@qduc/term2 0.32.2 → 0.34.0

package/dist/services/agent-runtime/application-run-loop.js

@@ -46,11 +46,16 @@ export class MaxTurnsExceededError extends Error {
     }
 }
 class EventQueue {
+    acceptBeforePublish;
+    constructor(acceptBeforePublish) {
+        this.acceptBeforePublish = acceptBeforePublish;
+    }
     #items = [];
     #waiters = [];
     #closed = false;
     #failure;
     push(item) {
+        this.acceptBeforePublish?.(item);
         const waiter = this.#waiters.shift();
         if (waiter)
             waiter.resolve({ done: false, value: item });
@@ -98,6 +103,7 @@ export class ApplicationRunLoop {
     #contextCompactionSessionState;
     #turnInputMailbox;
     #activeAbortController = null;
+    #activeTurnController = null;
     #segmentGeneration = 0;
     /** The run whose budget an out-of-band grant applies to, held across a pause. */
     #activeRunBudgetState;
@@ -143,6 +149,7 @@ export class ApplicationRunLoop {
         this.#turnInputMailbox.closeTurn();
     }
     abort() {
+        this.#activeTurnController?.abort();
         this.abortSegment();
         // An aborted turn will not resume, so nothing may keep waiting on it.
         this.#turnInputMailbox.abortTurn();
@@ -192,7 +199,6 @@ export class ApplicationRunLoop {
     #createToolCallExecution() {
         return new ToolCallExecution({
             toolLifecycle: this.#deps.toolLifecycle,
-            getOnToolDispatch: this.#deps.getOnToolDispatch,
             resolveMaxParallelToolCalls: this.#deps.resolveMaxParallelToolCalls,
             isMaxTurnsExceeded: (error) => error instanceof MaxTurnsExceededError,
             logDiagnostic: this.#deps.logDiagnostic,
@@ -256,12 +262,9 @@ export class ApplicationRunLoop {
      * the tool results of the round that was in flight when they typed.
      */
     startStream(agent, input, options = {}) {
-        // Without a declared turn scope, a fresh stream is the only evidence this
-        // loop gets that a new turn has begun: anything still waiting belonged to
-        // the previous one and must not leak into work the user did not aim it at.
-        // With one, this call may equally be the same turn restarting after a
-        // retry, which the loop cannot tell apart — so `openTurn`/`closeTurn` own
-        // the decision instead.
+        return this.#run(this.#createState(agent, input, options), options);
+    }
+    #createState(agent, input, options) {
         this.#turnInputMailbox.startStream();
         const state = {
             agent,
@@ -293,6 +296,7 @@ export class ApplicationRunLoop {
             approvals: options.approvals ?? new ApprovalLedger(),
             turnCount: 0,
             maxTurns: options.maxTurns,
+            completionRepairDeadline: options.completionRepairDeadline,
             ...(options.runBudget
                 ? {
                     runBudget: new RunBudget(options.runBudget),
@@ -329,24 +333,31 @@ export class ApplicationRunLoop {
             state.approvalDecisionInterruption = interruption;
             state.approvalMessage = approvalOptions?.message;
         };
-        return this.#run(state, options);
+        return state;
+    }
+    /** Root turns use the same loop; session operations never dispatch a run. */
+    async runTurn(session) {
+        const controller = new AbortController();
+        this.#activeTurnController = controller;
+        try {
+            await this.#runTurn(undefined, session, controller.signal);
+        }
+        finally {
+            if (this.#activeTurnController === controller)
+                this.#activeTurnController = null;
+        }
     }
     continueRunStream(handle, options = {}) {
         const state = unwrapContinuationHandle(handle);
         if (!state || typeof state !== 'object' || !('agent' in state) || !('input' in state)) {
             throw new Error('Invalid application continuation state');
         }
-        // Handles created before the ledger existed cannot resume meaningfully;
-        // give them a fresh ledger rather than crashing on `approvals` access.
-        if (!state.approvals)
-            state.approvals = new ApprovalLedger();
-        // A continuation handle from before approval batching carried only the
-        // current pending item. Normalize it into the queue used by current runs.
-        if (!state.toolExecution)
-            state.toolExecution = this.#createToolCallExecution();
-        // Handles created before cost accounting existed have no record list.
