@nodetool-ai/websocket 0.7.0-rc.22 → 0.8.1

package/dist/session/job-execution.js

@@ -0,0 +1,2089 @@
+/**
+ * Everything between "run_job arrived" and "terminal status persisted": the
+ * admission gates, the concurrency queue, run start-up, the message stream,
+ * reconnect, cancel, and terminal-status persistence.
+ *
+ * Extracted from `WebSocketClientSession` (T3 of
+ * docs/websocket-runner-refactor-plan.md). It knows the connection only
+ * through {@link ClientSession}; what the seam does not carry arrives as
+ * {@link JobExecutionDeps}.
+ */
+import { randomUUID } from "node:crypto";
+import { ApiErrorCode } from "../error-codes.js";
+import { admitSpend, releaseSpend, reserveSpend } from "../credit-gate.js";
+import { JobConcurrencyQueue } from "../job-queue.js";
+import { isBoolean, isFiniteNumber, isFunctionValue, isNonEmptyString, isNumber, isObjectLike, isRecord, isString } from "../lib/wire-values.js";
+import { Graph, withExplicitNodeFlags } from "@nodetool-ai/kernel";
+import { ExecutionSession, isExecutionPreflightError, isUnitBilledCapability, priceGeneration, toRawGraphInput } from "@nodetool-ai/execution";
+import { persistableOutputs, withoutInlineAssetBytes } from "@nodetool-ai/execution/service";
+import { createRunSupervisor } from "../run-supervisor.js";
+import { jobRunRegistry } from "../job-run-registry.js";
+import { Application, Asset, Job, listApplicationVersions, invocationIdInUse, releasedApplicationRelease, releasedApplicationVersion, reserveInvocation, settleInvocation, Workflow } from "@nodetool-ai/models";
+import { getInstanceId } from "../lib/instance-id.js";
+import { requestRemoteJobCancel } from "../job-control.js";
+import { estimateWorkflowCost, nodeExpectedQuantity } from "@nodetool-ai/node-sdk/cost-estimate";
+import { extractPricingParams } from "@nodetool-ai/node-sdk/pricing-params";
+import { getModelUnitPrice } from "@nodetool-ai/model-pricing";
+import { applicationReleaseResponse } from "@nodetool-ai/protocol/api-schemas/applications.js";
+import { confineRunRequest, isRunRefusal } from "../lib/app-session-scope.js";
+import { releaseBlockedReason } from "../lib/app-deployment-service.js";
+import { autoSaveAssets, primaryTextOutputName } from "./asset-autosave.js";
+import { createRuntimeContext } from "./model-interfaces.js";
+import { formatSanitizedError, sanitizeLargeText } from "./sanitize.js";
+import { createLogger } from "@nodetool-ai/config";
+const log = createLogger("nodetool.websocket.runner");
+const TERMINAL_JOB_STATUSES = ["completed", "failed", "cancelled", "error"];
+export const DEFAULT_RUN_JOB_EXECUTION_OPTIONS = Object.freeze({
+    persistence: "job",
+    eventDetail: "full",
+    assetPersistence: "auto"
+});
+export function resolveRunJobExecutionOptions(value, sdkDefaults = false) {
+    return {
+        persistence: value?.persistence === "session" ? "session" : "job",
+        eventDetail: value?.event_detail === "outputs" || value?.event_detail === "terminal"
+            ? value.event_detail
+            : "full",
+        assetPersistence: value?.asset_persistence === "temporary" ||
+            (value?.asset_persistence == null && sdkDefaults)
+            ? "temporary"
+            : "auto"
+    };
+}
+export function resolveRunJobUserId(requestUserId, connectionUserId) {
+    return requestUserId?.trim() || connectionUserId;
+}
+export function createRelayActivityWaiter(context, executionSettled, abortSignal) {
+    let pending = context.hasMessages();
+    let resolveWaiter = null;
+    let disposed = false;
+    const notify = () => {
+        pending = true;
+        const resolve = resolveWaiter;
+        resolveWaiter = null;
+        resolve?.();
+    };
+    const removeMessageListener = context.addMessageListener(notify);
+    const dispose = () => {
+        if (disposed)
+            return;
+        disposed = true;
+        removeMessageListener();
+        abortSignal?.removeEventListener("abort", onAbort);
+    };
+    const settle = () => {
+        notify();
+        dispose();
+    };
+    const onAbort = () => settle();
+    if (abortSignal?.aborted) {
+        settle();
+    }
+    else {
+        abortSignal?.addEventListener("abort", onAbort, { once: true });
+    }
+    void executionSettled.then(settle, settle);
+    return async () => {
+        if (pending) {
+            pending = false;
+            return;
+        }
+        await new Promise((resolve) => {
+            resolveWaiter = resolve;
+        });
+        pending = false;
+    };
+}
+/** The shape `crypto.randomUUID()` produces, on either side of the wire. */
+const UUID_PATTERN = /^[0-9a-f]{8}-[0-9a-f]{4}-[1-8][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$/i;
+export class JobExecutionManager {
+    session;
+    deps;
+    constructor(session, deps) {
+        this.session = session;
+        this.deps = deps;
+    }
+    /**
+     * The runs in flight on this connection. Private since T8's D2: chat's runs
+     * and the suites that stage one go through the methods below, so every
+     * mutation of the connection's slot accounting happens in this file.
+     */
+    activeJobs = new Map();
+    /**
+     * Runs that arrived while {@link MAX_CONCURRENT_JOBS} runs were already in
+     * flight. They start automatically (FIFO) as active jobs finish.
+     */
+    jobQueue = new JobConcurrencyQueue();
+    dequeuedJobs = new Set();
+    /**
+     * Count of jobs that have passed the concurrency gate but haven't been added
+     * to {@link activeJobs} yet (startJob awaits graph hydration first). Counted
+     * toward the cap synchronously so two run_job commands arriving back-to-back
+     * can't both slip past `activeJobs.size` and exceed MAX_CONCURRENT_JOBS.
+     */
+    startingJobs = 0;
+    /**
+     * WS slot accounting, for leak accounting: after every run finishes both
+     * must be back to zero. Read by the reliability harness's ws-server driver,
+     * whose `cleanup-leaks` invariant can only assert what it can measure.
+     */
+    get slotCounters() {
+        return {
+            activeJobs: this.activeJobs.size,
+            startingJobs: this.startingJobs
+        };
+    }
+    /** Ids of the runs in flight on this connection. Read for diagnostics. */
+    activeJobIds() {
+        return [...this.activeJobs.keys()];
+    }
+    /**
+     * Register a run against this connection's slots. A chat turn that starts a
+     * workflow shares the connection's concurrency accounting, so its run is
+     * registered here rather than in a map of chat's own.
+     */
+    registerJob(jobId, active) {
+        this.activeJobs.set(jobId, active);
+    }
+    /** Drop a run without draining the queue — the caller's `finally` does that. */
+    dropJob(jobId) {
+        this.activeJobs.delete(jobId);
+    }
+    /** Drop a run and start whatever was queued behind it. */
+    releaseJob(jobId) {
+        this.activeJobs.delete(jobId);
+        this.drainQueue();
+    }
+    /** The run registered under `jobId`, if it is still in flight. */
+    getActiveJob(jobId) {
+        return this.activeJobs.get(jobId);
+    }
+    /** Whether a run is registered under `jobId`. */
+    hasActiveJob(jobId) {
+        return this.activeJobs.has(jobId);
+    }
+    /**
+     * Job ids of runs still executing on a previous connection's runner that
+     * THIS connection reattached to via `reconnect_job`. Client commands for
+     * them (`cancel_job`, `stop`, `stream_input`, `end_input_stream`,
+     * `update_node_properties`) are forwarded to the executing runner through
+     * the session's hooks.
+     *
+     * Ids, not session objects: holding the session would pin its frame buffer
+     * for this connection's whole life, defeating the registry's retention
+     * drop. Resolved through the registry at every use, so a dropped session
+     * simply resolves to null and its memory goes with it.
+     */
+    adoptedJobIds = new Set();
+    /**
+     * This connection's identity as a job-run delivery target. Separate object
+     * from the host's chat delivery target so a job session's `detach(target)`
+     * can never be confused with a chat session's — both guard on identity.
+     */
+    jobDeliveryTarget = {
+        deliver: (message) => this.deps.sendToSocket(message)
+    };
+    inferOutputType(value) {
+        if (value === null || value === undefined)
+            return "any";
+        if (isString(value))
+            return "str";
+        if (isNumber(value))
+            return Number.isInteger(value) ? "int" : "float";
+        if (isBoolean(value))
+            return "bool";
+        if (Array.isArray(value))
+            return "list";
+        if (isObjectLike(value))
+            return "dict";
+        return "any";
+    }
+    resolveOutputNodeForKey(active, outputKey) {
+        let fallback = null;
+        for (const raw of active.graph.nodes) {
+            const node = raw;
+            const id = isString(node.id) ? node.id : null;
+            if (!id)
+                continue;
+            const name = isString(node.name) ? node.name : id;
+            const type = isString(node.type) ? node.type : "";
+            if (name === outputKey || id === outputKey)
+                return { id, name };
+            if (type === "nodetool.output.Output" && !fallback)
+                fallback = { id, name };
+        }
+        return fallback;
+    }
+    async sendOutputUpdates(active, outputs) {
+        for (const [outputKey, values] of Object.entries(outputs)) {
+            const nodeRef = this.resolveOutputNodeForKey(active, outputKey) ?? {
+                id: outputKey,
+                name: outputKey
+            };
+            const seq = Array.isArray(values) ? values : [];
+            for (const rawValue of seq) {
+                const value = await active.context.normalizeOutputValue(rawValue);
+                await this.session.send({
+                    type: "output_update",
+                    node_id: nodeRef.id,
+                    node_name: nodeRef.name,
+                    output_name: "output",
+                    value,
+                    output_type: this.inferOutputType(value),
+                    metadata: {},
+                    workflow_id: active.workflowId,
+                    job_id: active.jobId
+                });
+            }
+        }
+    }
+    async normalizeFinalOutputs(active, outputs) {
+        const normalized = {};
+        for (const [outputKey, values] of Object.entries(outputs)) {
+            normalized[outputKey] = [];
+            for (const value of Array.isArray(values) ? values : []) {
+                normalized[outputKey].push(await active.context.normalizeOutputValue(value));
+            }
+        }
+        return normalized;
+    }
+    /**
+     * The resilient session a frame's `job_id` belongs to: one this connection
+     * started, or one it adopted via `reconnect_job`.
