ops-warden/workplans/WARDEN-WP-0038-plan-mutation-intent.md
tegwick 11e5e8be0f chore: record hub ids for 2026-09-08 workplans
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
2026-09-08 14:57:42 +02:00

72 lines
3 KiB
Markdown

---
id: WARDEN-WP-0038
type: workplan
title: "warden plan must distinguish reading a secret from mutating one"
domain: infotech
repo: ops-warden
status: proposed
owner: ops-warden
topic_slug: netkingdom
planning_priority: P1
depends_on_workplans:
- WARDEN-WP-0029
created: "2026-09-08"
updated: "2026-09-08"
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 <lane> --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.