-        if (!state.costRecords)
-            state.costRecords = [];
+        if (!state.live?.resume)
+            throw new Error('Application turn is not waiting for continuation');
+        this.#refreshContinuation(state, options);
+        return this.#run(state, options);
+    }
+    #refreshContinuation(state, options) {
         if (!state.runBudget && options.runBudget) {
             state.runBudget = new RunBudget(options.runBudget);
         }
@@ -382,13 +393,7 @@ export class ApplicationRunLoop {
             state.turnCount = 0;
         if (state.maxTurns === undefined)
             state.maxTurns = options.maxTurns;
-        if (typeof state.supportsConversationChaining !== 'boolean') {
-            state.supportsConversationChaining = options.supportsConversationChaining === true;
-        }
-        else {
-            // AgentClient supplies the current provider capability on every resume.
-            state.supportsConversationChaining = options.supportsConversationChaining === true;
-        }
+        state.supportsConversationChaining = options.supportsConversationChaining === true;
         // A continuation receives a fresh immutable snapshot from the session;
         // refresh the preparation closure when supplied, while preserving the
         // root closure for callers that do not provide one again.
@@ -409,12 +414,12 @@ export class ApplicationRunLoop {
             state.turnId = options.turnId;
         if (options.hookScope !== undefined)
             state.hookScope = options.hookScope;
-        return this.#run(state, options);
     }
-    #run(state, options) {
-        const queue = new EventQueue();
+    #createView(state, options, acceptEvent) {
+        const queue = new EventQueue(acceptEvent);
         const segmentGeneration = ++this.#segmentGeneration;
-        this.abortSegment();
+        if (!state.live)
+            this.abortSegment();
         const controller = new AbortController();
         this.#activeAbortController = controller;
         this.#turnInputMailbox.startSegment();
@@ -465,46 +470,137 @@ export class ApplicationRunLoop {
                 return state.costRecords ?? [];
             },
         });
-        let exitError;
-        stream.completed = this.#execute(state, stream, queue, effectiveOptions, toolContext, controller)
-            .catch((error) => {
-            exitError = error;
-            stream.cancelled = error instanceof Error && error.name === 'AbortError';
+        let resolveSegment;
+        let rejectSegment;
+        stream.completed = new Promise((resolve, reject) => {
+            resolveSegment = resolve;
+            rejectSegment = reject;
+        });
+        // The iterator may report a failure before the caller observes completed.
+        void stream.completed.catch(() => undefined);
+        const current = {
+            state,
+            stream,
+            queue,
+            options: effectiveOptions,
+            toolContext,
+            requestAbortController: controller,
+            generation: segmentGeneration,
+            resolveSegment,
+            rejectSegment,
+        };
+        this.#activeRunBudgetState = state.runBudget ? state : undefined;
+        return current;
+    }
+    #run(state, options) {
+        const current = this.#createView(state, options);
+        if (state.live) {
+            const resume = state.live.resume;
+            Object.assign(state.live, current, { resume: undefined });
+            resume();
+        }
+        else {
+            state.live = current;
+            void this.#runTurn(current).then((value) => this.#settleSegment(current, value), (error) => this.#settleSegment(current, undefined, error));
+        }
+        return current.stream;
+    }
+    #observe(context) {
+        context.session.beginSegment(context.stream);
+        context.observation = context.session.consume(context.stream);
+        void context.observation.catch(() => undefined);
+    }
+    async #completeTurn(context) {
+        const value = finish(context.stream, context.state, context.queue);
+        if (context.session) {
+            this.#settleSegment(context, value);
+            await context.observation;
+            await context.session.completed(context.stream);
+        }
+        return value;
+    }
+    /** Complete an event view; a pause leaves the owning async task alive. */
+    #settleSegment(context, value, error, waiting = false) {
+        const { state, stream, queue, generation } = context;
+        if (error === undefined)
+            context.session?.acceptSegment(stream);
+        state.runBudget?.pause();
+        const cancelled = error instanceof Error && error.name === 'AbortError';
+        const paused = (waiting || context.resume !== undefined) && !cancelled && error === undefined;
+        if (error !== undefined) {
+            stream.cancelled = cancelled;
             queue.close(error);
-            throw error;
-        })
-            .finally(() => {
-            state.runBudget?.pause();
-            if (segmentGeneration !== this.#segmentGeneration)
-                return;
-            if (this.#activeAbortController === controller)
-                this.#activeAbortController = null;
-            // A segment that ends holding approvals has paused the turn, not ended
-            // it: the caller resumes through continueRunStream, which offers
-            // another request boundary. Injections wait for it. Only a segment
-            // that ends with nothing outstanding has truly finished the turn, and
-            // then anything still waiting has to be sent as its own turn instead.
