aravind changes

README.md

@@ -4,13 +4,22 @@ The backend for Krow — a Go API over PostgreSQL, with a Python Owliver service
 to follow. The Krow frontend lives in a **separate repository** (`krow-demo`)
 and is not vendored, copied or modified here.
 
-**Status: Phase 3C — session authentication.** The API serves the endpoints in
+**Status: authenticated, authorized, multi-tenant.** The API serves the
-`docs/api-contract.md` against PostgreSQL, loaded with the frontend's own demo
+endpoints in `docs/api-contract.md` against PostgreSQL, loaded with the
-dataset. Every endpoint except `GET /health` and the two `/api/v1/auth/*` routes
+frontend's own demo dataset. Every endpoint except `GET /health` and the two
-requires a session: sign in with a password, hold an HttpOnly cookie, and the
+`/api/v1/auth/*` routes requires a session: sign in with a password, hold an
-server resolves it to a real user on every request. Authorization — which roles
+HttpOnly cookie, and the server resolves it to a real user on every request.
-may do what — is Phase 3D and is not implemented. No Owliver, no RAG, no Redis,
+Roles are enforced — a deny-by-default policy table answers 403 before any query
-NATS or S3.
+runs, and organization scope and talent ownership are SQL predicates, so a row
+you may not see answers 404. On top of that sits the agent and skill definition
+surface: Markdown with YAML frontmatter, parsed and stored, with full CRUD.
+
+Not built: no Owliver, no RAG, no Redis, no object storage. The agent/skill
+runtime is an internal boundary with no executor behind it and no HTTP route
+into it.
+
+`docs/KROW_BACKEND_COMPLETE_SUMMARY.md` is the long-form technical account —
+architecture, history, and a frank list of the current limitations.
 
 ## Layout
 
@@ -19,27 +28,39 @@ krow-backend/
 ├── go-api/
 │   ├── cmd/
 │   │   ├── api/            the HTTP service entrypoint
-│   │   └── seed/           loads the demo dataset
+│   │   ├── seed/           loads the demo dataset
+│   │   └── setpassword/    the only way a password enters the database
 │   ├── internal/
 │   │   ├── config/         environment loading + validation
 │   │   ├── db/             pgx pool and the database health check
-│   │   ├── domain/         resource descriptors (generated from the schema)
+│   │   ├── domain/         resource descriptors (generated) + the policy table
 │   │   ├── repo/           pgx access layer — every statement built here
 │   │   ├── service/        validation, org scoping, contract semantics
+│   │   ├── auth/           argon2id passwords, sessions, the user store
+│   │   ├── authctx/        the authenticated identity a request carries
 │   │   ├── orgctx/         the organization a request runs as
+│   │   ├── definition/     the agent/skill Markdown + YAML frontmatter parser
+│   │   ├── runtime/        agent/skill load, validate, resolve, dispatch
 │   │   ├── httpserver/     router, handlers, /health, graceful shutdown
 │   │   ├── seeder/         fixture loading + the attendanceSeed.js port
 │   │   └── testutil/       disposable migrated + seeded test database
 │   └── go.mod
 ├── migrations/             SQL migrations — the source of truth for the schema
 ├── seed/fixtures/          seed.json, generated from the frontend repository
-├── docs/api-contract.md    the frozen client contract
+├── docs/
+│   ├── api-contract.md     the frozen client contract
+│   └── KROW_BACKEND_COMPLETE_SUMMARY.md   the long-form technical account
 ├── infrastructure/         deployment definitions (empty until later)
-├── scripts/                verify_schema.sql, gen_resources.py
+├── scripts/                verify_schema.sql, gen_resources.py, oracle.mjs
 ├── Makefile                developer, migration and seed tasks
 └── .env.example
 ```
 
+`migrations/`, `seed/`, `scripts/` and `infrastructure/` sit beside `go-api/`
+rather than inside it because none of them is Go: the migrations are run by the
+`golang-migrate` CLI, the generator is Python, the parser oracle is Node, and a
+second service (Owliver, Python) is expected to consume the same schema.
+
 ## Architecture
 
 ```
@@ -82,10 +103,16 @@ make run                  # http://127.0.0.1:8080/health
 
 ## The API
 
-38 endpoints, specified in `docs/api-contract.md`. Every one exists because a
+54 registered routes: 34 entity routes across 14 resources, 10 agent and skill
-frontend call site exists — a table in the database is never a reason for an
+definition routes, 4 `/me` routes, 2 `/auth` routes, 2 workflow routes, the
-endpoint. `DELETE /job-postings/{id}` and `POST /shift-records` return 405
+Owliver suggestion route and `GET /health`. The entity routes and the suggestion
-because nothing in the frontend deletes a posting or creates a shift record.
+route are specified in `docs/api-contract.md` (§2 and §2A); the definition and
+workflow routes are not yet in the contract and are documented in the source and
+its tests. Every
+entity route exists because a frontend call site exists — a table in the
+database is never a reason for an endpoint. `DELETE /job-postings/{id}` and
+`POST /shift-records` return 405 because nothing in the frontend deletes a
+posting or creates a shift record.
 
 ```bash
 curl localhost:8080/api/v1/job-postings
@@ -110,8 +137,19 @@ Set a password before signing in: the seeded user's `password_hash` is NULL unti
     curl -b jar localhost:8080/api/v1/me
     curl -b jar -X POST localhost:8080/api/v1/auth/logout
 
-Authorization is **not** implemented. `users.role` is carried on the identity and
+**Authorization is enforced.** `users.role` — `admin`, `employer` or `talent`,
-consulted nowhere: any signed-in user reaches every endpoint. That is Phase 3D.
+never the self-editable `account_type` — is the only authority. Entity routes
+are gated by the deny-by-default policy table in `internal/domain/policy.go`:
+a resource with no policy permits nothing to anyone, and `Server.authorize`
+answers 403 before any query runs. Row visibility is a SQL predicate rather than
+a filter — the organization scope always, plus an ownership clause for talent
+callers — so a row outside it is never fetched and answers 404, which does not
+distinguish "exists but not yours" from "does not exist".
+
+The ten definition routes are the exception: they are not `domain.Resource`
+values, so the descriptor machinery does not reach them and their role checks
+are written inline in `internal/service/definitions.go`. Tenancy and ownership
+*are* enforced for them, in the repository predicates.
 
 ## Seeding
 
@@ -135,16 +173,30 @@ records created through the API survive a re-seed.
 make test        # go test ./...
 ```
 
-55 tests. The database-backed ones build a disposable database — dropped,
+175 test functions. The database-backed ones build a disposable database —
-recreated, migrated and seeded per run, named `krow_backend_autotest_<pid>` so
+dropped, recreated, migrated and seeded per run, named
-concurrent test packages cannot collide. They skip rather than fail when
+`krow_backend_autotest_<pid>` so concurrent test packages cannot collide. They
-PostgreSQL is unreachable.
+skip rather than fail when PostgreSQL is unreachable.
 
 Seed assertions compare the database against the fixture field-by-field rather
 than against numbers typed into a test. The two ordering guarantees
 (`NULLS LAST` in both directions, and the `, id` tiebreaker) are covered by tests
 verified to fail when the guarantee is removed.
 
+**Parser conformance.** The Go definition parser must agree with the frontend's
+JavaScript one. `internal/definition/conformance_test.go` asserts against
+`internal/definition/testdata/oracle.json`, which is not written by hand: it is
+captured by running the real frontend module graph through Vite, so the fixture
+records what the JS parser actually does rather than what anyone believes it
+does. Regenerating it needs a `krow-demo` checkout, and is not part of
+`make test`:
+
+```bash
+node scripts/oracle.mjs go-api/internal/definition/testdata/oracle.json
+```
+
+`scripts/cases.mjs` holds the adversarial corpus that file is captured over.
+
 `GET /health` is public and says only whether traffic should be sent here:
 `{"status": "ok"}` with `200`, `{"status": "degraded"}` with `200` when the
 database is up but unmigrated or left dirty, and `{"status": "unavailable"}`

docs/KROW_BACKEND_COMPLETE_SUMMARY.md

@@ -1840,7 +1840,7 @@ Recorded in `docs/api-contract.md` §12 and still accurate against the current c
 | **GCS-compatible object storage** | — | **Future direction.** No reference in the repository; `infrastructure/README.md` names MinIO as a "later" candidate | — |
 | **Docker deployment** | — | **Future direction.** `infrastructure/` contains only a README; `Dockerfile.api` and `Dockerfile.owliver` are listed there as later work | — |
 | **Redis** | Shared state for rate limiting across instances | **Future direction.** `ratelimit.go` names Redis or the database as the seam; nothing is wired | Needed before multi-instance deployment for the limiter to mean anything |
-| **NATS is not part of the target architecture** | — | **Future direction / stated rule.** Note the discrepancy: `infrastructure/README.md` currently lists NATS as a later docker-compose candidate. See §17. | Remove NATS from that table if the rule stands |
+| **NATS is not part of the target architecture** | — | **Future direction / stated rule.** `infrastructure/README.md` previously listed NATS as a later docker-compose candidate; that table has been corrected. See §17.17. | Keep it out |
 | **RAG / pgvector** | — | **Future direction.** `pgvector` is not installed in the local database and is named only once, in `infrastructure/README.md` | Introduce when the architecture calls for it |
 
 ---
@@ -1850,35 +1850,31 @@ Recorded in `docs/api-contract.md` §12 and still accurate against the current c
 Everything below was checked against the current repository. Items the request asked
 about that could not be verified are marked as such rather than guessed at.
 
-### 17.1 Documentation is behind the code
+### 17.1 Documentation is behind the code — resolved
 
-**Issue.** `README.md` describes the repository as being at an earlier stage than the
+**Issue as recorded.** `README.md` described the repository as being at an earlier
-code is. It states "Authorization — which roles may do what — is Phase 3D and is not
+stage than the code is. It stated "Authorization — which roles may do what — is Phase
-implemented", "Authorization is **not** implemented. `users.role` is carried on the
+3D and is not implemented", "Authorization is **not** implemented. `users.role` is
-identity and consulted nowhere: any signed-in user reaches every endpoint", "38
+carried on the identity and consulted nowhere: any signed-in user reaches every
-endpoints" and "55 tests".
+endpoint", "38 endpoints" and "55 tests". The same staleness appeared in two source
+comments: the `internal/httpserver/server.go` package documentation ("Authorization is
+NOT here"), and `internal/authctx/authctx.go` ("It is NOT consulted anywhere in Phase
+3C: authentication only").
 
 **Cause.** Authorization, the definition tables, the CRUD surface and the runtime all
 landed after the README was last revised.
 
-**Impact.** A reader trusting the README would conclude that any signed-in user
-reaches every endpoint, which is not what the code does.
-
 **Current behaviour.** `internal/domain/policy.go`, `Server.authorize`,
 `builder.ownership` and `internal/httpserver/rbac_test.go` (730 lines) all exist and
-run. 51 routes are registered. 175 test functions run.
+run. 51 routes are registered. 175 test functions run. `docs/api-contract.md` §9A
-`docs/api-contract.md` §9A *is* current and documents the authorization contract
+*is* current and documents the authorization contract accurately.
-accurately.
 
-**Possible future handling.** Revise `README.md` against the code.
+**Resolution.** `README.md` was revised against the code: the status paragraph, the
-
+layout tree (which was missing `cmd/setpassword`, `internal/auth`, `internal/authctx`,
-The same staleness appears in two source comments:
+`internal/definition` and `internal/runtime`), the route count, the test count and the
-
+authorization section. The `server.go` and `authctx.go` package comments were
-- `internal/httpserver/server.go` package documentation: "Authorization is NOT here.
+corrected to describe the authorization that exists. `internal/orgctx/orgctx.go`,
-  A signed-in user reaches every endpoint they could reach before".
+which still framed itself as pre-authentication, was corrected in the same pass.
-- `internal/authctx/authctx.go`: "It is NOT consulted anywhere in Phase 3C:
-  authentication only" — `Identity.Role` is now consulted by `Server.authorize`,
-  `builder.ownership`, `guardInsert` and `service/definitions.go`.
 
 ### 17.2 The definitions endpoints are not in the API contract
 
@@ -2000,23 +1996,26 @@ the effective limit multiplies by instance count, resets on restart, and degrade
 a global budget behind a reverse proxy. **Possible future handling:** move the
 limiter behind shared state before deploying more than one instance.
 
-### 17.10 The repository has no commits
+### 17.10 The repository has almost no history
 
-**Issue.** `git log` reports `fatal: your current branch 'main' does not have any
+**Issue as recorded.** `git log` reported `fatal: your current branch 'main' does not
-commits yet`. Every file is untracked.
+have any commits yet`, and every file was untracked.
 
-**Impact.** There is no history, no recovery point, no blame, and no record of when
+**Current behaviour.** The tree is now committed: a single commit on `main`
-any of the work described in §2 happened. The chronology in this document was
+(`7d12ebe`, "first commit") holds the whole repository.
-reconstructed from migration headers, package documentation and the contract, not
-from version control.
 
-**Note.** This document does not change that; no commit was made.
+**Remaining impact.** One commit is not history. There is still no blame, no
+incremental recovery point, and no record of when any of the work described in §2
+happened. The chronology in this document was reconstructed from migration headers,
+package documentation and the contract, not from version control, and that remains
+the only source for it.
 
-### 17.11 A minor documentation defect in the source
+### 17.11 A minor documentation defect in the source — resolved
 
-`internal/httpserver/api.go` — the doc comment describing `decodeBody` sits
+`internal/httpserver/api.go` — the doc comment describing `decodeBody` sat
-immediately above `decodeInto`, so `decodeInto` carries two doc comments and
+immediately above `decodeInto`, so `decodeInto` carried two doc comments and
-`decodeBody` carries none.
+`decodeBody` carried none. The `decodeBody` comment has been moved to sit above
+`decodeBody`.
 
 ### 17.12 Badge endpoint mismatch — verified
 
@@ -2091,8 +2090,9 @@ among the things that do not exist yet.
 
 **Impact.** A reader of `infrastructure/README.md` would take NATS to be planned.
 
-**Possible future handling.** Amend that table if the rule stands. Nothing was changed
+**Resolution.** The rule stands, so the table was amended: `infrastructure/README.md`
-here.
+no longer lists NATS as a candidate and says explicitly that it is not part of the
+target architecture. `README.md` no longer names it either.
 
 ### 17.18 Deliberate contract behaviours that read as defects
 
@@ -2169,8 +2169,8 @@ deliberately and should be done knowingly.
 
 ### NATS
 
-**NATS is not part of the target architecture.** Note that `infrastructure/README.md`
+**NATS is not part of the target architecture.** `infrastructure/README.md` no longer
-currently lists it as a later candidate; see §17.17.
+lists it as a candidate; see §17.17.
 
 ### Nearest-term work implied by the code itself
 
@@ -2281,9 +2281,9 @@ disposable `krow_backend_autotest_<pid>` database per test process.
 **Current Known Limitations:** no AI execution and no public runtime endpoint; login
 rate limiting is per-process and in-memory; CORS does not permit credentials;
 definitions endpoints bypass the policy table; the runtime package is unreachable
-from the running service; `README.md` and two source comments are behind the code;
+from the running service; the definitions endpoints are absent from the API contract;
-the definitions endpoints are absent from the API contract; attendance data is empty
+attendance data is empty without a rostering source; the repository has a single
-without a rostering source; the repository has no commits.
+commit and so no usable history.
 
 **Current Development Focus:** the most recent work in the tree is the definition
 system — schema, parser conformance, CRUD — and the runtime boundary that sits on top
@@ -2392,13 +2392,12 @@ is how this codebase would stop being trustworthy.
 - **The main current limitations** are: no AI execution and no runtime endpoint;
   login rate limiting that does not survive scale-out; CORS that does not yet permit
   credentials; ten definition endpoints that sit outside the policy table and outside
-  the written contract; documentation that is behind the code; and a repository with
+  the written contract; and a repository with a single commit and so no usable
-  no commits.
+  history.
 - **The direction** is real execution behind the existing executor interfaces
   (Owliver / an LLM provider abstraction), then retrieval, then tooling, then
   deployment infrastructure — none of which exists here today. **NATS is not part of
-  the target architecture**, and the one place in the repository that still names it
+  the target architecture**, and the documentation no longer implies otherwise.
-  should be corrected.
 - **The rules that must not be broken** are in §21. The two that matter most in daily
   work: never trust a client-supplied identity or tenant, and never rewrite stored
   Markdown.
@@ -2409,3 +2408,9 @@ is how this codebase would stop being trustworthy.
 behaviour above was read from the current checkout or from read-only queries against
 the local development database. Nothing in the repository was modified to produce
 this document.*
+
+*Revised on 2026-08-24 by a structure-and-dead-code cleanup pass, which changed no
+schema, no migration, no database row and no runtime behaviour. What it did change is
+recorded in §17.1, §17.10, §17.11 and §17.17: documentation and source comments that
+had fallen behind the code were corrected. The counts above were re-verified against
+the checkout and are unchanged.*

docs/api-contract.md

@@ -129,6 +129,125 @@ these would leave the shim with methods that 404. See §11 (D6).
 
 ---
 
+## 2A. Owliver suggestions
+
+`GET /api/v1/owliver/suggestions?page={surface}&query={typed}`
+
+The one endpoint here that serves no resource. It answers "what could I usefully
+ask on this page?" for the Owliver panel, which calls it while the user types —
+so it reads no table, opens no transaction, calls no model, and its whole answer
+is computed from a static catalogue in `internal/owliver`.
+
+It is not on the public allowlist. Which readings exist depends on the caller's
+role, so there is no anonymous answer to give.
+
+### Request
+
+| Parameter | Required | Meaning |
+| --- | --- | --- |
+| `page` | yes | A surface id from the closed page vocabulary — the same one `internal/definition` validates a definition's `pages:` against. Aliases (`hired`, `forge`, `new-position`) and loose spellings (`Talent Pool`) resolve to the canonical id. |
+| `query` | no | What the user has typed so far. |
+
+Any other parameter is `invalid_query`. **Nothing about the caller is accepted
+here** — role, organization and user are read from the session, and a request
+that names one is refused rather than ignored.
+
+An unknown `page` is `invalid_query`, with the frontend's own wording:
+`Unsupported page: {value}. Supported pages: {…}.` An absent, blank or
+too-short `query` is **not** an error — there is simply nothing to rank yet.
+
+### Response
+
+```json
+{
+  "data": {
+    "suggestions": [
+      { "text": "Which position has the strongest pipeline?", "intent": "position-strength" },
+      { "text": "Show hiring activity as a flow", "intent": "hiring-operations", "capability": "flow" }
+    ]
+  }
+}
+```
+
+`data.suggestions` is always an array — `[]` when nothing matches, never `null`
+and never an error. There is no `meta`: the list is capped rather than paged.
+
+| Field | Meaning |
+| --- | --- |
+| `text` | The question, as the user reads it. |
+| `intent` | The **frontend capability id** the panel dispatches on, verbatim from the manifests in `src/components/ai-assistant/capabilities/`. Not a backend identifier, and never invented here. |
+| `capability` | The section type to draw the answer as, from the closed `OWLIVER_CAPABILITIES` vocabulary. Present only when the query asked for one ("as a flow", "summarize"); absent otherwise. |
+
+Nothing internal is exposed: no matching terms, no resource names, no scores, no
+policy detail. `text` is always catalogue wording — no part of the query is
+echoed back into a suggestion.
+
+### Selection
+
+Five stages, each of which only ever removes:
+
+```
+the page's catalogue → permission → relevance → deduplicate → top 3
+```
+
+- **Page.** Intents are keyed by surface, so a suggestion from another page
+  cannot appear. The same word answers differently per page by construction:
+  `pipeline` on `positions` is about which role converts, on `candidates` it is
+  the funnel the applicants are in.
+- **Permission.** Each intent declares the resources it reads, and those are
+  checked against the policy table in §9A — *before* ranking, so a refused
+  reading is never scored. Two gates, not one: the role must be allowed the
+  operation, and for an organization-wide reading the role's rows must not be
+  narrowed. Talent may list job applications; talent may not be offered "which
+  position has the strongest pipeline?", because their view of that resource is
+  their own rows. No role list is written down here — see §9A.1.
+- **Relevance.** Deterministic keyword ranking over the intent's own terms.
+  Exact token, then multi-word phrase, then prefix (so a half-typed word still
+  matches), then extension. Ties break on catalogue order, so the same request
+  always answers identically. A query naming only a section type ranks the
+  page's readings; once it names a subject, readings that merely *support* that
+  shape are dropped rather than used as padding.
+- **Deduplicate.** One suggestion per intent id, and no two with the same text.
+- **Cap.** Three. Nothing is added to reach three.
+
+A query with fewer than two letters or digits after normalization returns `[]`.
+
+### Normalization
+
+The query is truncated to 200 characters, lower-cased, and reduced to letters,
+digits and single spaces — every other character becomes a space rather than
+being stripped, so nothing can be glued into a token that was not typed as one.
+
+There is no injection surface to defend: the normalized text is compared against
+a fixed table of literals and never reaches SQL, a template, a shell or a log
+message. Hostile input is ranked like any other text, and can only ever produce
+entries the catalogue already holds.
+
+### Where the catalogue comes from
+
+The frontend owns the vocabulary. Owliver's capabilities are declared per page
+context in `src/components/ai-assistant/capabilities/`; `internal/owliver`
+transcribes the id and the page, and adds the two things a manifest does not
+carry — the words that mean a user is reaching for that reading, and the records
+it reads. Same pattern as `internal/definition/vocabulary.go`, and for the same
+reason: one vocabulary, named from both ends, with a test on this side that
+fails when an id here names no capability there.
+
+**No table, no migration.** The suggestions are derived from definitions that
+already exist. Persistence would only be warranted if suggestions became
+admin-managed, and nothing in the product asks for that today.
+
+### Not covered
+
+The seven workspace and configuration surfaces — `settings`, `workspace`,
+`workspace-agents`, `workspace-skills`, `workspace-skill-configure`,
+`skill-development`, `workspace-agent-configure` — are valid pages with no
+entries. Their panel answers from the registries rather than from workforce
+records, so there is nothing to rank a typed query against. They return `[]`,
+which is the honest answer, not a validation error.
+
+---
+
 ## 3. Request schemas
 
