@puku-ai/sdk 4.0.1 → 4.0.2

package/dist/vendor/lib/tools/SessionToolRunner.js

@@ -0,0 +1,737 @@
+import { PukuError } from '../../core/error.js';
+import { loggerFor } from '../../internal/utils/log.js';
+import { sleep } from '../../internal/utils/sleep.js';
+import { applyJitter, backoff, isFatal4xx } from '../../internal/utils/backoff.js';
+import { linkAbort } from '../../internal/utils/abort.js';
+import { AsyncQueue } from '../../internal/utils/async-queue.js';
+import { buildHeaders } from '../../internal/headers.js';
+import { helperHeader } from '../../internal/stainless-helper-header.js';
+import { runRunnableTool, toolName } from './BetaRunnableTool.js';
+/** Beta header for the managed-agents API. */
+export const MANAGED_AGENTS_BETA = 'managed-agents-2026-04-01';
+const STREAM_BACKOFF_START_MS = 500;
+const STREAM_BACKOFF_CAP_MS = 10_000;
+const TOOL_TIMEOUT_MS = 120_000;
+const DRAIN_TIMEOUT_MS = 30_000;
+const SEND_BACKOFF_START_MS = 1_000;
+const SEND_BACKOFF_CAP_MS = 30_000;
+/**
+ * How long a transiently failing tool-result send keeps retrying when the
+ * runner is used on its own: the server's default work-item lease TTL.
+ * `EnvironmentWorker` overrides it with the live TTL from each lease heartbeat
+ * (`_setSendRetryWindow`), so a send is only abandoned once the lease can no
+ * longer be ours.
+ */
+const SEND_RETRY_WINDOW_MS = 5 * 60_000;
+/** Default {@link SessionToolRunnerOptions.maxIdleMs}: 60 seconds. */
+export const DEFAULT_MAX_IDLE_MS = 60_000;
+/** Returns true if `ev` is a `session.status_idle` with `stop_reason` `end_turn`. */
+function isEndTurnIdle(ev) {
+    return ev.type === 'session.status_idle' && ev.stop_reason?.type === 'end_turn';
+}
+/**
+ * The `maxIdleMs` stop-countdown, including its deferral. {@link noteEvent}
+ * arms on `session.status_idle` with `stop_reason: end_turn` and disarms on
+ * anything else. Gated tool work registered via {@link block} — a call held for
+ * user confirmation, or a user-approved call still dispatching — keeps
+ * {@link arm} pending until {@link unblock} retires the last blocker, at which
+ * point the countdown starts. Event-driven — there is no polling watchdog.
+ */
+class IdleClock {
+    #maxIdleMs;
+    #onExpire;
+    #blockers = new Set();
+    // Set when arm() found blockers outstanding; the unblock that retires the
+    // last blocker applies it. Cleared by any disarm.
+    #armPending = false;
+    #timer;
+    constructor(maxIdleMs, onExpire) {
+        this.#maxIdleMs = maxIdleMs;
+        this.#onExpire = onExpire;
+    }
+    /**
+     * Arm on `status_idle{end_turn}`; disarm otherwise. `user.tool_confirmation`
+     * is neutral: it signals neither agent activity nor an idle, and its effect
+     * on the clock flows through {@link block} / {@link unblock} instead —
+     * disarming here would discard the pending arm the verdict is about to
+     * settle.
+     */
+    noteEvent(ev) {
+        if (ev.type === 'user.tool_confirmation')
+            return;
+        if (isEndTurnIdle(ev))
+            this.arm();
+        else
+            this.disarm();
+    }
+    /** Register gated work that must resolve before an idle countdown starts. */
+    block(toolUseId) {
+        this.#blockers.add(toolUseId);
+        if (this.#timer !== undefined) {
+            // Defensive: every caller disarms first (any event that routes a tool
+            // call is itself a disarming event), but a countdown running when gated
+            // work appears is stale evidence — the session is not idly waiting to
+            // stop. Convert it into a pending arm rather than let it fire over the
+            // gated call.