-            const pendingApprovals = state.toolExecution.pendingApprovals.length;
-            const pendingBudgetInteraction = state.pendingRunBudgetInteraction !== undefined;
-            const cancelled = stream.cancelled === true;
-            const turnPaused = (pendingApprovals > 0 || pendingBudgetInteraction) && !cancelled && exitError === undefined;
-            this.#turnInputMailbox.settleSegment({
-                paused: turnPaused,
-                reason: {
-                    pendingApprovals,
-                    pendingBudgetInteraction,
-                    cancelled,
-                    error: exitError instanceof Error ? exitError.name : exitError ? String(exitError) : undefined,
-                    turnCount: state.turnCount,
-                },
-            });
-            if (turnPaused)
+            context.rejectSegment(error);
+        }
+        else {
+            context.resolveSegment(value);
+        }
+        if (generation !== this.#segmentGeneration)
+            return;
+        if (!paused && this.#activeAbortController === context.requestAbortController)
+            this.#activeAbortController = null;
+        this.#turnInputMailbox.settleSegment({
+            paused,
+            reason: {
+                pendingApprovals: state.toolExecution.pendingApprovals.length,
+                pendingBudgetInteraction: state.pendingRunBudgetInteraction !== undefined,
+                cancelled,
+                error: error instanceof Error ? error.name : error ? String(error) : undefined,
+                turnCount: state.turnCount,
+            },
+        });
+        if (!paused && this.#activeRunBudgetState === state)
+            this.#activeRunBudgetState = undefined;
+    }
+    /** Publish a paused view and suspend here until its owner reconnects or cancels. */
+    async #waitForDecision(context) {
+        if (context.session) {
+            this.#settleSegment(context, finish(context.stream, context.state, context.queue), undefined, true);
+            await context.observation;
+            const options = await context.session.interrupted(context.stream);
+            this.#assertLive(context);
+            this.#refreshContinuation(context.state, options);
+            Object.assign(context, this.#createView(context.state, { ...options, signal: context.turnSignal }, (event) => context.session.acceptEvent(event)));
+            this.#observe(context);
+            return;
+        }
+        const { state, stream, queue, options } = context;
+        await new Promise((resolve, reject) => {
+            const signal = options.signal;
+            const abort = () => {
+                context.resume = undefined;
+                reject(Object.assign(new Error('Operation aborted'), { name: 'AbortError' }));
+            };
+            context.resume = () => {
+                signal?.removeEventListener('abort', abort);
+                resolve();
+            };
+            signal?.addEventListener('abort', abort, { once: true });
+            if (signal?.aborted) {
+                abort();
                 return;
-            if (this.#activeRunBudgetState === state)
-                this.#activeRunBudgetState = undefined;
+            }
+            this.#settleSegment(context, finish(stream, state, queue));
         });
-        this.#activeRunBudgetState = state.runBudget ? state : undefined;
-        return stream;
+    }
+    /** A budget decision resumes this exact boundary without recounting it. */
+    async #waitForBudget(context) {
+        const { state } = context;
+        while (state.pendingRunBudgetInteraction) {
+            if (state.runBudgetInteractionDecision === 'stop') {
+                state.pendingRunBudgetInteraction = undefined;
+                state.runBudgetInteractionDecision = undefined;
+                return true;
+            }
+            if (state.runBudgetInteractionDecision === 'continue') {
+                state.pendingRunBudgetInteraction = undefined;
+                state.runBudgetInteractionDecision = undefined;
+                return false;
+            }
+            context.stream.interruptions = [state.pendingRunBudgetInteraction];
+            await this.#waitForDecision(context);
+        }
+        return false;
     }
     /** Route boundary advice through the same admission lane as mode notices. */
     #queuePendingSystemNotice(text) {
@@ -512,746 +608,825 @@ export class ApplicationRunLoop {
     }
     #admitPendingSteers(state, stream, queue) {
         this.#turnInputMailbox.admitAtRequestBoundary((item) => {
+            const text = typeof item.content === 'string'
+                ? item.content
+                : Array.isArray(item.content)
+                    ? item.content.map((part) => (typeof part?.text === 'string' ? part.text : '')).join('')
+                    : '';
+            if (item.role === 'user' && !text.startsWith('[Mode Notice]'))
+                state.completionRepairDeadline = undefined;
             state.history.push(item);
             state.input.push(...normalizeApplicationInput([item]));
             outputPush(stream, queue, { type: 'item', item });
         }, state.turnCount);
     }
-    async #execute(state, stream, queue, options, toolContext, requestAbortController) {
+    /** The single model → tools → model cycle, shared by root and nested turns. */
+    async #runTurn(context, session, turnSignal) {
         while (true) {
-            if (options.signal?.aborted)
-                throw Object.assign(new Error('Operation aborted'), { name: 'AbortError' });
-            if (state.pendingRunBudgetInteraction) {
-                if (state.runBudgetInteractionDecision === 'stop') {
-                    state.pendingRunBudgetInteraction = undefined;
-                    state.runBudgetInteractionDecision = undefined;
-                    return finish(stream, state, queue);
-                }
-                if (state.runBudgetInteractionDecision === 'continue') {
-                    state.pendingRunBudgetInteraction = undefined;
-                    state.runBudgetInteractionDecision = undefined;
-                }
-                else {
-                    return this.#pauseForRunBudgetInteraction(state, stream, queue);
+            try {
+                turnSignal?.throwIfAborted();
+                if (!context) {
+                    const start = await session.prepare(turnSignal);
+                    turnSignal?.throwIfAborted();
+                    if (!start)
+                        return;
+                    const options = { ...start.options, signal: turnSignal };
+                    const state = this.#createState(start.agent, start.input, options);
+                    this.#refreshContinuation(state, options);
+                    context = this.#createView(state, options, (event) => session.acceptEvent(event));
+                    context.session = session;
+                    context.turnSignal = turnSignal;
+                    state.live = context;
+                    this.#observe(context);
                 }
+                this.#assertLive(context);
+                // New explicit user authority supersedes a harness repair continuation before its deadline is applied.
