@intx/hub-sessions 0.2.2 → 0.4.0

package/dist/hub-session-lookups.js

@@ -5,92 +5,80 @@
 // Each lookup is a stateless DB or repo call. They are gathered into a
 // single struct that the hub app passes to `createSidecarRouter` as
 // `lookups`.
-import { eq, and, isNull } from "drizzle-orm";
-import { agentInstance, sessionMail, workflowDeployment, } from "@intx/db/schema";
+import { eq, and, asc, inArray, isNotNull, isNull } from "drizzle-orm";
+import { createApprovalStore, createSignalCorrelationStore, createWorkflowRunDispatchStore, createWorkflowRunStore, } from "@intx/db";
+import { agentSession, liveWorkflowRunStatuses, principal, sessionMail, sidecarAllocation, workflowRun, } from "@intx/db/schema";
 import { getLogger } from "@intx/log";
-import { parseAgentAddress } from "@intx/types";
-import { isWorkflowDerivedAddress } from "@intx/workflow-deploy";
+import { parseRunAddress, signalName } from "@intx/types";
+import { SignalDeliverFrame } from "@intx/types/sidecar";
+import { deriveWorkflowRunRepoId } from "@intx/workflow-deploy";
 import { generateId } from "@intx/hub-common";
+import { listAcceptedWorkflowDispatches, listConsumedWorkflowDispatches, } from "./workflow-dispatch-settlement.js";
+import { readCommittedWorkflowRunLifecycle } from "./workflow-run-kind.js";
 const logger = getLogger(["hub", "lookups"]);
 export function createHubSessionLookups(deps) {
     const { db, agentRepoStore } = deps;
+    const signalCorrelationStore = createSignalCorrelationStore(db);
+    const approvalStore = createApprovalStore(db);
+    const workflowRunStore = createWorkflowRunStore(db);
+    const workflowRunDispatchStore = createWorkflowRunDispatchStore(db);
     return {
-        async lookupPublicKey(agentAddress) {
-            // Route by address space, not a blind two-table fallback: a
-            // workflow-derived address's key lives on its workflow_deployment
-            // row, a launched agent's on its agent_instance row, and the two
-            // spaces are disjoint. Routing (rather than falling back) means a
-            // launched agent that is missing its instance row returns null and
-            // fails its challenge visibly, instead of silently resolving against
-            // the wrong table.
-            if (isWorkflowDerivedAddress(agentAddress)) {
-                // Filter to a live ("deployed") deployment so a torn-down
-                // deployment's key can no longer satisfy a challenge. A null
-                // publicKey (deployed but not yet acked, or pre-migration) or an
-                // absent row returns null -- the challenge fails closed and the
-                // address stays unrouted rather than routing without ownership
-                // proof.
-                const row = await db
-                    .select({ publicKey: workflowDeployment.publicKey })
-                    .from(workflowDeployment)
-                    .where(and(eq(workflowDeployment.address, agentAddress), eq(workflowDeployment.status, "deployed")))
-                    .limit(1)
-                    .then((rows) => rows[0]);
-                return row?.publicKey ?? null;
-            }
-            const row = await db
-                .select({ publicKey: agentInstance.publicKey })
-                .from(agentInstance)
-                .where(and(eq(agentInstance.address, agentAddress), isNull(agentInstance.endedAt)))
-                .limit(1)
-                .then((rows) => rows[0]);
-            return row?.publicKey ?? null;
-        },
-        async lookupDeployRef(agentAddress) {
-            const agentId = parseAgentId(agentAddress);
-            return agentRepoStore.getDeployRef(agentId);
+        async lookupDeployRef() {
+            // A workflow run is a supervised workflow-process pinned forever like a
+            // native deployment: it keeps its deploy-time definition and never
+            // reconciles, so no address enrolls in the reconnect deploy-ref catch-up.
+            return null;
         },
         async persistMail({ senderAddress, recipients, raw }) {
-            const senderInstance = await requireInstance(db, senderAddress);
-            if (!senderInstance.sessionId) {
-                throw new Error(`Instance ${senderInstance.id} has no session for address "${senderAddress}"`);
+            // The sender and recipients are run addresses, each backed by its
+            // self-anchored workflow_run; resolve each through the resolver. A mail
+            // record's `runId` is always null for a run -- it keys on the run's
+            // session instead.
