Stop deleting live folded-run rows on post-deploy grants/clone failure

packages/folded-runs/src/launch.ts

@@ -641,33 +641,56 @@ export async function deployAtHead(
       : {}),
   });
 
-  await markRunDeployClone(
-    deps.db,
-    params.instanceId,
-    definitionIdBeforeDeploy,
-  );
+  // Everything from here on runs AFTER `deployAdoptedWorkflowFromSource`
+  // resolved, meaning a live child now exists on the sidecar. A failure
+  // in either step below is a leaked deploy, never a "nothing happened" —
+  // wrapped as `SessionLaunchError(..., leakedAgent: true)` so a caller's
+  // failure-path rollback (`launchFoldedRun`) preserves the run's rows
+  // instead of deleting them out from under a genuinely running agent
+  // (CL-7213). The distinction is structural (everything past this
+  // comment), not a classification any caller has to reconstruct from
+  // error type. Each step gets its own phase label — `SessionLaunchError.phase`
+  // is surfaced in cleanup diagnostics upstream (see
+  // `deployAgentUndeploy` failure logging in `session-service.ts`), so a
+  // shared label across both steps would misreport which one actually
+  // failed.
+  try {
+    await markRunDeployClone(
+      deps.db,
+      params.instanceId,
+      definitionIdBeforeDeploy,
+    );
+  } catch (err) {
+    throw new SessionLaunchError("clone", err, true);
+  }
 
-  // Produce the run's `run.grants` frame, the same contract upstream's hub
-  // fires on every run birth: the sidecar writes it to
-  // `runs/<runId>/grants.json` in the deployment's workflow-run repo, and
-  // the supervisor's `onRunStart` barrier reads that file to authorize the
-  // run. A folded run is self-anchored, so its run id IS its deployment id,
-  // and `grants` — the tool-pin and credential-binding set already deployed
-  // as `config.grants` — is the run's whole grant set.
+  // Produce the run's `run.grants` frame, the same contract upstream's
+  // hub fires on every run birth: the sidecar writes it to
+  // `runs/<runId>/grants.json` in the deployment's workflow-run repo,
+  // and the supervisor's `onRunStart` barrier reads that file to
+  // authorize the run. A folded run is self-anchored, so its run id IS
+  // its deployment id, and `grants` — the tool-pin and
+  // credential-binding set already deployed as `config.grants` — is
+  // the run's whole grant set.
   //
-  // Sent after the deploy resolves and before any trigger mail: the sidecar
-  // registers its grants handler during the deploy the ack acknowledges, and
-  // both frames ride the same per-address channel, so same-socket FIFO puts
-  // the grants on disk ahead of the mail that starts the run.
+  // Sent after the deploy resolves and before any trigger mail: the
+  // sidecar registers its grants handler during the deploy the ack
+  // acknowledges, and both frames ride the same per-address channel,
+  // so same-socket FIFO puts the grants on disk ahead of the mail that
+  // starts the run.
   if (
     !deps.sidecarRouter.sendRunGrants(
       params.triggerAddress,
       params.instanceId,
       grants,
     )
   ) {
-    throw new Error(
-      `${params.launchLabel}: deployment ${params.triggerAddress} is not routable for run ${params.instanceId}; cannot deliver its run grants, so the run would start under-authorized`,
+    throw new SessionLaunchError(
+      "grants",
+      new Error(
+        `${params.launchLabel}: deployment ${params.triggerAddress} is not routable for run ${params.instanceId}; cannot deliver its run grants, so the run would start under-authorized`,
+      ),
+      true,
     );
   }
   return { sourcesDigest: inferenceSourcesDigest(resolution) };

packages/folded-runs/test/launch.test.ts

@@ -14,7 +14,12 @@
 // a controllable stub — spreading through every other export unchanged
 // — so a real tenant catalog is never required to prove the wiring.
 import { describe, expect, mock, test } from "bun:test";
-import { agentSession, principal, workflowRun } from "@intx/db/schema";
+import {
+  agentSession,
+  principal,
+  workflowDefinition,
+  workflowRun,
+} from "@intx/db/schema";
 import { foldedRun } from "../src/schema";
 import { SessionLaunchError } from "@intx/hub-sessions";
 import type { EventCollectorRegistry, SidecarRouter } from "@intx/hub-sessions";
@@ -938,6 +943,121 @@ describe("launchFoldedRun", () => {
     expect(runUpdate?.values).toEqual({ status: "failed" });
   });
 