+                this.#admitPendingSteers(context.state, context.stream, context.queue);
+                if (await this.#withinRepairDeadline(context, () => this.#prepareRequest(context)))
+                    return await this.#completeTurn(context);
+                const response = await this.#withinRepairDeadline(context, () => this.#askModel(context));
+                this.#assertLive(context);
+                const calls = this.#recordResponse(context, response);
+                if (await this.#withinRepairDeadline(context, () => this.#executeTools(context, response, calls)))
+                    return await this.#completeTurn(context);
+                if (calls.length === 0 && (await this.#finishTurn(context)))
+                    return await this.#completeTurn(context);
             }
-            if (state.toolExecution.pendingApprovals.length > 0 && state.approvalDecision) {
-                state.toolExecution.resolveApproval({ callId: state.approvalDecisionCallId, interruption: state.approvalDecisionInterruption }, state.approvalDecision, state.approvalMessage, state.approvals);
-                state.approvalDecision = undefined;
-                state.approvalDecisionCallId = undefined;
-                state.approvalDecisionInterruption = undefined;
-                state.approvalMessage = undefined;
-                await state.toolExecution.settle(this.#toolCallContext(state, stream, queue, toolContext));
-                stream.interruptions = state.toolExecution.pendingApprovals.map((item) => item.interruption);
-                if (state.toolExecution.pendingApprovals.length > 0) {
-                    return finish(stream, state, queue);
+            catch (error) {
+                if (context?.completionRepairExpired) {
+                    context.state.terminalCause = 'completion_verification_blocked';
+                    this.#appendStopNotice(context, 'Completion verification blocked: the repair task deadline expired. The durable goal remains active.');
+                    return await this.#completeTurn(context);
                 }
-                // Nothing is outstanding, so this segment is terminal — see
-                // `stopAfterApprovalResolution`. Stopping here rather than at the
-                // request boundary below is what keeps the just-recorded tool result
-                // in the committed history.
-                if (options.stopAfterApprovalResolution)
-                    return finish(stream, state, queue);
-            }
-            // A budget interaction can park a response after its tool calls are
-            // recorded but before any body runs. Resume those retained calls only
-            // after processing an ordinary approval decision for that same plan.
-            if (state.toolExecution.hasPlan) {
-                await state.toolExecution.settle(this.#toolCallContext(state, stream, queue, toolContext));
-                if (state.toolExecution.pendingApprovals.length) {
-                    stream.interruptions = state.toolExecution.pendingApprovals.map((item) => item.interruption);
-                    return finish(stream, state, queue);
+                if (!session)
+                    throw error;
+                if (context) {
+                    this.#settleSegment(context, undefined, error);
+                    await context.observation?.catch(() => undefined);
                 }
+                turnSignal?.throwIfAborted();
+                if (!(await session.recover(error)))
+                    return;
+                context = undefined;
             }
-            // The request boundary: every tool result of the previous round is in
-            // history, and the next request has not been built. A user message
-            // admitted here reaches the model in sequence, mid-turn.