 ### 3.1 Create — `POST /{resource}`

go-api/internal/authctx/authctx.go

@@ -28,10 +28,12 @@ var ErrNoIdentity = errors.New("no authenticated identity in context")
 
 // Identity is who the request is, as resolved from the session row.
 //
-// Role is carried because Phase 3D will need it, and because carrying it now
+// Role is read once per request by the middleware, out of the user row, so an
-// means the middleware reads it once per request instead of every future
+// authorization check never has to re-query. It is the authorization authority:
-// authorization check re-querying the user. It is NOT consulted anywhere in
+// httpserver.Server.authorize gates operations on it, the repository's ownership
-// Phase 3C: authentication only.
+// predicate narrows a talent caller's rows by it, and service/definitions.go
+// checks it on every definition write. AccountType is NOT an authority — a user
+// can change their own through PATCH /me.
 type Identity struct {
 	UserID      string
 	OrgID       string

go-api/internal/config/config.go

@@ -259,8 +259,13 @@ func (c *Config) validate() error {
 // devCORSOrigins are the origins the Vite dev server can occupy. Vite binds
 // localhost by default and 127.0.0.1 when asked, and a browser treats those two
 // as different origins, so both are listed. 4173 is `vite preview`.
+// 5174 is where Vite lands when 5173 is already taken, which happens whenever a
+// second dev server is started; an origin missing from this list is refused at
+// the preflight with a bare 403 and no CORS headers, which reads as a server
+// fault rather than a misconfigured port.
 var devCORSOrigins = []string{
 	"http://localhost:5173", "http://127.0.0.1:5173",
+	"http://localhost:5174", "http://127.0.0.1:5174",
 	"http://localhost:4173", "http://127.0.0.1:4173",
 }
 

go-api/internal/httpserver/api.go

@@ -1,6 +1,7 @@
 package httpserver
 
 import (
+	"context"
 	"encoding/json"
 	"io"
 	"net/http"
@@ -130,7 +131,7 @@ func (s *Server) handleCreate(svc *service.Service) http.HandlerFunc {
 			writeError(w, s.log, err)
 			return
 		}
-		rec, err := svc.Create(r.Context(), ident, body)
+		rec, err := s.create(r.Context(), svc, ident, body)
 		if err != nil {
 			writeError(w, s.log, err)
 			return
@@ -139,6 +140,25 @@ func (s *Server) handleCreate(svc *service.Service) http.HandlerFunc {
 	}
 }
 
+// create inserts through the resource's own service, except where creating a
+// record has a consequence in another table.
+//
+// One resource has one: an AI interview is only half of completing an
+// interview, and the application it names has to be linked in the same
+// transaction — see internal/service/interviews.go for why the server performs
+// that write and the caller may not. Routing it here rather than registering a
+// second endpoint keeps POST /api/v1/ai-interviews the only way to write one,
+// which is what the client already calls and what the ownership guard already
+// covers.
+func (s *Server) create(ctx context.Context, svc *service.Service,
+	ident authctx.Identity, body domain.Record) (domain.Record, error) {
+
+	if svc.Resource().Path == service.InterviewsPath {
+		return s.workflows.CreateInterview(ctx, ident, body)
+	}
+	return svc.Create(ctx, ident, body)
+}
+
 func (s *Server) handleUpdate(svc *service.Service) http.HandlerFunc {
 	return func(w http.ResponseWriter, r *http.Request) {
 		ident, ok := s.authorize(w, r, svc, domain.OpUpdate)
@@ -174,11 +194,6 @@ func (s *Server) handleDelete(svc *service.Service) http.HandlerFunc {
 	}
 }
 
-// decodeBody reads a JSON object body.
-//
-// DisallowUnknownFields is not used — the target is a map, so every field is
-// "known" here. Unknown *columns* are rejected in the service, where the
-// resource's schema is available to say which those are.
 // decodeInto reads a JSON body into a typed struct.
 //
 // Beside decodeBody rather than replacing it: the resource handlers genuinely
@@ -200,6 +215,11 @@ func decodeInto(r *http.Request, dst any) error {
 	return nil
 }
 
+// decodeBody reads a JSON object body.
+//
+// DisallowUnknownFields is not used — the target is a map, so every field is
+// "known" here. Unknown *columns* are rejected in the service, where the
+// resource's schema is available to say which those are.
 func decodeBody(r *http.Request) (domain.Record, error) {
 	defer func() { _ = r.Body.Close() }()
 	raw, err := io.ReadAll(http.MaxBytesReader(nil, r.Body, maxBodyBytes))

go-api/internal/httpserver/auth.go

@@ -42,27 +42,30 @@ const sessionCookieName = "krow_session"
 // that never authenticate anything.
 func (s *Server) secureCookies() bool { return s.cfg.AppEnv != "development" }
 
-// sameSite resolves the configured SameSite mode.
+// sessionSameSite reports the SameSite mode the session cookie must carry.
 //
-// Lax remains the default and the recommendation. "none" exists for the one
+// Lax is the default and the safer value: it closes the CSRF hole by refusing
-// deployment shape that cannot work without it: a frontend on a different
+// to travel on cross-site subresource requests. That is exactly right when the
-// registrable domain from the API. In that case Lax withholds the cookie on
+// page and the API share an origin, which is the supported deployment.
-// every cross-site fetch, so the sign-in succeeds, the Set-Cookie arrives, and
-// the next request carries nothing — which reads as a broken session rather
-// than as a cookie policy.
 //
-// An unrecognised value falls back to Lax rather than to None. config.validate
+// When the API is configured with a CORS allowlist, the deployment is by
-// rejects those before startup, so this is only a belt-and-braces default in
+// definition the other one: a page on some other origin calls this API
-// the safe direction.
+// directly. A Lax cookie is never sent on those requests, so login would
-func (s *Server) sameSite() http.SameSite {
+// succeed once and every request after it would arrive anonymous. None is the
-	switch s.cfg.HTTP.CookieSameSite {
+// only mode a browser will send cross-site, and it requires Secure — which is
-	case "none":
+// why an origin allowlist forces Secure on regardless of AppEnv.
+func (s *Server) sessionSameSite() http.SameSite {
+	if len(s.cfg.HTTP.CORSOrigins) > 0 {
 		return http.SameSiteNoneMode
-	case "strict":
-		return http.SameSiteStrictMode
-	default:
-		return http.SameSiteLaxMode
 	}
+	return http.SameSiteLaxMode
+}
+
+// crossSiteCookies reports whether the cookie must be marked Secure because it
+// has to travel cross-site. SameSite=None without Secure is rejected outright
+// by every current browser.
+func (s *Server) crossSiteCookies() bool {
+	return s.sessionSameSite() == http.SameSiteNoneMode
 }
 
 // setSessionCookie writes the raw token to the browser.
@@ -82,15 +85,13 @@ func (s *Server) setSessionCookie(w http.ResponseWriter, token string, lifetime
 		Path:  "/",
 		// HttpOnly: script cannot read it.
 		HttpOnly: true,
-		// Lax by default, and Strict/None available through
+		// Lax, not Strict and not None. Strict would drop the cookie on any
-		// HTTP_COOKIE_SAMESITE. Strict would drop the cookie on any cross-site
+		// cross-site navigation, so following a link into the app would land on
-		// navigation, so following a link into the app would land on a login
+		// a login page despite a live session. None would require Secure and
-		// page despite a live session. None sends it on cross-site requests,
+		// would send the cookie on cross-site POSTs, which is the CSRF hole Lax
-		// which is the CSRF hole Lax exists to close — and is nonetheless the
+		// exists to close.
-		// only workable value when the frontend is on a different registrable
+		SameSite: s.sessionSameSite(),
-		// domain. See Server.sameSite.
+		Secure:   s.secureCookies() || s.crossSiteCookies(),
-		SameSite: s.sameSite(),
-		Secure:   s.secureCookies(),
 		MaxAge:   int(lifetime.Seconds()),
 	})
 }
@@ -107,10 +108,8 @@ func (s *Server) clearSessionCookie(w http.ResponseWriter) {
 		Value:    "",
 		Path:     "/",
 		HttpOnly: true,
-		// Must match the attributes it was set with, SameSite included, or the
+		SameSite: s.sessionSameSite(),
-		// browser treats this as a different cookie and leaves the original.
+		Secure:   s.secureCookies() || s.crossSiteCookies(),
-		SameSite: s.sameSite(),
-		Secure:   s.secureCookies(),
 		MaxAge:   -1,
 	})
 }

go-api/internal/httpserver/cors.go

@@ -72,6 +72,12 @@ func cors(origins []string) func(http.Handler) http.Handler {
 			}
 
 			w.Header().Set("Access-Control-Allow-Origin", origin)
+			// The frontend sends `credentials: "include"`, and a browser
+			// discards any response to such a request that does not carry this
+			// header — preflight included. Safe only because the origin was
+			// matched exactly above and is echoed back one at a time; "*" is
+			// never sent, which is the pairing the spec forbids.
+			w.Header().Set("Access-Control-Allow-Credentials", "true")
 
 			// Authentication is a cookie, so the browser will neither send it
 			// nor expose the response without this. It is set for allowlisted

go-api/internal/httpserver/interviews_test.go

@@ -0,0 +1,239 @@
+package httpserver_test
+
+import (
+	"context"
+	"net/http"
+	"testing"
+)
+
+// Completing an AI interview.
+//
+// The endpoint is unchanged — POST /api/v1/ai-interviews, the one the modal
+// already calls — but finishing an interview is two writes, and the second one
+// is a write the caller who most often makes the request may not perform. The
+// tests below are about that seam: the interview and the link land together,
+// they land for a talent user, and nothing about talent's own permissions has
+// widened to make it possible.
+
+// interviewCount counts the organization's interview rows.
+func interviewCount(t *testing.T, r *rbac) int {
+	t.Helper()
+	return countRows(t, r, "ai_interviews")
+}
+
+// talentApplication files an application through the API as the talent user, so
+// its email is whatever the server derived rather than what a test asked for.
+func talentApplication(t *testing.T, r *rbac, who actor) string {
+	t.Helper()
+	return mustCreate(t, r, who, "/api/v1/job-applications", map[string]any{
+		"job_posting_id": r.activePosting,
+		"applicant_name": who.name,
+	})
+}
+
+func interviewBody(applicationID, postingID string, score any) map[string]any {
+	body := map[string]any{
+		"application_id": applicationID,
+		"job_posting_id": postingID,
+		"job_title":      "Open Role",
+		"candidate_name": "Candidate",
+		"messages": []map[string]any{
+			{"role": "assistant", "content": "Tell me about a difficult shift."},
+			{"role": "user", "content": "We were two people short and I re-planned the passes."},
+		},
+		"verdict":             "hire",
+		"hire_recommendation": "Hire",
+		"summary":             "Composed under pressure.",
+	}
+	if score != nil {
+		body["overall_interview_score"] = score
+	}
+	return body
+}
+
+/* ── The RBAC break this fixes ──────────────────────────────────────────── */
+
+// A talent user completing their own interview is the whole talent flow, and it
+// could not finish: ai-interviews:Create is open to everyone, job-applications:
+// Update is operators only, so the interview was written and the application
+// never learned about it. Both writes now happen server-side, in one
+// transaction, on the row the interview already names.
+func TestTalentCompletesTheirOwnInterview(t *testing.T) {
+	r := newRBAC(t)
+	app := talentApplication(t, r, r.talA)
+
+	got := r.as(r.talA, "POST", "/api/v1/ai-interviews",
+		interviewBody(app, r.activePosting, 88))
+	if got.code != http.StatusCreated {
+		t.Fatalf("talent interview: got %d, want 201 (%v)", got.code, got.body)
+	}
+	interview := got.body["data"].(map[string]any)
+	interviewID, _ := interview["id"].(string)
+	if interviewID == "" {
+		t.Fatalf("the response carries no interview id: %v", got.body)
+	}
+
+	// The response is still the interview record, unchanged.
+	if interview["application_id"] != app {
+		t.Errorf("data.application_id = %v, want %s", interview["application_id"], app)
+	}
+
+	stored := applicationByID(t, r, app)
+	if stored["status"] != "interview" {
+		t.Errorf("application.status = %v, want interview — the analytics count "+
+			"status === 'interview' || interview_id", stored["status"])
+	}
+	if stored["interview_id"] != interviewID {
+		t.Errorf("application.interview_id = %v, want %s", stored["interview_id"], interviewID)
+	}
+	if score, ok := stored["ai_score"].(float64); !ok || int(score) != 88 {
+		t.Errorf("application.ai_score = %v, want the interview's 88", stored["ai_score"])
+	}
+}
+
+// And the permission itself has NOT widened. The server writes that one row on
+// the caller's behalf; the caller still cannot patch an application.
+func TestCompletingAnInterviewDoesNotWidenApplicationUpdate(t *testing.T) {
+	r := newRBAC(t)
+	app := talentApplication(t, r, r.talA)
+
+	if got := r.as(r.talA, "POST", "/api/v1/ai-interviews",
+		interviewBody(app, r.activePosting, 70)); got.code != http.StatusCreated {
+		t.Fatalf("talent interview: got %d, want 201 (%v)", got.code, got.body)
+	}
+	if got := r.as(r.talA, "PATCH", "/api/v1/job-applications/"+app,
+		map[string]any{"status": "hired"}); got.code != http.StatusForbidden {
+		t.Fatalf("talent PATCH of their own application: got %d, want 403 (%v)", got.code, got.body)
+	}
+}
+
+// An operator's interview links the same way. The atomicity half of the fix is
+// not talent-specific: a failure between the two writes left an interview
+// attached to an application that did not know about it, whoever ran it.
+func TestOperatorCompletingAnInterviewLinksTheApplication(t *testing.T) {
+	r := newRBAC(t)
+	app := applicationFor(t, r, r.activePosting, "Operator Candidate", "opcand@example.test")
+
+	got := r.as(r.empA, "POST", "/api/v1/ai-interviews",
+		interviewBody(app, r.activePosting, 64))
+	if got.code != http.StatusCreated {
+		t.Fatalf("operator interview: got %d, want 201 (%v)", got.code, got.body)
+	}
+	interviewID := got.body["data"].(map[string]any)["id"].(string)
+
+	stored := applicationByID(t, r, app)
+	if stored["status"] != "interview" || stored["interview_id"] != interviewID {
+		t.Errorf("application = status %v, interview_id %v; want interview / %s",
+			stored["status"], stored["interview_id"], interviewID)
+	}
+}
+
+/* ── What the link must not do ──────────────────────────────────────────── */
+
+// A body that says nothing about the score must not overwrite the screening
+// score with the interview column's default of 0. The field the caller never
+// mentioned is not a value they asked to store.
+func TestInterviewWithoutAScoreLeavesTheApplicationScore(t *testing.T) {
+	r := newRBAC(t)
+	app := applicationFor(t, r, r.activePosting, "Scored", "scored@example.test") // ai_score 77
+
+	got := r.as(r.admin, "POST", "/api/v1/ai-interviews",
+		interviewBody(app, r.activePosting, nil))
+	if got.code != http.StatusCreated {
+		t.Fatalf("interview: got %d, want 201 (%v)", got.code, got.body)
+	}
+
+	stored := applicationByID(t, r, app)
+	if score, ok := stored["ai_score"].(float64); !ok || int(score) != 77 {
+		t.Errorf("application.ai_score = %v, want the screening score 77 left alone",
+			stored["ai_score"])
+	}
+	// The status and the link still move — those are what completing an
+	// interview means.
+	if stored["status"] != "interview" || stored["interview_id"] == nil {
+		t.Errorf("application = status %v, interview_id %v; want interview and a link",
+			stored["status"], stored["interview_id"])
+	}
+}
+
+// Somebody else's application is not a subject a talent user may interview for,
+// and the refusal must leave nothing behind — not the interview, and not a
+// changed application.
+func TestInterviewForAnotherPersonsApplicationWritesNothing(t *testing.T) {
+	r := newRBAC(t)
+	app := talentApplication(t, r, r.talA)
+	before := interviewCount(t, r)
+
+	got := r.as(r.talB, "POST", "/api/v1/ai-interviews",
+		interviewBody(app, r.activePosting, 95))
+	if got.code != http.StatusNotFound {
+		t.Fatalf("interview for another person's application: got %d, want 404 (%v)",
+			got.code, got.body)
+	}
+	if after := interviewCount(t, r); after != before {
+		t.Errorf("ai_interviews: %d -> %d, want no row", before, after)
+	}
+
+	stored := applicationByID(t, r, app)
+	if stored["status"] != "applied" || stored["interview_id"] != nil {
+		t.Errorf("application = status %v, interview_id %v; want it untouched",
+			stored["status"], stored["interview_id"])
+	}
+}
+
+// An interview that cannot be written must not move the application either.
+// Both writes are in one transaction, so a refusal at the first is the whole
+// request rolled back rather than a partial completion.
+func TestARefusedInterviewLeavesTheApplicationAlone(t *testing.T) {
+	r := newRBAC(t)
+	app := applicationFor(t, r, r.activePosting, "Unfinished", "unfinished@example.test")
+	before := interviewCount(t, r)
+
+	body := interviewBody(app, r.activePosting, 80)
+	body["verdict"] = "definitely" // outside the interview_verdict enum
+
+	got := r.as(r.admin, "POST", "/api/v1/ai-interviews", body)
+	if got.code == http.StatusCreated {
+		t.Fatalf("an invalid verdict was accepted: %v", got.body)
+	}
+	if after := interviewCount(t, r); after != before {
+		t.Errorf("ai_interviews: %d -> %d, want no row", before, after)
+	}
+
+	stored := applicationByID(t, r, app)
+	if stored["status"] != "shortlisted" || stored["interview_id"] != nil {
+		t.Errorf("application = status %v, interview_id %v; want it untouched",
+			stored["status"], stored["interview_id"])
+	}
+}
+
+// Cross-tenant: the application is in another organization, so it is absent
+// rather than forbidden, and no interview is written for it.
+func TestInterviewCannotReachAnotherOrganizationsApplication(t *testing.T) {
+	r := newRBAC(t)
+	app := applicationFor(t, r, r.activePosting, "Ours", "ours-interview@example.test")
+
+	var before int
+	if err := r.h.Pool.QueryRow(context.Background(),
+		`SELECT count(*) FROM ai_interviews`).Scan(&before); err != nil {
+		t.Fatalf("count interviews: %v", err)
+	}
+
+	got := r.as(r.outsider, "POST", "/api/v1/ai-interviews",
+		interviewBody(app, r.activePosting, 90))
+	if got.code == http.StatusCreated {
+		t.Fatalf("an outsider wrote an interview for our application: %v", got.body)
+	}
+
+	var after int
+	if err := r.h.Pool.QueryRow(context.Background(),
+		`SELECT count(*) FROM ai_interviews`).Scan(&after); err != nil {
+		t.Fatalf("count interviews: %v", err)
+	}
+	if after != before {
+		t.Errorf("ai_interviews: %d -> %d, want no row", before, after)
+	}
+	if stored := applicationByID(t, r, app); stored["interview_id"] != nil {
+		t.Errorf("application.interview_id = %v, want it untouched", stored["interview_id"])
+	}
+}

go-api/internal/httpserver/owliver.go

@@ -0,0 +1,61 @@
+package httpserver
+
+import (
+	"net/http"
+
+	"github.com/krow/krow-backend/go-api/internal/authctx"
+	"github.com/krow/krow-backend/go-api/internal/domain"
+	"github.com/krow/krow-backend/go-api/internal/owliver"
+)
+
+// routeOwliver registers the Owliver panel's suggestion endpoint.
+//
+// GET, and a query string rather than a body, because the request is a read
+// with no side effect and the panel issues one per keystroke: a GET is what
+// makes it retryable, cancellable and cacheable by anything in front of it.
+//
+// It is deliberately NOT on the publicPaths allowlist in auth.go. Which
+// readings exist depends on the caller's role, so an anonymous suggestion has
+// no meaning — and the allowlist's failure mode is a route that refuses
+// everyone, which is the direction this should fall in.
+func (s *Server) routeOwliver(mux *http.ServeMux) int {
+	mux.HandleFunc("GET /api/v1/owliver/suggestions", s.handleOwliverSuggestions)
+	return 1
+}
+
+// suggestionsBody is the payload inside the standard data envelope.
+//
+// An object rather than a bare array, so the response has somewhere to grow — a
+// future `truncated` or `context` field would otherwise be a breaking change to
+// a client already reading `data` as a list.
+type suggestionsBody struct {
+	Suggestions []owliver.Suggestion `json:"suggestions"`
+}
+
+// handleOwliverSuggestions answers what the caller could usefully ask here.
+//
+// There is no s.authorize call and no policy lookup in this handler, and that
+// is the design rather than an omission: this endpoint exposes no resource, so
+// there is no operation to gate. Authorization happens per suggestion, inside
+// the catalogue, against the same domain.Policy table every other endpoint
+// consults — a reading the caller could not perform is never ranked, so it
+// cannot be returned. Authentication is upstream, in the middleware.
+func (s *Server) handleOwliverSuggestions(w http.ResponseWriter, r *http.Request) {
+	ident, err := authctx.MustFrom(r.Context())
+	if err != nil {
+		// Unreachable: the middleware refuses an unauthenticated request before
+		// the router sees it. A missing identity here is a wiring bug.
+		writeError(w, s.log, domain.Internal(err))
+		return
+	}
+
+	params, err := s.suggestions.ParseParams(r.URL.Query())
+	if err != nil {
+		writeError(w, s.log, err)
+		return
+	}
+
+	writeJSON(w, http.StatusOK, envelope{
+		Data: suggestionsBody{Suggestions: s.suggestions.Suggest(ident, params)},
+	})
+}