+     */
+    resolveJobSession(jobId) {
+        if (!isNonEmptyString(jobId))
+            return null;
+        const active = this.activeJobs.get(jobId);
+        if (active?.runSession)
+            return active.runSession;
+        if (!this.adoptedJobIds.has(jobId))
+            return null;
+        // No user means no connect(), so this session can own no registered run.
+        // "Nothing running" is the honest answer; main looked the job up under a
+        // hardcoded "1", which could return another user's run.
+        const userId = this.session.userId;
+        if (userId === null)
+            return null;
+        return jobRunRegistry.get(userId, jobId);
+    }
+    /**
+     * Where a client command for `jobId` should act: this connection's own
+     * ExecutionSession, or — for a run this client reconnected to — the hooks
+     * of the session whose runner still owns it. Null when nothing is running.
+     */
+    resolveJobControl(jobId) {
+        const active = this.activeJobs.get(jobId);
+        if (active) {
+            const session = active.session;
+            return {
+                workflowId: active.workflowId,
+                hooks: {
+                    cancel: () => session.cancel(),
+                    pushInput: (input, value, handle) => session.pushInput(input, value, handle),
+                    finishInputStream: (input, handle) => session.finishInputStream(input, handle),
+                    updateNodeProperties: (nodeId, properties) => session.updateNodeProperties(nodeId, properties)
+                }
+            };
+        }
+        const lookupUserId = this.session.userId;
+        if (lookupUserId === null)
+            return null;
+        const registered = jobRunRegistry.get(lookupUserId, jobId);
+        if (registered && registered.status === "running") {
+            return { hooks: registered.hooks, workflowId: registered.workflowId };
+        }
+        return null;
+    }
+    /**
+     * Normalize a raw graph so that the kernel's NodeDescriptor contract is met.
+     * The web-UI / Python serialisation stores node properties under `data`;
+     * the kernel expects them under `properties`.
+     */
+    normalizeGraph(graph) {
+        const nodes = graph.nodes.map((n) => {
+            if (n.properties === undefined && n.data !== undefined) {
+                const { data, ...rest } = n;
+                return { ...rest, properties: data };
+            }
+            return n;
+        });
+        const edges = graph.edges.map((edge) => {
+            const rawEdgeType = edge.edge_type ?? edge.type;
+            const edge_type = rawEdgeType === "control" ? "control" : "data";
+            const { type, ...rest } = edge;
+            return { ...rest, edge_type };
+        });
+        return { nodes, edges };
+    }
+    async hydrateGraph(graph) {
+        const normalized = this.normalizeGraph(graph);
+        if (!this.session.resolveNodeType) {
+            // No registry resolver configured — behavior flags can only come from
+            // the saved graph itself; absent ones are explicitly defaulted off.
+            // `normalizeGraph` above moved a saved node's `data` to `properties`
+            // and settled `edge_type`; what a saved record still lacks is the
+            // declared string type of the four identity fields, so read them out
+            // rather than assert them.
+            const asGraphData = {
+                nodes: normalized.nodes.map((n) => ({
+                    ...n,
+                    id: String(n.id ?? ""),
+                    type: String(n.type ?? "")
+                })),
+                edges: normalized.edges.map((e) => ({
+                    ...e,
+                    source: String(e.source ?? ""),
+                    sourceHandle: String(e.sourceHandle ?? ""),
+                    target: String(e.target ?? ""),
+                    targetHandle: String(e.targetHandle ?? "")
+                }))
+            };
+            return withExplicitNodeFlags(asGraphData);
+        }
+        const hydrated = await Graph.loadFromDict(normalized, {
+            resolver: this.session.resolveNodeType
+        });
+        return {
+            nodes: [...hydrated.nodes],
+            edges: [...hydrated.edges]
+        };
+    }
+    getRawGraph(req) {
+        if (req.graph) {
+            return this.normalizeGraph(req.graph);
+        }
+        if (req.workflow_id) {
+            const userId = this.session.requireUserId();
+            const workflowId = req.workflow_id;
+            return (async () => {
+                const workflow = await Workflow.find(userId, workflowId);
+                if (!workflow)
+                    throw new Error(`Workflow not found: ${workflowId}`);
+                return this.normalizeGraph(workflow.graph);
+            })();
+        }
+        throw new Error("workflow_id or graph is required");
+    }
+    /**
+     * Surface a clean terminal job_update when pre-run setup fails (typically
+     * because the Python bridge could not start). Without this the error would
+     * bubble up to handleCommand and be sent as a generic `invalid_command`
+     * envelope, which the UI does not associate with the job — the workflow
+     * appears to spin forever instead of failing.
+     */
+    async emitBeforeRunFailure(jobId, workflowId, err, persistJob) {
+        const errorMessage = err instanceof Error ? err.message : String(err);
+        this.session.logError("beforeRunJob failed", err);
+        await this.session.send({
+            type: "job_update",
+            status: "failed",
+            job_id: jobId,
+            workflow_id: workflowId,
+            error: errorMessage
+        });
+        if (!persistJob)
+            return;
+        try {
+            const job = (await Job.get(jobId));
+            if (job) {
+                job.markFailed(errorMessage);
+                await job.save();
+            }
+        }
+        catch (persistErr) {
+            this.session.logError("beforeRunJob failure persistence failed", persistErr);
+        }
+    }
+    /**
+     * Jobs occupying a concurrency slot: live + reserved-but-not-yet-registered.
+     *
+     * The cap counts a *user's* runs, not a socket's. A run detached from a
+     * dropped connection keeps executing, so counting only `activeJobs` would
+     * let a client reconnect its way past the cap — four abandoned sockets and
+     * a fresh one is eight concurrent runs. The registry is the cross-
+     * connection count; `activeJobs` contributes only the entries it does not
+     * already cover (a chat-triggered workflow run, which registers no
+     * session), or the same run would be counted twice.
+     */
+    get inFlightJobCount() {
+        let sessionless = 0;
+        for (const job of this.activeJobs.values()) {
+            if (!job.runSession)
+                sessionless += 1;
+        }
+        // A session with no user has no registered runs to count; the local
+        // counters are still the honest answer. Same reasoning as the registry
+        // lookups above.
+        const countUserId = this.session.userId;
+        return (this.startingJobs +
+            (countUserId === null ? 0 : jobRunRegistry.countRunning(countUserId)) +
+            sessionless);
+    }
+    /**
+     * Number of live (unfinished) runs currently executing for a workflow. Used
+     * to enforce the per-workflow concurrency limit so non-concurrent runs stay
+     * sequential and their live node updates don't clobber each other in the
+     * editor. Counted the same way as {@link inFlightJobCount}: the registry
+     * across connections, plus this connection's sessionless entries.
+     */
+    countActiveJobsForWorkflow(workflowId) {
+        if (!workflowId) {
+            return 0;
+        }
+        const workflowUserId = this.session.userId;
+        let count = workflowUserId === null
+            ? 0
+            : jobRunRegistry.countRunningForWorkflow(workflowUserId, workflowId);
+        for (const job of this.activeJobs.values()) {
+            if (!job.runSession && job.workflowId === workflowId && !job.finished) {
+                count++;
+            }
+        }
+        return count;
+    }
+    /**
+     * The per-workflow concurrency limit a given run is subject to: concurrency
+     * opt-in runs share the configurable {@link getMaxConcurrentRunsPerWorkflow}
+     * cap; everything else stays strictly sequential (one run per workflow) so
+     * live node updates don't clobber the editor.
+     */
+    perWorkflowLimitFor(req) {
+        return req.concurrent
+            ? this.deps.getMaxConcurrentRunsPerWorkflow()
+            : Promise.resolve(1);
+    }
+    /**
+     * Read-only view of the same admission counters used by runJob. This does
+     * not reserve a slot or mutate the queue; lifecycle preflight can therefore
+     * report likely queueing without changing current execution behavior.
+     */
+    async getSdkExecutionCapacitySnapshot(input) {
+        const [maxConcurrentJobs, maxConcurrentRunsForWorkflow] = await Promise.all([
+            this.deps.getMaxConcurrentJobs(),
+            this.perWorkflowLimitFor({ concurrent: input.concurrent })
+        ]);
+        const inFlightJobs = this.inFlightJobCount;
+        const workflowInFlightJobs = this.countActiveJobsForWorkflow(input.workflowId);
+        return {
+            inFlightJobs,
+            maxConcurrentJobs,
+            queuedJobs: this.jobQueue.size,
+            workflowInFlightJobs,
+            maxConcurrentRunsForWorkflow,
+            likelyQueued: inFlightJobs >= maxConcurrentJobs ||
+                workflowInFlightJobs >= maxConcurrentRunsForWorkflow
+        };
+    }
+    /**
+     * Best-effort pre-run cost estimate for a graph, in USD. Nodes the estimator
+     * cannot price contribute nothing, so the figure is a floor — good enough to
+     * stop a run that would obviously blow the budget, and never a reason to
+     * refuse one it cannot price.
+     */
+    estimateRunCost(req) {
+        try {
+            const estimate = this.estimateGraphCost(req);
+            if (!estimate)
+                return 0;
+            return Number.isFinite(estimate.total) ? estimate.total : 0;
+        }
+        catch (err) {
+            this.session.logError("run cost estimate failed", err);
+            return 0;
+        }
+    }
+    /**
+     * The graph estimate both gates read. One implementation so the budget gate,
+     * the credit gate and the editor's cost panel price a run the same way —
+     * including fan-out: a node asked for ten images is priced at ten, through
+     * the same `nodeExpectedQuantity` the panel uses.
+     */
+    estimateGraphCost(req) {
+        const nodes = req.graph?.nodes;
+        if (!nodes || !this.session.getNodeMetadata)
+            return null;
+        const priced = nodes.map((node) => ({
+            id: String(node.id),
+            type: String(node.type),
+            data: (node.data ?? {})
+        }));
+        return estimateWorkflowCost({
+            nodes: priced,
+            getMetadata: (nodeType) => this.session.getNodeMetadata?.(nodeType),
+            // Prices the model picked on a generic node (e.g. a FAL or kie model on
+            // nodetool.image.TextToImage), which node-type metadata alone cannot.
+            // Same lookup the editor's cost preview uses.
+            getModelPrice: getModelUnitPrice,
+            // A per-second model bills the clip it is asked for, so the duration
+            // and resolution the node states have to reach the price lookup.
+            getParams: (node) => extractPricingParams(node.data),
+            quantities: Object.fromEntries(priced.map((node) => [
+                node.id,
+                nodeExpectedQuantity(node.data)
+            ]))
+        });
+    }
+    /**
+     * The slice of a run that spends through NodeTool's managed provider —
+     * the only spend the credit balance meters. BYOK nodes are excluded on
+     * purpose: their cost rides the user's own keys.