+            const sender = await resolveRoutableAddress(db, senderAddress);
+            if (sender === undefined) {
+                throw new Error(`No active endpoint found for sender address "${senderAddress}"`);
+            }
+            if (sender.sessionId === null) {
+                throw new Error(`Endpoint ${sender.id} has no session for address "${senderAddress}"`);
             }
             const createdAt = new Date();
+            const senderRunId = null;
             // Outbound record on the sender's session.
             const outboundId = generateId("sessionMail");
             const outboundRecord = {
                 id: outboundId,
-                sessionId: senderInstance.sessionId,
-                instanceId: senderInstance.id,
-                tenantId: senderInstance.tenantId,
+                sessionId: sender.sessionId,
+                runId: senderRunId,
+                tenantId: sender.tenantId,
                 direction: "outbound",
                 status: "delivered",
                 raw,
                 createdAt,
             };
-            // Inbound records for each recipient that has an active agent
-            // instance. Recipients that are not agent instances (e.g. human
-            // user addresses) are skipped.
+            // Inbound records for each recipient that is a live endpoint.
+            // Recipients that are not (e.g. human user addresses) are skipped.
             const recipientResults = await Promise.all(recipients.map(async (addr) => {
-                const row = await findInstance(db, addr);
-                if (row === undefined) {
+                const endpoint = await resolveRoutableAddress(db, addr);
+                if (endpoint === undefined) {
                     return null;
                 }
-                if (row.sessionId === null) {
-                    logger.warn `Active instance ${row.id} for "${addr}" has no session; skipping inbound record`;
+                if (endpoint.sessionId === null) {
+                    logger.warn `Active endpoint ${endpoint.id} for "${addr}" has no session; skipping inbound record`;
                     return null;
                 }
-                return { addr, instance: row, sessionId: row.sessionId };
+                return { addr, endpoint, sessionId: endpoint.sessionId };
             }));
-            const recipientInstances = recipientResults.filter((r) => r !== null);
-            const inboundEntries = recipientInstances.map(({ addr, instance, sessionId }) => {
+            const recipientEndpoints = recipientResults.filter((r) => r !== null);
+            const inboundEntries = recipientEndpoints.map(({ addr, endpoint, sessionId }) => {
                 const id = generateId("sessionMail");
+                // A folded run is not an instance, so its mail records no runId.
+                const runId = null;
                 return {
                     record: {
                         id,
                         sessionId,
-                        instanceId: instance.id,
-                        tenantId: instance.tenantId,
+                        runId,
+                        tenantId: endpoint.tenantId,
                         direction: "inbound",
                         status: "delivered",
                         raw,
@@ -99,7 +87,7 @@ export function createHubSessionLookups(deps) {
                     result: {
                         id,
                         direction: "inbound",
-                        instanceId: instance.id,
+                        runId,
                         address: addr,
                         createdAt,
                     },
@@ -112,13 +100,111 @@ export function createHubSessionLookups(deps) {
                 {
                     id: outboundId,
                     direction: "outbound",
-                    instanceId: senderInstance.id,
-                    address: senderInstance.address,
+                    runId: senderRunId,
+                    address: sender.address,
                     createdAt,
                 },
                 ...inboundEntries.map((e) => e.result),
             ];
         },
+        async registerSignalCorrelation({ correlationId, runId, anchorRunId, agentAddress, kind, approvalSnapshot, }) {
+            // Resolve tenancy and co-write both rows in one transaction so a resolver
+            // never sees a correlation without its approval or vice versa. Both
+            // inserts are idempotent on their dedup key (the signal_correlation
+            // primary key and the approval's unique correlationId), so a redelivered
+            // frame -- sidecar reconnect, workflow-log replay, supervisor restart
+            // re-emitting -- is a no-op rather than a unique-violation. `timeoutAt` is
+            // null: an agent-step suspend holds indefinitely (`parkOnSignal` is called
+            // with no timeout), so no deadline reaches this co-write.
+            await db.transaction(async (tx) => {
+                // Resolve tenancy and the run's definition from the deployment's anchor
+                // run -- the workflow_run whose id is the deployment id, which the
+                // address names. The anchor is the tenancy origin every approval needs
+                // (an approval has no agent_instance/agent/principal referent). The
+                // lookup keys off `address` (the field the wire layer's ownership gate
+                // authorized), not the frame's `anchorRunId`: that is the workflow-run
+                // repo slug the supervisor derives from the address
+                // (`deriveWorkflowRunRepoId`), cross-checked below against the slug
+                // re-derived from `agentAddress` rather than against the row id. A
+                // mismatch fails loud instead of silently writing an inconsistent pair.
