYER VALUE IS CHANGED. INTENT.md still says Staff and layer.yaml still says
staff. Section 3's vocabulary is closed, four tokens, and case-insensitive: the
two forms were never in disagreement about a layer, and the ruling asked nobody
to re-spell anything. The comment marking the divergence is rewritten from
"unruled, do not touch" to "ruled, folding case is the checker's job".
check_layer_conformance.py would have rejected the conforming declaration this
ruling produces -- it listed standard_version as a required key. It now reads
INTENT.md as the governing form, ASCII-folds before comparing, validates both
values against the closed four-token vocabulary (Taxonomy included; omitting it
is the defect A9 records against the estate's other validator), requires the
derived marking, rejects a returning standard_version in either file, and
reports a post-fold disagreement between the forms as a finding rather than
resolving it away by precedence.
The test asserts the fold, not equality. An equality assertion here would be
this repository quietly performing the re-spelling the ruling declined to order;
the fold still fails on a real layer divergence.
pep-stance.yaml is untouched. A stance map is not a layer declaration, and the
sidecar schema beyond the derived marking and the version is explicitly not
ruled.
layer.yaml is the form seven repositories copied, so the adopter change set is
written out in wiki/playbooks/netkingdom-layer-declaration.md -- including the
trap that an adopter which also copied the checker turns a conforming
declaration into MALFORMED exit 2 by removing the field alone. No other
repository is edited here.
Still open: where the removed version lives. A12 says the derived conformance
record "already MUST" carry it; ops-warden has a re-runnable checker that emits
nothing durable. Asked of gate-house in message 4220413a, unanswered, and left
open rather than answered by choosing. Nothing above depends on it.
Carries WARDEN-WP-0034-T06 to done.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 63291@bnt-lap001
Assistant-Session: 8bd77868-ca68-4f49-bb1e-d539ecc0d703
Five inbox items worked, none of which changed a credential value or moved a
secret.
whynot-design-npm-publish: field reverted npm_token -> NPM_AUTH_TOKEN and the
path confirmed, on railiance-platform's attended, read-only, no-value field
enumeration (their docs/evidence/2026-09-10-npm-lane-field-resolution.json).
Exactly one field is present at the governed path. The 2026-09-09 change was
adopted from a coordination message and would have failed at the WP-0037-T03
rotation. The ungoverned second location is recorded as an explicit non-lane,
not deleted and not tidied away.
pep-stance coverage: published figures were stale by eight lanes (unknown
18->20, not_applicable 12->15) while resolved stayed at 3 — the denominator
moved, the classification did not. Caught by the test that asserts the published
block equals what report_coverage.py measures. tests/test_workload_join.py held
the same stale counts; both now measure the same populations.
rapp-qonto-keycape-client: blocker character updated — authority exists and is
unexercised by owner decision ("not yet", offer open), which is not the same as
no authority existing. Reopen triggers are events, never elapsed time.
flex-auth -> access-engine rename (WARDEN-IN-0003): access-engine added to the
policy-check lane's keywords so routing resolves under both names from today.
owner_repo deliberately not flipped — policy.py sends it as resource.system on
every /v1/check, and FLEX-DEC-2026-013 keeps runtime names as flex-auth.
layer declaration: INTENT.md says Staff, layer.yaml says staff, section 11 does
not say which governs. Neither changed; gate-house holds the ruling. Position in
docs/layer-declaration-precedence.md, wait in WARDEN-WP-0034-T06, and a comment
in layer.yaml telling the next session not to "fix" it — the divergence is the
evidence the ruling is made against.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 63291@bnt-lap001
Assistant-Session: 8bd77868-ca68-4f49-bb1e-d539ecc0d703
gate-house ruled the v0.8 assent round (GH-DEC-2026-011, net-kingdom@64394e9):
ask 1 declined, ask 2 adopted.