+     */
+    estimateNodetoolSpend(req) {
+        try {
+            const estimate = this.estimateGraphCost(req);
+            if (!estimate) {
+                return { usesNodetool: false, estimatedUsd: 0, models: [] };
+            }
+            const items = estimate.items.filter((item) => item.provider === "nodetool");
+            const total = items.reduce((sum, item) => sum +
+                (Number.isFinite(item.estimated_cost) ? item.estimated_cost : 0), 0);
+            const models = [
+                ...new Set(items
+                    .map((item) => item.model)
+                    .filter((model) => Boolean(model)))
+            ];
+            return { usesNodetool: items.length > 0, estimatedUsd: total, models };
+        }
+        catch (err) {
+            this.session.logError("nodetool spend estimate failed", err);
+            return { usesNodetool: false, estimatedUsd: 0, models: [] };
+        }
+    }
+    /** Tell the client a run was refused, in the shape a failed job takes. */
+    refuseRun(req, jobId, code, error) {
+        this.session.sendDetached({
+            type: "job_update",
+            status: "failed",
+            job_id: jobId,
+            workflow_id: req.workflow_id ?? null,
+            error,
+            error_code: code
+        });
+        return false;
+    }
+    /**
+     * The version number a stale client claimed, if the app really has it —
+     * otherwise null, and the claim is unsupportable.
+     *
+     * Version history is per app and short (one row per publish), so listing it
+     * is cheap; the ceiling exists only because the model helper takes a limit,
+     * and it has to be high enough that an old claim is never truncated away by
+     * the newest releases.
+     */
+    async claimedApplicationVersion(applicationId, claimed, userId) {
+        if (claimed == null)
+            return null;
+        const versions = await listApplicationVersions(applicationId, 10_000, userId);
+        return versions.some((v) => v.version === claimed) ? claimed : null;
+    }
+    /**
+     * Gate an app's run on its spend budget. Runs of a published app execute with
+     * the creator's secrets, so this refuses before the job exists rather than
+     * reporting an overspend afterwards. Returns false when the run was refused
+     * (the client has already been told why).
+     *
+     * `application_id` arrives on the wire, so before any of that the app has to
+     * be one this connection's user owns. Honouring the id as sent let a client
+     * name a stranger's app and spend their budget — and pollute their release
+     * telemetry, which is the same ledger.
+     */
+    async admitApplicationRun(req) {
+        const applicationId = req.application_id;
+        if (!applicationId)
+            return true;
+        const jobId = req.job_id ?? randomUUID();
+        req.job_id = jobId;
+        // The connection's authenticated user, not `req.user_id`: the request body
+        // is the thing being authorized, so it cannot supply the identity that
+        // authorizes it.
+        const userId = this.session.requireUserId();
+        // Authorization sits outside the try below, which swallows a ledger outage
+        // on purpose. Metering fails open; ownership fails closed — a lookup this
+        // never completed is not permission to bill the app it names.
+        let owned = false;
+        try {
+            const application = await Application.findById(applicationId);
+            owned = application?.user_id === userId;
+        }
+        catch (err) {
+            this.session.logError("application ownership check failed", err);
+        }
+        if (!owned) {
+            log.warn("Run refused: application not owned by this user", {
+                applicationId,
+                jobId,
+                userId
+            });
+            // Applications are owned by one user and there is no path today that
+            // serves someone else's app to run — `releasedApplicationDocument`
+            // itself requires ownership — so refusing cannot break a legitimate
+            // run, and it is the only answer that keeps the budget a hard stop.
+            return this.refuseRun(req, jobId, ApiErrorCode.NOT_FOUND, "Application not found");
+        }
+        try {
+            const estimatedUsd = this.estimateRunCost(req);
+            // The client says whether this is a release run or a draft run; the
+            // server decides which release the ledger records. A number taken on
+            // faith would let a run bill itself to a version it never executed, and
+            // the ledger is also the release telemetry — so a claim is only honoured
+            // below once the server has found that version in the app's history.
+            const released = req.application_version == null
+                ? null
+                : await releasedApplicationVersion(applicationId, userId);
+            if (req.application_version != null && !released) {
+                // A release run of an app that has released nothing. The claim is
+                // unsupportable rather than merely stale, and letting it through would
+                // file the run in the telemetry ledger as a release that never shipped.
+                return this.refuseRun(req, jobId, ApiErrorCode.INVALID_INPUT, "This app has no released version to run");
+            }
+            // Which version the ledger row belongs to. The release is the default,
+            // because that is what a current client runs.
+            let version = released?.version ?? null;
+            if (released && released.version !== req.application_version) {
+                log.warn("Run claimed a version other than the released one", {
+                    applicationId,
+                    jobId,
+                    claimed: req.application_version,
+                    released: released.version
+                });
+                // A client that loaded the app before the newest release still holds
+                // that older snapshot and is about to execute it. Filing the run under
+                // the current release would credit v2's metrics and budget with a run
+                // of v1, so the row follows what actually ran — but only once the
+                // server has confirmed the claimed version is a real version of this
+                // app. That check is what keeps the client from picking its own
+                // attribution: it can name a version it once had, not one it invents.
+                const claimed = await this.claimedApplicationVersion(applicationId, req.application_version, userId);
+                if (!claimed) {
+                    return this.refuseRun(req, jobId, ApiErrorCode.INVALID_INPUT, `This app has no version ${req.application_version} to run`);
+                }
+                version = claimed;
+                // The run proceeds, but the client is stale and has no other way to
+                // learn it: nothing in a job's updates mentions releases.
+                this.session.sendDetached({
+                    type: "notification",
+                    node_id: "",
+                    severity: "warning",
+                    workflow_id: req.workflow_id ?? null,
+                    content: `Running version ${claimed} of this app; version ${released.version} has since been released. Reload to get the latest.`
+                });
+            }
+            // Reserving claims the run against the budget in the same transaction
+            // that checks it, so concurrent runs of one app cannot each read a total
+            // that excludes the others and all be admitted.
+            // The ledger holds one row per (application, invocation), so two runs
+            // racing on one id lose the insert rather than double-spend. The loser
+            // is refused like any other rejected run instead of surfacing a driver
+            // error to the visitor.
+            let decision;
+            try {
+                decision = await reserveInvocation({
+                    applicationId,
+                    version,
+                    invocationId: jobId,
+                    operationId: req.operation_id ?? undefined,
+                    estimatedUsd,
+                    requireFiniteBudget: this.session.appSession !== null
+                });
+            }
+            catch (error) {
+                log.warn("Application run ledger rejected the reservation", {
+                    applicationId,
+                    jobId,
+                    error: String(error)
+                });
+                return this.refuseRun(req, jobId, ApiErrorCode.INVALID_INPUT, "This app run named a run id that is already in use");
+            }
+            if (!decision.allowed) {
+                log.warn("Run refused by application budget", {
+                    applicationId,
+                    jobId,
+                    reason: decision.reason
+                });
+                return this.refuseRun(req, jobId, ApiErrorCode.BUDGET_EXCEEDED, decision.reason);
+            }
+        }
+        catch (err) {
+            // A ledger that is unavailable must not take runs down with it; the
+            // refusals above are the only paths that block, and they return rather
+            // than throw so an outage can never swallow one.
+            this.session.logError("application budget check failed", err);
+        }
+        return true;
+    }
+    /**
+     * Gate a run on the user's credit balance — but only the part of the run
+     * that spends through NodeTool's managed provider. A graph with no
+     * `nodetool` models passes untouched (BYOK stays unmetered); one with them
+     * is refused when it names a model this server does not sell, or when the
+     * balance is empty or can't cover the estimate. Estimates are floors, so an
+     * empty balance refuses even a 0-estimate nodetool call. Fails open on gate
+     * errors, like the application-budget gate above.
+     */
+    async admitCreditRun(req) {
+        const { usesNodetool, estimatedUsd, models } = this.estimateNodetoolSpend(req);
+        if (!usesNodetool)
+            return true;
+        // Pin the job id now so the reservation taken here can be released at the
+        // run's terminal state (and on cancel-while-queued) under the same key.
+        req.job_id ??= randomUUID();
+        const decision = await admitSpend(this.session.requireUserId(), estimatedUsd, models);
+        if (!decision.allowed) {
+            return this.refuseRun(req, req.job_id, decision.refusal === "model_not_allowed"
+                ? ApiErrorCode.MODEL_NOT_AVAILABLE
+                : ApiErrorCode.BUDGET_EXCEEDED, decision.reason);
+        }
+        reserveSpend(this.session.requireUserId(), req.job_id, estimatedUsd);
+        return true;
+    }
+    /**
+     * Narrow a run started by a deployed app's visitor to the one run the
+     * session allows, or refuse it.
+     *
+     * The connection authenticates as the app's owner, so `admitApplicationRun`
+     * below would happily approve anything it names — ownership is what that
+     * check tests, and a visitor's session passes it. This is the check that
+     * actually confines them, and it runs first: what comes out is built from
+     * the signed session and the release, not from the request.
+     *
+     * The run executes only the release version the signed session names. A
+     * later publish invalidates the session for execution, so the visitor must
+     * reload and mint a session for the new release rather than run its graphs.
+     */
+    async confineAppSessionRun(req) {
+        const scope = this.session.appSession;
+        if (!scope)
+            return req;
+        // The client names the run so it can follow its own frames, but the id it
+        // names is also the ledger key a job command is authorized against. An id
+        // that is already a job or already in a ledger would let a visitor file a
+        // row under someone else's run and then command it, so a taken id — and a
+        // shape that is not a generated id at all — is refused rather than
+        // silently replaced with one the client would never recognize.
+        const requested = req.job_id ?? "";
+        const jobId = requested === "" ? randomUUID() : requested;
+        const deny = (code, error) => {
+            this.refuseRun(req, jobId, code, error);
+            return null;
+        };
+        if (requested !== "") {
+            if (!UUID_PATTERN.test(requested)) {
+                return deny(ApiErrorCode.INVALID_INPUT, "This app run named an invalid run id");
+            }
+            if ((await Job.get(requested)) !== null ||
+                (await invocationIdInUse(requested))) {
+                return deny(ApiErrorCode.INVALID_INPUT, "This app run named a run id that is already in use");
+            }
+        }
+        const release = await releasedApplicationRelease(scope.applicationId, this.session.requireUserId());
+        if (!release) {
+            return deny(ApiErrorCode.NOT_FOUND, "This app is not available");
+        }
+        // A document that no longer parses and one a public link may not serve get
+        // the same answer, so the visitor learns nothing about which it hit.
