@stigmer/server 3.23.0 → 3.24.0

package/src/domain/workflow/__tests__/agent-call-references.test.ts

@@ -0,0 +1,373 @@
+/**
+ * Pins the workflow domain's collector for the reference rule
+ * (../agent-call-references.ts): the classifier's four forms (a literal
+ * `org/slug`, a relative bare slug, a value holding `${` that is fixed
+ * only at run, anything else malformed) and the refusal copy; the walk
+ * over top-level, nested and compensate tasks in declaration order, the
+ * relative form filled with the workflow's organization and marked, a
+ * run-time value collected as nothing; a config that does not decode
+ * skipped rather than double-reported; the step over a real store
+ * answering the rule's verdicts — a missing agent with the rule's copy,
+ * another organization's org-visible agent with the one
+ * cross-organization sentence, a platform-visible one admitted, the floor
+ * for a literal, the floor capped at org for a bare slug, a run-time value
+ * saved with no agent behind it; and the workflow's escalation door
+ * applying the same cap when a workflow calling a bare slug is raised to
+ * platform.
+ */
+import { create } from "@bufbuild/protobuf";
+import { Code, ConnectError } from "@connectrpc/connect";
+import { afterEach, beforeEach, describe, expect, it } from "vitest";
+
+import { AgentSchema } from "@stigmer/protos/ai/stigmer/agentic/agent/v1/api_pb";
+import type { Workflow } from "@stigmer/protos/ai/stigmer/agentic/workflow/v1/api_pb";
+import { WorkflowSchema } from "@stigmer/protos/ai/stigmer/agentic/workflow/v1/api_pb";
+import { WorkflowTaskKind } from "@stigmer/protos/ai/stigmer/agentic/workflow/v1/enum_pb";
+import type { WorkflowSpec } from "@stigmer/protos/ai/stigmer/agentic/workflow/v1/spec_pb";
+import { WorkflowSpecSchema } from "@stigmer/protos/ai/stigmer/agentic/workflow/v1/spec_pb";
+import { ApiResourceKind } from "@stigmer/protos/ai/stigmer/commons/apiresource/apiresourcekind/api_resource_kind_pb";
+import { ApiResourceVisibility } from "@stigmer/protos/ai/stigmer/commons/apiresource/enum_pb";
+import { UpdateVisibilityInputSchema } from "@stigmer/protos/ai/stigmer/commons/apiresource/io_pb";
+
+import { RequestContext } from "../../../pipeline/request-context.js";
+import { testCallerIdentity } from "../../../pipeline/__tests__/support.js";
+import {
+  belowFloorMessage,
+  missingReferencesMessage,
+  newGuardReferenceFloorOnEscalationStep,
+  notAvailableReferenceMessage,
+  referenceTargetKind,
+} from "../../../pipeline/steps/references.js";
+import type { SqliteStore } from "../../../store/sqlite/store.js";
+import { tempStore } from "../../../store/sqlite/__tests__/support.js";
+import {
+  classifyAgentReference,
+  collectAgentCallReferences,
+  malformedAgentReferenceMessage,
+  newValidateAgentCallReferencesStep,
+} from "../agent-call-references.js";
+
+const V = ApiResourceVisibility;
+const AGENT = referenceTargetKind(ApiResourceKind.agent)!;
+
+function agentCall(name: string, agent: string) {
+  return {
+    name,
+    kind: WorkflowTaskKind.agent_call,
+    taskConfig: { agent, message: "do the thing" },
+  };
+}
+
+function specOf(tasks: ReadonlyArray<Record<string, unknown>>): WorkflowSpec {
+  return create(WorkflowSpecSchema, { tasks } as never);
+}
+
+/** The three shapes the runner evaluates, whole and embedded, as authors write them. */
+const RUN_TIME_VALUES = [
+  "${.env_vars.TEAM_ORG}/assistant",
+  "${.secrets.AGENT_REF}",
+  '${ "acme/reviewer" }',
+  "${ .input.agent }",
+  "acme/${ .input.slug }",
+  "${ .input.org }/a/b",
+];
+
+describe("classifyAgentReference", () => {
+  it("reads `org/slug` as a literal and a bare slug as relative", () => {