+            this.#armPending = true;
+            clearTimeout(this.#timer);
+            this.#timer = undefined;
+        }
+    }
+    /**
+     * Retire gated work (a no-op for ids never blocked); applies a pending arm —
+     * with a fresh full `maxIdleMs` window — once the last blocker retires.
+     */
+    unblock(toolUseId) {
+        this.#blockers.delete(toolUseId);
+        if (this.#blockers.size === 0 && this.#armPending)
+            this.arm();
+    }
+    /**
+     * (Re)start the idle countdown — or, while blockers are outstanding, hold
+     * the arm pending instead. Stopping then would drop a held call when its
+     * verdict later arrives, or cut the runner off before a released call's
+     * result can drive the next turn.
+     */
+    arm() {
+        if (this.#maxIdleMs <= 0)
+            return;
+        if (this.#blockers.size > 0) {
+            this.#armPending = true;
+            return;
+        }
+        this.#armPending = false;
+        if (this.#timer !== undefined)
+            clearTimeout(this.#timer);
+        this.#timer = setTimeout(this.#onExpire, this.#maxIdleMs);
+    }
+    /**
+     * Cancel the idle countdown and any pending arm. Blockers persist — they
+     * track real outstanding work, retired only by {@link unblock}.
+     */
+    disarm() {
+        this.#armPending = false;
+        if (this.#timer !== undefined) {
+            clearTimeout(this.#timer);
+            this.#timer = undefined;
+        }
+    }
+}
+/**
+ * The sessions-side counterpart to `client.beta.messages.toolRunner`: an
+ * async-iterable that attaches to a managed-agents session, executes every
+ * incoming `agent.tool_use` and `agent.custom_tool_use` event against a local
+ * tool registry, posts the matching result back (`user.tool_result` for the
+ * former, `user.custom_tool_result` for the latter), and yields one
+ * {@link DispatchedToolCall} per completed call. Server-side `agent.mcp_tool_use`
+ * calls are not dispatched. Internally drives event-stream reconnect and result
+ * posting.
+ *
+ * A call the server gated with `evaluated_permission: "ask"` (the `always_ask`
+ * policy — or any value this SDK doesn't recognize, which fails closed) is held
+ * until its `user.tool_confirmation` arrives: only an explicit `allow` runs it;
+ * `deny` — or any verdict this SDK doesn't recognize, failing closed — is never
+ * executed and posts nothing (the denial resolves the call server-side), but is
+ * still yielded (`confirmation="deny"`, `posted=false`, `result=undefined`) so
+ * the consumer can observe it. A held call — and a user-approved one still
+ * dispatching — defers the `maxIdleMs` countdown, so an `end_turn` idle
+ * observed in the meantime cannot stop the runner: it waits until the verdict
+ * arrives, the session terminates, or the abort signal fires — pass
+ * `AbortSignal.timeout(...)` for a wall-clock bound.
+ *
+ * Iteration ends when the session terminates (`session.status_terminated` /
+ * `session.deleted`), when the consumer `break`s out of the loop or aborts the
+ * supplied signal, or — once the session has gone idle with
+ * `stop_reason.type === "end_turn"` — when `maxIdleMs` elapses with no new
+ * event (any new event resets that countdown; it re-arms on the next `end_turn`
+ * idle; `maxIdleMs <= 0` disables it). The `finally` branch drains any in-flight
+ * tool calls and runs each tool's `close()` cleanup hook. It does *not* touch
+ * the work-item lease — wrap it in an `EnvironmentWorker` if you need
+ * heartbeating / force-stop.