+        const parsedRelease = applicationReleaseResponse.safeParse(release);
+        if (!parsedRelease.success || releaseBlockedReason(parsedRelease.data)) {
+            return deny(ApiErrorCode.NOT_FOUND, "This app is not available");
+        }
+        if (release.version !== scope.version) {
+            return deny(ApiErrorCode.INVALID_INPUT, "This app has been updated. Reload the page before running it.");
+        }
+        const confined = confineRunRequest(req, scope, release);
+        if (isRunRefusal(confined)) {
+            return deny(ApiErrorCode.INVALID_INPUT, confined.refused);
+        }
+        confined.job_id = jobId;
+        return confined;
+    }
+    /**
+     * Entry point for the "run_job" command. Starts the run immediately when the
+     * client is under its concurrency cap, otherwise queues it (FIFO) and emits a
+     * `queued` job update. Queued runs start automatically as active jobs finish.
+     */
+    async runJob(incoming) {
+        const req = await this.confineAppSessionRun(incoming);
+        if (!req)
+            return;
+        req._accepted_at_ms ??= performance.now();
+        if (!(await this.admitApplicationRun(req)))
+            return;
+        if (!(await this.admitCreditRun(req)))
+            return;
+        const max = await this.deps.getMaxConcurrentJobs();
+        const perWorkflowMax = await this.perWorkflowLimitFor(req);
+        // Queue the run when over the global cap, or when this workflow already has
+        // its per-workflow limit of runs in flight — 1 for normal runs, or the
+        // configurable MAX_CONCURRENT_RUNS_PER_WORKFLOW for runs that opt into
+        // concurrency. Reserve the slot synchronously (after the awaits above) so
+        // two run_job commands can't both observe a free slot before either registers.
+        if (this.inFlightJobCount >= max ||
+            this.countActiveJobsForWorkflow(req.workflow_id) >= perWorkflowMax) {
+            await this.enqueueJob(req);
+            return;
+        }
+        this.startingJobs++;
+        await this.startJob(req);
+    }
+    /** Queue a run that can't start yet, persist it, and notify the client. */
+    async enqueueJob(req) {
+        const jobId = req.job_id ?? randomUUID();
+        req.job_id = jobId;
+        const position = this.jobQueue.enqueue(req);
+        log.info("Job queued", { jobId, position });
+        // Persist the queued run so it shows in jobs.list (Queue panel, reload,
+        // other tabs). Best-effort, mirroring startJobInner's persistence. It flips
+        // to "running" in startJobInner when a slot frees.
+        if (resolveRunJobExecutionOptions(req.execution_options, req.require_terminal_result === true).persistence === "job") {
+            try {
+                const existing = await Job.get(jobId);
+                if (!existing) {
+                    await Job.create({
+                        id: jobId,
+                        workflow_id: req.workflow_id ?? "",
+                        user_id: resolveRunJobUserId(req.user_id, this.session.requireUserId()),
+                        status: "queued",
+                        name: req.job_name ?? "",
+                        params: req.params ?? {},
+                        graph: req.graph ?? { nodes: [], edges: [] }
+                    });
+                }
+            }
+            catch (err) {
+                this.session.logError("enqueue persistence failed", err);
+            }
+        }
+        this.session.sendDetached({
+            type: "job_update",
+            status: "queued",
+            job_id: jobId,
+            workflow_id: req.workflow_id ?? null,
+            queue_position: position,
+            message: `Queued (#${position})`
+        });
+    }
+    /**
+     * Start the next queued run (if any) after a job slot frees up, and refresh
+     * the reported positions of the runs still waiting.
+     */
+    drainQueue() {
+        void (async () => {
+            const max = await this.deps
+                .getMaxConcurrentJobs()
+                .catch(() => this.deps.defaultMaxConcurrentJobs);
+            const perWorkflowMax = await this.deps
+                .getMaxConcurrentRunsPerWorkflow()
+                .catch(() => this.deps.defaultMaxConcurrentRunsPerWorkflow);
+            // Fill free slots with the first queued run whose workflow is still under
+            // its per-workflow limit (1 for normal runs, perWorkflowMax for opted-in
+            // concurrent runs). startJob registers the job before it returns, so the
+            // next iteration sees it as in-flight.
+            while (this.inFlightJobCount < max) {
+                const candidate = this.jobQueue
+                    .positions()
+                    .find((p) => this.countActiveJobsForWorkflow(p.workflowId) <
+                    (p.concurrent ? perWorkflowMax : 1));
+                if (!candidate) {
+                    break;
+                }
+                const next = this.jobQueue.remove(candidate.jobId);
+                if (!next) {
+                    break;
+                }
+                // Reserve the slot synchronously, mirroring runJob, so a concurrent
+                // run_job/drain can't also claim it before startJob registers.
+                this.startingJobs++;
+                const nextId = next.job_id;
+                if (nextId) {
+                    this.dequeuedJobs.add(nextId);
+                }
+                try {
+                    await this.startJob(next);
+                }
+                catch (err) {
+                    // The dequeued job threw before it could register/stream. Don't
+                    // silently lose it: tell the client this run failed, then keep
+                    // draining so the rest of the queue still progresses.
+                    this.session.logError("startJob (from queue) failed", err);
+                    await this.session.send({
+                        type: "job_update",
+                        status: "failed",
+                        job_id: next.job_id ?? null,
+                        workflow_id: next.workflow_id ?? null,
+                        error: formatSanitizedError(err)
+                    });
+                }
+                finally {
+                    if (nextId) {
+                        this.dequeuedJobs.delete(nextId);
+                    }
+                }
+            }
+            this.broadcastQueuePositions();
+        })();
+    }
+    /** Push updated queue positions to every still-waiting run. */
+    broadcastQueuePositions() {
+        for (const { jobId, workflowId, position } of this.jobQueue.positions()) {
+            this.session.sendDetached({
+                type: "job_update",
+                status: "queued",
+                job_id: jobId,
+                workflow_id: workflowId,
+                queue_position: position,
+                message: `Queued (#${position})`
+            });
+        }
+    }
+    async startJob(req) {
+        // The caller (runJob/drainQueue) reserved a concurrency slot via
+        // startingJobs++. Release it exactly once here: the slot is handed off to
+        // activeJobs on successful registration, or freed on early return/throw.
+        let slotReleased = false;
+        const releaseSlot = () => {
+            if (!slotReleased) {
+                slotReleased = true;
+                this.startingJobs = Math.max(0, this.startingJobs - 1);
+            }
+        };
+        try {
+            await this.startJobInner(req, releaseSlot);
+        }
+        finally {
+            // Safety net: if startJobInner returned/threw without registering, the
+            // slot is freed so it doesn't leak and permanently shrink the cap.
+            releaseSlot();
+        }
+    }
+    async startJobInner(req, releaseSlot) {
+        const userId = resolveRunJobUserId(req.user_id, this.session.requireUserId());
+        const workflowId = req.workflow_id ?? null;
+        const jobId = req.job_id ?? randomUUID();
+        let projectId = req.project_id ?? null;
+        // Inline graphs do not require a database lookup. This matters for
+        // hermetic callers that provide a workflow_id only as a correlation id.
+        // Saved-workflow requests still resolve the persisted project owner here.
+        if (workflowId && !req.graph) {
+            const workflow = await Workflow.find(userId, workflowId);
+            if (workflow) {
+                if (projectId !== null && projectId !== workflow.project_id) {
+                    await this.emitBeforeRunFailure(jobId, workflowId, new Error("Workflow is owned by another project"), true);
+                    return;
+                }
+                projectId = workflow.project_id;
+            }
+        }
+        const executionOptions = resolveRunJobExecutionOptions(req.execution_options, req.require_terminal_result === true);
+        const acceptedAt = req._accepted_at_ms ?? performance.now();
+        const preparationStartedAt = performance.now();
+        let phaseStartedAt = preparationStartedAt;
+        const rawGraph = await this.getRawGraph(req);
+        const graphLoadedMs = performance.now() - phaseStartedAt;
+        // Hydrate the graph (resolves node types from the registry)
+        phaseStartedAt = performance.now();
+        const graph = await this.hydrateGraph(rawGraph);
+        const graphHydratedMs = performance.now() - phaseStartedAt;
+        // The kernel keys terminal outputs by node.name. For SDK runs, align output
+        // node names with their public interface names before execution so the
+        // authoritative terminal snapshot addresses the same pins as output_update.
+        if (req.require_terminal_result) {
+            for (const node of graph.nodes) {
+                if (!node.type.startsWith("nodetool.output."))
+                    continue;
+                const properties = node.properties;
+                const publicName = properties?.name;
+                if (isString(publicName) && publicName.trim().length > 0) {
+                    node.name = publicName;
+                }
+            }
+        }
+        phaseStartedAt = performance.now();
+        if (this.deps.beforeRunJob) {
+            try {
+                await this.deps.beforeRunJob(graph);
+            }
+            catch (err) {
+                await this.emitBeforeRunFailure(jobId, workflowId, err, executionOptions.persistence === "job");
+                return;
+            }
+        }
+        const preRunMs = performance.now() - phaseStartedAt;
+        const workspace = this.session.workspaceResolver
+            ? await this.session.workspaceResolver(workflowId ?? null, userId)
+            : null;
+        const context = createRuntimeContext({
+            jobId,
+            workflowId,
+            projectId,
+            userId,
+            workspace,
+            assetOutputMode: this.session.mode === "text" ? "data_uri" : "temp_url",
+            persistOutputAssets: executionOptions.assetPersistence === "auto"
+        });
+        // Expose executor/node-type resolution on the context so that
+        // sub-workflow nodes (WorkflowNode) can create child runners.
+        context.setResolveExecutor((node) => this.session.resolveExecutor(node));
+        if (this.session.resolveNodeType) {
+            const resolverObj = isFunctionValue(this.session.resolveNodeType)
+                ? { resolveNodeType: this.session.resolveNodeType }
+                : this.session.resolveNodeType;
+            context.setResolveNodeType((nodeType) => resolverObj.resolveNodeType(nodeType));
+        }
+        // Persistence runs BEFORE the session exists, because
+        // `ExecutionSession.create()` starts the kernel: a job cancelled while it
+        // sat in the queue must never execute, and the only way to guarantee that
+        // is to check before anything can run. (The DB-only cancel path — tRPC
+        // `jobs.cancel` — doesn't remove the job from `jobQueue`, so drainQueue
+        // can still hand us a cancelled job.)
+        phaseStartedAt = performance.now();
+        // Which machine holds this run's session, for owner-aware reconnects and
+        // cross-instance cancel. Null on a single-machine deployment.