Ask 1's refusal is accepted without reservation and the reason is better than
the ask -- a sanctioned transitional fail_open is indistinguishable at runtime
from the stance the rule forbids, and would make the rule optional at the only
moment it costs anything.
Ask 2 gave §13.1 a Coverage column with this repo's figures as its first
entries. Since we asked for the column, we owe it accuracy:
scripts/report_coverage.py measures both populations from the artifacts the
runtime uses (reusing the workload-join build rather than re-deriving it), and
a test asserts pep-stance.yaml's published block equals what it measures.
A hand-counted number in a register that explicitly does not recompute it
decays silently, and a stale figure beside a marked cell is worse than the
blank the other four rows carry.
pep-stance.yaml marks the unknown cell inline as a declared gap -- assent, the
measured reason for not flipping, the declined ask, WARDEN-WP-0040 as route --
and a second test keeps it marked while it is fail_open, failing when it is
flipped. standard_version stays 0.7 because that is what binds; v0.8 is
proposed, so it gains standard_version_reviewed rather than pre-adopting.
Separately, gate-house corrected GH-DEC-2026-008: the claim/decision digest
comparison it originally required is unimplementable and a fail-closed
consumer obeying it would have denied permanently. We had never copied the
wording, so nothing to unwind -- but everything they have sent about this lane
was living in an inbox thread, a bad home for a correction that only matters
when someone finally wires the consume. Now wiki/ApprovalConsumption.md,
leading with "nothing is wired", carrying the corrected target and the
attribution gap that digest matching does not discharge.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EPuTc18FjU5WFqoSEKH3C
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1276224@bnt-lap001
Assistant-Session: 426ec497-e1c4-4dd3-b417-dfce1ca1dbc3
`warden plan` scored needs by keyword overlap with no notion of what the
caller wanted to DO, so "generate a successor secret and CAS-write it to two
custodians" matched the lane that READS that path and inherited its
`autonomous` verdict -- answered with --out/--exec/--wrap.
Two counterparties reported it in two days. key-cape distrusted the output on
principle and was right to; railiance-platform, answering as the write
authority being wrongly bypassed, said plainly that `founder_required` is the
verdict it should have returned and that until it is fixed a plan result must
not stand in for the owner's answer.
A mutating need on a lane ops-warden does not permanently own can no longer
reach any branch returning `autonomous`: it becomes `founder_required` with
an approve act naming the write owner, or `unroutable` with a CCR stub when
the lane admits no rotation route. Commands carry no read transport either
way, which is the half that made the wrong verdict actionable.
The ownership test does the work a verb list cannot. SSH certificate
issuance is itself a mutating act, so `delegation.mode: permanent` -- not the
absence of a verb -- separates ops-warden's own front door from someone
else's custody. A regression asserts `warden sign` still proceeds; a guard
that refused our own lane would be worse than the defect it fixes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EPuTc18FjU5WFqoSEKH3C
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1276224@bnt-lap001
Assistant-Session: 426ec497-e1c4-4dd3-b417-dfce1ca1dbc3
flex-auth fixed enrichment so registry facts beat caller-supplied ones
(FLEX-DEC-2026-012) and asked each consumer to confirm the ceiling and
allowlist keys are actually declared -- the fix wins only where the registry
HAS a value, and a manifest omitting max_ttl_hours hands that ceiling back
to the caller.
Confirmed, and made durable rather than read once.
scripts/check_flex_auth_manifest_coverage.py audits both ways a ceiling
gets handed back: an actor with no manifest resource at all (warden sign
names ssh-cert:actor/<name> whether or not the snapshot was rebuilt --
an honour-system step in SCOPE.md), and a resource missing one of the
seven keys. A null is treated as absent, because for enrichment it is.
Also asserts the property their exploitability assessment rested on and
nothing here held: ops-warden sends no resource.attributes. It was true
when they read it, secrets-engine sends them on every request, and it was
one refactor from silently stopping being true.