go-api/internal/httpserver/owliver_test.go

@@ -0,0 +1,315 @@
+package httpserver_test
+
+import (
+	"net/http"
+	"net/url"
+	"testing"
+)
+
+// GET /api/v1/owliver/suggestions.
+//
+// The ranking itself is tested in internal/owliver, against no database and no
+// server. What is tested here is only what the HTTP boundary adds: the session
+// requirement, the query-string contract, the response envelope, and the fact
+// that the role deciding which readings exist is the session's rather than
+// anything the caller can set.
+
+const suggestPath = "/api/v1/owliver/suggestions"
+
+// suggestURL builds the endpoint's address, escaping as a browser would.
+func suggestURL(page, query string) string {
+	v := url.Values{}
+	if page != "" {
+		v.Set("page", page)
+	}
+	if query != "" {
+		v.Set("query", query)
+	}
+	return suggestPath + "?" + v.Encode()
+}
+
+// suggestions reads the list out of the data envelope, failing the test if the
+// response is not shaped as the contract says.
+func suggestions(t *testing.T, r response) []map[string]any {
+	t.Helper()
+	if r.code != http.StatusOK {
+		t.Fatalf("status %d, body %v", r.code, r.body)
+	}
+	data, ok := r.body["data"].(map[string]any)
+	if !ok {
+		t.Fatalf("data is not an object: %v", r.body)
+	}
+	raw, ok := data["suggestions"].([]any)
+	if !ok {
+		// json null decodes to nil, and an absent key to nothing at all. Both
+		// break a client that iterates the list without checking.
+		t.Fatalf("suggestions is not an array (got %#v)", data["suggestions"])
+	}
+	out := make([]map[string]any, len(raw))
+	for i, item := range raw {
+		entry, ok := item.(map[string]any)
+		if !ok {
+			t.Fatalf("suggestion %d is not an object: %#v", i, item)
+		}
+		out[i] = entry
+	}
+	return out
+}
+
+/* ── Authentication ─────────────────────────────────────────────────────── */
+
+// The endpoint is not on the public allowlist. Which readings exist depends on
+// who is asking, so an anonymous suggestion has no meaning.
+func TestOwliverSuggestionsRequireASession(t *testing.T) {
+	a := newAPI(t)
+
+	got := a.doAnon("GET", suggestURL("positions", "pipeline"), nil)
+	if got.code != http.StatusUnauthorized {
+		t.Fatalf("status %d, want 401", got.code)
+	}
+	if code := got.codeOrEmpty(); code != "unauthorized" {
+		t.Fatalf("error code %q, want unauthorized", code)
+	}
+	// A refusal must not describe the catalogue it refused to rank.
+	if _, present := got.body["data"]; present {
+		t.Fatalf("an unauthenticated refusal carried data: %v", got.body)
+	}
+}
+
+/* ── The query string ───────────────────────────────────────────────────── */
+
+func TestOwliverSuggestionsValidation(t *testing.T) {
+	a := newAPI(t)
+
+	cases := []struct {
+		name string
+		path string
+		want int
+	}{
+		{"no page", suggestPath, http.StatusBadRequest},
+		{"blank page", suggestPath + "?page=%20", http.StatusBadRequest},
+		{"unknown page", suggestURL("nowhere", "pipeline"), http.StatusBadRequest},
+		{"a route, not a surface", suggestURL("/admin/positions", "pipeline"), http.StatusBadRequest},
+		{"unknown parameter", suggestURL("positions", "pipeline") + "&role=admin", http.StatusBadRequest},
+
+		// A query is optional: with nothing typed there is nothing to rank, and
+		// that is an empty list rather than a refusal.
+		{"no query", suggestURL("positions", ""), http.StatusOK},
+		{"page alias", suggestURL("hired", "recent"), http.StatusOK},
+		{"page spelled loosely", suggestURL("Talent Pool", "availability"), http.StatusOK},
+	}
+
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got := a.do("GET", c.path, nil)
+			if got.code != c.want {
+				t.Fatalf("status %d, want %d (body %v)", got.code, c.want, got.body)
+			}
+			if c.want == http.StatusBadRequest && got.codeOrEmpty() != "invalid_query" {
+				t.Fatalf("error code %q, want invalid_query", got.codeOrEmpty())
+			}
+		})
+	}
+}
+
+// The mux answers anything but GET, so the endpoint cannot be reached with a
+// body that might carry a page, a role or an identity.
+func TestOwliverSuggestionsAreReadOnly(t *testing.T) {
+	a := newAPI(t)
+	for _, method := range []string{"POST", "PATCH", "DELETE", "PUT"} {
+		got := a.do(method, suggestURL("positions", "pipeline"), map[string]any{"page": "positions"})
+		if got.code != http.StatusMethodNotAllowed {
+			t.Errorf("%s: status %d, want 405", method, got.code)
+		}
+	}
+}
+
+/* ── The response ───────────────────────────────────────────────────────── */
+
+func TestOwliverSuggestionsResponseShape(t *testing.T) {
+	a := newAPI(t) // the seeded user is an admin
+
+	got := suggestions(t, a.do("GET", suggestURL("positions", "pipeline"), nil))
+	if len(got) == 0 {
+		t.Fatal("pipeline on positions returned nothing")
+	}
+	if len(got) > 3 {
+		t.Fatalf("%d suggestions, the cap is 3", len(got))
+	}
+
+	seenIntent, seenText := map[string]bool{}, map[string]bool{}
+	for i, s := range got {
+		text, _ := s["text"].(string)
+		intent, _ := s["intent"].(string)
+		if text == "" || intent == "" {
+			t.Fatalf("suggestion %d is incomplete: %v", i, s)
+		}
+		if seenIntent[intent] {
+			t.Fatalf("duplicate intent %q", intent)
+		}
+		if seenText[text] {
+			t.Fatalf("duplicate text %q", text)
+		}
+		seenIntent[intent], seenText[text] = true, true
+
+		// Nothing internal may ride along: no terms, no resource names, no
+		// scores, no page keys.
+		for key := range s {
+			switch key {
+			case "text", "intent", "capability":
+			default:
+				t.Fatalf("suggestion %d exposes %q: %v", i, key, s)
+			}
+		}
+	}
+}
+
+// Asking for a rendering names it in the answer — and only where the reading
+// can actually be drawn that way.
+func TestOwliverSuggestionsCarryARequestedShape(t *testing.T) {
+	a := newAPI(t)
+
+	got := suggestions(t, a.do("GET", suggestURL("positions", "show hiring activity as a flow"), nil))
+	if len(got) != 1 {
+		t.Fatalf("got %d suggestions, want 1: %v", len(got), got)
+	}
+	if got[0]["intent"] != "hiring-operations" || got[0]["capability"] != "flow" {
+		t.Fatalf("got %v", got[0])
+	}
+
+	// With no shape asked for, the field is absent rather than empty.
+	plain := suggestions(t, a.do("GET", suggestURL("positions", "draft"), nil))
+	if len(plain) == 0 {
+		t.Fatal("draft on positions returned nothing")
+	}
+	if _, present := plain[0]["capability"]; present {
+		t.Fatalf("capability was sent for an unshaped query: %v", plain[0])
+	}
+}
+
+// No match is an empty array, not an error and not null.
+func TestOwliverSuggestionsEmptyResults(t *testing.T) {
+	a := newAPI(t)
+
+	for _, c := range []struct{ name, query string }{
+		{"nothing typed", ""},
+		{"one character", "p"},
+		{"irrelevant", "sourdough starter recipe"},
+	} {
+		t.Run(c.name, func(t *testing.T) {
+			if got := suggestions(t, a.do("GET", suggestURL("positions", c.query), nil)); len(got) != 0 {
+				t.Fatalf("got %v, want none", got)
+			}
+		})
+	}
+
+	// A real surface the catalogue holds no readings for is the same answer.
+	if got := suggestions(t, a.do("GET", suggestURL("settings", "owliver"), nil)); len(got) != 0 {
+		t.Fatalf("settings returned %v", got)
+	}
+}
+
+// The page decides the answer, so the same word must not produce the same list
+// everywhere.
+func TestOwliverSuggestionsAreScopedToThePage(t *testing.T) {
+	a := newAPI(t)
+
+	read := func(page string) []string {
+		out := []string{}
+		for _, s := range suggestions(t, a.do("GET", suggestURL(page, "pipeline"), nil)) {
+			out = append(out, s["intent"].(string))
+		}
+		return out
+	}
+
+	positions, candidates := read("positions"), read("candidates")
+	if len(positions) == 0 || len(candidates) == 0 {
+		t.Fatalf("positions=%v candidates=%v", positions, candidates)
+	}
+	if len(positions) == len(candidates) {
+		same := true
+		for i := range positions {
+			if positions[i] != candidates[i] {
+				same = false
+				break
+			}
+		}
+		if same {
+			t.Fatalf("both pages answered pipeline with %v", positions)
+		}
+	}
+}
+
+/* ── Authorization ──────────────────────────────────────────────────────── */
+
+// Who is asking comes from the session, and it decides which readings exist.
+//
+// Talent may list job applications — but only their own, by a predicate in the
+// repository — so the organization-wide readings the operator console offers
+// are not theirs, and are absent rather than refused.
+func TestOwliverSuggestionsFollowTheCallersRole(t *testing.T) {
+	r := newRBAC(t)
+
+	// A query each page can actually answer, so an empty list means the role
+	// was filtered rather than that the words matched nothing.
+	operatorPages := []struct{ page, query string }{
+		{"control-center", "pipeline attention"},
+		{"positions", "pipeline attention"},
+		{"candidates", "candidate score"},
+		{"hired-history", "recent hires outcomes"},
+		{"talent-pool", "talent pool availability"},
+		{"activity", "audit unusual activity"},
+	}
+
+	for _, c := range operatorPages {
+		for _, act := range []actor{r.admin, r.empA} {
+			got := suggestions(t, r.as(act, "GET", suggestURL(c.page, c.query), nil))
+			if len(got) == 0 {
+				t.Errorf("%s was offered nothing on %s for %q", act.name, c.page, c.query)
+			}
+		}
+
+		if got := suggestions(t, r.as(r.talA, "GET", suggestURL(c.page, c.query), nil)); len(got) != 0 {
+			t.Errorf("talent was offered %v on %s", got, c.page)
+		}
+	}
+
+	// Still 200 with an empty list, never 403: refusing would tell a caller
+	// which pages hold readings they cannot have.
+	refused := r.as(r.talA, "GET", suggestURL("positions", "pipeline"), nil)
+	if refused.code != http.StatusOK {
+		t.Fatalf("talent got status %d, want 200 with an empty list", refused.code)
+	}
+}
+
+// The role filter is not a blanket refusal for talent: what they may genuinely
+// ask — about their own account — is still offered. Without this, the test
+// above would pass with the permission check stubbed out to deny everything.
+func TestOwliverSuggestionsStillServeTalentTheirOwnReadings(t *testing.T) {
+	r := newRBAC(t)
+
+	got := suggestions(t, r.as(r.talA, "GET", suggestURL("profile", "permission"), nil))
+	if len(got) == 0 {
+		t.Fatal("talent was offered nothing about their own account")
+	}
+	if got[0]["intent"] != "profile-permissions" {
+		t.Fatalf("got %v", got[0])
+	}
+}
+
+// A permission-sensitive reading: Hired History reads the staff table, which
+// policy.go grants to operators only. Nothing about the request differs — only
+// the session behind it.
+func TestOwliverSuggestionsHideReadingsARoleCannotPerform(t *testing.T) {
+	r := newRBAC(t)
+	const path = suggestPath + "?page=hired-history&query=recent+hires"
+
+	for _, act := range []actor{r.admin, r.empA} {
+		if got := suggestions(t, r.as(act, "GET", path, nil)); len(got) == 0 {
+			t.Errorf("%s was offered no hiring outcomes", act.name)
+		}
+	}
+	if got := suggestions(t, r.as(r.talA, "GET", path, nil)); len(got) != 0 {
+		t.Fatalf("talent was offered readings of the staff table: %v", got)
+	}
+}

go-api/internal/httpserver/rbac_test.go

@@ -314,6 +314,50 @@ func mustCreate(t *testing.T, r *rbac, act actor, path string, body map[string]a
 
 /* ── 3. Mass assignment ─────────────────────────────────────────────────── */
 
+// The other half of a talent-only derivation: what an OPERATOR must supply.
+//
+// The server fills these columns from the session for a talent caller and for
+// nobody else — an operator filing an application or logging evidence is
+// writing about somebody who is not them. Treating the column as
+// server-supplied for every role let an operator's request past validation and
+// into SQL, where it came back as a not-null violation instead of the
+// required-field message the contract promises. The two halves have to agree:
+// what the repository will derive, and what validation stops asking for.
+func TestTalentOnlyDerivedFieldsAreRequiredOfOperators(t *testing.T) {
+	r := newRBAC(t)
+
+	cases := []struct {
+		name, path, column string
+		body               map[string]any
+	}{
+		{"job_applications.email", "/api/v1/job-applications", "email",
+			map[string]any{"job_posting_id": r.activePosting, "applicant_name": "Nameless"}},
+		{"evidence.worker_email", "/api/v1/evidence", "worker_email",
+			map[string]any{"type": "photo_identify"}},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := r.as(r.admin, "POST", tc.path, tc.body)
+			if got.code != http.StatusUnprocessableEntity {
+				t.Fatalf("operator create without %s: got %d, want 422 (%v)",
+					tc.column, got.code, got.body)
+			}
+			details, _ := got.body["error"].(map[string]any)["details"].(map[string]any)
+			if details[tc.column] != "required" {
+				t.Errorf("details = %v, want %s: required", details, tc.column)
+			}
+
+			// The same body from a talent caller is complete, because the
+			// server is about to fill the column in from their session.
+			if got := r.as(r.talA, "POST", tc.path, tc.body); got.code != http.StatusCreated {
+				t.Errorf("talent create without %s: got %d, want 201 (%v)",
+					tc.column, got.code, got.body)
+			}
+		})
+	}
+}
+
 // Identity a caller supplies is ignored; identity the server derives wins.
 //
 // This is the test that makes the ownership predicates above mean anything. If

go-api/internal/httpserver/server.go

@@ -1,17 +1,24 @@
 // Package httpserver holds the HTTP surface.
 //
 // It serves /health, the sign-in endpoints, the entity endpoints described in
-// docs/api-contract.md, and the current-user endpoints.
+// docs/api-contract.md, the current-user endpoints, the agent and skill
+// definition endpoints, and the Owliver panel's suggestion endpoint.
 //
-// Phase 3C replaced the development identity with real authentication. Every
+// Authentication replaced the development identity: every request outside the
-// request outside the small public allowlist in auth.go must carry a session
+// small public allowlist in auth.go must carry a session cookie; the middleware
-// cookie; the middleware resolves it to a user row and puts that user, and
+// resolves it to a user row and puts that user, and their organization, on the
-// their organization, on the request context. Nothing downstream changed —
+// request context. Nothing downstream changed — every service and repository
-// every service and repository already took the organization as a parameter,
+// already took the organization as a parameter, which is what devOrgMiddleware
-// which is what devOrgMiddleware existed to make true.
+// existed to make true.
 //
-// Authorization is NOT here. A signed-in user reaches every endpoint they could
+// Authorization is here, in Server.authorize: it reads the role off the
-// reach before; deciding which roles may do what is Phase 3D.
+// authenticated identity, consults the deny-by-default policy table in
+// internal/domain/policy.go, and answers 403 before any query runs. Row
+// visibility — organization scope, and ownership for talent callers — is a SQL
+// predicate in internal/repo instead, so an invisible row answers 404 rather
+// than 403. The definition endpoints are the exception: they are not
+// domain.Resource values, so their role checks are written inline in
+// internal/service/definitions.go rather than in the policy table.
 package httpserver
 
 import (
@@ -38,6 +45,7 @@ type Server struct {
 	api         *service.Registry
 	definitions *service.DefinitionsService
 	workflows   *service.WorkflowService
+	suggestions *service.SuggestionsService
 	log         *slog.Logger
 	http        *http.Server
 	started     time.Time
@@ -126,6 +134,7 @@ func New(cfg *config.Config, database *db.DB, log *slog.Logger, opts ...Option)
 		api:          service.NewRegistry(database.Pool),
 		definitions:  service.NewDefinitions(database.Pool),
 		workflows:    service.NewWorkflows(database.Pool).WithClock(o.now),
+		suggestions:  service.NewSuggestions(),
 		started:      o.now(),
 		sessions:     sessions,
 		users:        users,
@@ -138,7 +147,7 @@ func New(cfg *config.Config, database *db.DB, log *slog.Logger, opts ...Option)
 	mux := http.NewServeMux()
 	mux.HandleFunc("GET /health", s.handleHealth)
 	s.endpoints = s.routeAuth(mux) + s.routeResources(mux) + s.routeMe(mux) +
-		s.routeDefinitions(mux) + s.routeWorkflows(mux)
+		s.routeDefinitions(mux) + s.routeWorkflows(mux) + s.routeOwliver(mux)
 
 	handler := jsonErrors(mux)
 	// Authentication sits where devOrgMiddleware used to, so every route below

go-api/internal/httpserver/workflows.go

@@ -128,6 +128,11 @@ func (s *Server) handleAssign(w http.ResponseWriter, r *http.Request) {
 	ident, ok := s.authorizeAll(w, r,
 		requirement{"assignments", domain.OpCreate},
 		requirement{"job-applications", domain.OpUpdate},
+		// The workflow may now FILE an application as well as patch one, for a
+		// worker placed on a posting they never applied to. A write the handler
+		// performs has to appear in the list it is authorized against, even
+		// when — as here — the resulting permission set is unchanged.
+		requirement{"job-applications", domain.OpCreate},
 		requirement{"user-activity", domain.OpCreate},
 	)
 	if !ok {

