state-hub/docs/decision-services-alignment-answer.md
tegwick 042f3c9b71
All checks were successful
CI Smoke / host-smoke (push) Successful in 0s
CI Smoke / container-smoke (push) Successful in 2s
Decision services alignment S3: owner answer and STATE-WP-0101 (proposed)
Accept S3 with adaptations: write-back via the dispatch worker to one
authenticated route, hub status is never execution authority, kept
portable for retirement. 6.4: typed columns for enforced fields plus
typed evidence_refs for the rest.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 5502@bnt-lap001
Assistant-Session: 68b3b02f-5d0f-4e71-8421-796f62e50196
2026-10-31 23:02:27 +01:00

14 KiB

Decision services alignment: State Hub owner answer (S3)

Date: 2026-10-31 Answers: hub message 59390db6 (net-kingdom), proposal net-kingdom/docs/proposals/2026-10-31-decision-services-alignment-review.md (commit b764561), sections 3, 4.4, 5 (S3) and open question 6.4. Status: owner answer; design only. No API, schema or deployment change is made with this document. Implementation is proposed as workplans/STATE-WP-0101-decision-evidence-links.md.

1. Answer in short

S3 is accepted, with adaptations.

  • Hub decisions become the index of open questions. Decisions of class C1 or C2 are resolved only by evidence from informed-decision, never by free text. C0 stays as today.
  • Adaptation 1, transport: the write-back is carried by State Hub's existing private memo dispatch worker (STATE-WP-0096), which pulls the decided outcome from informed-decision with its existing intake-audience token and posts it to one new authenticated hub route. informed-decision does not get a hub credential, and the hub authenticates exactly one writer.
  • Adaptation 2, authority: a hub decision's resolved status is a projection of evidence held by informed-decision, approval-engine and audit-core. Nothing may execute on the strength of a hub decision status. A C2 executor checks the approval object, never the hub.
  • Adaptation 3, retirement: State Hub is in retirement (INTENT.md, prj-state-hub-retirement). The change is kept small and portable so it can move to the decisions owner in hub-core without reshaping. The decisions capability needs a keep/move disposition in the retirement ledger; this design does not claim permanent ownership.

Answer to 6.4: both, split by purpose. Typed columns for the fields the hub enforces, constrains or queries on; one typed evidence_refs list for everything else. Details in section 3.

2. What exists today (state-hub evidence)

  • decisions (api/models/decision.py): status open/resolved/escalated/superseded, decided_by free text (100 chars), rationale, no link to a memo or approval.
  • POST /decisions/{id}/resolve accepts {rationale, decided_by, write_log} from any caller. The only gate is a required multi-owner review contract, if one is attached.
  • PATCH /decisions/{id} can set status, decided_by and decided_at directly, which bypasses even that gate. POST /decisions/ can create a decision already resolved.
  • api/services/memo_forwarding.py already contains a resolution ledger (observe_resolution, run_due_resolution). It posts the free-text payload above to /resolve and recognises its own earlier resolution by searching the rationale for memo <id> . No production caller invokes it; that is the "disabled" write-back.
  • The dispatch worker runs as system:serviceaccount:state-hub:state-hub, has no database credential, and its NetworkPolicy allows egress only to DNS and informed-decision intake on 8081.
  • The hub API has no authenticated-caller boundary (STATE-WP-0096 rationale). STATE-WP-0100 adds claimed write attribution for all writes; its token verification is observe-only by founder decision of 2026-10-25.

3. Data model (answer to 6.4)

3.1 Columns on decisions (enforced, queried)

Column Type Purpose
decision_class enum C0,C1,C2,C3, nullable NULL means legacy and behaves as C0. Raise-only once set
memo_id text, nullable informed-decision memo that answers this question. Replaces the rationale search for correlating a 409
approval_id uuid, nullable approval-engine object; required to resolve C2
resolution_source enum text,informed_decision, nullable how it was resolved
resolution_evidence jsonb, nullable the verified write-back record (section 5), stored as received

Constraints (DB CHECK and API):

  • status = resolved AND decision_class IN (C1,C2) requires resolution_source = informed_decision, memo_id and resolution_evidence not null.
  • status = resolved AND decision_class = C2 additionally requires approval_id.
  • Unique partial index on memo_id where not null: one memo answers at most one decision.