+                // The FK columns take the anchor run's id (= the deployment id), which
+                // is what `signal_correlation.anchor_run_id` and `approval.anchor_run_id`
+                // reference.
+                //
+                // The resolution takes a `FOR UPDATE` row lock and runs inside the
+                // co-write transaction, gated on a live anchor run ("deployed" or
+                // "running"), so the liveness check and the inserts are atomic against a
+                // concurrent teardown that flips the anchor run terminal. The lock order
+                // is workflow_run before signal_correlation and approval; a teardown
+                // path must take the anchor-run lock before touching those rows to keep
+                // the ordering acyclic.
+                const anchor = await tx
+                    .select({
+                    id: workflowRun.id,
+                    tenantId: workflowRun.tenantId,
+                    definitionId: workflowRun.definitionId,
+                })
+                    .from(workflowRun)
+                    .where(and(eq(workflowRun.address, agentAddress), inArray(workflowRun.status, [...liveWorkflowRunStatuses])))
+                    .for("update")
+                    .limit(1)
+                    .then((rows) => rows[0]);
+                if (anchor === undefined) {
+                    throw new Error(`No live workflow run for address "${agentAddress}"; cannot register signal correlation ${correlationId}`);
+                }
+                const addressSlug = deriveWorkflowRunRepoId(agentAddress);
+                if (addressSlug !== anchorRunId) {
+                    throw new Error(`Anchor run id mismatch registering signal correlation ${correlationId}: frame claims "${anchorRunId}" but address "${agentAddress}" derives the workflow-run repo slug "${addressSlug}"`);
+                }
+                const tenantId = anchor.tenantId;
+                const definitionId = anchor.definitionId;
+                // Lazily anchor the run before its correlation and approval reference
+                // it. A workflow-spawned internal run never crosses the external
+                // trigger route that mints a run principal, so its run row would
+                // otherwise not exist; ensure it here so the co-written rows have a
+                // referent. The principal is null: an internal run inherits its
+                // deployment's grants and has no principal of its own. The insert is
+                // idempotent on the run id, so a redelivered register frame -- the same
+                // redelivery the co-writes below tolerate -- is a no-op.
+                await workflowRunStore.createIfAbsent({
+                    id: runId,
+                    anchorRunId: anchor.id,
+                    definitionId,
+                    tenantId,
+                    principalId: null,
+                    status: "running",
+                }, tx);
+                await signalCorrelationStore.registerIfAbsent({
+                    correlationId,
+                    tenantId,
+                    anchorRunId: anchor.id,
+                    agentAddress,
+                    runId,
+                    signalName: signalName(correlationId),
+                    kind,
+                }, tx);
+                await approvalStore.createIfAbsent({
+                    id: generateId("approval"),
+                    tenantId,
+                    anchorRunId: anchor.id,
+                    runId,
+                    agentAddress,
+                    correlationId,
+                    status: "pending",
+                    // The register frame guarantees the snapshot (the ask rail is its
+                    // only producer), so the approver-facing columns are always
+                    // populated -- never null on this path.