-            if (!state.requestBoundaryReady) {
-                this.#admitPendingSteers(state, stream, queue);
-                this.#evaluateRunBudget(state, stream, queue);
-                if (state.pendingRunBudgetInteraction)
-                    return this.#pauseForRunBudgetInteraction(state, stream, queue);
-                const compactionHistoryResult = state.usesDeltaHistory
-                    ? state.compactionHistory
-                        ? mergeChainedCompactionHistory(state.compactionHistory, state.history)
-                        : { kind: 'skipped', reason: 'missing_snapshot' }
-                    : { kind: 'ready', history: state.history };
-                const boundaryDecision = options.onRequestBoundary?.(compactionHistoryResult.kind === 'ready' ? compactionHistoryResult.history : state.history, (text) => this.#queuePendingSystemNotice(text), {
-                    ...(state.lastCompletedInputTokens !== undefined
-                        ? { lastCompletedInputTokens: state.lastCompletedInputTokens }
-                        : {}),
-                    ...(compactionHistoryResult.kind === 'skipped'
-                        ? { compactionSkipReason: compactionHistoryResult.reason }
-                        : {}),
-                });
-                const compactionHistory = compactionHistoryResult.kind === 'ready' ? compactionHistoryResult.history : undefined;
-                if (options.boundaryCompaction && compactionHistory && !boundaryDecision?.deferCompaction) {
-                    const compactionStartedAt = Date.now();
-                    let costsRecordedLive = false;
-                    const compaction = await options.boundaryCompaction.compact({
-                        history: compactionHistory,
-                        instructions: state.agent.resolveInstructionsForRequest?.() ?? state.agent.instructions,
-                        tools: toModelTools(state.agent.tools),
-                        recordCostRecords: (records) => {
-                            if (records.length) {
-                                costsRecordedLive = true;
-                                state.costRecords ??= [];
-                                state.costRecords.push(...records);
-                                for (const record of records)
-                                    queue.push({ type: 'cost_update', record });
-                                this.#evaluateRunBudget(state, stream, queue);
-                            }
-                            return !state.pendingRunBudgetInteraction && !state.criticalWrapUpPending;
-                        },
-                        automaticCompactionsThisRun: state.automaticCompactionsThisRun ?? 0,
-                        lastCompletedInputTokens: state.lastCompletedInputTokens,
-                        signal: options.signal,
-                        onStarted: (provider) => outputPush(stream, queue, { type: 'context_compaction_started', provider, strategy: 'local' }),
-                    });
-                    if (compaction.kind !== 'unchanged' && !costsRecordedLive && compaction.costRecords?.length) {
+        }
+    }
+    async #withinRepairDeadline(context, operation) {
+        const deadline = context.state.completionRepairDeadline;
+        if (deadline === undefined)
+            return operation();
+        let timer;
+        try {
+            return await Promise.race([
+                new Promise((_resolve, reject) => {
+                    const expire = () => {
+                        if (context.state.completionRepairDeadline !== deadline)
+                            return;
+                        context.completionRepairExpired = true;
+                        context.requestAbortController.abort();
+                        reject(new Error('Completion repair deadline exceeded'));
+                    };
+                    if (Date.now() >= deadline)
+                        expire();
+                    else
+                        timer = setTimeout(expire, deadline - Date.now());
+                }),
+                Promise.resolve().then(() => {
+                    if (context.completionRepairExpired)
+                        throw new Error('Completion repair deadline exceeded');
+                    return operation();
+                }),
+            ]);
+        }
+        finally {
+            if (timer)
+                clearTimeout(timer);
+        }
+    }
+    #appendStopNotice(context, notice) {
+        const text = `\n\n${notice}`;
+        context.stream.finalOutput = (context.stream.finalOutput ?? '') + text;
+        context.queue.push({ type: 'text_delta', text });
+    }
+    #assertLive(context) {
+        if (context.options.signal?.aborted)
+            throw Object.assign(new Error('Operation aborted'), { name: 'AbortError' });
+    }
+    /** Admit input and prepare context once for the next logical model round. */
+    async #prepareRequest(context) {
+        let { state, stream, queue, options } = context;
+        // The request boundary: every tool result of the previous round is in
+        // history, and the next request has not been built. A user message
+        // admitted here reaches the model in sequence, mid-turn.