Current state: no gap. 4 actors, 4 resources, all seven keys declared.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EPuTc18FjU5WFqoSEKH3C
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1276224@bnt-lap001
Assistant-Session: 426ec497-e1c4-4dd3-b417-dfce1ca1dbc3
key-cape corrected two blockers that had stopped being true after our
2026-08-28 source-read:
- rapp-qonto-keycape-client: `keycape service-token` (2026-09-05) is the
native exchange the blocker recorded as absent, and `keycape verify-client`
(2026-09-08) is rotation step 3 as one command. Narrowed to steps 1-2 --
successor generation and the CAS write -- rather than cleared, as they
asked. rotation.automatable -> false so a future executable driver is not
told a lane with no admitted custody transport is drivable; the per-step
truth moves into the steps.
- key-cape-oidc-login: ownership ACCEPTED by key-cape, so verified moves from
asked-and-waiting to owner-confirmed. Lane stays interim -- acceptance
covers the identity half, while the fetch_command yields an OpenBao token
whose mount, role mapping and enforcement are not key-cape's.
Two tests pinned `key-cape-oidc-login` to sitting `asked-and-waiting`, so
answering the question broke them -- they failed on good news. Both now
assert the property instead: an unverified blocker is stale regardless of
date, over whatever lanes are in that state.
Verifying the routing answer against our own front door turned up a defect:
`warden plan` returns `autonomous` for a custody *write* and answers it with
read transports, because it has no read-versus-mutate intent. Recorded as
WARDEN-WP-0038 (proposed) -- the WP-0033-T06 shape, as a class this time.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EPuTc18FjU5WFqoSEKH3C
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1276224@bnt-lap001
Assistant-Session: 426ec497-e1c4-4dd3-b417-dfce1ca1dbc3
All three v0.4 findings were acted on — §9.1 split into pending/declared-gap and
§5's scope rule adopted as recommended and credited, and §9.6 ruled via the
load-bearing/attributive distinction with ops-warden's `# audit must not block
signing` named as the estate's live example.
Checked the favourable ruling rather than accepting it. §9.6's test is "no
control branches on its presence": the only consumer of audit.jsonl is `warden
activity`, which displays. Nothing gates on a signing record, so the lane is
genuinely attributive. AuditTrail.md now records the ruling instead of the open
question, and states that the trade must be revisited if a control ever gates on
the trail.
CONFORMANCE ACTION. §6.4 obligation 3 requires a stance map "published rather
than held in code", and requires every PEP-shaped consumer to publish one so the
maps can be inventoried — naming ADR-0009 as the reference shape. ops-warden was
not doing it: the map lived in PolicyConfig.failure_modes, a dataclass default.
Not a code comment, but not published either.
pep-stance.yaml publishes it, and the test asserts the published map EQUALS the
shipped default. A published map that may drift from the code is worse than no
map, because it invites reliance it cannot support.
Two findings sent to gate-house, in history/2026-08-29-layer-model-v06-review.md:
§6.4 obligation 1 (no side effect without a decision record) contradicts
obligation 3 and §9.3, with ops-warden's blessed fail-open stance as the
instance; and §6.4 mandates a stance-map inventory in §13 that §13 does not
implement — where ops-warden is currently the only PEP to have published one.
402 tests pass, ruff clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YWBMovyFoy9RRrfL7zKvPJ
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 4014535@bnt-lap001
Assistant-Session: d0036016-73e8-4da1-8e47-563e3ab39a3c
The security layer model moved v0.1 -> v0.4 (accepted) after ops-warden's
assent. Both §5 asks from ADR-0010 were adopted: §5.2 now sanctions the conduit
shape on the supplied-authority property, and §5.3 is the declared engine gap
amendment, carrying the four fields verbatim and crediting ops-warden's
delegation machinery as prior art.