go-api/internal/httpserver/workflows_test.go

@@ -329,3 +329,326 @@ func TestWorkflowEndpointsRequireASession(t *testing.T) {
 		}
 	}
 }
+
+/* ── Activity vocabulary ────────────────────────────────────────────────── */
+
+// activityTypes returns the event types written about one worker, newest first.
+//
+// Read straight from the table rather than through GET /user-activity so the
+// assertion is about what was STORED. The frontend's anomaly detection reads
+// these strings — PRIVILEGED_EVENTS is ['hire_candidate', 'create_position'] —
+// and a value the vocabulary does not contain is not a different label, it is
+// an event that silently stops counting.
+func activityTypes(t *testing.T, r *rbac, workerEmail string) []string {
+	t.Helper()
+	rows, err := r.h.Pool.Query(context.Background(),
+		`SELECT event_type FROM user_activity
+		  WHERE org_id = $1::uuid AND worker_email = $2::citext
+		  ORDER BY created_date DESC, id DESC`, r.orgID, workerEmail)
+	if err != nil {
+		t.Fatalf("read user_activity: %v", err)
+	}
+	defer rows.Close()
+	var out []string
+	for rows.Next() {
+		var s string
+		if err := rows.Scan(&s); err != nil {
+			t.Fatalf("scan user_activity: %v", err)
+		}
+		out = append(out, s)
+	}
+	if err := rows.Err(); err != nil {
+		t.Fatalf("read user_activity: %v", err)
+	}
+	return out
+}
+
+func TestHireWritesTheFrontendsActivityEvent(t *testing.T) {
+	r := newRBAC(t)
+	app := applicationFor(t, r, r.activePosting, "Evented", "evented@example.test")
+
+	if got := r.as(r.admin, "POST", "/api/v1/job-applications/"+app+"/hire",
+		map[string]any{}); got.code != http.StatusCreated {
+		t.Fatalf("hire: got %d, want 201 (%v)", got.code, got.body)
+	}
+
+	events := activityTypes(t, r, "evented@example.test")
+	if len(events) != 1 || events[0] != "hire_candidate" {
+		t.Errorf("activity = %v, want exactly [hire_candidate] — the vocabulary "+
+			"activitySignals.js reads, and the one the seed fixture uses", events)
+	}
+}
+
+func TestAssignWritesTheFrontendsActivityEvent(t *testing.T) {
+	r := newRBAC(t)
+	if got := r.as(r.admin, "POST", "/api/v1/job-postings/"+r.activePosting+"/assignments",
+		map[string]any{"workers": []map[string]any{
+			{"worker_email": "evented-assign@example.test", "worker_name": "Evented",
+				"starts_at": "2026-09-01T09:00:00Z"},
+		}}); got.code != http.StatusCreated {
+		t.Fatalf("assign: got %d, want 201 (%v)", got.code, got.body)
+	}
+
+	events := activityTypes(t, r, "evented-assign@example.test")
+	if len(events) != 1 || events[0] != "assign_employee" {
+		t.Errorf("activity = %v, want exactly [assign_employee]", events)
+	}
+}
+
+/* ── Assign: source ─────────────────────────────────────────────────────── */
+
+// An unspecified source must mean what the column says it means. The default in
+// 000001 is `owliver` and the frontend sends `owliver`; substituting `manual`
+// made a row written through this endpoint disagree with a row written through
+// POST /assignments about where the same action came from.
+func TestAssignDefaultsSourceToTheColumnDefault(t *testing.T) {
+	r := newRBAC(t)
+	got := r.as(r.admin, "POST", "/api/v1/job-postings/"+r.activePosting+"/assignments",
+		map[string]any{"workers": []map[string]any{
+			{"worker_email": "default-source@example.test", "starts_at": "2026-09-01T09:00:00Z"},
+			{"worker_email": "explicit-source@example.test", "starts_at": "2026-09-01T09:00:00Z",
+				"source": "manual"},
+		}})
+	if got.code != http.StatusCreated {
+		t.Fatalf("assign: got %d, want 201 (%v)", got.code, got.body)
+	}
+
+	created := assignmentsOf(t, got)
+	if created[0]["source"] != "owliver" {
+		t.Errorf("source = %v, want owliver", created[0]["source"])
+	}
+	// An explicit value is still the caller's.
+	if created[1]["source"] != "manual" {
+		t.Errorf("source = %v, want the supplied manual", created[1]["source"])
+	}
+}
+
+// assignmentsOf reads the assignment records out of an assign response.
+func assignmentsOf(t *testing.T, got response) []map[string]any {
+	t.Helper()
+	data, _ := got.body["data"].(map[string]any)
+	raw, _ := data["assignments"].([]any)
+	if raw == nil {
+		t.Fatalf("response carries no assignments: %v", got.body)
+	}
+	out := make([]map[string]any, 0, len(raw))
+	for _, rec := range raw {
+		out = append(out, rec.(map[string]any))
+	}
+	return out
+}
+
+// applicationByID reads one application as an operator, or fails.
+func applicationByID(t *testing.T, r *rbac, id string) map[string]any {
+	t.Helper()
+	list := r.as(r.admin, "GET", "/api/v1/job-applications?limit=500", nil)
+	if list.code != http.StatusOK {
+		t.Fatalf("list applications: %d (%v)", list.code, list.body)
+	}
+	for _, raw := range list.body["data"].([]any) {
+		rec := raw.(map[string]any)
+		if rec["id"] == id {
+			return rec
+		}
+	}
+	t.Fatalf("application %s not found", id)
+	return nil
+}
+
+/* ── Assign: the application a worker does not have yet ─────────────────── */
+
+// The behaviour the frontend had and the endpoint did not.
+//
+// An application is what puts a person in the pipeline for a role: the
+// candidate record is addressed by it and an interview takes one as its
+// subject. A worker assigned from the talent pool has none, so the endpoint has
+// to file one — in the same transaction as the assignment, which is the half
+// the frontend could not do.
+func TestAssignCreatesTheApplicationItNeeds(t *testing.T) {
+	r := newRBAC(t)
+	appsBefore := countRows(t, r, "job_applications")
+
+	got := r.as(r.admin, "POST", "/api/v1/job-postings/"+r.activePosting+"/assignments",
+		map[string]any{"workers": []map[string]any{{
+			"worker_email": "from-pool@example.test",
+			"worker_name":  "Pool Worker",
+			"starts_at":    "2026-09-01T09:00:00Z",
+			"match_score":  88,
+			"application": map[string]any{
+				"job_title":            "Open Role",
+				"phone":                "555-0100",
+				"years_experience":     4,
+				"skills":               []string{"service", "bar"},
+				"professional_summary": "Placed from the talent pool.",
+				"ai_score":             88,
+			},
+		}}})
+	if got.code != http.StatusCreated {
+		t.Fatalf("assign: got %d, want 201 (%v)", got.code, got.body)
+	}
+
+	if after := countRows(t, r, "job_applications"); after != appsBefore+1 {
+		t.Fatalf("job_applications: %d -> %d, want exactly one more", appsBefore, after)
+	}
+
+	assignment := assignmentsOf(t, got)[0]
+	linked, _ := assignment["application_id"].(string)
+	if linked == "" {
+		t.Fatal("the assignment was not linked to the application that was created for it")
+	}
+
+	app := applicationByID(t, r, linked)
+	if app["status"] != "assigned" {
+		t.Errorf("application.status = %v, want assigned", app["status"])
+	}
+	if app["email"] != "from-pool@example.test" {
+		t.Errorf("application.email = %v, want the worker's email", app["email"])
+	}
+	if app["job_posting_id"] != r.activePosting {
+		t.Errorf("application.job_posting_id = %v, want the posting being assigned to", app["job_posting_id"])
+	}
+	// applicant_name falls back to the worker's name rather than being blank —
+	// the column has a not-blank check.
+	if app["applicant_name"] != "Pool Worker" {
+		t.Errorf("application.applicant_name = %v, want the worker's name", app["applicant_name"])
+	}
+	if app["phone"] != "555-0100" {
+		t.Errorf("application.phone = %v, want the supplied phone", app["phone"])
+	}
+	if score, ok := app["ai_score"].(float64); !ok || int(score) != 88 {
+		t.Errorf("application.ai_score = %v, want 88", app["ai_score"])
+	}
+
+	// The audit entry names the application, so the feed can open it.
+	var activityApp *string
+	if err := r.h.Pool.QueryRow(context.Background(),
+		`SELECT application_id::text FROM user_activity
+		  WHERE org_id = $1::uuid AND worker_email = $2::citext`,
+		r.orgID, "from-pool@example.test").Scan(&activityApp); err != nil {
+		t.Fatalf("read the activity entry: %v", err)
+	}
+	if activityApp == nil || *activityApp != linked {
+		t.Errorf("activity.application_id = %v, want %s", activityApp, linked)
+	}
+}
+
+// (job_posting_id, email) is UNIQUE, so the second assign of the same person to
+// the same posting must find the application rather than try to file another —
+// and the comparison is case-insensitive, because the column is citext and the
+// frontend's own lookup lowercased both sides.
+func TestAssignLinksAnExistingApplicationInsteadOfDuplicating(t *testing.T) {
+	r := newRBAC(t)
+	existing := applicationFor(t, r, r.activePosting, "Already Applied", "Already.Applied@example.test")
+
+	appsBefore := countRows(t, r, "job_applications")
+	got := r.as(r.admin, "POST", "/api/v1/job-postings/"+r.activePosting+"/assignments",
+		map[string]any{"workers": []map[string]any{{
+			"worker_email": "already.applied@example.test",
+			"worker_name":  "Already Applied",
+			"starts_at":    "2026-09-01T09:00:00Z",
+			"application":  map[string]any{"job_title": "Open Role"},
+		}}})
+	if got.code != http.StatusCreated {
+		t.Fatalf("assign: got %d, want 201 (%v)", got.code, got.body)
+	}
+
+	if after := countRows(t, r, "job_applications"); after != appsBefore {
+		t.Errorf("job_applications: %d -> %d, want no new row for a person who already applied",
+			appsBefore, after)
+	}
+	if linked := assignmentsOf(t, got)[0]["application_id"]; linked != existing {
+		t.Errorf("assignment.application_id = %v, want the existing application %s", linked, existing)
+	}
+	if app := applicationByID(t, r, existing); app["status"] != "assigned" {
+		t.Errorf("application.status = %v, want assigned", app["status"])
+	}
+}
+
+// An id the caller already has still wins over the payload: it is a decision
+// they have made, and honouring the payload instead could file a second
+// application for the same placement.
+func TestAssignPrefersTheSuppliedApplicationID(t *testing.T) {
+	r := newRBAC(t)
+	app := applicationFor(t, r, r.activePosting, "Named", "named@example.test")
+	appsBefore := countRows(t, r, "job_applications")
+
+	got := r.as(r.admin, "POST", "/api/v1/job-postings/"+r.activePosting+"/assignments",
+		map[string]any{"workers": []map[string]any{{
+			"worker_email":   "named@example.test",
+			"worker_name":    "Named",
+			"starts_at":      "2026-09-01T09:00:00Z",
+			"application_id": app,
+			"application":    map[string]any{"applicant_name": "Ignored"},
+		}}})
+	if got.code != http.StatusCreated {
+		t.Fatalf("assign: got %d, want 201 (%v)", got.code, got.body)
+	}
+	if after := countRows(t, r, "job_applications"); after != appsBefore {
+		t.Errorf("job_applications: %d -> %d, want no new row", appsBefore, after)
+	}
+	if linked := assignmentsOf(t, got)[0]["application_id"]; linked != app {
+		t.Errorf("assignment.application_id = %v, want %s", linked, app)
+	}
+	if stored := applicationByID(t, r, app); stored["applicant_name"] != "Named" {
+		t.Errorf("applicant_name = %v — the payload overwrote a named application",
+			stored["applicant_name"])
+	}
+}
+
+// No payload, no application. A worker placed straight from the workforce is
+// legitimate, and one must not be invented for them.
+func TestAssignWithoutAnApplicationPayloadLinksNothing(t *testing.T) {
+	r := newRBAC(t)
+	appsBefore := countRows(t, r, "job_applications")
+
+	got := r.as(r.admin, "POST", "/api/v1/job-postings/"+r.activePosting+"/assignments",
+		map[string]any{"workers": []map[string]any{
+			{"worker_email": "unattached@example.test", "worker_name": "Unattached",
+				"starts_at": "2026-09-01T09:00:00Z"},
+		}})
+	if got.code != http.StatusCreated {
+		t.Fatalf("assign: got %d, want 201 (%v)", got.code, got.body)
+	}
+	if after := countRows(t, r, "job_applications"); after != appsBefore {
+		t.Errorf("job_applications: %d -> %d, want no application invented", appsBefore, after)
+	}
+	if linked := assignmentsOf(t, got)[0]["application_id"]; linked != nil {
+		t.Errorf("assignment.application_id = %v, want null", linked)
+	}
+}
+
+// The application payload is validated exactly as POST /job-applications would
+// validate it, and the batch element that produced the complaint is named.
+func TestAssignValidatesTheApplicationPayload(t *testing.T) {
+	r := newRBAC(t)
+	assignBefore := countRows(t, r, "assignments")
+	appsBefore := countRows(t, r, "job_applications")
+
+	got := r.as(r.admin, "POST", "/api/v1/job-postings/"+r.activePosting+"/assignments",
+		map[string]any{"workers": []map[string]any{
+			{"worker_email": "good@example.test", "worker_name": "Good",
+				"starts_at":   "2026-09-01T09:00:00Z",
+				"application": map[string]any{"job_title": "Open Role"}},
+			{"worker_email": "bad@example.test", "worker_name": "Bad",
+				"starts_at":   "2026-09-01T09:00:00Z",
+				"application": map[string]any{"english_level": "telepathic"}},
+		}})
+	if got.code != http.StatusUnprocessableEntity {
+		t.Fatalf("assign with an invalid application: got %d, want 422 (%v)", got.code, got.body)
+	}
+	details, _ := got.body["error"].(map[string]any)["details"].(map[string]any)
+	if details["workers[1].english_level"] == nil {
+		t.Errorf("details = %v, want the failure attributed to workers[1]", details)
+	}
+
+	// And the first worker — whose application WAS filed before the second
+	// failed — must be gone with it.
+	if after := countRows(t, r, "job_applications"); after != appsBefore {
+		t.Errorf("job_applications: %d -> %d — a rejected batch left an application behind",
+			appsBefore, after)
+	}
+	if after := countRows(t, r, "assignments"); after != assignBefore {
+		t.Errorf("assignments: %d -> %d — a rejected batch left an assignment behind",
+			assignBefore, after)
+	}
+}

go-api/internal/orgctx/orgctx.go

@@ -1,14 +1,17 @@
 // Package orgctx carries the organization a request operates on.
 //
-// Phase 2C has no authentication, so there is nothing to derive a tenant from.
+// It predates authentication. Before there was a session to derive a tenant
-// Rather than defaulting org_id deep inside the SQL — where it would have to be
+// from, org_id was put on the request context by one middleware rather than
-// unpicked from fourteen repositories once auth arrives — the value is put on
+// defaulted deep inside the SQL, so that it would not have to be unpicked from
-// the request context by one middleware and threaded explicitly through the
+// fourteen repositories later. That handover has since happened: the
-// service and repository boundaries.
+// authenticate middleware in httpserver reads the organization off the user's
+// row, and authctx.Identity.OrgID is what the service and repository layers
+// actually read.
 //
-// When authentication lands, DevMiddleware is replaced by one that reads the
+// What is still load-bearing here is DevOrgSlug and DevOrgName, which name the
-// organization off the authenticated session. Nothing below this package
+// organization the seeder creates. With and From remain the read/write pair for
-// changes: every caller already takes an org id as a parameter.
+// the context key the middleware still sets; retiring them in favour of
+// authctx.Identity.OrgID alone is a deliberate change, not a tidy-up.
 package orgctx
 