Why columns here: these fields carry rules (who may resolve, what must be present), the conflict check needs an indexed lookup instead of a text search, and the dashboard and needs_human views filter by class.

3.2 evidence_refs (descriptive, not enforced)

evidence_refs: jsonb list, default [], items {type, ref, digest?} with a closed type vocabulary validated by the API: presentation, audit_record, ccr, workplan_task, adr, commit, document. Agents may add references to any decision class (for example the CCR or the workplan task that a decision concerns). They do not resolve anything and have no constraint attached. The vocabulary grows by owner change, not free form.

3.3 Why not only one of them

  • Only typed refs: the C1/C2 rule becomes "a ref of type X must exist", which a DB constraint cannot express cleanly, and the 409 correlation stays a scan.
  • Only columns: every new evidence kind would be a migration in a retiring service.

4. Resolution rules

Path C0 / legacy (NULL) C1 C2
POST /decisions/{id}/resolve (text) allowed, as today 409 evidence_required 409 evidence_required
PATCH setting status=resolved, decided_by, decided_at, rationale, memo_id, approval_id, resolution_* allowed for C0 (as today) refused refused
PATCH changing decision_class may raise may raise, never lower never lower
POST /decisions/ with status=resolved and class C1/C2 allowed for C0 refused refused
Supersede (DELETE) allowed allowed (withdraws the question; records who and why) allowed; does not revoke the approval object
POST /decisions/{id}/resolve-evidence (new) refused (C0 needs no evidence) authenticated writer only authenticated writer only, approval_id required
MCP resolve_decision as today gets the same 409 from the API same

Who sets the class: the creator, at creation or later by raising it. Doctrine for which class a question needs comes from gate-house (S5). Until S5, the hub suggests a floor only for the existing financial/legal auto-escalation keywords (C1 minimum). Linking a decision to a task with needs_human: true does not by itself raise the class.

The existing review-contract gate keeps applying on both resolve routes.

5. Write-back: who writes, how it authenticates, what informed-decision sends

5.1 Flow (pull through the existing dispatch worker)

  1. A C1/C2 hub decision is posed as a memo through the existing STATE-WP-0096 path. The envelope already carries serves: {hub_decision_id, resolves: on-answer} and the stable idempotency key state-hub.<decision_id>.<memo_id>. At enqueue the worker records memo_id on the decision through the new route's link mode (section 5.3), so the hub shows what is waiting where.
  2. The founder decides in informed-decision (accept or decline, at the assurance the class needs).
  3. The worker reads the outcome of its own submitted memos from informed-decision with its existing intake-audience token. informed-decision answers only for memos this submitter posed.
  4. The worker validates the outcome record (section 5.2) and posts it to POST /decisions/{id}/resolve-evidence.
  5. The hub verifies the caller, checks the record against the decision (class, memo_id, approval_id, hub_decision_id), stores it, sets resolved, writes a decision_resolved progress event and publishes org.statehub.decision.resolved with the evidence fields added.

Why pull rather than informed-decision pushing: informed-decision then needs no hub credential and no egress to the hub; the hub admits one writer it already deploys; the worker already holds the retry, replay and conflict ledger. If informed-decision prefers push, the same route works with informed-decision's workload subject allowlisted instead; that is a founder choice (section 8).

5.2 What informed-decision must send (outcome record, informed-decision.decision-outcome v1)

Field Required Notes
kind, version yes informed-decision.decision-outcome, 1
hub_decision_id yes from the memo's serves; must equal the route's id
memo_id, memo_revision yes the revision that was shown
verb yes accept or decline (terminal only; other verbs do not resolve)
option C1 with options the chosen option id
subject yes authenticated human subject: issuer + sub from the KeyCape session, not a display name
assurance yes acr/aal and authentication time of the deciding session
decided_at yes RFC 3339 UTC
view_hash yes sha256:<64 hex> of what was presented
presentation_id yes informed-decision presentation record
approval_id, approval_entry_id C2 the approval-engine object and the human entry
audit_ref when available audit-core record id
display_name optional used only to fill decided_by for display
note optional bounded human text (max 2 KiB), secret-scanned like memo content

Rules: no secrets, no tokens, no packet content; the record is bounded at 8 KiB as the existing ledger already enforces. A decline resolves the decision with outcome declined; it is an answer.