+  // `deployAdoptedWorkflowFromSource` resolving means a live child now
+  // exists on the sidecar; `sendRunGrants` returning `false` (address not
+  // yet routable — a real hub-link drop between the deploy ack and the
+  // grants send) must never read as "nothing was deployed." CL-7213: this
+  // used to throw a plain `Error`, which the catch's `leaked` check missed
+  // entirely, deleting the row for a run that was actually live and now
+  // permanently unauthorized and untracked.
+  test("marks the run failed (not deleted) when sendRunGrants fails after a successful deploy", async () => {
+    resolveDefinitionSourcesResult = {
+      ok: true,
+      sources: [
+        {
+          id: "off_1",
+          provider: "anthropic",
+          baseURL: "https://inference.invalid",
+          apiKey: "placeholder",
+          model: "claude-sonnet-5",
+        },
+      ],
+      defaultSource: "off_1",
+    };
+
+    const db = createFakeDb();
+    const sessionService = createFakeSessionService();
+    const eventCollectors = createFakeEventCollectors();
+
+    await expect(
+      launchFoldedRun(
+        {
+          db: db as never,
+          sessionService,
+          assetService: createFakeAssetService(),
+          sidecarRouter: createFakeSidecarRouter(false),
+          toolGrantsForPins: () => [],
+          eventCollectors,
+        },
+        {
+          tenantId: "ten_1",
+          instanceId: "ins_workbench1",
+          triggerAddress: "ins_workbench1@ten1.workbench.test",
+          definitionId: "wfd_workbench1",
+          foldedBody: FOLDED_BODY,
+          launchLabel: "the workbench host",
+        },
+      ),
+    ).rejects.toThrow(SessionLaunchError);
+
+    expect(db.deleted).toEqual([]);
+    const runUpdate = db.updated.find((row) => row.table === workflowRun);
+    expect(runUpdate?.values).toEqual({ status: "failed" });
+  });
+
+  // Same class of bug as the `sendRunGrants` case above, but for the
+  // other post-deploy step: `markRunDeployClone` runs after the live
+  // deploy resolves too, so a plain `Error` out of it must be treated as
+  // a leaked deploy for the same reason.
+  test("marks the run failed (not deleted) when markRunDeployClone fails after a successful deploy", async () => {
+    resolveDefinitionSourcesResult = {
+      ok: true,
+      sources: [
+        {
+          id: "off_1",
+          provider: "anthropic",
+          baseURL: "https://inference.invalid",
+          apiKey: "placeholder",
+          model: "claude-sonnet-5",
+        },
+      ],
+      defaultSource: "off_1",
+    };
+
+    // A second definition read that differs from the first is what makes
+    // `markRunDeployClone` attempt the `workflowDefinition` update at all
+    // (see `createFakeDb`'s own doc) — that update is the call this test
+    // makes fail.
+    const db = createFakeDb("ast_definition1", [
+      "wfd_definition1",
+      "wfd_definition1_clone",
+    ]);
+    const originalUpdate = db.update.bind(db);
+    db.update = ((table: unknown) => {
+      if (table === workflowDefinition) {
+        throw new Error("clone marking failed");
+      }
+      return originalUpdate(table);
+    }) as typeof db.update;
+    const sessionService = createFakeSessionService();
+    const eventCollectors = createFakeEventCollectors();
+
+    await expect(
+      launchFoldedRun(
+        {
+          db: db as never,
+          sessionService,
+          assetService: createFakeAssetService(),
+          sidecarRouter: createFakeSidecarRouter(),
+          toolGrantsForPins: () => [],
+          eventCollectors,
+        },
+        {
+          tenantId: "ten_1",
+          instanceId: "ins_workbench1",
+          triggerAddress: "ins_workbench1@ten1.workbench.test",
+          definitionId: "wfd_workbench1",
+          foldedBody: FOLDED_BODY,
+          launchLabel: "the workbench host",
+        },
+      ),
+    ).rejects.toThrow(SessionLaunchError);
+
+    expect(db.deleted).toEqual([]);
+    const runUpdate = db.updated.find((row) => row.table === workflowRun);
+    expect(runUpdate?.values).toEqual({ status: "failed" });
+  });
+
   test("throws InferenceResolutionError when the tenant catalog has no launchable source", async () => {
     resolveDefinitionSourcesResult = {
       ok: false,