+        const instanceId = getInstanceId();
+        if (executionOptions.persistence === "job") {
+            try {
+                const existing = await Job.get(jobId);
+                if (existing) {
+                    if (existing.status === "cancelled") {
+                        log.info("Skipping start of cancelled job", { jobId });
+                        // Nothing was registered in activeJobs yet — free the reserved
+                        // slot and promote any queued run, matching every other slot
+                        // release (streamJobMessages finally, the chat-run finally).
+                        // Without it a cancelled-while-queued job leaves its slot idle
+                        // and the next queued run stalls.
+                        releaseSlot();
+                        this.drainQueue();
+                        this.session.sendDetached({
+                            type: "job_update",
+                            status: "cancelled",
+                            job_id: jobId,
+                            workflow_id: workflowId
+                        });
+                        return;
+                    }
+                    // Was persisted as "queued" while waiting for a slot — flip it to
+                    // running now that it's actually starting. The stamp goes on here
+                    // too: the queued row may have been written by another instance.
+                    if (existing.status !== "running" ||
+                        existing.runner_instance !== instanceId) {
+                        existing.markRunning();
+                        existing.runner_instance = instanceId;
+                        await existing.save();
+                    }
+                }
+                else {
+                    await Job.create({
+                        id: jobId,
+                        workflow_id: workflowId ?? "",
+                        user_id: userId,
+                        status: "running",
+                        name: req.job_name ?? "",
+                        started_at: new Date().toISOString(),
+                        params: req.params ?? {},
+                        graph,
+                        project_id: projectId ?? "default",
+                        runner_instance: instanceId
+                    });
+                }
+            }
+            catch (error) {
+                this.session.logError("runJob persistence failed", error);
+                // Persistence is best-effort in TS runtime mode.
+            }
+        }
+        const persistenceMs = performance.now() - phaseStartedAt;
+        // A5 (docs/RELIABILITY_TASKS.md Track A): the facade replaces the direct
+        // `new WorkflowRunner` + later `runner.run()` pair — `graph` is already
+        // hydrated/output-name-rewritten above, so this call re-hydrates it
+        // (idempotent: `withExplicitNodeFlags`, used because this class has no
+        // `NodeRegistry` of its own, passes every already-resolved field through
+        // unchanged via `...node`) and starts the run immediately. `resolveExecutor`
+        // is passed through as-is (not rebuilt from a registry+bridge) because
+        // this class only ever holds the bootstrap-injected closure, never a
+        // `NodeRegistry` instance.
+        // Opt-in per request; a run that asked for no supervisor gets none, and a
+        // supervisor that cannot be built leaves the run unsupervised rather than
+        // failing it.
+        const supervisor = await createRunSupervisor({
+            supervise: req.supervise,
+            supervisor: req.supervisor,
+            context,
+            defaultProvider: this.deps.defaults.provider,
+            defaultModel: this.deps.defaults.model
+        });
+        const sessionOptions = {
+            graph: toRawGraphInput(graph),
+            resolveExecutor: (node) => this.session.resolveExecutor(node),
+            bridgeFactory: async () => null,
+            // This runner owns a long-lived shared bridge, so `bridgeFactory` hands
+            // the session nothing to close. The run boundary still has to reach that
+            // bridge — pass it explicitly.
+            jobLifecycleBridge: this.session.pythonBridge ?? null,
+            jobId,
+            workflowId,
+            context,
+            params: req.params ?? {},
+            projectId,
+            validateNode: this.session.validateNode
+        };
+        if (supervisor) {
+            sessionOptions.supervisor = supervisor;
+        }
+        // A graph this runtime cannot honour (unknown model, unregistered
+        // provider, missing credential) is refused before the kernel starts.
+        // Route it through the same terminal `job_update` a failed pre-run hook
+        // uses: a bare throw reaches handleCommand as a generic `invalid_command`
+        // the UI never associates with the job, so the run appears to spin
+        // forever instead of failing with the reason.
+        let session;
+        try {
+            session = await ExecutionSession.create(sessionOptions);
+        }
+        catch (err) {
+            if (!isExecutionPreflightError(err))
+                throw err;
+            await this.emitBeforeRunFailure(jobId, workflowId, err, executionOptions.persistence === "job");
+            return;
+        }
+        const active = {
+            jobId,
+            workflowId,
+            context,
+            session,
+            graph,
+            finished: false,
+            status: "running",
+            requireTerminalResult: req.require_terminal_result === true,
+            executionOptions,
+            timings: {
+                acceptedAt,
+                queueMs: Math.max(0, preparationStartedAt - acceptedAt),
+                graphLoadedMs,
+                graphHydratedMs,
+                preRunMs,
+                persistenceMs,
+                kernelStartedAt: performance.now()
+            },
+            applicationId: req.application_id ?? null
+        };
+        // Decouple the run from this socket: from here on every frame carrying
+        // this job_id is stamped with `job_seq` and buffered, so a client that
+        // drops mid-run can `reconnect_job` from a fresh connection and replay
+        // the tail — including the terminal job_update.
+        //
+        // Keyed on the connection's identity, not `userId`: every lookup
+        // (`reconnect_job`, `cancel_job`, the slot counts) reads
+        // `this.session.requireUserId()`, and a run opened under an explicit differing
+        // `req.user_id` would be unreachable — no replay, and a cancel that
+        // reports the job as not found.
+        const runSession = jobRunRegistry.open(this.session.requireUserId(), jobId, workflowId, {
+            cancel: () => session.cancel(),
+            pushInput: (input, value, handle) => session.pushInput(input, value, handle),
+            finishInputStream: (input, handle) => session.finishInputStream(input, handle),
+            updateNodeProperties: (nodeId, properties) => session.updateNodeProperties(nodeId, properties)
+        });
+        runSession.attach(this.jobDeliveryTarget, runSession.lastSeq);
+        active.runSession = runSession;
+        this.activeJobs.set(jobId, active);
+        // Slot ownership transfers from startingJobs to the registry's running
+        // count now that the job is registered. Released before the DB check
+        // below so the reservation and the registry entry never both count.
+        releaseSlot();
+        log.info("Job started", { jobId, workflowId });
+        await this.settleRunAgainstLostConnection(runSession, jobId, executionOptions);
+        // The run itself already started inside `ExecutionSession.create()`
+        // above — `session.result` is the same never-rejecting terminal-result
+        // promise `runner.run()` used to return.
+        const executePromise = session.result;
+        // `streamJobMessages` handles its own failures, but nothing awaits this
+        // promise — attach a terminal handler so a bug there can never surface as
+        // an unhandled rejection (which Node 22 turns into a process exit).
+        active.streamTask = this.streamJobMessages(active, executePromise).catch((error) => {
+            this.session.logError("job stream task failed", error);
+        });
+    }
+    /**
+     * A run whose start was mid-flight when `disconnect()` fired registers
+     * AFTER that walk of `activeJobs`, so nothing detached it and nothing
+     * armed its grace timer — it would execute unattended forever, delivering
+     * into a dead target. Two fixes, both after registration:
+     *   - no live socket: detach, which starts the detach-grace countdown;
+     *   - the row already reads cancelled: `disconnect()`'s queue drain (or a
+     *     DB-only cancel) settled it while this run was starting, so cancel
+     *     the run rather than let it execute to completion under a row that
+     *     says it never did.
+     */
+    async settleRunAgainstLostConnection(runSession, jobId, executionOptions) {
+        const socketGone = !this.deps.isSocketConnected();
+        if (socketGone) {
+            runSession.detach(this.jobDeliveryTarget);
+        }
+        if (executionOptions.persistence !== "job")
+            return;
+        try {
+            const job = await Job.get(jobId);
+            if (job?.status === "cancelled") {
+                log.info("Job was cancelled while starting, cancelling the run", {
+                    jobId
+                });
+                runSession.cancel();
+            }
+        }
+        catch (error) {
+            this.session.logError("post-start cancellation check failed", error);
+        }
+    }
+    async streamJobMessages(active, executePromise) {
+        // The drain loop has awaited operations (Asset.paginate, normalizeOutputValue,
+        // sendMessage) that can throw. Guarantee the job slot is released and the
+        // queue drains no matter what — otherwise one throw permanently leaks a
+        // MAX_CONCURRENT_JOBS slot and stalls every queued run.
+        try {
+            await this._streamJobMessagesInner(active, executePromise);
+        }
+        catch (error) {
+            // Without this catch the rejection escaped entirely: the caller only
+            // assigns `active.streamTask` and never awaits it, so Node's default
+            // unhandledRejection behaviour terminated the process — and none of the
+            // terminal bookkeeping below ran, leaving the DB row stuck at "running",
+            // the client UI spinning, and the app ledger holding the estimate.
+            const message = error instanceof Error ? error.message : String(error);
+            this.session.logError("job message streaming failed", error);
+            active.finished = true;
+            active.status = "failed";
+            active.error = message;
+            try {
+                await this.session.send({
+                    type: "job_update",
+                    status: "failed",
+                    job_id: active.jobId,
+                    workflow_id: active.workflowId,
+                    error: message
+                });
+            }
+            catch (sendError) {
+                this.session.logError("terminal job_update send failed", sendError);
+            }
+            await this.persistTerminalJobStatus(active);
+            await this.settleApplicationInvocation(active);
+            releaseSpend(this.session.requireUserId(), active.jobId);
+        }
+        finally {
+            // Terminal: the session stops buffering and starts its retention
+            // window, so a client reconnecting shortly after still gets the tail
+            // (and the outcome) before the persisted row becomes the only source.
+            active.runSession?.finish(active.status);
+            this.activeJobs.delete(active.jobId);
+            this.drainQueue();
+        }
+    }
+    /**
+     * Write the run's terminal status onto the persisted Job row. Skipped for
+     * explicitly session-scoped runs, which own no row. Never throws —
+     * persistence is best-effort and must not mask the run's own outcome.
+     */
+    async persistTerminalJobStatus(active, outputs) {
+        if ((active.executionOptions?.persistence ??
+            DEFAULT_RUN_JOB_EXECUTION_OPTIONS.persistence) !== "job") {
+            return;
+        }
+        try {
+            const job = (await Job.get(active.jobId));
+            // A DB-only cancel (tRPC `jobs.cancel`) can finalize the row as cancelled
+            // while the job is still executing in memory. Don't overwrite that with a
+            // completed/failed status when the in-flight run finishes.