 import (

go-api/internal/owliver/catalog.go

@@ -0,0 +1,625 @@
+// Package owliver answers "what could I usefully ask here?".
+//
+// It is the suggestion side of the Owliver panel and nothing else. It does not
+// answer questions, does not reach a database, does not call a model, and holds
+// no state: a request names a page and what the user has typed so far, and this
+// package ranks a static catalogue against it. That is deliberate — the panel
+// calls the endpoint on every keystroke, so the work per call has to be a few
+// string comparisons over a table built once at process start.
+//
+// WHERE THE CATALOGUE COMES FROM. Owliver's capabilities are declared in the
+// frontend, one manifest per page context, in
+// src/components/ai-assistant/capabilities/. Each entry there is an id, a label
+// and the function that answers it. This file transcribes the id and the page
+// it is offered on, and adds the two things a manifest does not carry because
+// the browser never needed them: the words that mean a user is reaching for
+// that reading, and the records it reads.
+//
+// Transcription rather than a second registry, on the pattern already set by
+// internal/definition/vocabulary.go: the frontend owns the vocabulary, this
+// side names keys from it, and TestIntentIDsAreFrontendCapabilities keeps the
+// two honest. An Intent.ID that no manifest declares is a bug here, not a new
+// capability — the id in a response is the id the panel dispatches on.
+//
+// WHAT MAKES A SUGGESTION SAFE TO SHOW. Offering someone a question is a claim
+// that they could ask it. Reads is that claim written down: the resources the
+// capability actually reads, checked against the one authorization table in
+// internal/domain/policy.go. No role list is repeated here, and no rule is
+// re-derived — a capability that reads staff is unavailable to talent because
+// policy.go says talent may not list staff, and for no other reason.
+package owliver
+
+import "github.com/krow/krow-backend/go-api/internal/domain"
+
+// Need is one record set a capability reads, and how much of it.
+//
+// OrgWide is the half a role check alone cannot express. Every role may list
+// job applications; a talent caller sees only their own, because the policy
+// attaches a row predicate rather than a refusal. So "which position has the
+// strongest pipeline?" is not forbidden to talent — it is unanswerable for
+// them, and offering it would promise a reading across records they will never
+// be shown. OrgWide asks policy.ScopeFor for exactly that: is this caller's
+// view of this resource narrowed, or is it the whole organization?
+type Need struct {
+	Resource string // domain resource path, e.g. "job-applications"
+	Op       domain.Op
+	OrgWide  bool
+}
+
+// permitted reports whether a role may perform this reading.
+//
+// Three gates, all sourced from the descriptors: the API must expose the
+// operation at all, the policy must allow the role, and — for an org-wide
+// reading — the policy must not narrow the role's rows.
+func (n Need) permitted(role domain.Role) bool {
+	res, ok := domain.ResourceByPath[n.Resource]
+	if !ok || !res.Supports(n.Op) {
+		return false
+	}
+	if !res.Policy.Allows(n.Op, role) {
+		return false
+	}
+	if n.OrgWide && res.Policy.ScopeFor(role).Kind != domain.ScopeNone {
+		return false
+	}
+	return true
+}
+
+// Intent is one reading Owliver can offer on one page.
+type Intent struct {
+	// ID is the frontend capability id, verbatim. It is what the panel
+	// dispatches on, which is why it is not invented here.
+	ID string
+
+	// Text is the suggestion as the user reads it: a question in their words,
+	// not a label. Unique within a page.
+	Text string
+
+	// Subject names what the reading is about — "hiring activity", "the
+	// candidate pipeline" — and exists so a shaped variant can be phrased
+	// ("Summarize hiring activity", "Show hiring activity as a flow") without
+	// writing every combination out. An intent with no Subject is offered only
+	// as its Text.
+	Subject string
+
+	// Shapes are the section types this reading can be drawn as, named from the
+	// closed OWLIVER_CAPABILITIES vocabulary in src/lib/skills/surfaces.js.
+	// Empty means prose only. `summary` is never listed: it has no component,
+	// so it applies to anything with a Subject.
+	Shapes []string
+
+	// Terms are the words that mean a user is reaching for this reading. About
+	// the subject, never about the shape — "as a flow" belongs to the shape
+	// table, and putting it here would offer every page's intents to anyone who
+	// typed "chart".
+	Terms []string
+
+	// Reads is what the capability actually reads. Empty means it reads nothing
+	// but the caller's own account, and is therefore available to anyone signed
+	// in.
+	Reads []Need
+}
+
+// permitted reports whether a role may be offered this intent. Every reading
+// must be permitted: a suggestion that is half-answerable is not answerable.
+func (i Intent) permitted(role domain.Role) bool {
+	for _, n := range i.Reads {
+		if !n.permitted(role) {
+			return false
+		}
+	}
+	return true
+}
+
+/* ── The readings each resource stands for ──────────────────────────────── */
+//
+// Named rather than repeated so that "this capability reads the organization's
+// applications" is written once and reads the same everywhere it appears.
+
+var (
+	postings     = Need{Resource: "job-postings", Op: domain.OpList, OrgWide: true}
+	applications = Need{Resource: "job-applications", Op: domain.OpList, OrgWide: true}
+	interviews   = Need{Resource: "ai-interviews", Op: domain.OpList, OrgWide: true}
+	profiles     = Need{Resource: "worker-profiles", Op: domain.OpList, OrgWide: true}
+	staff        = Need{Resource: "staff", Op: domain.OpList, OrgWide: true}
+	orgActivity  = Need{Resource: "user-activity", Op: domain.OpList, OrgWide: true}
+	courses      = Need{Resource: "courses", Op: domain.OpList, OrgWide: true}
+	certs        = Need{Resource: "certifications", Op: domain.OpList, OrgWide: true}
+	evidence     = Need{Resource: "evidence", Op: domain.OpList, OrgWide: true}
+
+	// The caller's own audit trail rather than the organization's — the Profile
+	// page reads what *you* did, which every role may do for themselves.
+	ownActivity = Need{Resource: "user-activity", Op: domain.OpList}
+)
+
+/* ── The catalogue ──────────────────────────────────────────────────────── */
+
+// catalogue is every intent, keyed by canonical page id.
+//
+// Keys are canonical SKILL_SURFACES ids — the same vocabulary
+// internal/definition validates a definition's `pages:` against. A page that is
+// a real surface but has no entry here is not an error: it answers with an
+// empty list, which is the honest reply for a screen holding no readings.
+//
+// The seven workspace and configuration surfaces are deliberately absent.
+// Their manifests explain the screen rather than read workforce records — they
+// have no data behind them to rank a typed query against, and the panel there
+// answers from the registries directly.
+//
+// Declaration order is the tie-break when two intents score equally, so the
+// order within a page is the order the frontend manifest lists them in.
+var catalogue = map[string][]Intent{
+
+	/* ── Control Center — the platform read as a whole ─────────────────── */
+
+	"control-center": {
+		{
+			ID: "platform-health", Text: "How healthy is the platform right now?",
+			Subject: "platform health", Shapes: []string{"stats", "card", "insight"},
+			Terms: []string{"health", "healthy", "platform", "integrity", "data quality",
+				"status", "wrong", "broken", "degraded"},
+			Reads: []Need{postings, applications, interviews, profiles},
+		},
+		{
+			ID: "workforce-summary", Text: "Summarize the workforce across the platform",
+			Subject: "the workforce", Shapes: []string{"stats", "table", "progress", "card"},
+			Terms: []string{"workforce", "scale", "how big", "composition", "headcount",
+				"people", "how many"},
+			Reads: []Need{postings, profiles, staff},
+		},
+		{
+			ID: "hiring-operations", Text: "How is hiring operating overall?",
+			Subject: "hiring activity", Shapes: []string{"flow", "stats", "timeline", "table", "card"},
+			Terms: []string{"hiring", "operations", "activity", "velocity", "speed",
+				"throughput", "how fast", "time to hire"},
+			Reads: []Need{applications, postings},
+		},
+		{
+			ID: "pipeline-health", Text: "Where is the hiring pipeline getting stuck?",
+			Subject: "the hiring pipeline", Shapes: []string{"flow", "stats", "progress", "table", "card"},
+			Terms: []string{"pipeline", "funnel", "bottleneck", "stuck", "blocked",
+				"conversion", "stage", "stages", "drop off", "dropoff"},
+			Reads: []Need{applications},
+		},
+		{
+			ID: "attention-required", Text: "What needs attention right now?",
+			Subject: "what needs attention", Shapes: []string{"list", "table", "insight"},
+			Terms: []string{"attention", "urgent", "priority", "action", "unusual",
+				"anomaly", "anomalous", "risk", "risks", "problem", "problems", "issue"},
+			Reads: []Need{applications, postings},
+		},
+		{
+			ID: "recommendations", Text: "What should I do next?",
+			Subject: "the recommendations", Shapes: []string{"list", "insight"},
+			Terms: []string{"recommend", "recommendation", "recommendations", "suggest",
+				"should", "advice", "next step", "next"},
+			Reads: []Need{applications, postings},
+		},
+	},
+
+	/* ── Positions — the roles being filled ────────────────────────────── */
+
+	"positions": {
+		{
+			ID: "position-drafts", Text: "Which positions are still unfinished drafts?",
+			Subject: "the unfinished drafts", Shapes: []string{"list", "table"},
+			Terms: []string{"draft", "drafts", "unfinished", "incomplete", "unpublished",
+				"not posted", "half", "position", "positions", "role", "roles"},
+			Reads: []Need{postings},
+		},
+		{
+			ID: "position-strength", Text: "Which position has the strongest pipeline?",
+			Subject: "pipeline strength by position", Shapes: []string{"table", "stats", "list", "card"},
+			Terms: []string{"pipeline", "strength", "strongest", "healthiest", "best",
+				"conversion", "which position", "compare", "position", "positions",
+				"role", "roles"},
+			Reads: []Need{postings, applications},
+		},
+		{
+			ID: "positions-attention", Text: "Which positions need attention?",
+			Subject: "positions needing attention", Shapes: []string{"list", "table", "insight"},
+			Terms: []string{"attention", "risk", "risks", "at risk", "stalled", "stale",
+				"ageing", "aging", "neglected", "urgent", "slipping", "behind",
+				"position", "positions", "role", "roles"},
+			Reads: []Need{postings, applications},
+		},
+		{
+			ID: "hiring-priority", Text: "Which position should I fill first?",
+			Subject: "what to fill first", Shapes: []string{"list", "table"},
+			Terms: []string{"priority", "prioritise", "prioritize", "fill", "fill first",
+				"first", "most important", "which position", "vacancy", "vacancies",
+				"position", "positions", "role", "roles"},
+			Reads: []Need{postings, applications},
+		},
+		{
+			ID: "pipeline-health", Text: "Where are the hiring bottlenecks?",
+			Subject: "the hiring bottlenecks", Shapes: []string{"flow", "stats", "progress", "table"},
+			Terms: []string{"bottleneck", "bottlenecks", "funnel", "stuck", "blocked",
+				"stage", "stages", "waiting", "backlog", "pipeline"},
+			Reads: []Need{applications},
+		},
+		{
+			ID: "hiring-operations", Text: "Summarize hiring activity across all positions",
+			Subject: "hiring activity", Shapes: []string{"flow", "stats", "timeline", "table", "card"},
+			Terms: []string{"hiring", "activity", "operations", "throughput", "velocity",
+				"how many", "posting", "postings", "role", "roles", "position", "positions"},
+			Reads: []Need{applications, postings},
+		},
+		{
+			/* Deliberately NOT carrying "pipeline". This is the Positions page's
+			   reading of who is waiting on a decision, and it belongs here — the
+			   frontend manifest declares it on this page. But "pipeline" on
+			   Positions is a question about how the ROLES are converting, and
+			   answering it with a list of people is the page context being
+			   right and the ranking being wrong. Its own words are what it
+			   answers to. */
+			ID: "candidates-waiting", Text: "Which candidates are waiting on a decision?",
+			Subject: "the candidates waiting", Shapes: []string{"list", "table", "stats"},
+			Terms: []string{"waiting", "candidate", "candidates", "applicant", "applicants",
+				"review", "screen", "screening", "decision", "queue"},
+			Reads: []Need{applications},
+		},
+	},
+
+	/* ── Candidates — the people in the funnel, as a triage queue ──────── */
+
+	"candidates": {
+		{
+			ID: "candidates-attention", Text: "Which candidates need attention?",
+			Subject: "the candidates needing attention", Shapes: []string{"list", "table", "insight"},
+			Terms: []string{"attention", "waiting", "stalled", "overdue", "action",
+				"decision", "decide", "urgent", "candidate", "candidates", "applicant",
+				"applicants"},
+			Reads: []Need{applications},
+		},
+		{
+			ID: "top-candidates", Text: "Who are the strongest candidates?",
+			Subject: "the strongest candidates", Shapes: []string{"list", "table", "stats", "card"},
+			Terms: []string{"strongest", "top", "best", "compare", "shortlist", "rank",
+				"ranking", "highest", "score", "scores", "scored", "scoring", "who",
+				"candidate", "candidates", "applicant", "applicants"},
+			Reads: []Need{applications},
+		},
+		{
+			ID: "interview-ready", Text: "Who is ready to interview?",
+			Subject: "the interview-ready candidates", Shapes: []string{"list", "table", "stats"},
+			Terms: []string{"interview", "interviews", "interviewed", "ready", "schedule",
+				"next round", "shortlist", "candidate", "candidates", "applicant",
+				"applicants"},
+			Reads: []Need{applications, interviews},
+		},
+		{
+			ID: "screening-gaps", Text: "Which candidates have not been scored yet?",
+			Subject: "the screening gaps", Shapes: []string{"list", "table", "stats", "progress"},
+			Terms: []string{"unscored", "score", "scores", "scoring", "screening", "screen",
+				"gap", "gaps", "missing", "incomplete", "coverage", "candidate",
+				"candidates", "applicant", "applicants"},
+			Reads: []Need{applications},
+		},
+		{
+			ID: "pipeline-summary", Text: "Summarize the candidate pipeline",
+			Subject: "the candidate pipeline", Shapes: []string{"flow", "stats", "progress", "table", "card"},
+			Terms: []string{"pipeline", "funnel", "stage", "stages", "breakdown",
+				"how many", "where are", "conversion"},
+			Reads: []Need{applications},
+		},
+		{
+			ID: "candidate-risk", Text: "Which candidates carry risk flags?",
+			Subject: "the candidate risks", Shapes: []string{"list", "table", "insight"},
+			Terms: []string{"risk", "risks", "risky", "flag", "flags", "flagged", "concern",
+				"concerns", "integrity", "doubt", "decision", "candidate", "candidates",
+				"applicant", "applicants"},
+			Reads: []Need{applications, interviews},
+		},
+	},
+
+	/* ── Candidates Analysis — the same records read as a pool ─────────── */
+
+	"candidates-analysis": {
+		{
+			ID: "candidate-risk", Text: "Where is candidate risk concentrated?",
+			Subject: "candidate risk", Shapes: []string{"table", "stats", "insight"},
+			Terms: []string{"risk", "risks", "flag", "flags", "flagged", "concern",
+				"integrity", "concentrated"},
+			Reads: []Need{applications, interviews},
+		},
+		{
+			ID: "recruitment-insights", Text: "What do the recruitment numbers show?",
+			Subject: "the recruitment insights", Shapes: []string{"stats", "table", "insight", "card"},
+			Terms: []string{"insight", "insights", "recruitment", "quality", "pool",
+				"stands out", "numbers", "trend", "trends"},
+			Reads: []Need{applications, profiles},
+		},
+		{
+			ID: "screening-gaps", Text: "Where are the screening gaps?",
+			Subject: "the screening gaps", Shapes: []string{"table", "stats", "progress"},
+			Terms: []string{"screening", "screen", "gap", "gaps", "unscored", "coverage",
+				"missing", "incomplete", "score", "scores"},
+			Reads: []Need{applications},
+		},
+		{
+			ID: "hiring-recommendations", Text: "Who should we hire?",
+			Subject: "the hiring recommendations", Shapes: []string{"list", "table", "insight"},
+			Terms: []string{"recommend", "recommendation", "recommendations", "hire",
+				"hiring", "should", "advice", "decision", "who"},
+			Reads: []Need{applications, profiles},
+		},
+		{
+			ID: "top-candidates", Text: "Compare the top candidates",
+			Subject: "the top candidates", Shapes: []string{"table", "list", "stats"},
+			Terms: []string{"compare", "comparison", "top", "best", "strongest",
+				"shortlist", "rank", "ranking", "side by side"},
+			Reads: []Need{applications},
+		},
+	},
+
+	/* ── Analytics — performance over time ─────────────────────────────── */
+
+	"analytics": {
+		{
+			ID: "hiring-trend", Text: "How has hiring trended over time?",
+			Subject: "the hiring trend", Shapes: []string{"timeline", "flow", "stats", "table"},
+			Terms: []string{"trend", "trends", "trending", "over time", "month", "monthly",
+				"week", "weekly", "history", "growth", "change"},
+			Reads: []Need{applications, postings},
+		},
+		{
+			ID: "department-performance", Text: "How is each department performing?",
+			Subject: "department performance", Shapes: []string{"table", "stats", "progress"},
+			Terms: []string{"department", "departments", "team", "teams", "category",
+				"categories", "performance", "performing", "breakdown", "compare"},
+			Reads: []Need{applications, postings},
+		},
+		{
+			ID: "pipeline-health", Text: "Where are the hiring bottlenecks?",
+			Subject: "the hiring bottlenecks", Shapes: []string{"flow", "stats", "progress", "table"},
+			Terms: []string{"bottleneck", "bottlenecks", "funnel", "pipeline", "conversion",
+				"stuck", "stage", "stages"},
+			Reads: []Need{applications},
+		},
+		{
+			ID: "position-conversion", Text: "Which positions convert best?",
+			Subject: "position conversion", Shapes: []string{"table", "stats", "list"},
+			Terms: []string{"conversion", "convert", "converts", "position", "positions",
+				"role", "roles", "rate", "rates", "ratio", "yield"},
+			Reads: []Need{postings, applications},
+		},
+		{
+			ID: "hiring-operations", Text: "How is hiring performing overall?",
+			Subject: "hiring performance", Shapes: []string{"stats", "flow", "timeline", "card"},
+			Terms: []string{"hiring", "performance", "operations", "velocity", "speed",
+				"average", "averages", "time to hire", "throughput"},
+			Reads: []Need{applications, postings},
+		},
+		{
+			ID: "attention-required", Text: "What needs attention in the numbers?",
+			Subject: "what needs attention", Shapes: []string{"list", "insight", "table"},
+			Terms: []string{"attention", "outlier", "outliers", "anomaly", "unusual",
+				"risk", "risks", "worst", "falling"},
+			Reads: []Need{applications, postings},
+		},
+	},
+
+	/* ── Activity — the audit log ──────────────────────────────────────── */
+
+	"activity": {
+		{
+			ID: "audit-summary", Text: "Summarize the audit log",
+			Subject: "the audit log", Shapes: []string{"stats", "table", "timeline", "card"},
+			Terms: []string{"audit", "log", "logs", "event", "events", "activity",
+				"trail", "record", "records", "how many"},
+			Reads: []Need{orgActivity},
+		},
+		{
+			ID: "user-activity", Text: "Who has been most active?",
+			Subject: "activity by user", Shapes: []string{"table", "list", "stats"},
+			Terms: []string{"user", "users", "who", "account", "accounts", "busiest",
+				"most active", "behaviour", "behavior", "person"},
+			Reads: []Need{orgActivity},
+		},
+		{
+			ID: "unusual-activity", Text: "Has anything unusual happened?",
+			Subject: "the unusual activity", Shapes: []string{"list", "timeline", "insight"},
+			Terms: []string{"unusual", "anomaly", "anomalies", "anomalous", "suspicious",
+				"spike", "spikes", "burst", "odd", "strange", "out of hours"},
+			Reads: []Need{orgActivity},
+		},
+		{
+			ID: "security-insights", Text: "Are there any security concerns?",
+			Subject: "the security findings", Shapes: []string{"list", "insight", "table"},
+			Terms: []string{"security", "secure", "breach", "compliance", "compliant",
+				"integrity", "trace", "traceable", "attributable", "risk", "risks"},
+			Reads: []Need{orgActivity},
+		},
+	},
+
+	/* ── Talent Pool — supply, before anyone applies ───────────────────── */
+
+	"talent-pool": {
+		{
+			ID: "talent-priorities", Text: "Who should I prioritize in the talent pool?",
+			Subject: "the talent priorities", Shapes: []string{"list", "table", "stats"},
+			Terms: []string{"prioritise", "prioritize", "priority", "priorities", "who",
+				"best", "top", "strongest", "elite", "star", "shortlist", "score", "scores"},
+			Reads: []Need{profiles},
+		},
+		{
+			ID: "talent-summary", Text: "Summarize the talent pool",
+			Subject: "the talent pool", Shapes: []string{"stats", "table", "progress", "card"},
+			Terms: []string{"pool", "talent", "worker", "workers", "profile", "profiles",
+				"segment", "segments", "composition", "supply", "how many"},
+			Reads: []Need{profiles},
+		},
+		{
+			ID: "talent-verification", Text: "Which profiles are missing verification?",
+			Subject: "the verification gaps", Shapes: []string{"list", "table", "progress", "stats"},
+			Terms: []string{"verification", "verify", "verified", "unverified", "gap",
+				"gaps", "missing", "credential", "credentials", "proof", "evidence"},
+			Reads: []Need{profiles, evidence},
+		},
+		{
+			ID: "talent-availability", Text: "Who is available to start?",
+			Subject: "availability", Shapes: []string{"list", "table", "stats"},
+			Terms: []string{"available", "availability", "unavailable", "free", "start",
+				"capacity", "when", "now", "notice"},
+			Reads: []Need{profiles},
+		},
+	},
+
+	/* ── Hired History — what happened after the hire ───────────────────── */
+
+	"hired-history": {
+		{
+			ID: "hiring-outcomes", Text: "How have our hires worked out?",
+			Subject: "the hiring outcomes", Shapes: []string{"stats", "table", "progress", "card"},
+			Terms: []string{"outcome", "outcomes", "result", "results", "quality",
+				"retention", "worked out", "performance", "department", "departments"},
+			Reads: []Need{staff, applications},
+		},
+		{
+			ID: "hiring-strongest", Text: "Who are our strongest hires?",
+			Subject: "the strongest hires", Shapes: []string{"list", "table", "stats"},
+			Terms: []string{"strongest", "best", "top", "star", "highest", "score",
+				"scores", "standout"},
+			Reads: []Need{staff, applications},
+		},
+		{
+			ID: "hiring-patterns", Text: "What stands out about who we hire?",
+			Subject: "the hiring patterns", Shapes: []string{"table", "stats", "insight"},
+			Terms: []string{"pattern", "patterns", "stands out", "trend", "trends",
+				"common", "typical", "breakdown", "profile"},
+			Reads: []Need{staff, applications},
+		},
+		{
+			ID: "hiring-recent", Text: "Who did we hire recently?",
+			Subject: "the recent hires", Shapes: []string{"list", "timeline", "table"},
+			Terms: []string{"recent", "recently", "latest", "last", "new hire", "new hires",
+				"who did we hire", "this month", "hired"},
+			Reads: []Need{staff},
+		},
+	},
+
+	/* ── KROW Forge — the skill library and the proof behind it ────────── */
+
+	"krow-forge": {
+		{
+			ID: "forge-library", Text: "What is in the skill library?",
+			Subject: "the skill library", Shapes: []string{"stats", "table", "list", "card"},
+			Terms: []string{"library", "skill", "skills", "catalogue", "catalog", "course",
+				"courses", "challenge", "challenges", "what do we have", "how many"},
+			Reads: []Need{courses},
+		},
+		{
+			ID: "forge-published", Text: "Which skills are published?",
+			Subject: "the published skills", Shapes: []string{"list", "table", "stats"},
+			Terms: []string{"published", "publish", "live", "draft", "drafts", "archived",
+				"archive", "status", "in service"},
+			Reads: []Need{courses},
+		},
+		{
+			ID: "forge-evaluation", Text: "How is submitted proof evaluated?",
+			Subject: "the evaluation criteria", Shapes: []string{"list", "table", "insight"},
+			Terms: []string{"evaluate", "evaluated", "evaluation", "rubric", "rubrics",
+				"criteria", "criterion", "grading", "graded", "proof", "verify",
+				"verification", "assess"},
+			Reads: []Need{courses, evidence},
+		},
+		{
+			ID: "forge-workforce", Text: "How is the workforce using the Forge?",
+			Subject: "workforce usage", Shapes: []string{"stats", "progress", "table"},
+			Terms: []string{"workforce", "usage", "using", "uptake", "adoption", "progress",
+				"completion", "completed", "training", "learning"},
+			Reads: []Need{courses, evidence, profiles},
+		},
+		{
+			ID: "forge-gaps", Text: "Which skill gaps are still open?",
+			Subject: "the skill gaps", Shapes: []string{"list", "table", "progress"},
+			Terms: []string{"gap", "gaps", "missing", "uncovered", "close", "coverage",
+				"needed", "shortfall", "certification", "certifications"},
+			Reads: []Need{courses, certs, profiles},
+		},
+	},
+
+	/* ── Create Position — the authoring form ──────────────────────────── */
+
+	"create-position": {
+		{
+			ID: "vetting-weights", Text: "What do the vetting weights score?",
+			Subject: "the vetting weights", Shapes: []string{"table", "insight"},
+			Terms: []string{"weight", "weights", "weighting", "weightings", "vetting",
+				"criteria", "criterion", "scoring", "score", "balance", "importance"},
+			Reads: []Need{postings},
+		},
+		{
+			ID: "position-benchmarks", Text: "How does this compare with similar roles?",
+			Subject: "the benchmarks", Shapes: []string{"table", "stats", "insight"},
+			Terms: []string{"compare", "comparison", "benchmark", "benchmarks", "similar",
+				"typical", "average", "pay", "rate", "rates", "salary", "experience",
+				"market"},
+			Reads: []Need{postings},
+		},
+		{
+			ID: "position-requirements", Text: "Which credentials are already in use?",
+			Subject: "the credentials in use", Shapes: []string{"list", "table", "stats"},
+			Terms: []string{"credential", "credentials", "certification", "certifications",
+				"requirement", "requirements", "qualification", "qualifications",
+				"licence", "license", "skill", "skills"},
+			Reads: []Need{postings, certs},
+		},
+		{
+			ID: "position-spec-steps", Text: "How does this form work?",
+			Terms: []string{"form", "field", "fields", "step", "steps", "section",
+				"sections", "how do i", "what do i", "explain", "fill in", "required"},
+		},
+	},
+
+	/* ── Profile — the account, not the workforce ──────────────────────── */
+
+	"profile": {
+		{
+			ID: "profile-permissions", Text: "What am I permitted to do?",
+			Terms: []string{"permission", "permissions", "permitted", "allowed", "can i",
+				"scope", "access", "role", "rights", "privilege", "privileges"},
+		},
+		{
+			ID: "profile-identity", Text: "What are my account details?",
+			Terms: []string{"account", "details", "name", "email", "identity", "profile",
+				"who am i", "my details"},
+		},
+		{
+			ID: "profile-security", Text: "How is my account secured?",
+			Terms: []string{"security", "secure", "secured", "password", "two-factor",
+				"two factor", "2fa", "session", "sessions", "sign out", "log out",
+				"protect", "protected"},
+		},
+		{
+			ID: "profile-preferences", Text: "Which preferences are set?",
+			Terms: []string{"preference", "preferences", "setting", "settings", "digest",
+				"density", "owliver", "default", "defaults", "toggle"},
+		},
+		{
+			ID: "profile-activity", Text: "What have I done recently?",
+			Subject: "my recent activity", Shapes: []string{"timeline", "list", "table"},
+			Terms: []string{"activity", "recent", "recently", "history", "what have i",
+				"my actions", "audit", "did i"},
+			Reads: []Need{ownActivity},
+		},
+		{
+			ID: "profile-actions", Text: "What can I do on this page?",
+			Terms: []string{"do here", "what can i", "action", "actions", "edit", "change",
+				"change my", "update", "manage"},
+		},
+	},
+}
+
+// Pages is every page the catalogue holds intents for, for tests and
+// diagnostics. It is not the set of valid pages — that is
+// definition.SupportedPages, and a valid page absent from here answers with an
+// empty list rather than a validation error.
+func Pages() []string {
+	out := make([]string, 0, len(catalogue))
+	for page := range catalogue {
+		out = append(out, page)
+	}
+	return out
+}