+    expect(classifyAgentReference("acme/reviewer")).toEqual({
+      kind: "literal",
+      org: "acme",
+      slug: "reviewer",
+    });
+    expect(classifyAgentReference("reviewer")).toEqual({
+      kind: "relative",
+      slug: "reviewer",
+    });
+  });
+
+  it("reads any value holding `${` as fixed only at run, whatever its shape", () => {
+    for (const value of RUN_TIME_VALUES) {
+      expect(classifyAgentReference(value), value).toEqual({
+        kind: "run-time",
+      });
+    }
+  });
+
+  it("reads everything else as malformed", () => {
+    for (const malformed of ["", "/reviewer", "acme/", "a/b/c", "//"]) {
+      expect(classifyAgentReference(malformed), malformed).toEqual({
+        kind: "malformed",
+      });
+    }
+  });
+});
+
+describe("collectAgentCallReferences", () => {
+  it("collects top-level, nested and compensate agent_call tasks in declaration order, a bare slug filled with the workflow's organization and marked relative", () => {
+    const spec = specOf([
+      agentCall("first", "reviewer"),
+      {
+        name: "loop",
+        kind: WorkflowTaskKind.for_each,
+        taskConfig: {
+          each: "item",
+          in: "${ .items }",
+          // A nested task inside the Struct is JSON: proto field names.
+          do: [
+            {
+              name: "inner",
+              kind: "agent_call",
+              task_config: { agent: "globex/summarizer", message: "sum" },
+            },
+          ],
+        },
+        compensate: [agentCall("undo", "cleaner")],
+      },
+      {
+        name: "pause",
+        kind: WorkflowTaskKind.wait,
+        taskConfig: { duration: {} },
+      },
+    ]);
+    expect(collectAgentCallReferences(spec, "acme")).toEqual([
+      {
+        kind: ApiResourceKind.agent,
+        org: "acme",
+        slug: "reviewer",
+        resolvesIn: "running-organization",
+      },
+      { kind: ApiResourceKind.agent, org: "globex", slug: "summarizer" },
+      {
+        kind: ApiResourceKind.agent,
+        org: "acme",
+        slug: "cleaner",
+        resolvesIn: "running-organization",
+      },
+    ]);
+  });
+
+  it("collects nothing for a value fixed only at run, beside the references it does collect", () => {
+    const spec = specOf([
+      ...RUN_TIME_VALUES.map((value, i) => agentCall(`dynamic_${i}`, value)),
+      agentCall("fixed", "globex/summarizer"),
+    ]);
+    expect(collectAgentCallReferences(spec, "acme")).toEqual([
+      { kind: ApiResourceKind.agent, org: "globex", slug: "summarizer" },
+    ]);
+  });
+
+  it("refuses a malformed agent string naming the task", () => {
+    const spec = specOf([agentCall("broken", "a/b/c")]);
+    expect(() => collectAgentCallReferences(spec, "acme")).toThrow(
+      malformedAgentReferenceMessage("broken", "a/b/c"),
+    );
+    let error: unknown;
+    try {
+      collectAgentCallReferences(spec, "acme");
+    } catch (e) {
+      error = e;
+    }
+    expect((error as ConnectError).code).toBe(Code.InvalidArgument);
+  });
+
+  it("skips a config that does not decode — the spec validation step has already refused it", () => {
+    const spec = specOf([
+      {
+        name: "odd",
+        kind: WorkflowTaskKind.agent_call,
+        taskConfig: { agent: "reviewer", not_a_field: true },
+      },
+    ]);
+    expect(collectAgentCallReferences(spec, "acme")).toEqual([]);
+    expect(collectAgentCallReferences(undefined, "acme")).toEqual([]);
+  });
+});
+
+describe("ValidateAgentCallReferences over a store", () => {
+  let store: SqliteStore;
+  let cleanup: () => Promise<void>;
+
+  beforeEach(() => {
+    const temp = tempStore();