+                    toolDefinition: {
+                        name: approvalSnapshot.name,
+                        description: approvalSnapshot.description,
+                        inputSchema: approvalSnapshot.inputSchema,
+                    },
+                    toolArguments: approvalSnapshot.arguments,
+                    scope: null,
+                    timeoutAt: null,
+                }, tx);
+            });
+        },
         async receiveAgentStatePack(repoId, pack, ref, commitSha) {
             if (repoId.kind !== "agent-state") {
                 throw new Error(`hub-session lookups receiveAgentStatePack received unsupported repo kind ${JSON.stringify(repoId.kind)}`);
@@ -150,55 +236,424 @@ export function createHubSessionLookups(deps) {
             }
             return { accepted: true };
         },
-        async receiveWorkflowRunPack(repoId, pack, ref, commitSha) {
+        async receiveWorkflowRunPack(repoId, pack, ref, commitSha, source) {
             if (repoId.kind !== "workflow-run") {
                 throw new Error(`hub-session lookups receiveWorkflowRunPack received unsupported repo kind ${JSON.stringify(repoId.kind)}`);
             }
-            const deploymentId = repoId.id;
+            const workflowRunRepoId = repoId.id;
+            if (deriveWorkflowRunRepoId(source.agentAddress) !== workflowRunRepoId) {
+                logger.warn `Workflow-run pack rejected for ${workflowRunRepoId}: source address does not own the repository`;
+                return { accepted: false, reason: "path_violation" };
+            }
+            const [anchor] = await db
+                .select({
+                id: workflowRun.id,
+                address: workflowRun.address,
+                anchorRunId: workflowRun.anchorRunId,
+                tenantId: workflowRun.tenantId,
+                definitionId: workflowRun.definitionId,
+            })
+                .from(workflowRun)
+                .where(and(eq(workflowRun.address, source.agentAddress), inArray(workflowRun.status, [...liveWorkflowRunStatuses]), isNotNull(workflowRun.definitionId)))
+                .limit(1);
+            if (anchor === undefined ||
+                anchor.anchorRunId !== anchor.id ||
+                anchor.address === null) {
+                logger.warn `Workflow-run pack rejected for ${workflowRunRepoId}: source address has no live deployment anchor`;
+                return { accepted: false, reason: "path_violation" };
+            }
+            const anchorAddress = anchor.address;
+            let newlyTerminalRuns;
             try {
-                await agentRepoStore.receiveWorkflowRunPack({ kind: "workflow-run", id: deploymentId }, pack, ref, commitSha);
+                newlyTerminalRuns = await db.transaction(async (tx) => {
+                    const [allocation] = await tx
+                        .select()
+                        .from(sidecarAllocation)
+                        .where(eq(sidecarAllocation.anchorRunId, anchor.id))
+                        .limit(1)
+                        .for("update");
+                    if (allocation === undefined ||
+                        allocation.id !== source.allocationId ||
+                        allocation.anchorRunId !== source.anchorRunId ||
+                        source.anchorRunId !== anchor.id ||
+                        allocation.status !== "allocated" ||
+                        allocation.generation !== source.generation ||
+                        allocation.ensureAcceptedGeneration !== source.generation) {
+                        return null;
+                    }
+                    // Replacement advances this same row. Keep its lock until the
+                    // repository ref has advanced so ownership cannot change after
+                    // validation but before the old worker's pack becomes authoritative.
+                    return agentRepoStore.receiveWorkflowRunPack({ kind: "workflow-run", id: workflowRunRepoId }, pack, ref, commitSha);
+                });
+                if (newlyTerminalRuns === null) {
+                    logger.warn `Workflow-run pack rejected for ${workflowRunRepoId}: source connection does not own the deployment's current allocation`;
+                    return { accepted: false, reason: "path_violation" };
+                }
             }
             catch (err) {
                 const msg = err instanceof Error ? err.message : String(err);
                 if (msg.startsWith("path_violation")) {
-                    logger.warn `Workflow-run pack rejected for ${deploymentId}: ${msg}`;
+                    logger.warn `Workflow-run pack rejected for ${workflowRunRepoId}: ${msg}`;
                     return { accepted: false, reason: "path_violation" };
                 }
-                // Mirror the agent-state branch's catch-all: any other failure
-                // from the repo subsystem (filesystem races, kind-handler
-                // diagnostics surfaced as Error messages, etc.) becomes a
-                // structured `corrupt` rejection so the sender can re-push,
-                // and the underlying error is logged so the cause stays
-                // traceable on the hub side.
-                logger.error `Workflow-run pack receive failed for ${deploymentId}: ${msg}`;
+                // Mirror the agent-state branch's catch-all: any other failure from
+                // the fenced receive (transaction failures, filesystem races,
+                // kind-handler diagnostics surfaced as Error messages, etc.) becomes
+                // a structured `corrupt` rejection so the sender can re-push, and the
+                // underlying error is logged so the cause stays traceable on the hub
+                // side.