go-api/internal/owliver/suggest.go

@@ -0,0 +1,359 @@
+package owliver
+
+import (
+	"sort"
+	"strings"
+	"unicode"
+
+	"github.com/krow/krow-backend/go-api/internal/domain"
+)
+
+// MaxSuggestions is the most a response may carry.
+//
+// Three, because the panel shows them under a composer the user is still typing
+// into. A fourth line pushes the input off a phone screen, and a ranked list
+// nobody reads to the bottom is a longer list, not a better one.
+const MaxSuggestions = 3
+
+// MinQueryChars is the shortest query that is worth ranking.
+//
+// Counted in letters and digits after normalization, so "  a  " and "?!" are
+// both too short. One character matches a prefix of almost every term in the
+// catalogue, which would make the first keystroke return three arbitrary
+// readings and the second replace all three — noise that reads as a bug.
+const MinQueryChars = 2
+
+// MaxQueryChars bounds the work one request can ask for. The panel sends what
+// is in the composer, and nothing about a suggestion improves past a couple of
+// sentences. Beyond this the query is truncated, never rejected: a long paste
+// should rank on its opening words, not fail.
+const MaxQueryChars = 200
+
+// Suggestion is one offered question.
+//
+// Text is what the user reads. Intent is the frontend capability id the panel
+// dispatches on. Capability is the section type the answer should be drawn as,
+// present only when the query asked for one — nothing internal is exposed here:
+// no terms, no resource names, no policy detail, no scores.
+type Suggestion struct {
+	Text       string `json:"text"`
+	Intent     string `json:"intent"`
+	Capability string `json:"capability,omitempty"`
+}
+
+/* ── Shapes ─────────────────────────────────────────────────────────────── */
+
+// shape is one section type an answer can be drawn as.
+//
+// Transcribed from OWLIVER_CAPABILITIES in src/lib/skills/surfaces.js: the ids
+// and the terms are that table's, and `phrase` is how the id reads inside a
+// sentence. `summary` is first and has no component — it is prose, so it
+// applies to any intent with a subject.
+//
+// Terms here describe the SHAPE and never a subject, which is what keeps a
+// shape from dragging in another page's readings: "as a flow" belongs here,
+// "hiring activity" belongs to an intent.
+type shape struct {
+	id     string
+	phrase string
+	terms  []string
+}
+
+var shapes = []shape{
+	{id: "summary", phrase: "", terms: []string{
+		"summary", "summarise", "summarize", "summarised", "summarized",
+		"summarising", "summarizing", "sum up", "recap", "overview", "brief me",
+		"in short", "tell me about"}},
+	{id: "flow", phrase: "as a flow", terms: []string{
+		"flow", "as a flow", "chart", "graph", "diagram", "funnel", "visual",
+		"visualise", "visualize", "step by step"}},
+	{id: "stats", phrase: "as stats", terms: []string{
+		"stats", "statistics", "figures", "numbers", "counts"}},
+	{id: "list", phrase: "as a list", terms: []string{
+		"list", "which ones", "show me the records"}},
+	{id: "table", phrase: "as a table", terms: []string{
+		"table", "as a table", "rows", "grid", "spreadsheet"}},
+	{id: "timeline", phrase: "as a timeline", terms: []string{
+		"timeline", "chronology", "over time", "what happened"}},
+	{id: "progress", phrase: "as progress bars", terms: []string{
+		"progress", "bars", "completion", "how far"}},
+	{id: "weights", phrase: "as weights", terms: []string{
+		"weighting", "weightings", "set the weights", "adjust the weights",
+		"screening weight", "vetting weight"}},
+	{id: "insight", phrase: "as an insight", terms: []string{
+		"insight", "finding", "takeaway", "headline"}},
+	{id: "card", phrase: "as a card", terms: []string{
+		"card", "panel", "at a glance"}},
+}
+
+/* ── Normalization ──────────────────────────────────────────────────────── */
+
+// normalize reduces a raw query to the one form everything downstream matches
+// against: lower case, letters and digits only, single-spaced.
+//
+// Every other character — punctuation, quotes, brackets, control characters,
+// emoji, an SQL fragment, a script tag — becomes a space rather than being
+// stripped, so nothing can be glued into a token that was not typed as one.
+// The result is compared against a fixed table of literals and never reaches a
+// query, a template or a log message, so there is no construction to inject
+// into; this is about matching sanely, not about escaping.
+func normalize(raw string) (phrase string, tokens []string, meaningful int) {
+	runes := []rune(raw)
+	if len(runes) > MaxQueryChars {
+		runes = runes[:MaxQueryChars]
+	}
+
+	var b strings.Builder
+	b.Grow(len(runes))
+	for _, r := range runes {
+		switch {
+		case unicode.IsLetter(r) || unicode.IsDigit(r):
+			b.WriteRune(unicode.ToLower(r))
+			meaningful++
+		default:
+			b.WriteByte(' ')
+		}
+	}
+
+	tokens = strings.Fields(b.String())
+	return strings.Join(tokens, " "), tokens, meaningful
+}
+
+/* ── Scoring ────────────────────────────────────────────────────────────── */
+
+// Scores are small integers with a deliberate order:
+//
+//	exact      a token is the term                     — the user typed it
+//	phrase     a multi-word term appears in the query  — the most specific hit
+//	prefix     the term begins with a token            — mid-typing: "pipel"
+//	extension  a token begins with the term            — "pipelines"
+//	shaped     the query named a section type          — weakest on its own
+//
+// A shape hit is worth less than any subject hit, so typing "overview" can
+// surface a page's readings but can never outrank a reading the user named.
+const (
+	scoreExact     = 10
+	scorePhrase    = 12
+	scorePrefix    = 6
+	scoreExtension = 5
+	scoreShaped    = 4
+
+	// minPrefixToken keeps one- and two-letter tokens from matching a term by
+	// prefix. "a" begins nothing usefully; "at" would match "attention",
+	// "audit" and "activity" at once.
+	minPrefixToken = 3
+	// minExtensionTerm keeps a short term from being found inside a longer
+	// word: without it "list" matches "listen" and "score" matches "scoreboard".
+	minExtensionTerm = 4
+)
+
+// tokenScore is how well one typed token matches one single-word term.
+func tokenScore(token, term string) int {
+	switch {
+	case token == term:
+		return scoreExact
+	case len(token) >= minPrefixToken && strings.HasPrefix(term, token):
+		return scorePrefix
+	case len(term) >= minExtensionTerm && strings.HasPrefix(token, term):
+		return scoreExtension
+	default:
+		return 0
+	}
+}
+
+// termsScore ranks a whole term list against the query.
+//
+// Multi-word terms are matched against the phrase, because "how many" means
+// something its two words do not. Single-word terms are scored per TYPED TOKEN,
+// taking that token's best term — so a query is rewarded for how much of what
+// the user typed the intent accounts for, and an intent cannot climb the
+// ranking by listing eight synonyms of one word.
+func termsScore(terms []string, tokens []string, phrase string) int {
+	total := 0
+	for _, term := range terms {
+		if !strings.Contains(term, " ") {
+			continue
+		}
+		if strings.Contains(phrase, term) {
+			total += scorePhrase + 2*strings.Count(term, " ")
+		}
+	}
+	for _, token := range tokens {
+		best := 0
+		for _, term := range terms {
+			if strings.Contains(term, " ") {
+				continue
+			}
+			if s := tokenScore(token, term); s > best {
+				best = s
+			}
+		}
+		total += best
+	}
+	return total
+}
+
+// matchShape is the section type the query asked for, if it asked for one.
+// Highest scoring wins; ties go to declaration order, which puts `summary`
+// first.
+func matchShape(tokens []string, phrase string) (shape, bool) {
+	best, bestScore := shape{}, 0
+	for _, s := range shapes {
+		if score := termsScore(s.terms, tokens, phrase); score > bestScore {
+			best, bestScore = s, score
+		}
+	}
+	return best, bestScore > 0
+}
+
+// supportsShape reports whether an intent can be drawn as a section type.
+// `summary` needs only a subject, having no component of its own; every other
+// shape must be one the intent declares.
+func (i Intent) supportsShape(id string) bool {
+	if i.Subject == "" {
+		return false
+	}
+	if id == "summary" {
+		return true
+	}
+	for _, s := range i.Shapes {
+		if s == id {
+			return true
+		}
+	}
+	return false
+}
+
+// shaped is the suggestion text for an intent asked for in a given shape.
+func (i Intent) shaped(s shape) string {
+	if s.id == "summary" {
+		return "Summarize " + i.Subject
+	}
+	return "Show " + i.Subject + " " + s.phrase
+}
+
+/* ── The pipeline ───────────────────────────────────────────────────────── */
+
+// filterOnTopic keeps the candidates the query actually named, or nothing if
+// it named none.
+func filterOnTopic(candidates []scored) []scored {
+	out := make([]scored, 0, len(candidates))
+	for _, c := range candidates {
+		if c.onTopic {
+			out = append(out, c)
+		}
+	}
+	return out
+}
+
+// scored is one candidate on its way through ranking.
+type scored struct {
+	suggestion Suggestion
+	score      int
+	order      int // declaration index, the tie-break
+
+	// onTopic records that the query matched this reading's own terms, rather
+	// than only naming a section type it happens to support. See Suggest.
+	onTopic bool
+}
+
+// Suggest ranks the page's catalogue against what the user has typed.
+//
+// The order is fixed and each stage only ever removes: page context, then
+// permission, then relevance, then duplicates, then the cap. Permission comes
+// before relevance so a reading the caller cannot perform is never scored, and
+// therefore cannot be leaked by an ordering bug later.
+//
+// It returns an empty slice, never nil and never a filler suggestion: a query
+// that matches nothing on this page has no answer here, and saying so is more
+// useful than three questions the user did not ask.
+func Suggest(page, query string, role domain.Role) []Suggestion {
+	out := []Suggestion{}
+
+	// Deny by default, as policy.go does. The service resolves the role before
+	// calling, so an unrecognised one should be unreachable — but an intent
+	// that reads nothing is permitted by every role it is asked about, so
+	// without this line a caller whose role failed to parse would be offered
+	// the account readings. The check belongs where the answer is decided.
+	if _, known := domain.ParseRole(string(role)); !known {
+		return out
+	}
+
+	intents, ok := catalogue[page]
+	if !ok {
+		return out
+	}
+
+	phrase, tokens, meaningful := normalize(query)
+	if meaningful < MinQueryChars {
+		return out
+	}
+
+	requested, wantsShape := matchShape(tokens, phrase)
+
+	candidates := make([]scored, 0, len(intents))
+	for order, intent := range intents {
+		if !intent.permitted(role) {
+			continue
+		}
+
+		score := termsScore(intent.Terms, tokens, phrase)
+		onTopic := score > 0
+
+		suggestion := Suggestion{Text: intent.Text, Intent: intent.ID}
+		if wantsShape && intent.supportsShape(requested.id) {
+			score += scoreShaped
+			suggestion.Text = intent.shaped(requested)
+			suggestion.Capability = requested.id
+		}
+
+		if score == 0 {
+			continue
+		}
+		candidates = append(candidates, scored{
+			suggestion: suggestion, score: score, order: order, onTopic: onTopic,
+		})
+	}
+
+	// A shape on its own is a weak signal, and what it means depends on what
+	// else matched. "as a table" typed alone is a real request — draw this
+	// page's readings that way — but the same words after "hiring activity" are
+	// how the user asked for ONE reading, and offering two more that merely
+	// support tables is the padding this endpoint is supposed to refuse.
+	//
+	// So the two are kept in separate tiers: if anything matched the query's
+	// subject, only those compete. Shape-only matches answer for the whole page
+	// or not at all.
+	if onTopic := filterOnTopic(candidates); len(onTopic) > 0 {
+		candidates = onTopic
+	}
+
+	// Highest score first; declaration order breaks every tie, so the same
+	// request always produces the same three in the same sequence.
+	sort.SliceStable(candidates, func(a, b int) bool {
+		if candidates[a].score != candidates[b].score {
+			return candidates[a].score > candidates[b].score
+		}
+		return candidates[a].order < candidates[b].order
+	})
+
+	// One suggestion per intent, and no two reading the same. The catalogue is
+	// unique per page by construction — TestCatalogueIsWellFormed holds it that
+	// way — so this guards the shaped rewrite, which can phrase two intents
+	// identically only if two subjects ever collide.
+	seenIntent := make(map[string]bool, MaxSuggestions)
+	seenText := make(map[string]bool, MaxSuggestions)
+	for _, c := range candidates {
+		if len(out) == MaxSuggestions {
+			break
+		}
+		key := strings.ToLower(c.suggestion.Text)
+		if seenIntent[c.suggestion.Intent] || seenText[key] {
+			continue
+		}
+		seenIntent[c.suggestion.Intent] = true
+		seenText[key] = true
+		out = append(out, c.suggestion)
+	}
+	return out
+}