+    store = temp.store;
+    cleanup = temp.cleanup;
+  });
+
+  afterEach(async () => {
+    await cleanup();
+  });
+
+  async function seedAgent(
+    id: string,
+    org: string,
+    slug: string,
+    visibility: ApiResourceVisibility,
+  ): Promise<void> {
+    await store.saveResource(
+      ApiResourceKind.agent,
+      id,
+      AgentSchema,
+      create(AgentSchema, {
+        apiVersion: "agentic.stigmer.ai/v1",
+        kind: "Agent",
+        metadata: { id, name: slug, slug, org, visibility },
+        spec: { instructions: "a conformant instruction body" },
+      }),
+    );
+  }
+
+  function workflowCalling(
+    visibility: ApiResourceVisibility,
+    ...agents: string[]
+  ): Workflow {
+    return create(WorkflowSchema, {
+      apiVersion: "agentic.stigmer.ai/v1",
+      kind: "Workflow",
+      metadata: { id: "wfl_1", name: "Pipeline", org: "acme", visibility },
+      spec: specOf(agents.map((agent, i) => agentCall(`call_${i}`, agent))),
+    });
+  }
+
+  async function run(workflow: Workflow): Promise<unknown> {
+    const ctx = new RequestContext(
+      WorkflowSchema,
+      workflow,
+      testCallerIdentity(),
+      ApiResourceKind.workflow,
+    );
+    try {
+      await newValidateAgentCallReferencesStep(store).execute(ctx);
+      return undefined;
+    } catch (error) {
+      return error;
+    }
+  }
+
+  it("a missing agent is refused with the rule's copy; an existing one passes", async () => {
+    const error = await run(workflowCalling(V.visibility_org, "reviewer"));
+    expect(error).toBeInstanceOf(ConnectError);
+    expect((error as ConnectError).code).toBe(Code.FailedPrecondition);
+    expect((error as ConnectError).rawMessage).toBe(
+      missingReferencesMessage(AGENT, [{ slug: "reviewer", org: "acme" }]),
+    );
+    await seedAgent("agt_1", "acme", "reviewer", V.visibility_org);
+    expect(
+      await run(workflowCalling(V.visibility_org, "reviewer")),
+    ).toBeUndefined();
+  });
+
+  it("another organization's agent is admitted only at platform visibility, with the one sentence otherwise", async () => {
+    await seedAgent("agt_shared", "globex", "shared", V.visibility_platform);
+    await seedAgent("agt_internal", "globex", "internal", V.visibility_org);
+    expect(
+      await run(workflowCalling(V.visibility_org, "globex/shared")),
+    ).toBeUndefined();
+    const error = await run(
+      workflowCalling(V.visibility_org, "globex/internal"),
+    );
+    expect((error as ConnectError).rawMessage).toBe(
+      notAvailableReferenceMessage(AGENT, {
+        kind: ApiResourceKind.agent,
+        org: "globex",
+        slug: "internal",
+      }),
+    );
+  });
+
+  it("the floor holds for the agent a workflow calls: an org-visible workflow may not call a private agent", async () => {
+    await seedAgent("agt_mine", "acme", "mine", V.visibility_private);
+    const error = await run(workflowCalling(V.visibility_org, "mine"));
+    expect((error as ConnectError).code).toBe(Code.FailedPrecondition);
+    expect((error as ConnectError).rawMessage).toContain(
+      "referenced agent 'acme/mine' is visibility_private while this resource is visibility_org",
+    );
+  });
+
+  it("a platform-visible workflow may call its organization's org-visible agent by bare slug, never by literal, and never a private one", async () => {
+    await seedAgent("agt_team", "acme", "team", V.visibility_org);
+    await seedAgent("agt_mine", "acme", "mine", V.visibility_private);
+    expect(
+      await run(workflowCalling(V.visibility_platform, "team")),
+    ).toBeUndefined();
+    // The literal names a fixed row that every organization's runs read.