+                logger.error `Workflow-run pack receive failed for ${workflowRunRepoId}: ${msg}`;
                 return { accepted: false, reason: "corrupt" };
             }
+            // The substrate has already durably advanced the git ref by the time it
+            // returns, so the pack is accepted regardless of what happens below. The
+            // per-run status flip and principal deactivation are a best-effort
+            // downstream side effect of that durable advance, not part of accepting
+            // the pack. A failure here leaves the run "running" in the DB with its
+            // principal still active; there is no automatic re-fire, because a
+            // redelivery of the same durable tip produces no newly-terminal signal
+            // (the substrate's per-commit walk short-circuits on an already-present
+            // tip). The failure is therefore logged at ERROR as the only record that
+            // the row needs a manual flip, and the pack verdict stays accepted so the
+            // sidecar is acked and does not wedge re-pushing a pack that already
+            // landed.
+            const now = new Date();
+            for (const { runId, status } of newlyTerminalRuns) {
+                try {
+                    await db.transaction(async (tx) => {
+                        // Lazily anchor the run before settling it. An internal run that
+                        // parks only on a plain signal gate never reaches
+                        // `registerSignalCorrelation`, the sole other path that mints an
+                        // internal run row, so its terminal event can be the first the hub
+                        // sees of the run. A never-minted row is ordinary bookkeeping, not
+                        // a deployment-boundary violation, so mint it here against this
+                        // deployment's anchor rather than letting the ownership guard below
+                        // mistake absence for foreignness. The insert no-ops when any row
+                        // already exists, which keeps that guard authoritative for a row
+                        // that exists and anchors elsewhere. The principal is null: an
+                        // internal run inherits its deployment's grants and has none of its
+                        // own.
+                        //
+                        // The mint necessarily precedes the ownership guard, so an id the
+                        // hub has never seen is claimed under THIS anchor before anything
+                        // establishes it belongs here. That ordering is required -- the
+                        // guard reads the row the mint may have to create -- and it is
+                        // bounded rather than unbounded: internal run ids are supplied by
+                        // the sidecar and accepted verbatim, so the value is
+                        // caller-influenced, but it is a different population from the
+                        // anchor ids the hub mints itself, and nothing resolves an
+                        // internal id without also constraining the anchor or the tenant.
+                        // The insert cannot take a row away from another deployment; the
+                        // worst it does is create one for an id that deployment would
+                        // otherwise have created later.
+                        await workflowRunStore.createIfAbsent({
+                            id: runId,
+                            anchorRunId: anchor.id,
+                            definitionId: anchor.definitionId,
+                            tenantId: anchor.tenantId,
+                            principalId: null,
+                            status: "running",
+                        }, tx);
+                        const [ownedRun] = await tx
+                            .select({ anchorRunId: workflowRun.anchorRunId })
+                            .from(workflowRun)
+                            .where(eq(workflowRun.id, runId))
+                            .limit(1);
+                        if (ownedRun?.anchorRunId !== anchor.id) {
+                            logger.error `Ignoring terminal event for run ${runId}: it does not belong to source deployment ${anchor.id}`;
+                            return;
+                        }
+                        const won = await workflowRunStore.markTerminal(runId, status, now, tx);
+                        if (won === null) {
+                            // The row exists (the mint above guarantees it) and belongs to
+                            // this deployment (the guard above), so no running row matched
+                            // only because the run is already terminal -- a benign replay
+                            // against an already-settled row. Leave its settled status and
+                            // `endedAt` alone.
+                            return;
+                        }
+                        // Deactivate the run's own principal, if it has one. Externally-
+                        // triggered runs carry a principal; internal, workflow-spawned runs
+                        // have `principalId = null` and inherit the deployment's grants, so
+                        // there is nothing to deactivate. Deactivation is gated on winning
+                        // the flip -- the single claim point -- not on the principal's own
+                        // status.
+                        if (won.principalId !== null) {
+                            await tx
+                                .update(principal)
+                                .set({ status: "deactivated", updatedAt: now })
+                                // The `refId` clause is a defensive mirror of the per-instance
+                                // teardown in instances.ts: `won.principalId` is already this
+                                // run's own principal, and `principal.id` is the primary key,
+                                // so the `refId` match is belt-and-suspenders that the id we
+                                // won belongs to this run.
+                                .where(and(eq(principal.id, won.principalId), eq(principal.refId, runId)));
+                        }
+                    });
+                }
+                catch (err) {
+                    // Per-run isolation: a failed flip for one run must not abort the
+                    // rest of the batch, and must not throw out of this method -- a throw
+                    // would leave the sidecar with neither an ack nor a reject for a pack
+                    // the substrate already accepted. This ERROR is the only signal that
+                    // the run is stuck "running" in the DB with its principal active, so
+                    // it carries enough to find and flip the row by hand.