+ *
+ * @example
+ * ```ts
+ * import { betaAgentToolset20260401 } from '@puku-ai/sdk/tools/agent-toolset/node';
+ *
+ * for await (const call of client.beta.sessions.events.toolRunner(work.data.id, {
+ *   tools: [...betaAgentToolset20260401({ workdir }), myTool],
+ * })) {
+ *   console.log(`${call.name} -> ${call.isError ? 'error' : 'ok'}`);
+ * }
+ * ```
+ */
+export class SessionToolRunner {
+    client;
+    sessionId;
+    tools;
+    maxIdleMs;
+    #consumed = false;
+    #controller;
+    #detachExternal;
+    #requestOpts;
+    #toolByName;
+    #logger;
+    #seen = new Set();
+    #answered = new Set();
+    // Confirmation gating (`always_ask` tools): `#confirmationVerdicts` records
+    // every `user.tool_confirmation` verdict by `tool_use_id`;
+    // `#awaitingConfirmation` holds the tool-call events whose
+    // `evaluated_permission` is `ask` and whose verdict has not arrived —
+    // released (or resolved as denied) by `#noteConfirmation` / the next
+    // reconcile pass. Like `#seen` and `#answered`, `#confirmationVerdicts` is
+    // per-session O(tool calls): recorded verdicts persist for the life of the run.
+    #confirmationVerdicts = new Map();
+    #awaitingConfirmation = new Map();
+    #results = new AsyncQueue();
+    #inFlightCount = 0;
+    #sendRetryWindowMs = SEND_RETRY_WINDOW_MS;
+    #onIdle = null;
+    #idleClock;
+    constructor(sessionId, opts) {
+        this.client = opts.client;
+        this.sessionId = sessionId;
+        this.tools = opts.tools;
+        this.maxIdleMs = opts.maxIdleMs ?? DEFAULT_MAX_IDLE_MS;
+        this.#logger = loggerFor(opts.client);
+        this.#toolByName = new Map(opts.tools.map((t) => [toolName(t), t]));
+        this.#controller = new AbortController();
+        this.#detachExternal = linkAbort(opts.signal, this.#controller);
+        this.#requestOpts = opts.requestOptions;
+        this.#idleClock = new IdleClock(this.maxIdleMs, () => {
+            this.#logger.info('session idle after end_turn; stopping', {
+                component: 'session-tool-runner',
+                session_id: this.sessionId,
+                max_idle_ms: this.maxIdleMs,
+            });
+            this.#controller.abort();
+        });
+    }
+    /** Read-only view of this runner's abort signal. */
+    get signal() {
+        return this.#controller.signal;
+    }
+    /** Abort the runner. Background tasks will wind down and `for await` will exit cleanly. */
+    abort() {
+        this.#controller.abort();
+    }
+    /**
+     * @internal
+     * `EnvironmentWorker` keeps this equal to the lease TTL each heartbeat
+     * reports; applies to a send already retrying.
+     */
+    _setSendRetryWindow(ms) {
+        this.#sendRetryWindowMs = ms;
+    }
+    async *[Symbol.asyncIterator]() {
+        if (this.#consumed) {
+            throw new PukuError('Cannot iterate over a consumed SessionToolRunner');
+        }
+        this.#consumed = true;
+        this.#logger.info('session tool runner starting', {
+            component: 'session-tool-runner',
+            session_id: this.sessionId,
+        });
+        // The one background promise: drives the event stream and dispatches tools.
+        // Its `.catch` aborts the controller so the main loop unwinds.
+        const streamPromise = this.#streamLoop().catch((e) => {
+            if (!this.#controller.signal.aborted) {
+                this.#logger.error('stream loop failed', { error: String(e) });
+            }
+            this.#controller.abort();
+        });
+        try {
+            // Phase 1: yield results as they arrive. `next(signal)` resolves
+            // `done: true` when the controller aborts — cancellation is handled in
+            // the queue read, no outer `Promise.race` needed.
+            while (true) {
+                const next = await this.#results.next(this.#controller.signal);
+                if (next.done)
+                    break;
+                yield next.value;
+            }
+            // Phase 2: let the stream loop settle (and push any final results), then
+            // drain whatever is still queued before closing.