+        this.#admitPendingSteers(state, stream, queue);
+        this.#evaluateRunBudget(state, stream, queue);
+        if (await this.#waitForBudget(context))
+            return true;
+        ({ state, stream, queue, options } = context);
+        const compactionHistoryResult = state.usesDeltaHistory
+            ? state.compactionHistory
+                ? mergeChainedCompactionHistory(state.compactionHistory, state.history)
+                : { kind: 'skipped', reason: 'missing_snapshot' }
+            : { kind: 'ready', history: state.history };
+        const boundaryDecision = options.onRequestBoundary?.(compactionHistoryResult.kind === 'ready' ? compactionHistoryResult.history : state.history, (text) => this.#queuePendingSystemNotice(text), {
+            ...(state.lastCompletedInputTokens !== undefined
+                ? { lastCompletedInputTokens: state.lastCompletedInputTokens }
+                : {}),
+            ...(compactionHistoryResult.kind === 'skipped' ? { compactionSkipReason: compactionHistoryResult.reason } : {}),
+        });
+        const compactionHistory = compactionHistoryResult.kind === 'ready' ? compactionHistoryResult.history : undefined;
+        if (options.boundaryCompaction && compactionHistory && !boundaryDecision?.deferCompaction) {
+            const compactionStartedAt = Date.now();
+            let costsRecordedLive = false;
+            const compaction = await options.boundaryCompaction.compact({
+                history: compactionHistory,
+                instructions: state.agent.resolveInstructionsForRequest?.() ?? state.agent.instructions,
+                tools: toModelTools(state.agent.tools),
+                recordCostRecords: (records) => {
+                    if (records.length) {
+                        costsRecordedLive = true;
                         state.costRecords ??= [];
-                        state.costRecords.push(...compaction.costRecords);
-                    }
-                    if (compaction.kind === 'compacted') {
-                        state.inputUsageInvalidatedByCompaction = true;
-                        state.lastCompletedInputTokens = undefined;
-                        state.compactedDuringRun = true;
-                        state.history.splice(0, state.history.length, ...compaction.history);
-                        state.input.splice(0, state.input.length, ...normalizeApplicationInput(compaction.modelInput));
-                        state.responseId = undefined;
-                        state.responseProviderId = undefined;
-                        state.usesDeltaHistory = false;
-                        state.automaticCompactionsThisRun = (state.automaticCompactionsThisRun ?? 0) + 1;
-                        outputPush(stream, queue, {
-                            type: 'context_compaction_completed',
-                            provider: state.currentProviderId ?? 'unknown',
-                            strategy: 'local',
-                            durationMs: Math.max(0, Date.now() - compactionStartedAt),
-                        });
-                    }
-                    else if (compaction.kind === 'failed') {
-                        outputPush(stream, queue, {
-                            type: 'context_compaction_failed',
-                            provider: compaction.provider,
-                            strategy: 'local',
-                            durationMs: Math.max(0, Date.now() - compactionStartedAt),
-                        });
-                        if (compaction.error)
-                            throw compaction.error;
-                    }
-                    else if (compaction.notice && compaction.notice !== state.compactionNotice) {
-                        state.compactionNotice = compaction.notice;
-                        this.#queuePendingSystemNotice(compaction.notice);
+                        state.costRecords.push(...records);
+                        for (const record of records)
+                            queue.push({ type: 'cost_update', record });
+                        this.#evaluateRunBudget(state, stream, queue);
                     }
-                    // Summarization is an asynchronous part of this same request
-                    // boundary. Admit anything that arrived while it was in flight after
-                    // applying a replacement, so the steer cannot be overwritten by the
-                    // checkpoint or miss a terminal model request.
-                    this.#admitPendingSteers(state, stream, queue);
-                    this.#evaluateRunBudget(state, stream, queue);
-                    if (state.pendingRunBudgetInteraction)
-                        return this.#pauseForRunBudgetInteraction(state, stream, queue);
-                }
-                this.#admitPendingSteers(state, stream, queue);
-                state.turnCount += 1;
-                this.#evaluateRunBudget(state, stream, queue);
-                if (state.pendingRunBudgetInteraction) {
-                    state.requestBoundaryReady = true;
-                    return this.#pauseForRunBudgetInteraction(state, stream, queue);
-                }
-                // Retain the legacy error only for callers that have not adopted staged
-                // budgets. A configured run budget turns the same count into critical
-                // evidence and lets the decision surface choose what happens next.