+                    const msg = err instanceof Error ? err.message : String(err);
+                    logger.error `Terminal DB flip failed for run ${runId} (deployment ${anchor.id}, target status ${status}); run left running in the DB: ${msg}`;
+                }
+            }
+            // A sidecar-local mail ack is only receipt. The raw trigger remains in
+            // workflow_run_dispatch until the accepted Git tip proves either that
+            // the run recorded it (RunStarted / SignalReceived), or that the
+            // supervisor consumed it with an explicit rejection. Rescan the bounded
+            // retained claim-check index after every accepted pack, and search the
+            // stable run log newest-first only for currently-unsettled ids. Settlement
+            // is idempotent, and a later pack naturally retries a transient database
+            // failure here.
+            let topLevelTerminalSettlementProjected = false;
+            try {
+                const reads = await agentRepoStore.repoStore.openCommittedReads({ kind: "hub" }, repoId, ref);
+                if (reads !== null) {
+                    const unsettledDispatches = await workflowRunDispatchStore.listUnsettled(anchor.id);
+                    const unsettledByMessageId = new Map(unsettledDispatches.map((dispatch) => [
+                        dispatch.messageId,
+                        dispatch,
+                    ]));
+                    for (const consumed of await listConsumedWorkflowDispatches(reads)) {
+                        if (consumed.address !== anchorAddress)
+                            continue;
+                        const persisted = unsettledByMessageId.get(consumed.messageId);
+                        if (persisted?.kind !== "mail")
+                            continue;
+                        if (consumed.rejection === undefined) {
+                            await workflowRunDispatchStore.settle(anchor.id, consumed.messageId, now);
+                        }
+                        else {
+                            await workflowRunDispatchStore.fail({
+                                anchorRunId: anchor.id,
+                                messageId: consumed.messageId,
+                                code: consumed.rejection.code,
+                                message: consumed.rejection.message,
+                                now,
+                            });
+                        }
+                        unsettledByMessageId.delete(consumed.messageId);
+                    }
+                    // Mail is recorded on the stable deployment run, while a signal is
+                    // recorded on the exact run named by its durable delivery frame.
+                    // Group retained dispatches by that Git run before scanning so an
+                    // internal run's SignalReceived evidence settles its own dispatch.
+                    // The `runs/<runId>/` log keys on the run id (the address local
+                    // part), NOT the full address: post-collapse the top-level run's id
+                    // IS `anchor.id`, so mail keys on `anchor.id`; a signal keys on its
+                    // frame's own run id. (The `addresses/<address>/` consumed subtree
+                    // above keys on the full address -- a different subtree.)
+                    const messageIdsByRun = new Map();
+                    for (const dispatch of unsettledByMessageId.values()) {
+                        const runId = dispatch.kind === "mail"
+                            ? anchor.id
+                            : SignalDeliverFrame.assert(JSON.parse(new TextDecoder().decode(dispatch.rawMessage))).runId;
+                        const messageIds = messageIdsByRun.get(runId) ?? new Set();
+                        messageIds.add(dispatch.messageId);
+                        messageIdsByRun.set(runId, messageIds);
+                    }
+                    for (const [runId, messageIds] of messageIdsByRun) {
+                        for (const accepted of await listAcceptedWorkflowDispatches(reads, runId, messageIds)) {
+                            const persisted = unsettledByMessageId.get(accepted.messageId);
+                            if (persisted?.kind !== accepted.kind)
+                                continue;
+                            await workflowRunDispatchStore.settle(anchor.id, accepted.messageId, now);
+                            unsettledByMessageId.delete(accepted.messageId);
+                        }
+                    }
+                    topLevelTerminalSettlementProjected =
+                        (await readCommittedWorkflowRunLifecycle(reads, anchor.id)) ===
+                            "terminal";
+                }
+            }
+            catch (error) {
+                logger.error `Workflow dispatch settlement failed for ${anchor.id}; accepted Git state remains authoritative and the retained payload will be retried: ${error instanceof Error ? error.message : String(error)}`;
+            }
+            if (topLevelTerminalSettlementProjected) {
+                try {
+                    await workflowRunDispatchStore.failUnsettled(anchor.id, "workflow_run_terminal", `Workflow run ${anchorAddress} is terminal and cannot accept this dispatch`, now);
+                }
+                catch (error) {
+                    logger.error `Failed to close unsettled workflow dispatches for terminal run ${anchorAddress}: ${error instanceof Error ? error.message : String(error)}`;
+                }
+            }
             return { accepted: true };
         },
     };
 }
 /**
- * Extract the instance id from an `<instanceId>@<domain>` agent address.