+    const literal = await run(
+      workflowCalling(V.visibility_platform, "acme/team"),
+    );
+    expect((literal as ConnectError).rawMessage).toContain(
+      "referenced agent 'acme/team' is visibility_org while this resource is visibility_platform",
+    );
+    const below = await run(workflowCalling(V.visibility_platform, "mine"));
+    expect((below as ConnectError).rawMessage).toContain(
+      "referenced agent 'acme/mine' is visibility_private while this resource is visibility_platform",
+    );
+  });
+
+  it("a value fixed only at run saves with no agent behind it, at every level", async () => {
+    for (const visibility of [
+      V.visibility_private,
+      V.visibility_org,
+      V.visibility_platform,
+    ]) {
+      expect(
+        await run(workflowCalling(visibility, ...RUN_TIME_VALUES)),
+        V[visibility],
+      ).toBeUndefined();
+    }
+  });
+
+  describe("the workflow's escalation door", () => {
+    const TARGET = "loadedWorkflow";
+
+    async function escalate(
+      stored: Workflow,
+      requested: ApiResourceVisibility,
+    ): Promise<unknown> {
+      const ctx = new RequestContext(
+        UpdateVisibilityInputSchema,
+        create(UpdateVisibilityInputSchema, {
+          resourceId: "wfl_1",
+          visibility: requested,
+        }),
+        testCallerIdentity(),
+        ApiResourceKind.workflow,
+      );
+      ctx.set(TARGET, stored);
+      try {
+        await newGuardReferenceFloorOnEscalationStep(store, TARGET, [
+          (row) =>
+            collectAgentCallReferences(
+              (row as Workflow).spec,
+              (row as Workflow).metadata?.org ?? "",
+            ),
+        ]).execute(ctx);
+        return undefined;
+      } catch (error) {
+        return error;
+      }
+    }
+
+    it("raising a workflow that calls a bare slug to platform asks the capped floor", async () => {
+      await seedAgent("agt_team", "acme", "team", V.visibility_org);
+      await seedAgent("agt_mine", "acme", "mine", V.visibility_private);
+      expect(
+        await escalate(
+          workflowCalling(V.visibility_org, "team"),
+          V.visibility_platform,
+        ),
+      ).toBeUndefined();
+      const error = await escalate(
+        workflowCalling(V.visibility_private, "mine"),
+        V.visibility_platform,
+      );
+      expect((error as ConnectError).code).toBe(Code.FailedPrecondition);
+      expect((error as ConnectError).rawMessage).toBe(
+        belowFloorMessage(
+          AGENT,
+          { kind: ApiResourceKind.agent, org: "acme", slug: "mine" },
+          V.visibility_private,
+          V.visibility_platform,
+        ),
+      );
+    });
+  });
+});

package/src/domain/workflow/agent-call-references.ts

@@ -0,0 +1,204 @@
+/**
+ * ValidateAgentCallReferences — the second collector for the reference rule
+ * (pipeline/steps/references.ts): a workflow's `agent_call` tasks name their
+ * agent as a STRING inside a `google.protobuf.Struct` config the generic
+ * spec walker cannot see, so the rule would never meet them. This step
+ * parses each task's config through the converter's own strict decoder
+ * (`unmarshalTaskConfig`, the one reader of a task config), walks nested
+ * control-flow tasks exactly as the constraint validator does, turns every
+ * `agent` string the write can judge into a reference with the workflow as
+ * parent, and asks the rule's one function — same clauses, same copy. It
+ * lives in the workflow domain and not inside `InProcessValidator` because
+ * the validator is pure over the spec by design and a rule that reads the
+ * target's row cannot be.