go-api/internal/owliver/suggest_test.go

@@ -0,0 +1,695 @@
+package owliver
+
+import (
+	"fmt"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/krow/krow-backend/go-api/internal/definition"
+	"github.com/krow/krow-backend/go-api/internal/domain"
+)
+
+// These tests need no database and no server: the catalogue is static and the
+// ranking is a pure function of (page, query, role). That is the property worth
+// protecting — an endpoint the panel calls on every keystroke should be
+// testable at the speed of a string comparison.
+
+// intents is the ids Suggest returned, in order.
+func intents(got []Suggestion) []string {
+	out := make([]string, len(got))
+	for i, s := range got {
+		out[i] = s.Intent
+	}
+	return out
+}
+
+// ask is Suggest for an admin, the role the Owliver panel is placed for.
+func ask(page, query string) []Suggestion {
+	return Suggest(page, query, domain.RoleAdmin)
+}
+
+/* ── Control Center ─────────────────────────────────────────────────────── */
+
+func TestControlCenterSuggestions(t *testing.T) {
+	cases := []struct {
+		name  string
+		query string
+		want  []string
+	}{
+		{"attention", "attention", []string{"attention-required"}},
+		{"pipeline", "pipeline", []string{"pipeline-health"}},
+		{"health", "health", []string{"platform-health"}},
+		{"recommendation", "what should i do", []string{"recommendations"}},
+
+		// A question about a subject this page does not hold. The Control
+		// Center reads the platform, not the training library.
+		{"irrelevant", "forklift certification renewal", nil},
+		{"empty", "", nil},
+	}
+
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got := intents(ask("control-center", c.query))
+			if len(c.want) == 0 {
+				if len(got) != 0 {
+					t.Fatalf("query %q: want no suggestions, got %v", c.query, got)
+				}
+				return
+			}
+			if !reflect.DeepEqual(got, c.want) {
+				t.Fatalf("query %q: got %v, want %v", c.query, got, c.want)
+			}
+		})
+	}
+}
+
+/* ── Positions ──────────────────────────────────────────────────────────── */
+
+func TestPositionsSuggestions(t *testing.T) {
+	cases := []struct {
+		name  string
+		query string
+		want  []string
+	}{
+		// The page's own noun offers the page's readings, in declaration order.
+		{"position", "position", []string{"position-drafts", "position-strength", "positions-attention"}},
+
+		// Pipeline on Positions is a question about the ROLES: which one is
+		// converting, and where it is stuck. Two answers, not three — the third
+		// slot is left empty rather than filled with the page's list of people
+		// waiting, which is a different question wearing a nearby word.
+		{"pipeline", "pipeline", []string{"position-strength", "pipeline-health"}},
+
+		{"attention", "attention", []string{"positions-attention"}},
+		{"risk", "risk", []string{"positions-attention"}},
+		{"drafts", "draft", []string{"position-drafts"}},
+		{"fill first", "what should i fill first", []string{"hiring-priority"}},
+
+		// Attendance is a workforce reading and belongs to another surface. It
+		// must not fall through to this page's default report.
+		{"irrelevant", "attendance last week", nil},
+		{"empty", "", nil},
+	}
+
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got := intents(ask("positions", c.query))
+			if len(c.want) == 0 {
+				if len(got) != 0 {
+					t.Fatalf("query %q: want no suggestions, got %v", c.query, got)
+				}
+				return
+			}
+			if !reflect.DeepEqual(got, c.want) {
+				t.Fatalf("query %q: got %v, want %v", c.query, got, c.want)
+			}
+		})
+	}
+}
+
+/* ── Candidates ─────────────────────────────────────────────────────────── */
+
+func TestCandidatesSuggestions(t *testing.T) {
+	cases := []struct {
+		name  string
+		query string
+		want  []string
+	}{
+		{"candidate", "candidate", []string{"candidates-attention", "top-candidates", "interview-ready"}},
+		{"score", "score", []string{"top-candidates", "screening-gaps"}},
+		{"interview", "interview", []string{"interview-ready"}},
+		{"decision", "decision", []string{"candidates-attention", "candidate-risk"}},
+		{"pipeline", "pipeline", []string{"pipeline-summary"}},
+		{"risk", "risk", []string{"candidate-risk"}},
+
+		{"irrelevant", "payroll export", nil},
+		{"empty", "", nil},
+	}
+
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got := intents(ask("candidates", c.query))
+			if len(c.want) == 0 {
+				if len(got) != 0 {
+					t.Fatalf("query %q: want no suggestions, got %v", c.query, got)
+				}
+				return
+			}
+			if !reflect.DeepEqual(got, c.want) {
+				t.Fatalf("query %q: got %v, want %v", c.query, got, c.want)
+			}
+		})
+	}
+}
+
+/* ── Page context is the first filter ───────────────────────────────────── */
+
+// One keyword, every page: the answers must differ, and none may name a
+// reading belonging to another surface.
+func TestSameKeywordDiffersByPage(t *testing.T) {
+	const query = "pipeline"
+
+	seen := map[string][]string{}
+	for _, page := range Pages() {
+		got := intents(ask(page, query))
+		if len(got) == 0 {
+			continue
+		}
+		seen[page] = got
+
+		for _, id := range got {
+			if !declaredOn(page, id) {
+				t.Fatalf("page %q returned %q, which it does not declare", page, id)
+			}
+		}
+	}
+
+	if len(seen) < 2 {
+		t.Fatalf("%q matched on %d pages; the comparison needs at least two", query, len(seen))
+	}
+	if reflect.DeepEqual(seen["positions"], seen["candidates"]) {
+		t.Fatalf("positions and candidates both answered %q with %v", query, seen["positions"])
+	}
+	// The pipeline reading on Candidates is the candidates' own.
+	if !reflect.DeepEqual(seen["candidates"], []string{"pipeline-summary"}) {
+		t.Fatalf("candidates answered %q with %v", query, seen["candidates"])
+	}
+}
+
+// An unknown page is not this package's error to raise — the service refuses it
+// during parsing. Reached directly it answers empty rather than borrowing
+// another page's readings.
+func TestUnknownPageIsEmpty(t *testing.T) {
+	for _, page := range []string{"", "nowhere", "POSITIONS", "settings"} {
+		if got := ask(page, "pipeline"); len(got) != 0 {
+			t.Fatalf("page %q: got %v, want none", page, got)
+		}
+	}
+}
+
+func declaredOn(page, id string) bool {
+	for _, i := range catalogue[page] {
+		if i.ID == id {
+			return true
+		}
+	}
+	return false
+}
+
+/* ── Query handling ─────────────────────────────────────────────────────── */
+
+func TestQueryNormalization(t *testing.T) {
+	want := intents(ask("positions", "pipeline"))
+	if len(want) == 0 {
+		t.Fatal("the baseline query matched nothing")
+	}
+
+	// Every one of these is the same question typed differently: case,
+	// surrounding whitespace, punctuation and control characters carry no
+	// meaning, so all of them must rank identically.
+	for _, query := range []string{
+		"  pipeline  ", "PIPELINE", "PiPeLiNe", "\tpipeline\n",
+		"pipeline?", "\"pipeline\"", "pipeline!!!", "…pipeline…",
+		"(pipeline)", "**pipeline**", "pipeline\x00\x01", "\u200bpipeline",
+	} {
+		if got := intents(ask("positions", query)); !reflect.DeepEqual(got, want) {
+			t.Errorf("query %q: got %v, want %v", query, got, want)
+		}
+	}
+}
+
+// Hostile input is data like any other. There is no query to inject into — the
+// normalized text is compared against a fixed table of literals and never
+// reaches SQL, a template or a shell — so the property under test is that such
+// a query is ranked rather than refused, and that it can only ever produce
+// entries this page declares.
+func TestHostileInputIsJustText(t *testing.T) {
+	for _, query := range []string{
+		"pipeline'; DROP TABLE job_postings; --",
+		"pipeline\" OR 1=1 --",
+		"<script>alert('pipeline')</script>",
+		"{{7*7}} pipeline ${jndi:ldap://x/y}",
+		"../../etc/passwd pipeline",
+		"pipeline%00%0d%0aSet-Cookie:+x=1",
+		strings.Repeat("' OR ''='", 40),
+	} {
+		for _, s := range ask("positions", query) {
+			if !declaredOn("positions", s.Intent) {
+				t.Errorf("query %q produced %q, which positions does not declare", query, s.Intent)
+			}
+			if !declaredText("positions", s) {
+				t.Errorf("query %q produced unrecognised text %q", query, s.Text)
+			}
+		}
+	}
+}
+
+// declaredText reports whether a suggestion's wording came from the catalogue —
+// either an intent's own Text, or its subject phrased in a shape it declares.
+// Nothing the caller typed may appear in a response.
+func declaredText(page string, got Suggestion) bool {
+	for _, i := range catalogue[page] {
+		if i.ID != got.Intent {
+			continue
+		}
+		if got.Capability == "" {
+			return got.Text == i.Text
+		}
+		for _, s := range shapes {
+			if s.id == got.Capability {
+				return got.Text == i.shaped(s)
+			}
+		}
+	}
+	return false
+}
+
+// Fewer than two meaningful characters is not a question yet. Punctuation and
+// whitespace are not meaningful.
+func TestShortQueriesAreEmpty(t *testing.T) {
+	for _, query := range []string{"", " ", "\n\t ", "p", " p ", "?", "!!!", "-", "€", "  ,  "} {
+		if got := ask("positions", query); len(got) != 0 {
+			t.Errorf("query %q: got %v, want none", query, got)
+		}
+	}
+}
+
+// The panel calls this on every keystroke, so a half-typed word has to match.
+func TestPrefixMatchingWhileTyping(t *testing.T) {
+	full := intents(ask("positions", "pipeline"))
+	for _, query := range []string{"pip", "pipe", "pipel", "pipelin", "pipeline", "pipelines"} {
+		got := intents(ask("positions", query))
+		if len(got) == 0 {
+			t.Fatalf("query %q matched nothing; the panel would blank mid-word", query)
+		}
+		if query != "pipelines" && !reflect.DeepEqual(got, full) {
+			t.Errorf("query %q: got %v, want %v", query, got, full)
+		}
+	}
+}
+
+// A long paste ranks on its opening words rather than being refused.
+func TestOverlongQueryIsTruncatedNotRejected(t *testing.T) {
+	query := "pipeline " + strings.Repeat("x", 5000)
+	got := intents(ask("positions", query))
+	if len(got) == 0 {
+		t.Fatal("an overlong query was refused instead of truncated")
+	}
+	if !reflect.DeepEqual(got, intents(ask("positions", "pipeline"))) {
+		t.Fatalf("an overlong query ranked differently: %v", got)
+	}
+}
+
+/* ── Shapes ─────────────────────────────────────────────────────────────── */
+
+// Asking for a section type names it in the answer and phrases the suggestion
+// in those terms — the brief's "Show hiring activity as a flow".
+func TestShapedSuggestions(t *testing.T) {
+	got := ask("positions", "show hiring activity as a flow")
+	if len(got) != 1 {
+		t.Fatalf("got %d suggestions, want 1: %+v", len(got), got)
+	}
+	want := Suggestion{
+		Text:       "Show hiring activity as a flow",
+		Intent:     "hiring-operations",
+		Capability: "flow",
+	}
+	if got[0] != want {
+		t.Fatalf("got %+v, want %+v", got[0], want)
+	}
+
+	summarized := ask("positions", "summarize hiring activity")
+	if len(summarized) != 1 || summarized[0].Capability != "summary" ||
+		summarized[0].Text != "Summarize hiring activity" {
+		t.Fatalf("got %+v", summarized)
+	}
+}
+
+// A shape alone is a question about the page. A shape after a subject is a
+// question about that subject, and the other readings that merely support the
+// shape are padding — which this endpoint does not do.
+func TestShapeAloneAnswersThePageButNeverPads(t *testing.T) {
+	alone := ask("control-center", "summarize")
+	if len(alone) != MaxSuggestions {
+		t.Fatalf("a bare shape returned %d suggestions, want %d: %+v",
+			len(alone), MaxSuggestions, alone)
+	}
+	for _, s := range alone {
+		if s.Capability != "summary" {
+			t.Fatalf("got capability %q, want summary: %+v", s.Capability, s)
+		}
+	}
+
+	// "attention" names one reading; nothing else may ride along on the shape.
+	withSubject := ask("control-center", "summarize what needs attention")
+	if len(withSubject) != 1 || withSubject[0].Intent != "attention-required" {
+		t.Fatalf("got %+v, want only attention-required", withSubject)
+	}
+}
+
+// A shape an intent cannot be drawn as leaves its wording alone.
+func TestUnsupportedShapeIsNotClaimed(t *testing.T) {
+	for _, s := range ask("create-position", "adjust the weights") {
+		if s.Capability == "weights" {
+			t.Fatalf("offered a weights rendering nothing declares: %+v", s)
+		}
+	}
+}
+
+/* ── Response limits ────────────────────────────────────────────────────── */
+
+func TestNeverMoreThanThreeAndNeverDuplicated(t *testing.T) {
+	// A query broad enough to match everything the page has.
+	queries := []string{
+		"position pipeline attention risk draft hiring activity waiting candidates",
+		"candidate score interview decision risk pipeline screening",
+		"summarize", "attention risk", "who what how many",
+	}
+
+	for _, page := range Pages() {
+		for _, query := range queries {
+			got := Suggest(page, query, domain.RoleAdmin)
+			if len(got) > MaxSuggestions {
+				t.Fatalf("page %q query %q: %d suggestions, cap is %d",
+					page, query, len(got), MaxSuggestions)
+			}
+
+			seenIntent, seenText := map[string]bool{}, map[string]bool{}
+			for _, s := range got {
+				if s.Text == "" || s.Intent == "" {
+					t.Fatalf("page %q query %q: incomplete suggestion %+v", page, query, s)
+				}
+				if seenIntent[s.Intent] {
+					t.Fatalf("page %q query %q: duplicate intent %q", page, query, s.Intent)
+				}
+				if seenText[strings.ToLower(s.Text)] {
+					t.Fatalf("page %q query %q: duplicate text %q", page, query, s.Text)
+				}
+				seenIntent[s.Intent], seenText[strings.ToLower(s.Text)] = true, true
+			}
+		}
+	}
+}
+
+// The same request must answer the same way every time — the panel re-issues it
+// on every keystroke, and a list that reshuffles under the cursor is unusable.
+func TestSuggestIsDeterministic(t *testing.T) {
+	for _, page := range Pages() {
+		first := Suggest(page, "attention risk pipeline summary", domain.RoleAdmin)
+		for i := 0; i < 20; i++ {
+			again := Suggest(page, "attention risk pipeline summary", domain.RoleAdmin)
+			if !reflect.DeepEqual(first, again) {
+				t.Fatalf("page %q: run %d differed\n first: %+v\n again: %+v",
+					page, i, first, again)
+			}
+		}
+	}
+}
+
+// Empty, never nil: `{"suggestions": []}` and not `{"suggestions": null}`.
+func TestNoMatchIsAnEmptySliceNotNil(t *testing.T) {
+	for _, c := range []struct{ page, query string }{
+		{"positions", "sourdough"}, {"positions", ""}, {"nowhere", "pipeline"},
+	} {
+		got := ask(c.page, c.query)
+		if got == nil {
+			t.Fatalf("page %q query %q: got nil, want an empty slice", c.page, c.query)
+		}
+		if len(got) != 0 {
+			t.Fatalf("page %q query %q: got %v", c.page, c.query, got)
+		}
+	}
+}
+
+/* ── Authorization ──────────────────────────────────────────────────────── */
+
+// Talent may list job applications, but only their own — so a reading across
+// the organization's pipeline is not theirs to be offered, even though the
+// operation itself is permitted. The same holds for postings, profiles, staff,
+// evidence and the audit log.
+//
+// The exception is stated rather than hidden: `courses` is the one resource in
+// the policy table that talent lists unscoped, because the training library is
+// shared platform-wide and everybody learns from it. So the two Forge readings
+// that ask only what the library holds survive, and every other Forge reading —
+// evaluation, workforce usage, gaps, all of which read evidence or profiles —
+// does not. If that ever widens, this test says exactly what widened.
+func TestTalentIsOfferedOnlyUnscopedReadings(t *testing.T) {
+	pages := []string{
+		"control-center", "positions", "candidates", "candidates-analysis",
+		"analytics", "activity", "talent-pool", "hired-history", "krow-forge",
+		"create-position",
+	}
+	queries := []string{
+		"pipeline", "attention", "candidate", "position", "risk", "summarize",
+		"hiring", "score", "activity", "who", "how many", "library", "skill",
+		"published", "gaps", "evaluation", "weights", "credential",
+	}
+
+	allowed := map[string]bool{"krow-forge/forge-library": true, "krow-forge/forge-published": true}
+
+	for _, page := range pages {
+		for _, query := range queries {
+			for _, s := range Suggest(page, query, domain.RoleTalent) {
+				if !allowed[page+"/"+s.Intent] {
+					t.Errorf("talent was offered %q on %q for %q", s.Intent, page, query)
+				}
+			}
+		}
+	}
+
+	// And the operator's own Forge readings stay the operator's.
+	for _, id := range []string{"forge-evaluation", "forge-workforce", "forge-gaps"} {
+		for _, s := range Suggest("krow-forge", "evaluation workforce gaps", domain.RoleTalent) {
+			if s.Intent == id {
+				t.Errorf("talent was offered the operator reading %q", id)
+			}
+		}
+	}
+	if len(Suggest("krow-forge", "evaluation workforce gaps", domain.RoleAdmin)) == 0 {
+		t.Error("admin was offered none of them either; the query no longer matches")
+	}
+}
+
+// The filter is not a blanket refusal: what a talent caller may genuinely ask —
+// about their own account — is still offered. Otherwise the test above would
+// pass with the role check stubbed out to "deny".
+func TestTalentIsStillOfferedTheirOwnReadings(t *testing.T) {
+	for _, query := range []string{"permission", "password", "my recent activity"} {
+		if got := Suggest("profile", query, domain.RoleTalent); len(got) == 0 {
+			t.Fatalf("talent was offered nothing on profile for %q", query)
+		}
+	}
+}
+
+// A role the API does not recognise authorizes nothing, matching policy.go.
+func TestUnknownRoleIsOfferedNothing(t *testing.T) {
+	for _, role := range []domain.Role{"", "root", "superuser", "Admin"} {
+		for _, page := range Pages() {
+			if got := Suggest(page, "attention pipeline permission", role); len(got) != 0 {
+				t.Fatalf("role %q was offered %v on %q", role, intents(got), page)
+			}
+		}
+	}
+}
+
+// Every permission decision must come from the policy table, not from a list
+// kept here. This asserts the mechanism rather than a particular outcome: an
+// intent is offered exactly when policy.go allows every reading it declares.
+func TestPermissionsComeFromThePolicyTable(t *testing.T) {
+	for _, role := range []domain.Role{domain.RoleAdmin, domain.RoleEmployer, domain.RoleTalent} {
+		for page, list := range catalogue {
+			for _, intent := range list {
+				want := true
+				for _, need := range intent.Reads {
+					res, ok := domain.ResourceByPath[need.Resource]
+					if !ok || !res.Supports(need.Op) || !res.Policy.Allows(need.Op, role) {
+						want = false
+						break
+					}
+					if need.OrgWide && res.Policy.ScopeFor(role).Kind != domain.ScopeNone {
+						want = false
+						break
+					}
+				}
+				if got := intent.permitted(role); got != want {
+					t.Errorf("%s/%s for %s: permitted=%v, policy says %v",
+						page, intent.ID, role, got, want)
+				}
+			}
+		}
+	}
+}
+
+/* ── The catalogue itself ───────────────────────────────────────────────── */
+
+func TestCatalogueIsWellFormed(t *testing.T) {
+	for page, list := range catalogue {
+		if definition.CanonicalPage(page) != page {
+			t.Errorf("page key %q is not a canonical surface", page)
+		}
+		if len(list) == 0 {
+			t.Errorf("page %q has no intents; omit the key instead", page)
+		}
+
+		ids, texts := map[string]bool{}, map[string]bool{}
+		for _, intent := range list {
+			where := fmt.Sprintf("%s/%s", page, intent.ID)
+
+			if intent.ID == "" || intent.Text == "" {
+				t.Errorf("%s: an intent needs both an id and a text", where)
+			}
+			if ids[intent.ID] {
+				t.Errorf("%s: duplicate intent id on this page", where)
+			}
+			if texts[strings.ToLower(intent.Text)] {
+				t.Errorf("%s: duplicate suggestion text on this page", where)
+			}
+			ids[intent.ID], texts[strings.ToLower(intent.Text)] = true, true
+
+			if len(intent.Terms) == 0 {
+				t.Errorf("%s: no terms, so it can never be suggested", where)
+			}
+			for _, term := range intent.Terms {
+				if term != strings.ToLower(strings.TrimSpace(term)) || term == "" {
+					t.Errorf("%s: term %q must be lower case and trimmed", where, term)
+				}
+				if _, _, meaningful := normalize(term); meaningful < MinQueryChars {
+					t.Errorf("%s: term %q is shorter than the shortest query", where, term)
+				}
+			}
+
+			if len(intent.Shapes) > 0 && intent.Subject == "" {
+				t.Errorf("%s: declares shapes but no subject to phrase them with", where)
+			}
+			for _, id := range intent.Shapes {
+				if id == "summary" {
+					t.Errorf("%s: `summary` applies to every subject and is never declared", where)
+				}
+				if !knownShape(id) {
+					t.Errorf("%s: shape %q is not in the Owliver vocabulary", where, id)
+				}
+			}
+
+			for _, need := range intent.Reads {
+				res, ok := domain.ResourceByPath[need.Resource]
+				if !ok {
+					t.Errorf("%s: reads %q, which is not a resource", where, need.Resource)
+					continue
+				}
+				if !res.Supports(need.Op) {
+					t.Errorf("%s: reads %q with an operation it does not serve", where, need.Resource)
+				}
+				// An intent nobody can be offered is dead weight, and usually a
+				// typo in the resource path rather than a deliberate lockout.
+				if !res.Policy.Allows(need.Op, domain.RoleAdmin) {
+					t.Errorf("%s: reads %q, which not even admin may list", where, need.Resource)
+				}
+			}
+		}
+	}
+}
+
+// Every id in the catalogue must be a capability the frontend actually
+// declares, because the id in a response is what the panel dispatches on. The
+// list is the union of the manifests in
+// src/components/ai-assistant/capabilities/, transcribed alongside the
+// catalogue; an id here that is absent there would be a suggestion the panel
+// cannot run.
+func TestIntentIDsAreFrontendCapabilities(t *testing.T) {
+	frontend := map[string]bool{}
+	for _, id := range []string{
+		// CONTROL_CENTER_CAPABILITIES
+		"platform-health", "workforce-summary", "hiring-operations", "pipeline-health",
+		"attention-required", "recommendations",
+		// POSITIONS_CAPABILITIES
+		"position-drafts", "position-strength", "positions-attention", "hiring-priority",
+		"candidates-waiting",
+		// CANDIDATE_LIST_CAPABILITIES
+		"candidates-attention", "top-candidates", "interview-ready", "screening-gaps",
+		"pipeline-summary", "candidate-risk",
+		// ADMIN_CANDIDATE_CAPABILITIES
+		"recruitment-insights", "hiring-recommendations",
+		// ADMIN_ANALYTICS_CAPABILITIES
+		"hiring-trend", "department-performance", "position-conversion",
+		// ACTIVITY_CAPABILITIES
+		"audit-summary", "user-activity", "unusual-activity", "security-insights",
+		// TALENT_POOL_CAPABILITIES
+		"talent-priorities", "talent-summary", "talent-verification", "talent-availability",
+		// HIRED_HISTORY_CAPABILITIES
+		"hiring-outcomes", "hiring-strongest", "hiring-patterns", "hiring-recent",
+		// FORGE_CAPABILITIES
+		"forge-library", "forge-published", "forge-evaluation", "forge-workforce", "forge-gaps",
+		// CREATE_POSITION_CAPABILITIES
+		"vetting-weights", "position-benchmarks", "position-requirements", "position-spec-steps",
+		// PROFILE_CAPABILITIES
+		"profile-permissions", "profile-identity", "profile-security", "profile-preferences",
+		"profile-activity", "profile-actions",
+	} {
+		frontend[id] = true
+	}
+
+	used := map[string]bool{}
+	for page, list := range catalogue {
+		for _, intent := range list {
+			used[intent.ID] = true
+			if !frontend[intent.ID] {
+				t.Errorf("%s/%s names no frontend capability", page, intent.ID)
+			}
+		}
+	}
+	for id := range frontend {
+		if !used[id] {
+			t.Errorf("capability %q is declared here but suggested on no page", id)
+		}
+	}
+}
+
+func knownShape(id string) bool {
+	for _, s := range shapes {
+		if s.id == id {
+			return true
+		}
+	}
+	return false
+}
+
+// The shape vocabulary is closed and mirrors OWLIVER_CAPABILITIES.
+func TestShapeVocabularyMatchesTheFrontend(t *testing.T) {
+	want := []string{"summary", "flow", "stats", "list", "table", "timeline",
+		"progress", "weights", "insight", "card"}
+
+	got := make([]string, len(shapes))
+	for i, s := range shapes {
+		got[i] = s.id
+		if len(s.terms) == 0 {
+			t.Errorf("shape %q has no terms", s.id)
+		}
+		if s.id != "summary" && s.phrase == "" {
+			t.Errorf("shape %q has no phrase to read inside a sentence", s.id)
+		}
+	}
+	if !reflect.DeepEqual(got, want) {
+		t.Fatalf("shapes are %v, want %v", got, want)
+	}
+}
+
+// Nothing internal may reach a response: no terms, no resource names, no
+// scores. The struct is the whole contract, so this asserts its shape.
+func TestSuggestionExposesNothingInternal(t *testing.T) {
+	fields := reflect.VisibleFields(reflect.TypeOf(Suggestion{}))
+	if len(fields) != 3 {
+		t.Fatalf("Suggestion has %d fields; the response contract is text, intent, capability", len(fields))
+	}
+	want := map[string]string{
+		"Text":       `json:"text"`,
+		"Intent":     `json:"intent"`,
+		"Capability": `json:"capability,omitempty"`,
+	}
+	for _, f := range fields {
+		if string(f.Tag) != want[f.Name] {
+			t.Errorf("field %s has tag %q, want %q", f.Name, f.Tag, want[f.Name])
+		}
+	}
+}

go-api/internal/service/interviews.go

@@ -0,0 +1,119 @@
+package service
+
+// Completing an AI interview, in one transaction.
+//
+// THE PROBLEM THIS SOLVES
+//
+// Finishing an interview is two writes: the interview record, and the
+// application it was for — which has to carry the verdict forward as
+// `status: interview`, `interview_id` and the score, because that is what the
+// funnel and the analytics read. The frontend performed them as two independent
+// requests (AIInterviewModal.finishInterview), and that had two consequences.
+//
+// The first is authorization. `ai-interviews:Create` is open to everyone and
+// `job-applications:Update` is operators only, so a talent user sitting their
+// own interview — which is the whole talent flow — got a 201 for the interview
+// and a 403 for the link. The interview existed, the application still said
+// `applied`, `interview_id` was never set, and every consumer that counts
+// `status === 'interview' || interview_id` could not see it.
+//
+// The second is atomicity: even for an operator, a failure between the two left
+// an interview attached to an application that did not know about it.
+//
+// WHY THE SERVER MAY WRITE WHAT THE CALLER MAY NOT
+//
+// The link is not a widening of `job-applications:Update`. A talent caller
+// still cannot PATCH an application — the policy table is unchanged, and the
+// role gate on that route still refuses them. What happens here is that the
+// server updates the one row the interview it just wrote already names, and
+// only after repo.guardInsert has proved that row belongs to the caller: a
+// talent caller creating an interview for an application that is not theirs is
+// answered 404 before anything is written. That guard is exactly the ownership
+// proof this update needs.
+//
+// The alternative — adding `talent` to `job-applications:Update` with a
+// per-column allowlist — was rejected in the plan for the reason the workflows
+// file header gives: it would put a second authorization mechanism beside the
+// per-operation one, and the two would eventually disagree.
+
+import (
+	"context"
+
+	"github.com/jackc/pgx/v5"
+
+	"github.com/krow/krow-backend/go-api/internal/authctx"
+	"github.com/krow/krow-backend/go-api/internal/domain"
+	"github.com/krow/krow-backend/go-api/internal/repo"
+)
+
+// InterviewsPath is the resource whose Create is routed through here.
+// Exported so the HTTP layer names the same resource this file special-cases,
+// rather than repeating a string literal that could drift.
+const InterviewsPath = "ai-interviews"
+
+// CreateInterview inserts an interview and links its application, atomically.
+//
+// The response is the interview record, unchanged: POST /api/v1/ai-interviews
+// answered 201 with the created interview before this existed and answers 201
+// with the created interview now. The application update is a consequence of
+// the request, not a second thing in it.
+func (s *WorkflowService) CreateInterview(ctx context.Context, ident authctx.Identity,
+	body domain.Record) (domain.Record, error) {
+
+	interviews, err := resourceByPath(InterviewsPath)
+	if err != nil {
+		return nil, err
+	}
+	apps, err := resourceByPath("job-applications")
+	if err != nil {
+		return nil, err
+	}
+
+	var out domain.Record
+	err = s.inTx(ctx, func(tx pgx.Tx) error {
+		// Through the resource's own service over the transaction, so the body
+		// is validated and the ownership guard runs exactly as they do on the
+		// plain create path. Nothing about the interview itself changes here.
+		created, err := New(interviews, tx).Create(ctx, ident, body)
+		if err != nil {
+			return err
+		}
+		out = created
+
+		applicationID, _ := created["application_id"].(string)
+		if applicationID == "" {
+			// Unreachable: application_id is NOT NULL and Required, so the
+			// create above would have refused. Checked rather than assumed
+			// because the alternative is an Update against an empty id.
+			return nil
+		}
+
+		patch := domain.Record{
+			"status":       "interview",
+			"interview_id": created["id"],
+		}
+		// The score moves onto the application only when the caller said
+		// something about it. Copying the column unconditionally would write
+		// the interview's default 0 over a real screening score, which is a
+		// loss caused by a field the request never mentioned.
+		if _, said := body["overall_interview_score"]; said {
+			patch["ai_score"] = created["overall_interview_score"]
+		}
+
+		updated, err := repo.New(apps, tx).Update(ctx, ident, applicationID, patch)
+		if err != nil {
+			return err
+		}
+		if updated == nil {
+			// Unreachable for the same reason the guard above passed: the row
+			// is in this organization and, for a talent caller, theirs. Kept so
+			// a silent no-op cannot pass for a completed interview.
+			return domain.NotFound(apps.Name, applicationID)
+		}
+		return nil
+	})
+	if err != nil {
+		return nil, err
+	}
+	return out, nil
+}

go-api/internal/service/owliver.go

@@ -0,0 +1,103 @@
+package service
+
+import (
+	"fmt"
+	"net/url"
+	"strings"
+
+	"github.com/krow/krow-backend/go-api/internal/authctx"
+	"github.com/krow/krow-backend/go-api/internal/definition"
+	"github.com/krow/krow-backend/go-api/internal/domain"
+	"github.com/krow/krow-backend/go-api/internal/owliver"
+)
+
+// allowedSuggestionParams names the accepted query parameters, on the pattern
+// of allowedDefinitionFilters: anything else is a caller mistake worth saying
+// out loud rather than a filter to ignore. It also keeps the endpoint from
+// quietly accepting a `role`, `org` or `user` parameter should one ever be
+// added by a client — who is asking is read from the session and nowhere else.
+var allowedSuggestionParams = map[string]bool{
+	"page":  true,
+	"query": true,
+}
+
+// maxEchoedPage bounds how much of a rejected page value is quoted back. Long
+// enough to name every real surface, short enough that the error cannot be used
+// to reflect a payload.
+const maxEchoedPage = 64
+
+// SuggestionQuery is a validated suggestion request.
+//
+// Page is canonical: aliases are resolved here so nothing downstream has to
+// know that `hired` and `hired-history` are the same surface.
+type SuggestionQuery struct {
+	Page  string
+	Query string
+}
+
+// SuggestionsService answers "what could I usefully ask on this page?".
+//
+// It holds no pool, opens no transaction and reads no table. That is not an
+// omission — the panel calls it while the user types, and everything it needs
+// is the static catalogue in internal/owliver plus the caller's role. It is a
+// service rather than a function in the handler so that validation and
+// authorization sit where every other endpoint's do.
+type SuggestionsService struct{}
+
+// NewSuggestions builds the suggestion service.
+func NewSuggestions() *SuggestionsService { return &SuggestionsService{} }
+
+// ParseParams validates the query string.
+//
+// `page` is required and must name a real surface — the same closed vocabulary
+// internal/definition validates a definition's `pages:` against, so there is
+// one answer to "is that a page" in this process. `query` is optional: an
+// absent or too-short one is not an error, it is a request that has nothing to
+// rank yet, and Suggest answers it with an empty list.
+func (s *SuggestionsService) ParseParams(q url.Values) (SuggestionQuery, error) {
+	var out SuggestionQuery
+
+	for name := range q {
+		if !allowedSuggestionParams[name] {
+			return out, domain.Invalid(fmt.Sprintf("unknown parameter %q", name))
+		}
+	}
+
+	raw := strings.TrimSpace(q.Get("page"))
+	if raw == "" {
+		return out, domain.Invalid("page is required")
+	}
+	page := definition.CanonicalPage(raw)
+	if page == "" {
+		echoed := raw
+		if len(echoed) > maxEchoedPage {
+			echoed = echoed[:maxEchoedPage]
+		}
+		return out, domain.Invalid(fmt.Sprintf(
+			"Unsupported page: %s. Supported pages: %s.",
+			echoed, strings.Join(definition.SupportedPages, ", ")))
+	}
+
+	out.Page = page
+	out.Query = q.Get("query")
+	return out, nil
+}
+
+// Suggest ranks the page's readings for this caller.
+//
+// The role comes off the session-resolved identity, exactly as Server.authorize
+// reads it, and an unrecognised role is offered nothing — the same deny-by-
+// default the policy table applies. Nothing else about the caller is consulted:
+// there is no branch here on organization, account type or anything a request
+// could set.
+//
+// No error case beyond parsing. A page with no readings for this caller, and a
+// query that matches none of them, both answer with an empty list — an empty
+// result is an answer, not a failure.
+func (s *SuggestionsService) Suggest(ident authctx.Identity, q SuggestionQuery) []owliver.Suggestion {
+	role, known := domain.ParseRole(ident.Role)
+	if !known {
+		return []owliver.Suggestion{}
+	}
+	return owliver.Suggest(q.Page, q.Query, role)
+}

go-api/internal/service/service.go

@@ -150,7 +150,7 @@ func (s *Service) Get(ctx context.Context, ident authctx.Identity, id string) (d
 