Which creates an obligation. §5.3 requires those fields MACHINE-READABLY, and
§11 makes "every direct Tooling client maps to a declared §5.1/§5.2/§5.3 entry"
a mechanical check. ops-warden's declaration was prose in INTENT.md — the repo
that proposed the shape was not implementing it.
layer.yaml is the map: 5 contacts (2 declared gaps, 1 read-only observation,
2 conduits) plus the non-Tooling clients recorded explicitly so the check is
total rather than silently selective.
scripts/check_layer_conformance.py enforces it and found three undeclared
modules on its first run — all false positives (help text, a docstring, and the
doubles library that SIMULATES bao rather than calling it), which is why the
scan now matches invocation shapes instead of the word: an httpx call built
against the configured OpenBao address, or an argv whose first element is the
bao binary.
tests/test_layer_conformance.py adds the §5.2 test the standard says SHOULD
exist: _caller_env() returns the caller's environment unchanged, and proxy.py
is asserted not to reference X-Vault-Token, approle login, or token create — a
conduit that presents its own token is not a conduit.
No assertion on review dates, deliberately: a date-triggered failure breaks the
build on a calendar day with no code change, the same reasoning WP-0033-T05
recorded for blocker staleness.
398 tests pass, ruff clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YWBMovyFoy9RRrfL7zKvPJ
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 4014535@bnt-lap001
Assistant-Session: d0036016-73e8-4da1-8e47-563e3ab39a3c
c374d41 added net-kingdom-lldap-bind-credential and
net-kingdom-privacyidea-admin-token as `risk: high` and did not re-run the
emitter, so registry/generated/high-risk-data-paths.yaml still described the
catalog at 0fae090. railiance-platform consumes that file instead of
hand-maintaining its deny list, and it has been reading a census two lanes short
since 2026-08-23.
This is precisely the drift WARDEN-WP-0033-T03 built the guard for — a lane
graded high after the last emit silently failing to reach the consumer. The
guard fired; nothing had acted on it.
The deny list itself does not move: both lanes are blocked on their OpenBao path
being published, so they land in `no_concrete_path` and concrete_path_count
stays 14. What changes is the count the consumer sees — 23 high-risk lanes, two
of which have no address yet. That is the honest signal and the reason the
bucket is listed rather than omitted.
check_agent_read_boundary.py still reports 0 uncovered.
The workload-join census moves 9 -> 11 not-applicable: both lanes are
provider/control-plane credentials rather than workload delivery lanes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YWBMovyFoy9RRrfL7zKvPJ
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 4014535@bnt-lap001
Assistant-Session: d0036016-73e8-4da1-8e47-563e3ab39a3c
risk-nexus accepted the offer to match their convention rather than grow a second
one, and published it: 14d critical/high, 30d medium, 60d low, nothing auto-closing
on staleness alone. Their preference — point warden route gaps at those windows and
the two registers agree without a shared mechanism — is better than a joint tool.
blocker_stale_days() now maps lane risk onto those windows. A flat 14 would have
been wrong in both directions: too aggressive for a low-risk pointer, and it treated
an admin PAT lane the same as one.
ungraded takes the shortest window, not the longest. ADR-0007 makes an absent grade
a defect and ADR-0008 makes a grade cover the whole path, so a lane nobody graded is
the one whose blocker deserves least trust. Encoding that as 60 days would have been
the fail-open default this repo already fixed once.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The 90-day --stale-days default on `warden route gaps` was not a loose threshold,
it was an inert one: the delegation register was created 2026-08-15, so it could
not have fired before November. It was inherited from the catalog pointer cadence
and applied to a claim with a completely different half-life.
Two changes. DEFAULT_BLOCKER_STALE_DAYS = 14 now governs interim blockers, while
DEFAULT_STALE_DAYS = 90 keeps governing pointer freshness -- "is this the right
owner and page" is quarterly, "has the owner answered" is not. 14 is calibrated
on blockers that actually cost something: ten days for the secrets-engine lanes,
one for RISK-F-0001, roughly fifty for FLEX-WP-0007.