+            await streamPromise;
+            let pending;
+            while ((pending = this.#results.tryShift()) !== undefined) {
+                yield pending;
+            }
+        }
+        finally {
+            this.#controller.abort();
+            this.#idleClock.disarm();
+            // Re-await defensively in case the consumer broke out of phase 1 before
+            // phase 2 ran — a no-op if it already settled.
+            await streamPromise;
+            try {
+                await this.#drain();
+            }
+            catch (e) {
+                this.#logger.warn('drain failed', { error: String(e) });
+            }
+            this.#results.close();
+            for (const t of this.tools) {
+                try {
+                    // `close` is typed `() => Promisable<void>`, so a single `await`
+                    // covers both the sync and async return.
+                    await t.close?.();
+                }
+                catch (e) {
+                    this.#logger.warn('tool.close failed', { tool: toolName(t), error: String(e) });
+                }
+            }
+            // Detach from the external signal so the consumer can drop their signal
+            // reference without leaking this iterator instance.
+            this.#detachExternal();
+        }
+    }
+    // ===== request options =====
+    /**
+     * Request options for every helper-issued call: the caller's `requestOptions`
+     * (custom proxy headers etc.) with the helper telemetry header stamped on and
+     * the runner's own abort signal forced last so it always owns cancellation.
+     */
+    #requestOptions() {
+        return {
+            ...this.#requestOpts,
+            headers: buildHeaders([helperHeader('session-tool-runner'), this.#requestOpts?.headers]),
+            signal: this.#controller.signal,
+        };
+    }
+    // ===== event stream =====
+    async #streamLoop() {
+        const ctrl = this.#controller;
+        let backoff = STREAM_BACKOFF_START_MS;
+        while (!ctrl.signal.aborted) {
+            try {
+                // Establish the event stream *before* reconciling history, so an event
+                // emitted in the gap between listing and attaching is buffered on the
+                // stream rather than lost. `seen`/`answered` dedup any event that shows
+                // up both in the reconcile pass and on the live stream.
+                const stream = await this.client.beta.sessions.events.stream(this.sessionId, {}, this.#requestOptions());
+                await this.#reconcile();
+                for await (const ev of stream) {
+                    backoff = STREAM_BACKOFF_START_MS;
+                    if (await this.#handleStreamEvent(ev))
+                        return;
+                }
+            }
+            catch (e) {
+                // An abort throws to unwind the caller (the iterator's `streamPromise`
+                // `.catch`) rather than returning early and letting it carry on.
+                ctrl.signal.throwIfAborted();
+                if (isFatal4xx(e)) {
+                    this.#logger.error('permanent stream failure, shutting down', { error: String(e) });
+                    ctrl.abort();
+                    throw e;
+                }
+                this.#logger.warn('stream disconnected, reconnecting', {
+                    error: String(e),
+                    backoff_ms: backoff,
+                });
+            }
+            ctrl.signal.throwIfAborted();
+            await sleep(backoff, ctrl.signal);
+            backoff = Math.min(backoff * 2, STREAM_BACKOFF_CAP_MS);
+        }
+    }
+    /**
+     * Read full history before dispatching so a `tool_use` whose result appears
+     * later in the same history is not re-executed. Runs after the live stream is
+     * already attached (see {@link SessionToolRunner.#streamLoop}).
+     */
+    async #reconcile() {
+        const ctrl = this.#controller;
+        const pending = [];
+        let lastWasEndTurn = false;
+        try {
+            for await (const ev of this.client.beta.sessions.events.list(this.sessionId, { limit: 1000 }, this.#requestOptions())) {
+                this.#ingestHistory(ev, pending);
+                lastWasEndTurn = isEndTurnIdle(ev);
+            }
+        }
+        catch (e) {
+            // An abort throws to unwind the caller; a real list failure is
+            // non-fatal — undo the speculative `seen` entries and let `#streamLoop`
+            // carry on with the live stream.
+            ctrl.signal.throwIfAborted();
+            this.#logger.warn('reconcile list failed', { error: String(e) });
+            // If list itself failed, undo the speculative `seen` entries so the next
+            // reconcile pass (or the live stream) can pick them up. Leave the idle
+            // timer untouched — the history we read may be incomplete.