+ *
+ * The string has three forms (the `agent` field's contract in
+ * agent_call.proto), told apart by `classifyAgentReference`:
+ *
+ *   - `org/slug`, a LITERAL: judged now, like any stored reference;
+ *   - a bare `slug`, RELATIVE: the runner resolves it in the organization
+ *     the workflow runs in (`__stigmer_org_id`, the execution's), which is
+ *     not the workflow's when a person runs it from another organization by
+ *     id. It is judged now in the workflow's own organization with its
+ *     floor capped at org (references.ts, "RELATIVE");
+ *   - a value holding `${`, fixed only AT RUN: the runner's jq phase
+ *     (`${ expr }`, whole or embedded) or its activity phase
+ *     (`${.env_vars.KEY}`, `${.secrets.KEY}`) produces the reference when
+ *     the task runs. The write collects nothing for it; the run reads the
+ *     resolved agent as the person who ran the workflow, as it reads every
+ *     id-bound binding. The slug grammar (metadata.proto) can never contain
+ *     `${`, so the test is exact. It deliberately does not recognise only
+ *     the runner's shapes: that would be a second copy of the runner's
+ *     grammar in a package that shares only the protos with it, and a
+ *     mistyped expression already fails at run with the read's not-found.
+ *
+ * Anything else is refused at write naming the task. The runner's own
+ * parser (`parseAgentReference` in runner/src/activities/call-agent.ts) is
+ * lenient — it splits on the first `/` and drops a third segment — and the
+ * CLI's `parseReference` (client-apps/cli/src/resources/reference.ts)
+ * refuses a leading slash; refusing every malformed literal here means
+ * neither ever meets one. The three cannot share code; the conformance
+ * suites are where their agreement is proven.
+ *
+ * Runs after NormalizeReferences and ValidateReferences on the workflow
+ * create and update chains, and feeds the workflow's escalation door
+ * (controller.ts). Proven by __tests__/agent-call-references.test.ts, the
+ * workflow conformance suite's agent_call reference arm, and the
+ * runner-as-subject execution arm for a reference fixed only at run.
+ */
+import type { WorkflowSchema } from "@stigmer/protos/ai/stigmer/agentic/workflow/v1/api_pb";
+import type { AgentCallTaskConfig } from "@stigmer/protos/ai/stigmer/agentic/workflow/v1/tasks/agent_call_pb";
+import type {
+  WorkflowSpec,
+  WorkflowTask,
+} from "@stigmer/protos/ai/stigmer/agentic/workflow/v1/spec_pb";
+import { WorkflowTaskKind } from "@stigmer/protos/ai/stigmer/agentic/workflow/v1/enum_pb";
+import { ApiResourceKind } from "@stigmer/protos/ai/stigmer/commons/apiresource/apiresourcekind/api_resource_kind_pb";
+
+import type { Store } from "../../store/interface.js";
+import { internalError, invalidArgumentError } from "../../pipeline/errors.js";
+import type { PipelineStep } from "../../pipeline/pipeline.js";
+import type { RequestContext } from "../../pipeline/request-context.js";
+import type {
+  ReferenceParent,
+  SpecReference,
+} from "../../pipeline/steps/references.js";
+import { checkReferences } from "../../pipeline/steps/references.js";
+import { metadataOf } from "../../pipeline/steps/shapes.js";
+import { unmarshalTaskConfig } from "./converter/unmarshal.js";
+import { nestedTasks } from "./validation/task-config-constraints.js";
+
+/** The one sentence for an `agent` string that is none of the three forms. */
+export function malformedAgentReferenceMessage(
+  taskName: string,
+  agent: string,
+): string {
+  return `task '${taskName}': agent '${agent}' is not a reference; write 'slug' for an agent of this organization or 'org/slug' for another organization's.`;
+}
+
+/** What an `agent` string is before the rule reads it (the module header). */
+export type AgentReferenceForm =
+  | { readonly kind: "literal"; readonly org: string; readonly slug: string }
+  | { readonly kind: "relative"; readonly slug: string }
+  | { readonly kind: "run-time" }
+  | { readonly kind: "malformed" };
+
+/** Opens every runtime expression and placeholder the runner evaluates; never part of a slug. */
+const RUNTIME_EXPRESSION_MARKER = "${";
+
+export function classifyAgentReference(agent: string): AgentReferenceForm {
+  if (agent.includes(RUNTIME_EXPRESSION_MARKER)) {
+    return { kind: "run-time" };
+  }
+  const parts = agent.split("/");
+  if (parts.length === 1 && parts[0] !== "") {
+    return { kind: "relative", slug: parts[0]! };
+  }
+  if (parts.length === 2 && parts[0] !== "" && parts[1] !== "") {
+    return { kind: "literal", org: parts[0]!, slug: parts[1]! };
+  }
+  return { kind: "malformed" };
+}
+
+/**
+ * Every `agent_call` task's agent the write can judge, as a reference,
+ * nested tasks included in declaration order: a literal as written, a
+ * relative slug in `parentOrg` and marked relative, a value fixed only at
+ * run not at all. Throws INVALID_ARGUMENT on the first malformed string. A
+ * config that does not decode is skipped: the spec validation step has
+ * already refused the workflow for it, and this collector never
+ * double-reports.