The second change matters more. `reviewed` records when someone touched an entry,
which is indistinguishable from re-checking it -- six lanes read as freshly
reviewed today because I typed in them. `verified:` now says how the claim was
established, and asked-and-waiting explicitly does NOT count: that is the state
the secrets-engine blocker sat in for ten days while looking current. A lane in
that state is stale at zero days old, and key-cape-oidc-login proves it works.
8 of 14 interim lanes are honestly marked unverified rather than given a fresh
date they did not earn.
--fail-on-stale exits 3 for a cron or gate. No CI test on age: a date-triggered
failure breaks the build for whoever commits next instead of whoever owns the
blocker. The CI test is structural -- every interim lane must record how it was
verified -- so it fails on the commit that introduces the omission.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
test_catalog_gaps_lists_only_interim asserted openbao-api-key was an interim
cover. It encoded the same classification the entry did, so it defended the
wrong answer rather than catching it -- the second time this session a test has
held a judgement still (see ADR-0008).
The staleness test added in T03 did its job on the first real change: it failed
the moment the catalog moved without the generated artifact being re-emitted.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
railiance-platform asked for a generated list to consume instead of hand-
maintaining agent-high-risk-boundary. Hand-maintaining it is what let the two
lists drift for four lanes in RISK-F-0009.
19 high-risk lanes, 14 concrete data paths, 5 without a single KV address listed
separately so absence does not read as omission. Carries catalog_revision and a
dirty flag. fields is null where unestablished, never a one-element guess.
The header states plainly that this is an input and not a policy: railiance-
platform owns the deny set and may deny more, less, or dispute a grade. ADR-0002
survives the handoff.
Two CI tests guard staleness, because a consumer applies this to a live control.
Note the immediate consequence of T02: 2 uncovered against a policy they closed
to 0 yesterday.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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>
The operator token was not expired after all -- `bao policy read` succeeded, so
the deployed policy is now compared directly instead of the file. Three
corrections to RISK-F-0009, which was filed static:
1. Six high-risk lanes are uncovered, not eight. The finding counted
openbao-api-key (a path pattern) and ops-warden-warden-sign-token (a broker
grant, not KV) among the concrete uncovered paths, while its own prose said
the first was a pattern. Five lanes have no address for a policy to deny.
2. Coverage holds at 6 of 17 against the live policy.
3. The deployed policy has drifted from the file: the file denies
platform/workloads/core-hub/runtime, the server does not. No ops-warden lane
maps there so our numbers are unchanged, but it proves the file was never a
safe proxy for the server -- which is what the finding flagged as unconfirmed.
scripts/check_agent_read_boundary.py is the invariant RISK-F-0009 asked for
rather than a one-off audit: it fails when a high-risk lane has no corresponding
deny. Capabilities-only by construction -- it reads the policy document, never a
secret value, and never mints a token.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The mechanism behind RISK-F-0003 was sharper than the finding described.
is_high_risk was risk == "high", but risk was never absent at the model layer:
RouteEntry.risk carried a dataclass default of "standard". An omitted grade was
not unhandled, it was actively resolved to the permissive value — fail-open by
construction, which is why nothing ever warned.
The default is now "ungraded" and is_high_risk returns true for anything outside
an explicit low-risk vocabulary (standard / low / accepted). An omitted grade and
an unrecognised grade from a newer catalog both resolve to high, so the boundary
fails safe in both directions rather than reading an unknown value as permission.
test_every_repo_catalog_lane_is_explicitly_graded is the CI gate that stops an
ungraded lane being committed, per ADR-0007: absence is not a grade.