+            for (const ev of pending)
+                this.#seen.delete(ev.id);
+            return;
+        }
+        const unanswered = pending.filter((ev) => !this.#answered.has(ev.id));
+        // Disarm before routing: `#execute` runs inline here, so a timer left armed
+        // from before the reconnect could fire over an in-flight tool.
+        this.#idleClock.disarm();
+        for (const ev of unanswered)
+            await this.#routeToolEvent(ev);
+        // A held call's verdict is normally applied by the routing pass above; if
+        // its tool_use fell outside the listed window the pass never saw it, so
+        // apply the verdict to the held copy here.
+        for (const held of [...this.#awaitingConfirmation.values()]) {
+            const verdict = this.#confirmationVerdicts.get(held.id);
+            if (verdict !== undefined)
+                await this.#applyVerdict(held, verdict);
+        }
+        // Routing resolves denied calls in place (marking them answered) and holds
+        // ask-gated calls for their `user.tool_confirmation`. If the most recent
+        // event in history is an `end_turn` idle and no tool work is outstanding,
+        // the session is done — arm the idle clock so the runner stops even if that
+        // `end_turn` arrived during a disconnect. A held call is not outstanding
+        // here: it blocks the clock, so this arm stays pending until the verdict
+        // (and, for an allow, the dispatch it releases) resolves it.
+        const outstanding = unanswered.filter((ev) => !this.#answered.has(ev.id) && !this.#awaitingConfirmation.has(ev.id));
+        if (lastWasEndTurn && outstanding.length === 0)
+            this.#idleClock.arm();
+        else
+            this.#idleClock.disarm();
+    }
+    #ingestHistory(ev, pending) {
+        if (ev.type === 'agent.tool_use' || ev.type === 'agent.custom_tool_use') {
+            // Mark the event seen so a replay on the live stream is not dispatched
+            // twice, but decide whether it still needs executing from `answered`, not
+            // `seen`: a call whose result post failed is seen-but-unanswered, and must
+            // be retried on the next reconcile pass rather than silently dropped.
+            this.#seen.add(ev.id);
+            if (!this.#answered.has(ev.id))
+                pending.push(ev);
+        }
+        else if (ev.type === 'user.tool_result') {
+            this.#answered.add(ev.tool_use_id);
+        }
+        else if (ev.type === 'user.custom_tool_result') {
+            this.#answered.add(ev.custom_tool_use_id);
+        }
+        else if (ev.type === 'user.tool_confirmation') {
+            // Record the verdict only, before the pending pass, so a call whose
+            // confirmation appears later in the same history routes with its verdict
+            // already known. Releasing a held call here as well would dispatch it a
+            // second time when the routing pass reaches its tool_use event.
+            if (!this.#answered.has(ev.tool_use_id))
+                this.#confirmationVerdicts.set(ev.tool_use_id, ev.result);
+        }
+    }
+    /** Returns true when the runner should exit. */
+    async #handleStreamEvent(ev) {
+        this.#idleClock.noteEvent(ev);
+        switch (ev.type) {
+            case 'agent.tool_use':
+            case 'agent.custom_tool_use':
+                if (!this.#seen.has(ev.id)) {
+                    this.#seen.add(ev.id);
+                    await this.#routeToolEvent(ev);
+                }
+                return false;
+            case 'user.tool_confirmation':
+                await this.#noteConfirmation(ev);
+                return false;
+            case 'user.tool_result':
+                this.#answered.add(ev.tool_use_id);
+                return false;
+            case 'user.custom_tool_result':
+                this.#answered.add(ev.custom_tool_use_id);
+                return false;
+            case 'session.status_terminated':
+            case 'session.deleted':
+                this.#logger.info('session terminated', {
+                    component: 'session-tool-runner',
+                    session_id: this.sessionId,
+                });
+                this.#controller.abort();
+                return true;
+            default:
+                return false;
+        }
+    }
+    // ===== confirmation gating (always_ask tools) =====
+    /**
+     * Dispatch `ev`, honoring its evaluated permission. A call the server gated
+     * (`evaluated_permission == "ask"`) is held until its `user.tool_confirmation`
+     * arrives. Fails closed: only an explicit `allow` verdict releases a gated
+     * call; a server-side `deny` overrides any recorded verdict; an unrecognized
+     * permission is held like `ask` and an unrecognized verdict is denied.