 // Create validates and inserts, returning the complete stored record.
 func (s *Service) Create(ctx context.Context, ident authctx.Identity, body domain.Record) (domain.Record, error) {
-	clean, err := s.validate(body, true)
+	clean, err := s.validate(ident, body, true)
 	if err != nil {
 		return nil, err
 	}
@@ -162,7 +162,7 @@ func (s *Service) Update(ctx context.Context, ident authctx.Identity, id string,
 	if !isUUID(id) {
 		return nil, domain.NotFound(s.res.Name, id)
 	}
-	clean, err := s.validate(patch, false)
+	clean, err := s.validate(ident, patch, false)
 	if err != nil {
 		return nil, err
 	}
@@ -201,7 +201,11 @@ func (s *Service) Delete(ctx context.Context, ident authctx.Identity, id string)
 // exactly how `interview_id`, `training_outline` and `score_breakdown` would
 // have been lost: the frontend would have written them, the API would have
 // accepted the request, and the data would never have arrived.
-func (s *Service) validate(in domain.Record, isCreate bool) (domain.Record, error) {
+//
+// The identity is a parameter because "the server will supply this column"
+// is not a property of the column alone: a talent-only derivation supplies it
+// for a talent caller and for nobody else. See serverSupplies.
+func (s *Service) validate(ident authctx.Identity, in domain.Record, isCreate bool) (domain.Record, error) {
 	details := map[string]string{}
 	out := make(domain.Record, len(in))
 
@@ -237,7 +241,7 @@ func (s *Service) validate(in domain.Record, isCreate bool) (domain.Record, erro
 			if !col.Required {
 				continue
 			}
-			if s.serverSupplies(col.Name) {
+			if s.serverSupplies(col.Name, ident) {
 				// The repository fills this from the session, so demanding it
 				// from the caller would reject a request the server is about to
 				// complete correctly. evidence.worker_email is the live case.
@@ -262,13 +266,24 @@ func (s *Service) validate(in domain.Record, isCreate bool) (domain.Record, erro
 }
 
 // serverSupplies reports whether a column is filled in from the authenticated
-// session rather than from the request body.
+// session rather than from the request body, FOR THIS CALLER.
-func (s *Service) serverSupplies(name string) bool {
+//
+// The caller matters. A TalentOnly derivation records who the row is ABOUT, and
+// the repository fills it for a talent caller only — when an operator files an
+// application or logs evidence on somebody else's behalf, the subject is not
+// the operator, so nothing is derived and the value has to come from the body.
+// Treating those columns as server-supplied for every role was how an operator
+// creating a job application without an email got as far as SQL and came back
+// with a not-null violation instead of the required-field message the contract
+// promises. It must agree with repo.derivedValues, which decides the same thing
+// on the write path.
+func (s *Service) serverSupplies(name string, ident authctx.Identity) bool {
 	if s.res.Policy == nil {
 		return false
 	}
+	isTalent := ident.Role == string(domain.RoleTalent)
 	for _, d := range s.res.Policy.Derived {
-		if d.Column == name {
+		if d.Column == name && (!d.TalentOnly || isTalent) {
 			return true
 		}
 	}

go-api/internal/service/service_test.go

@@ -1,12 +1,20 @@
 package service
 
 import (
+	"errors"
 	"net/url"
 	"testing"
 
+	"github.com/krow/krow-backend/go-api/internal/authctx"
 	"github.com/krow/krow-backend/go-api/internal/domain"
 )
 
+// The two callers validation distinguishes. Only the role is read.
+var (
+	asOperator = authctx.Identity{Role: string(domain.RoleAdmin)}
+	asTalent   = authctx.Identity{Role: string(domain.RoleTalent)}
+)
+
 // These exercise query parsing and validation without a database, so the
 // contract's defaults are pinned even when PostgreSQL is not available.
 
@@ -139,24 +147,24 @@ func TestReservedParametersAreNotFilters(t *testing.T) {
 func TestValidateRequiredAndUnknownAndEnum(t *testing.T) {
 	svc := New(resource(t, "job-postings"), nil)
 
-	if _, err := svc.validate(domain.Record{}, true); err == nil {
+	if _, err := svc.validate(asOperator, domain.Record{}, true); err == nil {
 		t.Error("a create with no title was accepted")
 	}
-	if _, err := svc.validate(domain.Record{"title": "  "}, true); err == nil {
+	if _, err := svc.validate(asOperator, domain.Record{"title": "  "}, true); err == nil {
 		t.Error("a blank title was accepted")
 	}
-	if _, err := svc.validate(domain.Record{"title": "X", "bogus": 1}, true); err == nil {
+	if _, err := svc.validate(asOperator, domain.Record{"title": "X", "bogus": 1}, true); err == nil {
 		t.Error("an unknown field was accepted")
 	}
-	if _, err := svc.validate(domain.Record{"title": "X", "status": "archived"}, true); err == nil {
+	if _, err := svc.validate(asOperator, domain.Record{"title": "X", "status": "archived"}, true); err == nil {
 		t.Error("an invalid enum value was accepted")
 	}
-	if _, err := svc.validate(domain.Record{"title": "X", "status": "active"}, true); err != nil {
+	if _, err := svc.validate(asOperator, domain.Record{"title": "X", "status": "active"}, true); err != nil {
 		t.Errorf("a valid payload was rejected: %v", err)
 	}
 
 	// Server-owned fields are stripped, not rejected.
-	out, err := svc.validate(domain.Record{"title": "X", "id": "abc", "org_id": "def"}, true)
+	out, err := svc.validate(asOperator, domain.Record{"title": "X", "id": "abc", "org_id": "def"}, true)
 	if err != nil {
 		t.Fatalf("server-owned fields caused a rejection: %v", err)
 	}
@@ -168,11 +176,62 @@ func TestValidateRequiredAndUnknownAndEnum(t *testing.T) {
 	}
 
 	// An update needs no required fields — it is a partial by definition.
-	if _, err := svc.validate(domain.Record{"location": "Here"}, false); err != nil {
+	if _, err := svc.validate(asOperator, domain.Record{"location": "Here"}, false); err != nil {
 		t.Errorf("a partial update was rejected: %v", err)
 	}
 }
 
+// A talent-only derivation is only server-supplied for a talent caller.
+//
+// The column records who the row is ABOUT, and the repository fills it from the
+// session for talent and for nobody else (repo.derivedValues). Validation has
+// to agree: an operator filing an application for somebody else must be told
+// `email` is required, rather than being let through to a not-null violation
+// from SQL — and a talent caller must not be asked for the value the server is
+// about to override anyway.
+func TestValidateHonoursTalentOnlyDerivation(t *testing.T) {
+	apps := New(resource(t, "job-applications"), nil)
+	body := domain.Record{
+		"job_posting_id": "00000000-0000-0000-0000-000000000000",
+		"applicant_name": "Someone",
+	}
+
+	if _, err := apps.validate(asTalent, body, true); err != nil {
+		t.Errorf("a talent create without email was rejected: %v", err)
+	}
+
+	_, err := apps.validate(asOperator, body, true)
+	if err == nil {
+		t.Fatal("an operator create without email was accepted")
+	}
+	var apiErr *domain.Error
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("error is not an API error: %v", err)
+	}
+	if apiErr.Details["email"] != "required" {
+		t.Errorf("details = %v, want email: required", apiErr.Details)
+	}
+
+	// evidence.worker_email is the same shape, and the case the original
+	// comment in serverSupplies was written for.
+	ev := New(resource(t, "evidence"), nil)
+	if _, err := ev.validate(asTalent, domain.Record{"type": "photo_identify"}, true); err != nil {
+		t.Errorf("a talent evidence create without worker_email was rejected: %v", err)
+	}
+	if _, err := ev.validate(asOperator, domain.Record{"type": "photo_identify"}, true); err == nil {
+		t.Error("an operator evidence create without worker_email was accepted")
+	}
+
+	// A derivation that is NOT talent-only stays server-supplied for everyone:
+	// user_activity records who acted, whoever that is.
+	act := New(resource(t, "user-activity"), nil)
+	for name, ident := range map[string]authctx.Identity{"operator": asOperator, "talent": asTalent} {
+		if _, err := act.validate(ident, domain.Record{"event_type": "x"}, true); err != nil {
+			t.Errorf("%s: an activity create was rejected: %v", name, err)
+		}
+	}
+}
+
 func TestIsUUID(t *testing.T) {
 	valid := []string{
 		"00000000-0000-0000-0000-000000000000",

go-api/internal/service/workflows.go

@@ -209,7 +209,7 @@ func (s *WorkflowService) Hire(ctx context.Context, ident authctx.Identity,
 		// happened without a log entry, or a log entry for a hire that rolled
 		// back, are both worse than neither.
 		if _, err := repo.New(activity, tx).Insert(ctx, ident, domain.Record{
-			"event_type":     "candidate_hired",
+			"event_type":     "hire_candidate",
 			"details":        fmt.Sprintf("%s was hired", name),
 			"application_id": applicationID,
 			"position_id":    app["job_posting_id"],
@@ -230,6 +230,11 @@ func (s *WorkflowService) Hire(ctx context.Context, ident authctx.Identity,
 /* ── Assign ─────────────────────────────────────────────────────────────── */
 
 // AssignWorker is one worker in an assign request.
+//
+// Three ways of naming the application this placement belongs to, in order of
+// precedence: an id the caller already has, an `application` payload to find or
+// file one from, and neither — a worker placed straight from the talent pool,
+// who has no application and is not given an invented one.
 type AssignWorker struct {
 	WorkerEmail     string  `json:"worker_email"`
 	WorkerName      string  `json:"worker_name"`
@@ -239,6 +244,17 @@ type AssignWorker struct {
 	WorkerProfileID *string `json:"worker_profile_id"`
 	MatchScore      *int    `json:"match_score"`
 	Source          string  `json:"source"`
+
+	// Application is the application to attach this worker to when no
+	// ApplicationID is supplied: the same fields POST /job-applications takes.
+	//
+	// It exists because an application is what puts a person in the pipeline
+	// for a role, and every downstream step keys on it — the candidate record
+	// is addressed by it, and an AI interview takes one as its subject. A
+	// worker assigned without one is unreachable: nothing to open, and nobody
+	// to interview. The frontend was creating it in a second, untransacted
+	// request; this carries it into the same transaction as the assignment.
+	Application *domain.Record `json:"application"`
 }
 
 // AssignRequest is the body of POST /job-postings/{id}/assignments.
@@ -258,6 +274,11 @@ type AssignResult struct {
 // request and one transaction. All-or-nothing across the whole batch: assigning
 // six workers and having the fourth fail should not leave three assigned, three
 // not, and the caller unsure which.
+//
+// Per worker the flow is: settle the application (see settleApplication —
+// patch the one named, find-or-file the one described, or neither), insert the
+// assignment linked to whatever that produced, and write the audit entry. The
+// application comes first because the assignment references it.
 func (s *WorkflowService) Assign(ctx context.Context, ident authctx.Identity,
 	postingID string, req AssignRequest) (*AssignResult, error) {
 
@@ -324,22 +345,43 @@ func (s *WorkflowService) Assign(ctx context.Context, ident authctx.Identity,
 
 		assignRepo := repo.New(assignments, tx)
 		appRepo := repo.New(apps, tx)
+		// The application is filed through the service rather than the
+		// repository so that a payload the caller sent is validated exactly as
+		// POST /job-applications would validate it — unknown fields rejected,
+		// enums checked, required fields demanded — instead of reaching SQL and
+		// coming back as a constraint violation.
+		appSvc := New(apps, tx)
 		activityRepo := repo.New(activity, tx)
 
 		for i, w := range req.Workers {
+			// The application is settled BEFORE the assignment is written,
+			// because the assignment carries the reference to it and a row
+			// cannot point at one that does not exist yet. An empty id means
+			// this worker legitimately has no application.
+			applicationID, err := s.settleApplication(ctx, ident, appRepo, appSvc, postingID, w)
+			if err != nil {
+				return annotate(err, i)
+			}
+
 			record := domain.Record{
 				"job_posting_id": postingID,
 				"worker_email":   w.WorkerEmail,
 				"worker_name":    w.WorkerName,
 				"starts_at":      w.StartsAt,
 				"status":         "active",
-				"source":         orDefault(w.Source, "manual"),
+				// `owliver` is the column's own default (000001:511) and what
+				// the frontend sends. Substituting `manual` here made the API
+				// and the schema disagree about what an unspecified source
+				// means, so a row written through this endpoint and a row
+				// written through POST /assignments recorded different origins
+				// for the same action.
+				"source": orDefault(w.Source, "owliver"),
 			}
 			if w.EndsAt != nil {
 				record["ends_at"] = *w.EndsAt
 			}
-			if w.ApplicationID != nil {
+			if applicationID != "" {
-				record["application_id"] = *w.ApplicationID
+				record["application_id"] = applicationID
 			}
 			if w.WorkerProfileID != nil {
 				record["worker_profile_id"] = *w.WorkerProfileID
@@ -354,29 +396,14 @@ func (s *WorkflowService) Assign(ctx context.Context, ident authctx.Identity,
 			}
 			out.Assignments = append(out.Assignments, created)
 
-			// The application moves to `assigned` only when one was named. A
-			// worker can be placed without having applied — that is what the
-			// talent pool is for — and inventing an application for them would
-			// be worse than leaving the link absent.
-			if w.ApplicationID != nil {
-				updated, err := appRepo.Update(ctx, ident, *w.ApplicationID,
-					domain.Record{"status": "assigned"})
-				if err != nil {
-					return annotate(err, i)
-				}
-				if updated == nil {
-					return domain.NotFound("JobApplication", *w.ApplicationID)
-				}
-			}
-
 			entry := domain.Record{
-				"event_type":   "worker_assigned",
+				"event_type":   "assign_employee",
 				"details":      fmt.Sprintf("%s was assigned", orDefault(w.WorkerName, w.WorkerEmail)),
 				"position_id":  postingID,
 				"worker_email": w.WorkerEmail,
 			}
-			if w.ApplicationID != nil {
+			if applicationID != "" {
-				entry["application_id"] = *w.ApplicationID
+				entry["application_id"] = applicationID
 			}
 			if _, err := activityRepo.Insert(ctx, ident, entry); err != nil {
 				return annotate(err, i)
@@ -391,6 +418,122 @@ func (s *WorkflowService) Assign(ctx context.Context, ident authctx.Identity,
 	return out, nil
 }
 
+// settleApplication resolves the application an assignment belongs to and
+// returns its id, or "" when this worker has none.
+//
+// Three cases, and the middle one is why this exists:
+//
+//   - an id was supplied — move that application to `assigned`. Unchanged
+//     behaviour, and it still wins over any payload, because an id is a
+//     decision the caller has already made.
+//   - an `application` payload was supplied — find the application on this
+//     posting for this email, and move it to `assigned` if there is one or file
+//     it if there is not. (job_posting_id, email) is UNIQUE
+//     (job_applications_posting_email_key, 000001), so there is at most one to
+//     find and the insert cannot produce a second.
+//   - neither — nothing. A worker placed straight from the talent pool has no
+//     application, and inventing one for them would be worse than leaving the
+//     link absent.
+//
+// Everything here runs inside the caller's transaction, which is what makes the
+// find-or-create safe: a concurrent assign of the same person to the same
+// posting blocks on the unique index and is then reported as a conflict, rather
+// than racing past the lookup and writing a duplicate.
+func (s *WorkflowService) settleApplication(ctx context.Context, ident authctx.Identity,
+	appRepo *repo.Repo, appSvc *Service, postingID string, w AssignWorker) (string, error) {
+
+	if w.ApplicationID != nil {
+		updated, err := appRepo.Update(ctx, ident, *w.ApplicationID,
+			domain.Record{"status": "assigned"})
+		if err != nil {
+			return "", err
+		}
+		if updated == nil {
+			return "", domain.NotFound("JobApplication", *w.ApplicationID)
+		}
+		return *w.ApplicationID, nil
+	}
+	if w.Application == nil {
+		return "", nil
+	}
+
+	existing, err := findApplication(ctx, ident, appRepo, postingID, w.WorkerEmail)
+	if err != nil {
+		return "", err
+	}
+	if existing != nil {
+		id, _ := existing["id"].(string)
+		updated, err := appRepo.Update(ctx, ident, id, domain.Record{"status": "assigned"})
+		if err != nil {
+			return "", err
+		}
+		if updated == nil {
+			return "", domain.NotFound("JobApplication", id)
+		}
+		return id, nil
+	}
+
+	// The posting and the person come from the assignment, not from the
+	// payload. A body naming a different posting or a different email would
+	// file an application about somebody other than the worker being placed,
+	// and the lookup above would never find it again.
+	record := domain.Record{}
+	for k, v := range *w.Application {
+		record[k] = v
+	}
+	record["job_posting_id"] = postingID
+	record["email"] = w.WorkerEmail
+	if _, ok := record["applicant_name"]; !ok {
+		record["applicant_name"] = w.WorkerName
+	}
+	if _, ok := record["status"]; !ok {
+		record["status"] = "assigned"
+	}
+
+	created, err := appSvc.Create(ctx, ident, record)
+	if err != nil {
+		return "", err
+	}
+	id, _ := created["id"].(string)
+	return id, nil
+}
+
+// findApplication returns this posting's application for this email, or nil.
+//
+// The read goes through the repository so it carries the same organization
+// scope and ownership predicate every other read does — an application in
+// another tenant is not found, rather than found and then refused. The email
+// column is citext, so the comparison is case-insensitive: the same equality
+// the frontend performed with toLowerCase before it had this endpoint.
+func findApplication(ctx context.Context, ident authctx.Identity, appRepo *repo.Repo,
+	postingID, email string) (domain.Record, error) {
+
+	res := appRepo.Resource()
+	postingCol, ok := res.Column("job_posting_id")
+	if !ok {
+		return nil, domain.Internal(errors.New("service: job_applications has no job_posting_id column"))
+	}
+	emailCol, ok := res.Column("email")
+	if !ok {
+		return nil, domain.Internal(errors.New("service: job_applications has no email column"))
+	}
+
+	page, err := appRepo.List(ctx, ident, domain.ListParams{
+		Limit: 1,
+		Filters: []domain.Filter{
+			{Column: postingCol, Values: []string{postingID}},
+			{Column: emailCol, Values: []string{email}},
+		},
+	})
+	if err != nil {
+		return nil, err
+	}
+	if len(page.Records) == 0 {
+		return nil, nil
+	}
+	return page.Records[0], nil
+}
+
 /* ── Helpers ────────────────────────────────────────────────────────────── */
 
 // pick takes the caller's value for a key when they supplied a usable one, and

infrastructure/README.md

@@ -12,10 +12,13 @@ What lands here in later phases, once each is actually approved:
 | File | Phase | Contents |
 | --- | --- | --- |
 | `docker-compose.dev.yml` | 2 | PostgreSQL + pgvector, so the dev database stops being a machine-local install |
-| `docker-compose.dev.yml` (extended) | later | Redis, NATS, MinIO — each only when the phase that needs it starts |
+| `docker-compose.dev.yml` (extended) | later | Redis, MinIO — each only when the phase that needs it starts |
 | `Dockerfile.api` | later | Multi-stage build for `go-api` |
 | `Dockerfile.owliver` | later | The Python service |
 | `otel-collector.yaml` | later | OpenTelemetry collector config |
 
+NATS is deliberately absent from that table: it is not part of the target
+architecture, and nothing here should reintroduce it.
+
 Adding any of these before its phase would be speculative, so the directory
 holds only this note for now.

scripts/oracle.mjs

@@ -3,8 +3,9 @@
  *
  * Runs the REAL frontend module graph through Vite — `import.meta.glob`, the
  * `@/` alias and raw Markdown loading behave exactly as they do in the app, the
- * same technique `scripts/skill-check.mjs` uses. A mock of the registry would
+ * same technique the frontend's own `scripts/skill-check.mjs` uses (that file
- * reproduce none of the behaviour this file exists to capture.
+ * lives in krow-demo, not here). A mock of the registry would reproduce none of
+ * the behaviour this file exists to capture.
  *
  * Emits one JSON document: for every shipped definition and every adversarial
  * case, what the JS parser did with it. That document is the fixture the Go