"accepted" is in the low-risk vocabulary deliberately, ready for the
maturity-derived default — an experimental-context lane may be explicitly
accepted, which is a graded decision rather than an omission.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
flex-auth's flex-auth-ops-warden pin (FLEX-WP-0016) TokenReviews the caller and
binds resource.system: ops-warden to system:serviceaccount:ops-warden:ops-warden.
policy.py posted /v1/check with no Authorization header, so the pin logs
"caller authentication warning" and can only run callerAuth.mode: warn — which,
under ADHOC-2026-08-17-T01, is exactly what blocks policy.enabled: true.
- policy.caller_auth (none | file | env | command) + src/warden/caller_identity.py:
token resolved per call, never cached, written, or logged (ADR-0002)
- both check_sign_policy and check_fetch_policy attach the bearer header; an
unobtainable token fails closed rather than retrying anonymously
- scripts/check_policy_caller_identity.py: read-only gate, prints length and a
truncated fingerprint only, distinguishes 401 (audience/binding) from 403
- example config: caller_auth block, and flex_auth_url corrected — it pointed at
flex-auth.flex-auth.svc, a Service that does not exist
- WARDEN-WP-0031, PolicyGatedSigning caller-identity section and flip sequence
Default stays mode: none, so behaviour is unchanged until an operator opts in.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every execution position is now explicit. Catalog entries carry
delegation.mode (permanent / native / interim) with intended owner and
blocker. warden route gaps lists the interim set. Promotion requires
the ownership question. Doctrine lives in AccessRouting.md; the
register was published to the named owner repos.
Ship posture-aware access planning: organization_posture=build (axis C),
catalog freshness warnings, warden plan verdicts, localhost founder desk,
and playbook/agent guidance that retire /tmp file-drop patterns.
Compose route catalog + handoff rather than a second routing layer.
- warden/mask.py: fingerprint()/mask_value() — presence, length, 8-char sha256
prefix; never the value.
- proxy.proxy_fetch_fingerprint + `warden access --fingerprint`: masked status view
(presence/length/hash) that emits no value, so it bypasses the T02 stdout guard.
Lets two parties compare sha256 prefixes to confirm a shared value without seeing
it (e.g. rotation landed).
- documented as defense-in-depth (raw bao bypasses it) in OperatorAccessAssist.md
and the module docstring.
- tests: tests/test_mask.py + CLI fingerprint test. 299 pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
RAILIANCE-WP-0011-T03: point reuse-surface-hub-write-token catalog,
playbook, and tests at bao kv get on platform/workloads/reuse/reuse-surface/runtime-secrets;
kubectl documented as break-glass only.
resolve_fetch_command used shlex.split, which treated `|` as a literal
argument — breaking reuse-surface-hub-write-token (kubectl | base64 -d).
Piped commands now run via shell=True with inherited stdio for --fetch.
REUSE_SURFACE_TOKEN is custody of the Railiance01 K8s secret reuse-surface-env,
not OpenBao. Workers blocked on hub register can now discover the lane via
warden route find and obtain it through the documented kubectl handoff.
Load coordination memory by default via ensure_memory_context on app bootstrap
and route/access flows; invalidate cache after episode writes. WARDEN_MEMORY=0
remains the opt-out. Document that warden memory activate is optional only.
RAILIANCE-WP-0009 T06 / RAILIANCE-WP-0010 T06 (CCR-2026-0002, CCR-2026-0003):
both OpenBao KV paths are live, ESO delivers the Secrets in cluster, and
positive/negative access verification is audit-logged. Catalog entries gain
concrete zero-placeholder handoffs (exec_capable, resolvable); draft tables
and playbook gates updated; routing tests repointed to still-draft lanes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add unified metadata-only audit.jsonl with secret-material guard, instrument
sign/access/worker paths, and expose warden activity CLI. Surface broker hint
when VAULT_TOKEN is unset, refresh INTENT/SCOPE docs, and add production
integration checklists plus catalog lane promotion playbook.
Document the railiance-platform credential broker as the owner-native path
for scoped VAULT_TOKEN needs. Add catalog entry, playbook, and doc updates
so warden route find ranks the broker lane first; manual export remains a
documented fallback only.