+     */
+    async #routeToolEvent(ev) {
+        // `getattr`-style read: today only `agent.tool_use` carries
+        // `evaluated_permission`, but if it ever lands on `agent.custom_tool_use`
+        // the gate must keep failing closed rather than dispatch by event type.
+        const permission = ev.evaluated_permission;
+        // A server-side `deny` overrides any (stray) recorded verdict.
+        const verdict = permission === 'deny' ? 'deny' : this.#confirmationVerdicts.get(ev.id);
+        if (verdict === undefined) {
+            if (permission === undefined || permission === 'allow') {
+                await this.#execute(ev, undefined);
+            }
+            else if (!this.#awaitingConfirmation.has(ev.id)) {
+                // "ask" — or a permission this SDK does not recognize, which must not
+                // dispatch unconfirmed — waits for the user's verdict. (Already-held: a
+                // reconcile after reconnect re-routes the call; keep the existing hold.)
+                this.#logger.info('tool call awaiting confirmation; holding', {
+                    component: 'session-tool-runner',
+                    session_id: this.sessionId,
+                    tool: ev.name,
+                    tool_use_id: ev.id,
+                });
+                this.#awaitingConfirmation.set(ev.id, ev);
+                this.#idleClock.block(ev.id);
+            }
+            return;
+        }
+        await this.#applyVerdict(ev, verdict);
+    }
+    /** Record an allow/deny verdict and release the held call it gates, if any. */
+    async #noteConfirmation(ev) {
+        this.#confirmationVerdicts.set(ev.tool_use_id, ev.result);
+        const held = this.#awaitingConfirmation.get(ev.tool_use_id);
+        // Nothing held: the verdict gates a call this runner has not seen yet (or
+        // one it never gates, e.g. an `agent.mcp_tool_use`). Keeping it in
+        // `#confirmationVerdicts` lets a later route of that call resolve instantly.
+        if (held === undefined)
+            return;
+        await this.#applyVerdict(held, ev.result);
+    }
+    /**
+     * Dispatch or resolve a gated call according to its verdict.
+     *
+     * The idle-clock blocker accounting lives here: a denial retires the held
+     * call's blocker, while an allow keeps one on the call — taking it now if the
+     * verdict was already known when the call was routed, so it was never held —
+     * until `#execute` has finished with it. The countdown must not run over
+     * gated work that is still in flight.
+     */
+    async #applyVerdict(ev, verdict) {
+        const wasHeld = this.#awaitingConfirmation.delete(ev.id);
+        if (verdict === 'allow') {
+            this.#logger.info('tool call confirmed', {
+                component: 'session-tool-runner',
+                session_id: this.sessionId,
+                tool: ev.name,
+                tool_use_id: ev.id,
+            });
+            if (!wasHeld)
+                this.#idleClock.block(ev.id);
+            try {
+                await this.#execute(ev, 'allow');
+            }
+            finally {
+                // The approved call is fully disposed of (executed, or moot because it
+                // was answered elsewhere) — the sole place an allow's blocker retires.
+                this.#idleClock.unblock(ev.id);
+            }
+            return;
+        }
+        // "deny" — or any value other than an explicit "allow" (fail closed). The
+        // denial resolves the call server-side, so mark it answered and yield it
+        // (nothing ran, nothing posted).