+            if (job) {
+                if (job.status !== "cancelled") {
+                    if (active.status === "completed") {
+                        job.markCompleted();
+                        if (outputs !== undefined) {
+                            job.metadata_json = {
+                                ...(job.metadata_json ?? {}),
+                                outputs: persistableOutputs(withoutInlineAssetBytes(outputs))
+                            };
+                        }
+                    }
+                    else if (active.status === "failed") {
+                        job.markFailed(active.error ?? "Unknown error");
+                    }
+                    else if (active.status === "cancelled") {
+                        job.markCancelled();
+                    }
+                }
+                job.cost = this.runMeasuredCost(active);
+                await job.save();
+            }
+        }
+        catch (error) {
+            this.session.logError("job persistence (final status) failed", error);
+        }
+    }
+    /**
+     * Close the app's ledger row at what the run actually cost. Until this lands
+     * the run keeps counting against the budget at its estimate, which is the
+     * conservative direction: a crash cannot free spend it may have incurred.
+     * The total is what anything measured: charges a node reported for itself
+     * plus generation the cost ledger priced. An absent total still means
+     * "nothing measured this run", not "this run was free" — passing null keeps
+     * the estimate standing rather than handing the spend back. Never throws.
+     */
+    async settleApplicationInvocation(active) {
+        if (!active.applicationId)
+            return;
+        try {
+            await settleInvocation(active.applicationId, active.jobId, this.runMeasuredCost(active), active.status === "failed"
+                ? "failed"
+                : active.status === "cancelled"
+                    ? "cancelled"
+                    : "completed");
+        }
+        catch (error) {
+            this.session.logError("application invocation settlement failed", error);
+        }
+    }
+    async _streamJobMessagesInner(active, executePromise) {
+        let terminalSeen = false;
+        let terminalWithResultSeen = false;
+        let outputUpdateSeen = false;
+        let finalOutputs = {};
+        // Per-node arrival counter for `generation_complete.index` within this job.
+        // The function is scoped to one job, so keying by node_id alone yields a
+        // per-(job_id, node_id) monotonic index. DB-ordering reconciliation is a
+        // later step (RFC Decision 8); this is the in-memory arrival order.
+        const generationIndexByNode = new Map();
+        // Arrival positions already autosaved in THIS run, keyed `${nodeId} ${index}`.
+        // A single generation_complete can persist several assets (a `list[image]`
+        // output, or media + text), so dedupe by the event's arrival index — NOT by
+        // a total asset count, which would under-save the next event (RFC D8).
+        const autosavedSlots = new Set();
+        // Cross-run replay dedupe: the `generation_index` values already persisted
+        // for a node by a PRIOR run. Warmed with ONE `Asset.paginate` on a node's
+        // first generation_complete, then reused for every later variant — so an
+        // N-variant run does one query per node, not one per variant (RFC D8).
+        const persistedIndexByNode = new Map();
+        // The kernel opens every run with the same running update. Authoritative
+        // SDK runs relay that one and avoid an otherwise duplicate WebSocket
+        // frame. Legacy clients keep the eager acknowledgement.
+        let runningSeen = false;
+        if (!active.requireTerminalResult) {
+            await this.session.send({
+                type: "job_update",
+                status: "running",
+                job_id: active.jobId,
+                workflow_id: active.workflowId
+            });
+            runningSeen = true;
+        }
+        // Guard the framing contract for authoritative runs: a terminal update
+        // must never be the first job_update a client sees. The kernel emits
+        // running before it can fail, but a run that dies before the kernel
+        // starts — or a future reordering there — must not silently drop the
+        // acknowledgement.
+        const ensureRunningFrame = async () => {
+            if (runningSeen)
+                return;
+            runningSeen = true;
+            await this.session.send({
+                type: "job_update",
+                status: "running",
+                job_id: active.jobId,
+                workflow_id: active.workflowId
+            });
+        };
+        const executionSettled = executePromise
+            .then((result) => {
+            active.status = result.status;
+            active.error = result.error;
+            finalOutputs = result.outputs ?? {};
+        })
+            .catch((err) => {
+            this.session.logError("job execution failed", err);
+            active.status = "failed";
+            active.error = formatSanitizedError(err);
+        })
+            .finally(() => {
+            active.finished = true;
+        });
+        const waitForActivity = createRelayActivityWaiter(active.context, executionSettled);
+        const graphNodes = active.graph.nodes ?? [];
+        const graphNodeMap = new Map();
+        for (const n of graphNodes) {
+            if (isString(n.id)) {
+                graphNodeMap.set(n.id, n);
+            }
+        }
+        while (!active.finished || active.context.hasMessages()) {
+            while (active.context.hasMessages()) {
+                const msg = active.context.popMessage();
+                if (!msg)
+                    break;
+                const outbound = { ...msg };
+                outbound.job_id ??= active.jobId;
+                outbound.workflow_id ??= active.workflowId;
+                // Leave a nullish error untouched (the kernel stamps `error: null` on
+                // every update) — only sanitize a real error value. Formatting null here
+                // would ship the literal string "null" to clients.
+                if (outbound.error != null) {
+                    outbound.error = formatSanitizedError(outbound.error);
+                }
+                if (outbound.type === "notification" && isString(outbound.content)) {
+                    outbound.content = sanitizeLargeText(outbound.content);
+                }
+                // Every message, not just node updates: a `prediction` is where
+                // ledger-priced generation spend (Replicate, Gemini, OpenAI, …)
+                // reports itself, and this is the path whose terminal status and app
+                // settlement read the run's measured cost.
+                this._handleNodeProviderCost(active, outbound);
+                if (outbound.type === "node_update" && outbound.status === "error") {
+                    log.error("Node error", {
+                        jobId: active.jobId,
+                        nodeId: outbound.node_id,
+                        error: outbound.error
+                    });
+                }
+                else if (outbound.type === "job_update" &&
+                    outbound.status === "failed") {
+                    log.error("Job failed", {
+                        jobId: active.jobId,
+                        error: outbound.error
+                    });
+                }
+                // Skip messages for constant/input nodes — they produce trivial
+                // outputs that don't need to be relayed to the frontend.
+                if (outbound.type === "output_update" ||
+                    outbound.type === "node_update" ||
+                    outbound.type === "generation_complete") {
+                    const nodeId = String(outbound.node_id ?? "");
+                    const node = graphNodeMap.get(nodeId);
+                    const nodeType = isString(node?.type) ? node.type : "";
+                    // Skip constant and input nodes entirely
+                    if (nodeType.startsWith("nodetool.constant.") ||
+                        nodeType.startsWith("nodetool.input.")) {
+                        continue;
+                    }
+                    const meta = this.session.getNodeMetadata?.(nodeType);
+                    // Stamp an arrival-order `index` on generation_complete, keyed per
+                    // (job_id, node_id) (the function is job-scoped, so node_id alone
+                    // suffices). job_id/workflow_id were already backfilled by the
+                    // outbound spread above.
+                    if (outbound.type === "generation_complete") {
+                        const arrivalIndex = generationIndexByNode.get(nodeId) ?? 0;
+                        outbound.index = arrivalIndex;
+                        generationIndexByNode.set(nodeId, arrivalIndex + 1);
+                        // Autosave one generation per generation_complete on the RAW outputs
+                        // (before the normalize at the bottom of this block strips inline
+                        // bytes), tagged { jobId, nodeId, index }. This is the autosave
+                        // cutover (RFC §7, D3): persistence is driven per generation event,
+                        // not by the terminal node_update{completed} — so an N-execution
+                        // run persists N distinct generations.
+                        //
+                        // Replay dedupe (D8) is keyed on the event's arrival `index`, NOT on
+                        // a total asset count: a single generation_complete can persist
+                        // several assets (a `list[image]` output, or media + a text asset),
+                        // so a count-vs-index gate would under-save the very first run. Two
+                        // guards, both keyed by (nodeId, index):
+                        //   - in-run: skip if this arrival slot was already saved this run;
+                        //   - cross-run: skip if an asset for (jobId, nodeId) already carries
+                        //     metadata.generation_index === arrivalIndex (a reconnect replay
+                        //     re-streams the same events with arrivalIndex back at 0..N-1).
+                        // Server-only (D9): this is the websocket runner; the browser never
+                        // reaches runJob, so no browser autosave is introduced here.
+                        if (active.executionOptions.assetPersistence === "auto" &&
+                            meta?.auto_save_asset &&
+                            outbound.outputs != null) {
+                            const userId = this.session.requireUserId();
+                            const slotKey = `${nodeId} ${arrivalIndex}`;
+                            // Warm the cross-run replay set once per node (on its first
+                            // generation_complete), then reconcile every later variant
+                            // against the in-memory set — one DB read per node, not per slot.
+                            let persistedIndices = persistedIndexByNode.get(nodeId);
+                            if (persistedIndices === undefined) {
+                                persistedIndices = new Set();
+                                // Best-effort like every other persistence on this path: a
+                                // DB-free run (session persistence in the reliability harness,
+                                // a misconfigured deployment) must degrade to skipping the
+                                // replay dedupe, not kill the drain loop and fail the job.
+                                try {
+                                    const [persisted] = await Asset.paginate(userId, {
+                                        jobId: active.jobId,
+                                        nodeId,
+                                        limit: 1000
+                                    });
+                                    for (const a of persisted) {
+                                        const gi = a.metadata?.generation_index;
+                                        if (isNumber(gi))
+                                            persistedIndices.add(gi);
+                                    }
+                                }
+                                catch (err) {
+                                    log.warn("generation replay-dedupe read failed", {
+                                        nodeId,
+                                        error: err instanceof Error ? err.message : String(err)
+                                    });
+                                }
+                                persistedIndexByNode.set(nodeId, persistedIndices);
+                            }
+                            if (!autosavedSlots.has(slotKey) &&
+                                !persistedIndices.has(arrivalIndex)) {
+                                autosavedSlots.add(slotKey);
+                                try {
+                                    await autoSaveAssets(outbound.outputs, {
+                                        userId,
+                                        workflowId: active.workflowId,
+                                        jobId: active.jobId,
+                                        nodeId,
+                                        textOutputName: primaryTextOutputName(meta),
+                                        generationIndex: arrivalIndex,
+                                        properties: outbound.properties ?? undefined,
+                                        nodeType: isString(outbound.node_type)
+                                            ? outbound.node_type
+                                            : undefined
+                                    });
+                                }
+                                catch (err) {
+                                    log.warn("autoSaveAssets error", { error: String(err) });
+                                }
+                            }
+                        }
+                    }
+                    // Relay output_update for display-sink nodes (Output, Preview) and
+                    // for streaming or auto-saving generative nodes (FAL / Replicate /
+                    // Kie / …) so the client receives one event per yielded item — the
+                    // UI accumulates and renders each generation as it arrives. The
+                    // Preview node re-emits each chunk it receives on its own terminal
+                    // `output` handle; relaying those is what lets the preview stream
+                    // incrementally instead of collapsing to the final value.