T4 (review→send loop): conservative tick persists structured drafts to
state_dir/worker-drafts.json; `warden worker drafts` lists them, `warden worker approve
<id> [--body …]` sends the reviewed draft as the reply + marks read + drops it. Escalated
plans persist no draft. Live-verified end-to-end.
T3 (visibility): `warden worker status` (pending drafts, triage count, last digest, timer
state); best-effort notify-send nudge in the tick when drafts are pending.
T5: wiki/playbooks/scheduled-worker.md (enable/disable, the approve loop, failure modes,
conservative-only posture) + SCOPE note.
WARDEN-WP-0021 finished: the conservative worker now runs on a systemd --user timer
(enabled, every 15 min), triages new inbox messages into drafts you approve with one
command, degrades gracefully, and stops with one command. 249 tests, lint clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Per Bernd's call: the guardrails prevent security harm but not LLM content errors, so the
worker should triage + draft, not auto-send, until reply quality is proven (matches the
build-stage/recoverability posture).
run_conservative triages NEW messages into a reviewed digest (state_dir/worker-digest.md)
with drafted replies, posts ONE progress note, tracks seen message ids (schedule-safe
dedup), and sends NOTHING to other agents / marks nothing read. `warden worker run
--execute` now runs this conservative tier; `--full-auto` opts into the auto-send path.
Live-verified with the LLM brain on the real inbox: produced a high-quality draft reply to
a secrets-engine coordination message and correctly flagged the llm-connect custody request
as NEEDS YOU. Conservative mode is safe to schedule (T4). 244 tests, lint clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
HubClient gains writes (mark_read, send_reply, add_progress). execute_plan/execute_plans
run the safe, allowlisted actions autonomously: route_answer (reply with the computed
answer + auto mark-read), reply (LLM-drafted body), progress_note, mark_read. Escalated
plans and non-auto-executable kinds are left for a human; every action is metadata-only
(no secret value read/sent/logged).
Deliberate guardrail: propose_catalog_diff and any code/routing change is NOT auto-executed
even under full-auto — a bad catalog commit could misroute credentials, so it goes to human
review (recoverability over convenience). AUTO_EXECUTABLE is the messaging/hub tier only.
`warden worker run --execute` runs the executor (dry-run still default). 7 executor tests
(reply+mark, with/without body, escalated skip, catalog-diff-left-for-human, progress,
failure-without-crash); 243 pass, lint clean. First live --execute shakedown is the
operator's (staged rollout); T4 schedules it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
llm-connect is operational (operator set OPENROUTER_API_KEY). Contract discovered from
the running service: POST /execute {"prompt":...} -> {"content":...}.
LlmConnectBrain embeds the fixed charter + the inbox message as untrusted data, calls
/execute, and parses a JSON action plan (_extract_json tolerates fences/prose), escalating
defensively on malformed/empty/transport errors. The build_plans guardrail still enforces
the allowlist + no-secret invariant on whatever the model returns — the LLM cannot widen
ops-warden's authority. `warden worker run --brain rule|llm` selects the planner.
Live-verified on the real inbox: the LLM brain planned a sensible reply+mark_read for a
secrets-engine coordination message and correctly escalated a secret-custody request as
out-of-lane — better classification than the deterministic RuleBrain.
6 new tests, 236 pass, lint clean. T3 (guarded executor) and T4 (scheduling) remain.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
build_plans now computes the concrete routing answer for each route_answer action
in-process (reuses the catalog; read-only, no subprocess/network) and render_plans
shows it as a `draft:` line. The dry-run demonstrates the actual answer the executor
(T3) will send, not just an intent. RuleBrain stays the default; the llm-connect brain
(T2) is gated on llm-connect being operational + its /execute contract.
230 tests, lint clean. Live dry-run verified against the real inbox.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