5.3 How the route authenticates (who may resolve with evidence)

  • POST /decisions/{id}/resolve-evidence requires a Kubernetes projected ServiceAccount token with a dedicated audience (state-hub-decision-writeback), verified by the API through TokenReview.
  • The verified subject must equal an allowlist of exactly one entry: system:serviceaccount:state-hub:state-hub (the dispatch worker's identity, selected by the founder in STATE-WP-0096). No other caller, no agent and no human browser session can use it.
  • The route has two modes: link (records memo_id, and approval_id for C2, while the decision stays open) and resolve (section 5.1 step 5). Both are idempotent on (decision_id, memo_id); a second resolve with the same memo returns 200, a different memo returns 409 with the linked memo_id.
  • decided_by is filled from display_name (or subject) for compatibility, but the authoritative fields are resolution_evidence.subject and assurance. The hub records the human subject as attested by informed-decision through the authenticated worker; it does not authenticate the human itself.
  • This is an enforced route, distinct from STATE-WP-0100's claimed, observe-only attribution of ordinary writes. When STATE-WP-0100 phase 3 lands, the same token verification can be shared.

Prerequisites outside this repository: TokenReview needs the API's ServiceAccount bound to system:auth-delegator (cluster RBAC, railiance-cluster); the worker needs a second projected token audience and NetworkPolicy egress to the hub API Service; the API's ingress policy must admit the worker. If TokenReview RBAC is refused, the alternative is verifying the token signature against the cluster's service-account issuer JWKS.

6. Migration and backward compatibility

  • One additive Alembic migration: nullable columns, evidence_refs defaulting to [], the CHECK constraints and the partial unique index. Existing rows get decision_class = NULL and behave as C0; no backfill, no reclassification of past decisions.
  • DecisionRead gains optional fields; existing clients (MCP, dashboard, statehub, consistency checks, recently-on-scope) keep working. DecisionResolve is unchanged for C0.
  • New refusals apply only to decisions explicitly classed C1 or C2. Agents that resolve C0/legacy decisions by text see no change.
  • memo_forwarding resolution ledger: the text payload (decided_by + rationale search) is replaced by the outcome record and the memo_id column. Its tests are rewritten accordingly. The rationale-text correlation is retained only for reading pre-migration rows.
  • Rollout order: schema and refusals first (inert until a decision is classed); then the route with its token check; then worker pull and write-back; then classing the first real decisions. Rollback leaves the columns and stored evidence in place.
  • Event envelope org.statehub.decision.resolved gains decision_class, memo_id, approval_id, resolution_source; existing attributes are kept.
  • For retirement: the column set and the outcome record are the portable contract; hub-core's decisions owner should adopt them unchanged.

7. What is not changed

  • No public enqueue route and no automatic scanning of decisions: posing a memo for a decision remains an explicit act, as in STATE-WP-0096.
  • The hub does not create, read or consume approvals and gains no approval-engine scopes.
  • The hub does not decide class doctrine (gate-house, S5) or present anything (informed-decision).

8. Founder decisions needed

  1. Pull or push: the dispatch worker pulls outcomes (recommended) or informed-decision pushes to the hub with its own workload identity.
  2. First enforced hub route: admit TokenReview-based enforcement on this one route, ahead of STATE-WP-0100's observe-only phase 3, including the system:auth-delegator binding.
  3. Class floor before S5: whether the existing financial/legal keyword escalation should set C1 as a minimum, or classes stay purely declared until gate-house doctrine exists.
  4. Retirement disposition: confirm the decisions capability is "keep, move to hub-core" so this work is not stranded.

9. Asks of other owners

  • informed-decision: an outcome read for a submitter's own memos returning the record in section 5.2; treat serves.hub_decision_id as the correlation key; keep resolves: on-answer semantics (accept/decline only).
  • approval-engine: none for S3. The hub only stores approval_id.
  • gate-house: S5 class doctrine, including who may raise a class.
  • net-kingdom / railiance-cluster: RBAC and NetworkPolicy prerequisites in section 5.3.