+                    if (outbound.type === "output_update") {
+                        const isDisplaySink = nodeType.includes("Output") || nodeType.endsWith(".Preview");
+                        const isStreamingLeaf = Boolean(meta?.is_streaming_output) ||
+                            Boolean(meta?.auto_save_asset);
+                        if (!isDisplaySink && !isStreamingLeaf)
+                            continue;
+                        outputUpdateSeen = true;
+                    }
+                    const isNodeError = outbound.type === "node_update" && outbound.status === "error";
+                    if (!isNodeError &&
+                        (active.executionOptions.eventDetail === "terminal" ||
+                            (active.executionOptions.eventDetail === "outputs" &&
+                                (outbound.type === "node_update" ||
+                                    outbound.type === "generation_complete")))) {
+                        continue;
+                    }
+                    // Materialize binary assets to temp URLs before sending over WebSocket
+                    if (outbound.type === "node_update" && outbound.result != null) {
+                        outbound.result = await active.context.normalizeOutputValue(outbound.result);
+                    }
+                    if (outbound.type === "output_update" && outbound.value != null) {
+                        outbound.value = await active.context.normalizeOutputValue(outbound.value);
+                    }
+                    // Normalize generation_complete.outputs the same way node_update.result
+                    // is treated (raw bytes → temp URLs) before sending over the wire.
+                    if (outbound.type === "generation_complete" &&
+                        outbound.outputs != null) {
+                        outbound.outputs = await active.context.normalizeOutputValue(outbound.outputs);
+                    }
+                }
+                if (outbound.type === "edge_update" &&
+                    active.executionOptions.eventDetail !== "full") {
+                    continue;
+                }
+                const status = outbound.type === "job_update" ? String(outbound.status ?? "") : "";
+                const suppressProvisionalCompletion = active.requireTerminalResult &&
+                    status === "completed" &&
+                    outbound.result === undefined;
+                if (!suppressProvisionalCompletion) {
+                    if (outbound.type === "job_update") {
+                        if (status === "running") {
+                            runningSeen = true;
+                        }
+                        else if (TERMINAL_JOB_STATUSES.includes(status)) {
+                            await ensureRunningFrame();
+                        }
+                    }
+                    await this.session.send(outbound);
+                }
+                if (outbound.type === "job_update" && !suppressProvisionalCompletion) {
+                    if (TERMINAL_JOB_STATUSES.includes(status)) {
+                        terminalSeen = true;
+                        if (outbound.result !== undefined) {
+                            terminalWithResultSeen = true;
+                        }
+                    }
+                }
+            }
+            if (!active.finished) {
+                await waitForActivity();
+            }
+        }
+        // The authoritative terminal snapshot is consumed in every event-detail
+        // mode. Keep it just as client-safe as streamed output_update values;
+        // otherwise Outputs/Full can replace a working temp URL with raw media.
+        if (Object.keys(finalOutputs).length > 0) {
+            finalOutputs = await this.normalizeFinalOutputs(active, finalOutputs);
+        }
+        if (active.executionOptions.eventDetail !== "terminal" &&
+            !outputUpdateSeen &&
+            Object.keys(finalOutputs).length > 0) {
+            await this.sendOutputUpdates(active, finalOutputs);
+        }
+        const relayCompletedAt = performance.now();
+        if (!terminalSeen ||
+            (!terminalWithResultSeen && Object.keys(finalOutputs).length > 0)) {
+            await ensureRunningFrame();
+            await this.session.send({
+                type: "job_update",
+                status: active.status,
+                job_id: active.jobId,
+                workflow_id: active.workflowId,
+                error: active.error,
+                result: { outputs: finalOutputs }
+            });
+        }
+        const terminalDeliveredAt = performance.now();
+        log.info("Job completed", {
+            jobId: active.jobId,
+            status: active.status,
+            executionOptions: active.executionOptions,
+            timings: {
+                queueMs: active.timings.queueMs,
+                graphLoadedMs: active.timings.graphLoadedMs,
+                graphHydratedMs: active.timings.graphHydratedMs,
+                preRunMs: active.timings.preRunMs,
+                persistenceMs: active.timings.persistenceMs,
+                executionAndRelayMs: Math.max(0, relayCompletedAt - active.timings.kernelStartedAt),
+                terminalDeliveryMs: Math.max(0, terminalDeliveredAt - relayCompletedAt),
+                totalMs: Math.max(0, terminalDeliveredAt - active.timings.acceptedAt)
+            }
+        });
+        await this.persistTerminalJobStatus(active, finalOutputs);
+        await this.settleApplicationInvocation(active);
+        releaseSpend(this.session.requireUserId(), active.jobId);
+        // Slot release + queue drain happen in the streamJobMessages wrapper's
+        // finally, so they run even if the drain loop above throws.
+    }
+    async reconnectJob(jobId, workflowId, lastSeq = 0) {
+        // A resilient session is the authoritative answer: the run may be
+        // executing on another connection's runner right now, or have finished
+        // while this client was away — either way the seq-stamped buffer holds
+        // exactly what was missed. Adopt it so this connection's `cancel_job` /
+        // `stream_input` / `stop` reach the runner that owns the ExecutionSession.
+        const registered = jobRunRegistry.get(this.session.requireUserId(), jobId);
+        if (registered) {
+            const { replay, incomplete } = registered.attach(this.jobDeliveryTarget, lastSeq);
+            if (registered.status === "running") {
+                this.adoptedJobIds.add(jobId);
+            }
+            // Header first, then the missed tail; live frames queue behind them on
+            // the session's ordered delivery chain.
+            await registered.deliverReplay(this.jobDeliveryTarget, [
+                {
+                    type: "job_resumed",
+                    job_id: jobId,
+                    workflow_id: workflowId ?? registered.workflowId ?? null,
+                    status: registered.status,
+                    last_seq: registered.lastSeq,
+                    replay_count: replay.length,
+                    replay_incomplete: incomplete
+                },
+                ...replay
+            ]);
+            return;
+        }
+        const active = this.activeJobs.get(jobId);
+        if (!active) {
+            // No session and no in-memory job: the run ended long enough ago that
+            // retention elapsed, or this process never had it. A row that already
+            // reached a settled outcome is echoed verbatim — a completed run stays
+            // completed, and the replay-unavailable note rides alongside as an
+            // `error` string explaining only the missing events.
+            //
+            // Every other row status (queued, scheduled, running) is reported as
+            // failed: nothing is left that could
+            // ever send this client another frame, and reporting the row as-is
+            // parks the UI in a state that never settles — a `queued` row from a
+            // dead connection's drained queue reads as "running" with a live Stop
+            // button forever.
+            // Ownership rule as in the jobs router: another user's row is
+            // indistinguishable from a missing one — it must be neither reported
+            // nor settled below.
+            const row = (await Job.get(jobId));
+            const job = row && row.user_id === (this.session.requireUserId()) ? row : null;
+            const settled = job != null &&
+                (job.status === "completed" ||
+                    job.status === "failed" ||
+                    job.status === "cancelled");
+            const replayUnavailable = job != null && job.status !== "failed" && job.status !== "cancelled";
+            // A non-settled row with no session and no in-memory job is a zombie:
+            // nothing is left that could ever finish it. Persist the failure when
+            // this instance owns the row (or nothing claims it), so the row stops
+            // advertising an in-flight run — otherwise every reload rediscovers it,
+            // reattaches, and re-reports the same loss forever. A row claimed by
+            // another instance is left alone: on a multi-instance deployment this
+            // connection may simply have been balanced away from a run that is
+            // still executing.
+            if (job && !settled) {
+                const instanceId = getInstanceId();
+                const ownedHere = !job.runner_instance ||
+                    !instanceId ||
+                    job.runner_instance === instanceId;
+                if (ownedHere) {
+                    try {
+                        job.markFailed("Run was lost after the execution connection went away.");
+                        await job.save();
+                    }
+                    catch (error) {
+                        this.session.logError("stale job row cleanup failed", error);
+                    }
+                }
+            }
+            await this.session.send({
+                type: "job_update",
+                status: settled ? job.status : "failed",
+                job_id: jobId,
+                workflow_id: workflowId ?? job?.workflow_id ?? null,
+                ...(job
+                    ? replayUnavailable
+                        ? {
+                            error: "Job event replay is unavailable after the execution connection was lost."
+                        }
+                        : job.error
+                            ? { error: job.error }
+                            : {}
+                    : { error: `Job ${jobId} not found` })
+            });
+            return;
+        }
+        await this.session.send({
+            type: "job_update",
+            status: active.status,
+            job_id: jobId,
+            workflow_id: workflowId ?? active.workflowId
+        });
+        for (const status of Object.values(active.context.getNodeStatuses())) {
+            await this.session.send({
+                ...status,
+                job_id: jobId,
+                workflow_id: workflowId ?? active.workflowId
+            });
+        }
+        for (const status of Object.values(active.context.getEdgeStatuses())) {
+            await this.session.send({
+                ...status,
+                job_id: jobId,
+                workflow_id: workflowId ?? active.workflowId
+            });
+        }
+    }
+    async cancelJob(jobId, workflowId) {
+        if (!jobId) {
+            return { error: "No job_id provided" };
+        }
+        // A run that's still queued has no ActiveJob yet — drop it from the queue
+        // and tell the client it's cancelled before it ever starts.
+        const queued = this.jobQueue.remove(jobId);
+        if (queued) {
+            releaseSpend(this.session.requireUserId(), jobId);
+            const cancelledWorkflowId = queued.workflow_id ?? workflowId ?? null;
+            // Mark the persisted queued row cancelled so it leaves the queue in
+            // jobs.list too (not just the in-memory queue).
+            if (resolveRunJobExecutionOptions(queued.execution_options, queued.require_terminal_result === true).persistence === "job") {
+                try {
+                    const job = await Job.get(jobId);
+                    if (job) {
+                        job.markCancelled();
+                        await job.save();
+                    }
+                }
+                catch (err) {
+                    this.session.logError("cancel persistence failed", err);
+                }
+            }
+            await this.session.send({
+                type: "job_update",
+                status: "cancelled",
+                job_id: jobId,
+                workflow_id: cancelledWorkflowId
+            });
+            this.broadcastQueuePositions();
+            return {
+                message: "Queued job cancelled",
+                job_id: jobId,
+                workflow_id: cancelledWorkflowId
+            };
+        }
+        const active = this.activeJobs.get(jobId);
+        if (!active) {
+            // Not ours, but possibly still running on the connection that started
+            // it (this client reconnected after a drop). Cancel through the
+            // session's hooks and persist the row here — the owning runner's own
+            // terminal bookkeeping still runs, and its `job_update` reaches this
+            // client over the session it just adopted.