-                if (!state.runBudget && state.maxTurns !== undefined && state.turnCount > state.maxTurns) {
-                    throw new MaxTurnsExceededError(state.maxTurns);
-                }
+                    return !state.pendingRunBudgetInteraction && !state.criticalWrapUpPending;
+                },
+                automaticCompactionsThisRun: state.automaticCompactionsThisRun ?? 0,
+                lastCompletedInputTokens: state.lastCompletedInputTokens,
+                signal: options.signal,
+                onStarted: (provider) => outputPush(stream, queue, { type: 'context_compaction_started', provider, strategy: 'local' }),
+            });
+            if (compaction.kind !== 'unchanged' && !costsRecordedLive && compaction.costRecords?.length) {
+                state.costRecords ??= [];
+                state.costRecords.push(...compaction.costRecords);
             }
-            else {
-                state.requestBoundaryReady = undefined;
+            if (compaction.kind === 'compacted') {
+                state.inputUsageInvalidatedByCompaction = true;
+                state.lastCompletedInputTokens = undefined;
+                state.compactedDuringRun = true;
+                state.history.splice(0, state.history.length, ...compaction.history);
+                state.input.splice(0, state.input.length, ...normalizeApplicationInput(compaction.modelInput));
+                state.responseId = undefined;
+                state.responseProviderId = undefined;
+                state.usesDeltaHistory = false;
+                state.automaticCompactionsThisRun = (state.automaticCompactionsThisRun ?? 0) + 1;
+                outputPush(stream, queue, {
+                    type: 'context_compaction_completed',
+                    provider: state.currentProviderId ?? 'unknown',
+                    strategy: 'local',
+                    durationMs: Math.max(0, Date.now() - compactionStartedAt),
+                });
             }
-            const model = await this.#deps.resolveModel(state.agent.model);
-            const maxRetries = state.agent.modelSettings?.retry?.maxRetries ?? 0;
-            let attempt = 0;
-            let streamedToolCalls = [];
-            let completion;
-            let sawToolCall = false;
-            let activeRequestId = '';
-            let activeRequest;
-            let pendingNativeReasoning;
-            let deferredNativeReasoning = [];
-            let criticalWrapUp = false;
-            while (true) {
-                streamedToolCalls = [];
-                completion = undefined;
-                sawToolCall = false;
-                pendingNativeReasoning = undefined;
-                deferredNativeReasoning = [];
-                // Stable process-unique request id, allocated immediately before dispatch
-                // so a cost record can be settled exactly once (success or failure).
-                activeRequestId = this.#nextRequestId();
-                const requestId = activeRequestId;
-                const outputLengthBefore = stream.output.length;
-                const newItemsLengthBefore = stream.newItems.length;
-                const inputLengthBefore = state.input.length;
-                const historyLengthBefore = state.history.length;
-                const criticalWrapUpDispatchedBefore = state.criticalWrapUpDispatched;
-                let provisionalTextCharacters = 0;
-                let provisionalReasoningCharacters = 0;
-                let provisionalTextDeltas = 0;
-                let provisionalReasoningDeltas = 0;
-                let lastTextPartId;
-                criticalWrapUp = state.criticalWrapUpPending === true && state.criticalWrapUpDispatched !== true;
-                if (criticalWrapUp)
-                    state.criticalWrapUpDispatched = true;
-                const disableChaining = state.disableChainingForAttempt === true;
-                if (disableChaining) {
-                    state.disableChainingForAttempt = false;
-                    state.responseId = undefined;
-                    state.responseProviderId = undefined;
-                }
-                // A per-request signal lets a recoverable guard trip cancel only this
-                // provider request; the segment signal still cancels everything.
-                const requestSignal = new AbortController();
-                const abortRequest = () => requestSignal.abort();
-                if (options.signal?.aborted)
-                    requestSignal.abort();
-                else
-                    options.signal?.addEventListener('abort', abortRequest, { once: true });
-                const requestInstructions = state.agent.resolveInstructionsForRequest?.() ?? state.agent.instructions;
-                // Same synchronous step as the instructions above, so the snapshot is
-                // exactly the state the model is shown in this request.