+ */
+export function collectAgentCallReferences(
+  spec: WorkflowSpec | undefined,
+  parentOrg: string,
+): SpecReference[] {
+  const refs: SpecReference[] = [];
+  if (spec === undefined) {
+    return refs;
+  }
+  const visit = (task: WorkflowTask): void => {
+    if (task.taskConfig !== undefined) {
+      let config;
+      try {
+        config = unmarshalTaskConfig(task.kind, task.taskConfig);
+      } catch {
+        config = undefined;
+      }
+      if (config !== undefined) {
+        if (task.kind === WorkflowTaskKind.agent_call) {
+          const agent = (config as AgentCallTaskConfig).agent;
+          const form = classifyAgentReference(agent);
+          switch (form.kind) {
+            case "literal":
+              refs.push({
+                kind: ApiResourceKind.agent,
+                org: form.org,
+                slug: form.slug,
+              });
+              break;
+            case "relative":
+              refs.push({
+                kind: ApiResourceKind.agent,
+                org: parentOrg,
+                slug: form.slug,
+                resolvesIn: "running-organization",
+              });
+              break;
+            case "run-time":
+              break;
+            case "malformed":
+              throw invalidArgumentError(
+                malformedAgentReferenceMessage(task.name, agent),
+              );
+            default: {
+              const exhaustive: never = form;
+              throw new Error(
+                `unknown agent reference form: ${JSON.stringify(exhaustive)}`,
+              );
+            }
+          }
+        }
+        for (const nested of nestedTasks(config)) {
+          visit(nested);
+        }
+      }
+    }
+    for (const compensate of task.compensate) {
+      visit(compensate);
+    }
+  };
+  for (const task of spec.tasks) {
+    visit(task);
+  }
+  return refs;
+}
+
+export function newValidateAgentCallReferencesStep(
+  store: Store,
+): PipelineStep<typeof WorkflowSchema> {
+  return {
+    name: "ValidateAgentCallReferences",
+    async execute(ctx: RequestContext<typeof WorkflowSchema>): Promise<void> {
+      const metadata = metadataOf(ctx.newState);
+      if (metadata === undefined) {
+        throw internalError(
+          new Error("resource metadata is nil"),
+          "validate agent_call references",
+        );
+      }
+      const refs = collectAgentCallReferences(ctx.newState.spec, metadata.org);
+      if (refs.length === 0) {
+        return;
+      }
+      const parent: ReferenceParent = {
+        org: metadata.org,
+        visibility: metadata.visibility,
+      };
+      const refusal = await checkReferences(store, parent, refs);
+      if (refusal !== undefined) {
+        throw refusal;
+      }
+    },
+  };
+}

package/src/domain/workflow/controller.ts

@@ -78,10 +78,6 @@ import {
 } from "../../pipeline/steps/authorize-resolved-target.js";
 import { versionHistoryTarget } from "../../pipeline/steps/version-history.js";
 import { newGuardReservedLabelsStep } from "../../pipeline/steps/guard-reserved-labels.js";
-import {
-  newAuthorizeVisibilityTransitionStep,
-  newGuardPublicVisibilityStep,
-} from "../../pipeline/steps/visibility-gates.js";