- * Throws on any input the `@intx/types`-owned `parseAgentAddress`
- * rejects: missing or leading `@`, empty domain, or an instance id
- * without the canonical `ins_` prefix.
+ * Extract the run id from an `<runId>@<domain>` run address.
+ * Throws on any input the `@intx/types`-owned `parseRunAddress`
+ * rejects: missing or leading `@`, empty domain, or a run id
+ * without the canonical `run_` prefix.
  */
 export function parseAgentId(agentAddress) {
-    const parsed = parseAgentAddress(agentAddress);
+    const parsed = parseRunAddress(agentAddress);
     if (parsed === null) {
-        throw new Error(`Invalid agent address: "${agentAddress}"`);
+        throw new Error(`Invalid run address: "${agentAddress}"`);
     }
-    return parsed.instanceId;
+    return parsed.runId;
 }
-export async function findInstance(db, agentAddress) {
-    return db.query.agentInstance.findFirst({
-        where: and(eq(agentInstance.address, agentAddress), isNull(agentInstance.endedAt)),
-    });
+/**
+ * Resolve a run address to the `workflow_run` endpoint backing it, keyed by
+ * the run's `address`. Every routable address names one self-anchored run --
+ * the deployment's anchor -- so this resolves the run's own address (the
+ * source `persistMail` depends on to record a triggered deployment's mail).
+ */
+export async function resolveRoutableAddress(db, address) {
+    const runRow = await db
+        .select({
+        id: workflowRun.id,
+        tenantId: workflowRun.tenantId,
+        publicKey: workflowRun.publicKey,
+        status: workflowRun.status,
+        principalId: workflowRun.principalId,
+    })
+        .from(workflowRun)
+        .where(and(eq(workflowRun.address, address), isNull(workflowRun.endedAt)))
+        .limit(1)
+        .then((rows) => rows[0]);
+    if (runRow === undefined) {
+        return undefined;
+    }
+    return {
+        id: runRow.id,
+        tenantId: runRow.tenantId,
+        address,
+        publicKey: runRow.publicKey,
+        status: runRow.status,
+        sessionId: await resolveRunSessionId(db, runRow.principalId),
+    };
 }
-export async function requireInstance(db, agentAddress) {
-    const row = await findInstance(db, agentAddress);
-    if (!row) {
-        throw new Error(`No active instance found for address "${agentAddress}"`);
+/**
+ * A folded run has no session column; its session is the `agent_session` keyed
+ * by the run's principal. By default this is the live (not-yet-ended) session,
+ * matching routing semantics; `includeEnded` also resolves a stopped run's
+ * ended session, which mail history needs. Returns null when the run has no
+ * principal or no matching session. Transitional, alongside
+ * `RoutableEndpoint.sessionId`.
+ */
+export async function resolveRunSessionId(db, principalId, opts = {}) {
+    if (principalId === null) {
+        return null;
+    }
+    // One session per run principal (invariant), so limit(1) returns the whole
+    // history; order deterministically so a hypothetical second row cannot make
+    // the pick flap. If a run ever grows multiple sessions per principal this
+    // becomes a union and limit(1) silently truncates.
+    const conditions = [eq(agentSession.principalId, principalId)];
+    if (opts.includeEnded !== true) {
+        conditions.push(isNull(agentSession.endedAt));
+    }
+    const row = await db
+        .select({ id: agentSession.id })
+        .from(agentSession)
+        .where(and(...conditions))
+        .orderBy(asc(agentSession.createdAt))
+        .limit(1)
+        .then((rows) => rows[0]);
+    return row?.id ?? null;
+}
+/**
+ * The folded run that owns a session, or null when the session belongs to no
+ * run. This is the inverse of `resolveRunSessionId`: a mail-read path holds a
+ * `sessionMail.sessionId` and no address, so it recovers the owning run by
+ * joining `workflow_run` to `agent_session` on their shared principal (a folded
+ * run, its session, and its launch all key on the same `instancePrincipalId`).