+        if (wasHeld)
+            this.#idleClock.unblock(ev.id);
+        this.#answered.add(ev.id);
+        this.#logger.info('tool call denied; not executing', {
+            component: 'session-tool-runner',
+            session_id: this.sessionId,
+            tool: ev.name,
+            tool_use_id: ev.id,
+        });
+        this.#surfaceCall({
+            event: ev,
+            toolUseId: ev.id,
+            name: ev.name,
+            isError: false,
+            posted: false,
+            confirmation: 'deny',
+        });
+    }
+    /**
+     * Yield `call` to the consumer. A closed queue — the consumer broke out of
+     * the iterator — makes `push` a no-op returning false; the underlying work
+     * already happened, so only the observability event is lost.
+     */
+    #surfaceCall(call) {
+        this.#results.push(call);
+    }
+    // ===== tool execution =====
+    async #execute(ev, confirmation) {
+        if (this.#answered.has(ev.id))
+            return;
+        this.#logger.info('executing tool', {
+            component: 'session-tool-runner',
+            session_id: this.sessionId,
+            tool: ev.name,
+            tool_use_id: ev.id,
+        });
+        this.#inFlightCount++;
+        try {
+            const tool = this.#toolByName.get(ev.name);
+            if (!tool) {
+                // Skip (split-client partial fulfilment): a name this runner
+                // is not registered for belongs to the other client servicing this
+                // session (typically the customer's app backend handling custom tools).
+                // Post NO result, do not mark it answered, and leave the tool_use_id
+                // pending for its owner — claiming it would corrupt the conversation.
+                // Still yield the call so the consumer can observe the unowned
+                // dispatch; nothing was sent, so `posted`/`isError` stay false and no
+                // `result` event is populated. The id stays unanswered, so reconcile
+                // keeps it out of the idle/end-turn accounting and re-surfaces it after
+                // a reconnect until its owner answers it.
+                this.#logger.info('tool not owned by this runner; leaving the tool_use_id pending for its owner', {
+                    component: 'session-tool-runner',
+                    session_id: this.sessionId,
+                    tool: ev.name,
+                    tool_use_id: ev.id,
+                });
+                this.#surfaceCall({
+                    event: ev,
+                    toolUseId: ev.id,
+                    name: ev.name,
+                    isError: false,
+                    posted: false,
+                    confirmation,
+                });
+                return;
+            }
+            let content;
+            let isError;
+            // Per-tool controller: aborts on the runner's own signal *or* the
+            // per-tool timeout, so an in-flight tool stops promptly when the runner
+            // is aborted instead of running until the timeout.
+            const toolCtrl = new AbortController();
+            const detachTool = linkAbort(this.#controller.signal, toolCtrl);
+            const timer = setTimeout(() => toolCtrl.abort(), TOOL_TIMEOUT_MS);
+            try {
+                // Pass the source `agent.tool_use` / `agent.custom_tool_use` event
+                // straight through as the run context's `toolUse` — it is a union
+                // member of `BetaToolUse`, no Messages-block adapter needed.
+                const outcome = await runRunnableTool(tool, ev.input, {
+                    toolUse: ev,
+                    toolUseBlock: ev,
+                    signal: toolCtrl.signal,
+                });
+                content = outcome.content;
+                isError = outcome.isError;
+            }
+            finally {
+                clearTimeout(timer);
+                detachTool();
+            }
+            // Answer with the result event that matches the call kind: a
+            // `user.tool_result` for an `agent.tool_use`, a `user.custom_tool_result`
+            // for an `agent.custom_tool_use`. Posting the wrong one leaves the call
+            // unanswered and the session stuck.
+            const result = buildResultEvent(ev, isError, toSessionContent(content));
+            const posted = await this.#sendResult(result, ev.id);
+            this.#surfaceCall({
+                event: ev,
+                result,
+                toolUseId: ev.id,
+                name: ev.name,
+                isError,
+                posted,
+                confirmation,
+            });
+        }
+        finally {
+            this.#inFlightCount--;
+            if (this.#inFlightCount === 0)
+                this.#onIdle?.();
+        }
+    }
+    async #sendResult(result, toolUseId) {
+        const ctrl = this.#controller;
+        const start = Date.now();
+        let lastErr;
+        let attempt = 0;
+        while (true) {
+            attempt++;
+            // An abort throws to unwind the caller rather than returning a
+            // `posted: false` result the iterator would carry on past.