 import { newBuildUpdateStateStep } from "../../pipeline/steps/build-update-state.js";
 import {
   setAuditFieldsForUpdate,
@@ -112,7 +108,16 @@ import {
   TARGET_RESOURCE_KEY,
   newLoadTargetStep,
 } from "../../pipeline/steps/load-target.js";
-import { newNormalizeReferencesStep } from "../../pipeline/steps/references.js";
+import {
+  collectSpecReferences,
+  newGuardReferenceFloorOnEscalationStep,
+  newNormalizeReferencesStep,
+  newValidateReferencesStep,
+} from "../../pipeline/steps/references.js";
+import {
+  collectAgentCallReferences,
+  newValidateAgentCallReferencesStep,
+} from "./agent-call-references.js";
 import {
   newCleanupIamPoliciesStep,
   newCreateAuthorizationTuplesStep,
@@ -238,9 +243,10 @@ async function createWorkflow(
     .addStep(newResolveSlugStep())
     .addStep(newCheckDuplicateStep(deps.store))
     .addStep(newBuildNewStateStep())
-    .addStep(newGuardPublicVisibilityStep(deps.authorizer))
     .addStep(newGuardReservedLabelsStep(deps.authorizer))
     .addStep(newNormalizeReferencesStep())
+    .addStep(newValidateReferencesStep(deps.store))
+    .addStep(newValidateAgentCallReferencesStep(deps.store))
     .addStep(newPopulateServerlessValidationStep(deps.logger))
     .addStep(newComputeVersionHashStep(deps.logger))
     .addStep(newPopulateVersionHashStep(true))
@@ -305,6 +311,8 @@ async function update(
     .addStep(newBuildUpdateStateStep())
     .addStep(newGuardReservedLabelsStep(deps.authorizer))
     .addStep(newNormalizeReferencesStep())
+    .addStep(newValidateReferencesStep(deps.store))
+    .addStep(newValidateAgentCallReferencesStep(deps.store))
     .addStep(newPopulateServerlessValidationStepForUpdate(deps.logger))
     .addStep(newComputeVersionHashStep(deps.logger))
     .addStep(newCheckVersionChangedStep(deps.logger))
@@ -411,8 +419,8 @@ async function deleteWorkflow(
 // updateVisibility — update_visibility.go: a targeted metadata update (only
 // metadata.visibility changes; spec/status untouched). The level check runs
 // AFTER load, preserving the cross-edition error precedence: unknown id +
-// bad level = NOT_FOUND on both editions. CW-9's suite block is the wire
-// pin (clientPublicVisibilityWrites, true on local targets).
+// bad level = NOT_FOUND on both editions. The workflow conformance suite's
+// visibility block is the wire pin.
 // ---------------------------------------------------------------------------
 
 const UPDATE_VISIBILITY_WORKFLOW_KEY = "updateVisibilityWorkflow";
@@ -452,10 +460,20 @@ async function updateVisibility(
       newRecordVisibilityBeforeUpdateStep(UPDATE_VISIBILITY_WORKFLOW_KEY),
     )
     .addStep(newValidateVisibilityUpdateStep())
+    // The reference floor's second door: a workflow may not be raised above
+    // the agents its agent_call tasks run.
     .addStep(
-      newAuthorizeVisibilityTransitionStep(
+      newGuardReferenceFloorOnEscalationStep(
+        deps.store,
         UPDATE_VISIBILITY_WORKFLOW_KEY,
-        deps.authorizer,
+        [
+          (row) => collectSpecReferences(WorkflowSchema, row),
+          (row) =>
+            collectAgentCallReferences(
+              (row as Workflow).spec,
+              (row as Workflow).metadata?.org ?? "",
+            ),
+        ],
       ),
     )
     .addStep(newSetWorkflowVisibilityStep())