+ * Scoped to the tenant and routed through `workflow_run` so the returned id is
+ * proven to name a real run of this tenant -- callers key an authorization
+ * subject on it, so a session held by a non-run principal must fail closed to
+ * null rather than resolve to a fabricated subject.
+ */
+export async function resolveRunIdForSession(db, sessionId, tenantId) {
+    // No `endedAt` filter: a stopped run's mail must stay fetchable, so the
+    // session resolves whether or not it has ended. The run principal is minted
+    // per launch and shared 1:1 by the run and its session, so at most one row
+    // matches; order deterministically anyway so a hypothetical second row cannot
+    // make the pick flap, mirroring `resolveRunSessionId`.
+    const row = await db
+        .select({ id: workflowRun.id })
+        .from(workflowRun)
+        .innerJoin(agentSession, eq(agentSession.principalId, workflowRun.principalId))
+        .where(and(eq(agentSession.id, sessionId), eq(workflowRun.tenantId, tenantId)))
+        .orderBy(asc(workflowRun.createdAt))
+        .limit(1)
+        .then((rows) => rows[0]);
+    return row?.id ?? null;
+}
+/**
+ * Shape a run row and its already-resolved routing address into the run
+ * record. Callers decide whether the run resolves at all -- only a top-level
+ * run (`isTopLevelRun`) does -- and pass the address they have narrowed; this
+ * only maps the columns, including the run's `endedAt ?? createdAt` stand-in
+ * for the absent `updatedAt`.
+ */
+export function runRowToRoutableRecord(run, address) {
+    return {
+        id: run.id,
+        tenantId: run.tenantId,
+        address,
+        publicKey: run.publicKey,
+        status: run.status,
+        createdAt: run.createdAt,
+        updatedAt: run.endedAt ?? run.createdAt,
+        endedAt: run.endedAt,
+        definitionId: run.definitionId,
+        principalId: run.principalId,
+        kernelId: run.kernelId,
+        sidecarId: run.sidecarId,
+    };
+}
+/**
+ * A run is a top-level run -- the addressable head of a deployment -- when it
+ * owns a routing address AND self-anchors (`anchorRunId === id`). A lazy child
+ * park row anchors on its parent (`anchorRunId !== id`) and carries no address;
+ * either condition excludes it. This is the single predicate the run read
+ * surface classifies on, so the resolver and the run list cannot drift.
+ */
+export function isTopLevelRun(row) {
+    return row.address !== null && row.anchorRunId === row.id;
+}
+/**
+ * Resolve a run id to its record. A run resolves only when it is a top-level
+ * run (`isTopLevelRun`): it owns a routing address and self-anchors. A child
+ * park row (address-null, anchored on its parent) is not served here.
+ */
+export async function findRoutableById(db, id, tenantId) {
+    const runRow = await db
+        .select({
+        id: workflowRun.id,
+        tenantId: workflowRun.tenantId,
+        address: workflowRun.address,
+        anchorRunId: workflowRun.anchorRunId,
+        publicKey: workflowRun.publicKey,
+        status: workflowRun.status,
+        createdAt: workflowRun.createdAt,
+        endedAt: workflowRun.endedAt,
+        principalId: workflowRun.principalId,
+        kernelId: workflowRun.kernelId,
+        sidecarId: workflowRun.sidecarId,
+        definitionId: workflowRun.definitionId,
+    })
+        .from(workflowRun)
+        .where(and(eq(workflowRun.id, id), eq(workflowRun.tenantId, tenantId)))
+        .limit(1)
+        .then((rows) => rows[0]);
+    // The `address === null` arm is redundant with `isTopLevelRun` (which already
+    // requires a non-null address) but narrows `address` from `string | null` to
+    // `string` for `runRowToRoutableRecord`, which `isTopLevelRun`'s boolean
+    // return cannot do.
+    if (runRow === undefined ||
+        !isTopLevelRun(runRow) ||
+        runRow.address === null ||
+        runRow.definitionId === null) {
+        return undefined;
     }
-    return row;
+    return runRowToRoutableRecord({ ...runRow, definitionId: runRow.definitionId }, runRow.address);
 }