--- id: WARDEN-WP-0038 type: workplan title: "warden plan must distinguish reading a secret from mutating one" domain: infotech repo: ops-warden status: finished owner: ops-warden topic_slug: netkingdom planning_priority: P1 depends_on_workplans: - WARDEN-WP-0029 created: "2026-09-08" updated: "2026-09-09" state_hub_workstream_id: "751ad530-e44b-52c4-b70e-cb47568a0179" --- # WARDEN-WP-0038 — Read versus mutate intent in the policy front door ## The defect `warden plan` scores a need against catalog lanes by keyword overlap and returns a verdict. It has no notion of what the caller wants to *do*. So a need that says *generate a successor secret and CAS-write it to two custodians* matches the lane that **reads** that path, scores well on the shared nouns, and returns: ``` verdict: autonomous commands: warden access --exec / --out / --wrap ``` Three read transports offered as the answer to a write, under the verdict that means "proceed without the founder". Reproduced 2026-09-08 against `rapp-qonto-keycape-client` (`ADHOC-2026-09-08-T04`). ## Why this is worse than a bad suggestion `warden plan` is the documented first step before drafting any founder credential step (`.claude/rules/credential-routing.md`), and `autonomous` is consumed as an authorization signal by agents that have no other check. `WARDEN-WP-0029` shipped the verdict *because* instinct was not auditable; a verdict that is confidently wrong in the permissive direction is worse than no verdict. key-cape hit this and refused to rely on it — "we are not treating a generic match as authorization" — which is the only reason it surfaced as a design question rather than as an unadmitted custody mutation. The control that caught it was a counterparty's caution. That is not a control. `WARDEN-WP-0033-T06` is the same shape (rapp-postgres, generic `openbao-api-key` match → `paste_once_provision` for an act that could not work) and was fixed by adding one exact lane plus one regression. That fixed the instance. This is the class. ## Shape of the fix (to be designed, not assumed) - Classify need intent — read / mutate / establish — before scoring, and never answer a mutate need with a read lane's transports. - A mutate need against a lane ops-warden does not own is `unroutable` with a CCR stub naming the custody owner, not `autonomous`. `ADR-0003`: cover gaps, never silently own them; a permissive verdict on someone else's custody act is a silent claim of ownership. - Consider whether `rotation.steps` should be reachable from `plan` at all, given they are advisory prose and several are explicitly not automatable. - Regression from the reproduced case, in the manner of T06. ## Open question for design Whether intent belongs in the matcher or in the lane. A lane could declare which acts it admits (`read`, `rotate`, `provision`) and `plan` could refuse any act a lane does not declare — stricter, and it fails closed on lanes that say nothing, which is the majority today. That is a catalog schema change and needs its own argument. ## Confirmed independently, then fixed — 2026-09-09 `railiance-platform` hit the same defect from the other side while answering the Qonto custody routing question (msg `7c7228ac`). Having stated that the correct verdict is `founder_required` — attended OIDC via `netkingdom` `role=platform-admin`, through the governed `openbao-platform-admin-login` lane — they added: > Note this is the same verdict your `warden plan` should have returned; it > returning `autonomous` and offering read transports for a write is a real > defect and WARDEN-WP-0038 is the right place for it. Until that is fixed, do > not let a plan result stand in for this answer. Two counterparties reporting the same wrong verdict in two days, one of whom is the write authority being wrongly bypassed, moved this from proposed to shipped. ### What shipped `_need_intent()` classifies a need as `mutate` or `read` by token match, and `_owns_write_authority()` tests `delegation.mode == "permanent"`. A mutating need on a lane ops-warden does not permanently own can no longer reach any branch that returns `autonomous`: - lane declares a `rotation` → `founder_required`, with an `approve` act naming the write owner, the rotation method, whether it is automatable, and `warden rotate-guide` as guidance; - lane declares none → `unroutable` with a CCR stub whose `owner_hint` names the write authority. **Commands are empty of read transports either way.** That is the half that made the wrong verdict actionable rather than merely wrong: `--out`, `--exec`, `--wrap` and a bare `bao kv get` were being offered as the answer to a write, and a test now asserts none of them appears. ### The design question in the plan, answered by use The plan asked whether intent belongs in the matcher or in the lane. It went in the matcher, with the *ownership* test doing the work that a verb list cannot: SSH certificate issuance is itself a mutating act, and `warden sign` must stay `autonomous`. `delegation.mode: permanent` is what separates ops-warden's own front door from someone else's custody — not the absence of a verb from a list. A regression asserts the SSH lane still proceeds, because a guard that refuses our own lane would be a worse defect than the one it fixes. The stricter alternative — lanes declaring which acts they admit, `plan` refusing anything undeclared — is not built. It is a catalog schema change that would fail closed on the majority of lanes, which say nothing today, and it needs its own argument rather than arriving as a side effect of this fix. ### Not claimed This narrows what `warden plan` will assert; it does not make the tool an authority on custody. railiance-platform's instruction stands and is the right standing posture: a plan result does not substitute for the owner's answer.