+            const registered = jobRunRegistry.get(this.session.requireUserId(), jobId);
+            if (registered && registered.status === "running") {
+                registered.cancel();
+                try {
+                    const job = await Job.get(jobId);
+                    if (job && job.status !== "cancelled") {
+                        job.markCancelled();
+                        await job.save();
+                    }
+                }
+                catch (err) {
+                    this.session.logError("cancel persistence failed", err);
+                }
+                return {
+                    message: "Job cancellation requested",
+                    job_id: jobId,
+                    workflow_id: workflowId ?? registered.workflowId ?? ""
+                };
+            }
+            // Nothing local holds it. With more than one instance the run may be
+            // executing on another machine: write the cancellation to its row, which
+            // the owning instance's poller picks up.
+            const remote = await requestRemoteJobCancel(this.session.requireUserId(), jobId);
+            if (remote.cancelled) {
+                return {
+                    message: "Job cancellation requested",
+                    job_id: jobId,
+                    workflow_id: workflowId ?? remote.workflowId ?? ""
+                };
+            }
+            return {
+                error: "Job not found or already completed",
+                job_id: jobId,
+                workflow_id: workflowId ?? ""
+            };
+        }
+        if (active.session) {
+            active.session.cancel();
+        }
+        active.status = "cancelled";
+        // Persist the cancellation to the DB right away, mirroring the queued
+        // branch and the tRPC jobs.cancel path. The runner's own cleanup can lag
+        // (it drains in-flight messages before its .finally() persists), so
+        // without this the persisted row stays "running" and jobs.list — which the
+        // Queue panel reads from — keeps reporting the job as running even though
+        // the toolbar Stop already fired.
+        //
+        // Deliberately NO eager `job_update cancelled` frame here: the kernel's
+        // own terminal frame relays through the drain loop AFTER the node-level
+        // terminal updates, and an out-of-band frame ahead of them tells the
+        // client the job is over while nodes still read "running" — the exact
+        // lifecycle violation the reliability harness's mid-run-cancel journeys
+        // pin (`lifecycle.running-after-job-terminal`), and what left canvas
+        // nodes stuck spinning after a Stop. The `cancel_job` RPC response is the
+        // immediate acknowledgement; the ordered terminal arrives a beat later.
+        if ((active.executionOptions?.persistence ??
+            DEFAULT_RUN_JOB_EXECUTION_OPTIONS.persistence) === "job") {
+            try {
+                const job = await Job.get(jobId);
+                if (job && job.status !== "cancelled") {
+                    job.markCancelled();
+                    await job.save();
+                }
+            }
+            catch (err) {
+                this.session.logError("cancel persistence failed", err);
+            }
+        }
+        const cancelledWorkflowId = workflowId ?? active.workflowId ?? null;
+        // Do NOT set active.finished = true here. Let the runner's cancellation
+        // propagate through executePromise's .finally() callback so that
+        // streamJobMessages can drain remaining messages and persist job state.
+        return {
+            message: "Job cancellation requested",
+            job_id: jobId,
+            workflow_id: cancelledWorkflowId ?? ""
+        };
+    }
+    /**
+     * `stop` for one run. The run is cancelled here when this connection owns
+     * it; otherwise it is executing on the connection that started it — this
+     * client only reconnected to it — so it is cancelled through its registered
+     * session, or, when nothing local holds it, through its row.
+     */
+    async stopJob(jobId) {
+        const active = this.activeJobs.get(jobId);
+        if (active) {
+            active.session.cancel();
+            active.status = "cancelled";
+            return;
+        }
+        const registered = jobRunRegistry.get(this.session.requireUserId(), jobId);
+        if (registered && registered.status === "running") {
+            registered.cancel();
+            return;
+        }
+        await requestRemoteJobCancel(this.session.requireUserId(), jobId);
+    }
+    getStatus(jobId) {
+        if (jobId) {
+            const active = this.activeJobs.get(jobId);
+            if (!active) {
+                return { status: "not_found", job_id: jobId };
+            }
+            return {
+                status: active.status,
+                job_id: active.jobId,
+                workflow_id: active.workflowId
+            };
+        }
+        return {
+            active_jobs: Array.from(this.activeJobs.values()).map((job) => ({
+                job_id: job.jobId,
+                workflow_id: job.workflowId,
+                status: job.status
+            }))
+        };
+    }
+    /**
+     * Accumulate provider cost from a completed node_update into the job total.
+     *
+     * The ledger row is *not* written here. `ExecutionSession` attaches
+     * `attachRunCostLedger` to the same context, so every surface — this server,
+     * the CLI, the debug harness — records the charge once, from one
+     * implementation. Writing it again here would double-count every FAL and kie
+     * generation.
+     */
+    _handleNodeProviderCost(active, outbound) {
+        if (outbound.type === "prediction") {
+            this._handlePredictionCost(active, outbound);
+            return;
+        }
+        if (outbound.type !== "node_update" ||
+            outbound.status !== "completed" ||
+            outbound.provider_cost == null) {
+            return;
+        }
+        const providerCost = outbound.provider_cost;
+        const amount = providerCost.amount;
+        if (isFiniteNumber(amount)) {
+            (active.selfReportedCostNodeIds ??= new Set()).add(String(outbound.node_id ?? ""));
+            active.providerCostTotal = (active.providerCostTotal ?? 0) + amount;
+        }
+        else {
+            // A non-finite amount (NaN/Infinity from a buggy provider call) can't
+            // be accumulated above, and JSON can't even represent it faithfully
+            // (`JSON.stringify(NaN)` silently becomes `null`). Rather than ship a
+            // `provider_cost` the wire contract calls a real number, drop it — the
+            // rest of the `node_update` still reports normally.
+            delete outbound.provider_cost;
+        }
+    }
+    /**
+     * Accumulate a completed unit-billed `prediction` into the run total, priced
+     * the way the cost ledger prices it. Replicate, Gemini, OpenAI, MiniMax and
+     * ElevenLabs generation reports no `provider_cost` of its own, so without
+     * this a run on those providers settled as "nothing measured" and stayed
+     * booked at its estimate.
+     */
+    _handlePredictionCost(active, outbound) {
+        if (outbound.status !== "completed")
+            return;
+        const capability = isString(outbound.capability)
+            ? outbound.capability
+            : null;
+        if (!isUnitBilledCapability(capability))
+            return;
+        const provider = isString(outbound.provider) ? outbound.provider : "";
+        const model = isString(outbound.model) ? outbound.model : "";
+        if (!provider || !model)
+            return;
+        const priced = priceGeneration({
+            // Pricing is a pure function of provider, model, quantity and params —
+            // `priceGeneration` never reads `userId`, so this must not demand one.
+            // `requireUserId()` here turned cost accounting into a throw on any
+            // session that has not connected. The guard stays on the paths that do
+            // key on identity: registry lookups and spend reservation.
+            userId: this.session.userId ?? "",
+            provider,
+            model,
+            capability,
+            quantity: 1,
+            params: isRecord(outbound.params) ? outbound.params : {}
+        });
+        if (!priced || !isFiniteNumber(priced.cost))
+            return;
+        const nodeId = String(outbound.node_id ?? "");
+        const byNode = (active.predictionCostByNode ??= new Map());
+        byNode.set(nodeId, (byNode.get(nodeId) ?? 0) + priced.cost);
+    }
+    /**
+     * What this run cost, as far as anything measured it: node-reported provider
+     * charges plus ledger-priced generation on nodes that reported none. Null
+     * when nothing measured — which keeps an app invocation standing at its
+     * estimate rather than handing the spend back.
+     */
+    runMeasuredCost(active) {
+        const selfReported = active.selfReportedCostNodeIds;
+        let total = active.providerCostTotal ?? 0;
+        let measured = active.providerCostTotal != null;
+        for (const [nodeId, cost] of active.predictionCostByNode ?? []) {
+            // The provider's own number wins over the catalog's estimate for the
+            // same node — counting both would double-bill a FAL or kie generation.
+            if (selfReported?.has(nodeId))
+                continue;
+            total += cost;
+            measured = true;
+        }
+        return measured && total > 0 ? total : null;
+    }
+    /**
+     * Release everything this connection still holds on the job side: detach the
+     * resilient runs (they keep executing and buffering), cancel the ones nobody
+     * is left to receive, and mark the queued rows cancelled. Called from the
+     * host's `disconnect`.
+     */
+    async cancelAll() {
+        // A run with a resilient session survives the socket: detach it (frames
+        // keep buffering for replay, and `streamJobMessages` keeps draining on
+        // this now-socketless runner) instead of cancelling. The session's
+        // detach-grace timer bounds how long it may run unattended. A run without
+        // one — a chat-triggered workflow, whose owning turn is cancelled anyway
+        // — is cancelled as before: nobody is left to receive its output.
+        for (const [jobId, job] of this.activeJobs) {
+            if (job.runSession) {
+                job.runSession.detach(this.jobDeliveryTarget);
+                continue;
+            }
+            job.session?.cancel();
+            this.activeJobs.delete(jobId);
+        }
+        for (const adoptedId of this.adoptedJobIds) {
+            jobRunRegistry
+                .get(this.session.requireUserId(), adoptedId)
+                ?.detach(this.jobDeliveryTarget);
+        }
+        this.adoptedJobIds.clear();
+        // Drain runs that were still queued (never started): the client is gone,
+        // so they will never run. Mark their persisted rows cancelled instead of
+        // leaving them as orphaned "scheduled" jobs in jobs.list.
+        for (let queued = this.jobQueue.dequeue(); queued; queued = this.jobQueue.dequeue()) {
+            const queuedId = queued.job_id;
+            if (!queuedId)
+                continue;
+            try {
+                const job = await Job.get(queuedId);
+                if (job) {
+                    job.markCancelled();
+                    await job.save();
+                }
+            }
+            catch (err) {
+                this.session.logError("disconnect queue cancellation failed", err);
+            }
+        }
+        for (const dequeuedId of this.dequeuedJobs) {
+            try {
+                const job = await Job.get(dequeuedId);
+                if (job) {
+                    job.markCancelled();
+                    await job.save();
+                }
+            }
+            catch (err) {
+                this.session.logError("disconnect dequeued-job cancellation failed", err);
+            }
+        }
+    }
+}
+//# sourceMappingURL=job-execution.js.map