+            ctrl.signal.throwIfAborted();
+            try {
+                await this.client.beta.sessions.events.send(this.sessionId, { events: [result] }, this.#requestOptions());
+                this.#answered.add(toolUseId);
+                return true;
+            }
+            catch (e) {
+                lastErr = e;
+                // Only short-circuit on a permanent 4xx; 408/409/429 deserve the
+                // remaining retries (aligned with the core client's retry policy).
+                if (isFatal4xx(e))
+                    break;
+                const remainingMs = this.#sendRetryWindowMs - (Date.now() - start);
+                if (remainingMs <= 0)
+                    break;
+                const waitMs = Math.min(applyJitter(backoff(attempt - 1, SEND_BACKOFF_START_MS, SEND_BACKOFF_CAP_MS)), remainingMs);
+                this.#logger.warn('tool result send failed; retrying', {
+                    tool_use_id: toolUseId,
+                    attempt,
+                    backoff_ms: waitMs,
+                    error: String(e),
+                });
+                await sleep(waitMs, ctrl.signal);
+            }
+        }
+        this.#logger.error('failed to send tool result', {
+            tool_use_id: toolUseId,
+            attempts: attempt,
+            error: String(lastErr),
+        });
+        return false;
+    }
+    /** Wait (bounded) for in-flight tool executions to finish during teardown. */
+    async #drain() {
+        if (this.#inFlightCount === 0)
+            return;
+        await Promise.race([new Promise((r) => (this.#onIdle = r)), sleep(DRAIN_TIMEOUT_MS)]);
+        this.#onIdle = null;
+        if (this.#inFlightCount > 0) {
+            this.#logger.warn('drain timeout exceeded');
+        }
+    }
+}
+/**
+ * Build the result event that answers `ev`: a `user.tool_result` for a builtin
+ * `agent.tool_use`, a `user.custom_tool_result` for a custom
+ * `agent.custom_tool_use`. The two `(use, result)` pairs are distinct API event
+ * types and must be matched exactly — a `user.tool_result` does not answer a
+ * custom tool call.
+ */
+function buildResultEvent(ev, isError, content) {
+    if (ev.type === 'agent.custom_tool_use') {
+        return { type: 'user.custom_tool_result', custom_tool_use_id: ev.id, is_error: isError, content };
+    }
+    return { type: 'user.tool_result', tool_use_id: ev.id, is_error: isError, content };
+}
+// The Messages-API tool-result block union is wider than the Sessions-API
+// tool_result content union; pass through text/image/document and stringify
+// anything else so a BetaRunnableTool authored for toolRunner still works here.
+function toSessionContent(content) {
+    if (typeof content === 'string')
+        return [{ type: 'text', text: content || '(no output)' }];
+    const out = content.map((b) => {
+        if (b.type === 'text')
+            return { type: 'text', text: b.text || '(no output)' };
+        if (b.type === 'image' || b.type === 'document')
+            return b;
+        if (b.type === 'search_result') {
+            // The Messages `search_result` block param maps field-for-field onto the
+            // Sessions `BetaManagedAgentsSearchResultBlock`; map it explicitly rather
+            // than letting it fall through to the JSON.stringify branch (which would
+            // bury a structured result inside a text block). `citations` is required
+            // on the Sessions side and optional on the Messages side — default the
+            // flag to `false` when the producer left it unset.
+            return {
+                type: 'search_result',
+                source: b.source,
+                title: b.title,
+                content: b.content.map((c) => ({ type: 'text', text: c.text })),
+                citations: { enabled: b.citations?.enabled ?? false },
+            };
+        }
+        return { type: 'text', text: JSON.stringify(b) };
+    });
+    return out.length > 0 ? out : [{ type: 'text', text: '(no output)' }];
+}
+//# sourceMappingURL=SessionToolRunner.js.map