-                state.requestSnapshot = state.agent.resolveRequestSnapshot?.();
-                const requestCatalog = getCatalogModel(state.currentProviderId ?? 'openai', state.agent.model);
-                const effectiveOutput = resolveRequestOutput(requestCatalog, state.agent.modelSettings?.maxTokens);
-                const request = {
-                    instructions: criticalWrapUp
-                        ? `${requestInstructions}\n\nBudget containment is terminal. Do not call tools. In this one final response, summarize what you completed, the evidence you have, and what remains.`
-                        : requestInstructions,
-                    ...(state.supportsConversationChaining &&
-                        !disableChaining &&
-                        state.responseId &&
-                        state.responseProviderId !== undefined &&
-                        state.responseProviderId === state.currentProviderId
-                        ? { previousResponseId: state.responseId }
-                        : {}),
-                    ...(disableChaining ? { disableChaining: true } : {}),
-                    input: projectModelRequestHistory(state.input),
-                    tools: toModelTools(criticalWrapUp ? [] : state.agent.tools),
-                    applicationTools: criticalWrapUp ? [] : state.agent.tools,
-                    ...(state.agent.modelSettings?.temperature !== undefined
-                        ? { temperature: state.agent.modelSettings.temperature }
-                        : {}),
-                    ...(state.agent.modelSettings?.reasoning ? { reasoning: state.agent.modelSettings.reasoning } : {}),
-                    ...(effectiveOutput !== undefined ? { maxTokens: effectiveOutput } : {}),
-                    ...(state.agent.outputType !== undefined ? { outputType: state.agent.outputType } : {}),
-                    ...(state.agent.modelSettings?.codex ? { codex: state.agent.modelSettings.codex } : {}),
-                    ...(state.agent.modelSettings?.providerData
-                        ? {
-                            providerOptions: this.#contextCompactionSessionState.disabled
-                                ? Object.fromEntries(Object.entries(state.agent.modelSettings.providerData).filter(([key]) => key !== 'contextCompaction'))
-                                : state.agent.modelSettings.providerData,
-                        }
-                        : {}),
-                    signal: requestSignal.signal,
-                    ...(state.recoveryBudget ? { recoveryBudget: state.recoveryBudget } : {}),
-                };
-                activeRequest = request;
-                const generationGuard = new GenerationGuard({
-                    maxOutputCharacters: state.agent.modelSettings?.maxStreamOutputChars,
-                    maxTextCharacters: state.agent.modelSettings?.maxStreamOutputChars,
-                    maxReasoningCharacters: state.agent.modelSettings?.maxStreamOutputChars,
-                    maxToolArgumentCharacters: state.agent.modelSettings?.maxStreamOutputChars,
-                    maxCumulativeToolArgumentCharacters: state.agent.modelSettings?.maxStreamOutputChars,
-                    requestDeadlineMs: state.agent.modelSettings?.maxModelRequestDurationMs,
-                    maxStreamIdleMs: state.agent.modelSettings?.maxModelStreamIdleMs,
-                    ...options.generationGuard,
+            else if (compaction.kind === 'failed') {
+                outputPush(stream, queue, {
+                    type: 'context_compaction_failed',
+                    provider: compaction.provider,
+                    strategy: 'local',
+                    durationMs: Math.max(0, Date.now() - compactionStartedAt),
                 });
-                const consume = async () => {
-                    const deadline = new GenerationStreamDeadlines({ totalMs: generationGuard.requestDeadlineMs, idleMs: generationGuard.maxStreamIdleMs }, () => requestAbortController.abort(), () => generationGuard.progress);
-                    const toolArgumentRunaway = isLunaChainedRequest(state, request)
-                        ? new ToolArgumentRunawayGuard({
-                            ...DEFAULT_TOOL_ARGUMENT_RUNAWAY_SIGNATURE,
-                            timeoutMs: generationGuard.toolArgumentRunawayMs,
-                        }, () => requestAbortController.abort())
-                        : undefined;
-                    let iterator;
-                    let iteratorFinished = false;
-                    try {
-                        iterator = model.stream(request)[Symbol.asyncIterator]();
-                        while (true) {
-                            const nextOperation = iterator.next();
-                            const next = await deadline.wait(toolArgumentRunaway ? toolArgumentRunaway.wait(nextOperation) : nextOperation);
-                            if (next.done) {
-                                iteratorFinished = true;
-                                return;
-                            }
-                            // Any streamed event is transport activity: re-arm the inactivity
-                            // window so long legitimate reasoning is never mistaken for a stall.
-                            deadline.recordActivity();
-                            const event = next.value;
-                            if (event.type === 'completion') {
-                                toolArgumentRunaway?.observeToolCallCompleted();
-                                generationGuard.observeCompletion(event.output);
-                                completion = event;
-                                continue;
-                            }
-                            if (event.type === 'text_delta') {
-                                const startsNewPart = event.partId !== undefined &&
-                                    lastTextPartId !== undefined &&
-                                    event.partId !== lastTextPartId &&
-                                    event.text.length > 0;
-                                if (event.partId !== undefined && event.text.length > 0)
-                                    lastTextPartId = event.partId;
-                                const text = startsNewPart ? `${ASSISTANT_TEXT_PART_SEPARATOR}${event.text}` : event.text;
-                                toolArgumentRunaway?.observeText();
-                                generationGuard.observeText(text);
-                                provisionalTextCharacters += text.length;
-                                provisionalTextDeltas++;
-                                outputPush(stream, queue, { type: 'text_delta', text });
-                                continue;
-                            }
-                            if (event.type === 'codex_rate_limits') {
-                                outputPush(stream, queue, { type: 'codex_rate_limits', rateLimits: event.rateLimits });
-                                continue;
-                            }
-                            if (event.type === 'reasoning_delta') {
-                                const accepted = generationGuard.observeReasoning(event.text);
-                                if (!accepted)
-                                    continue;
-                                const forwarded = accepted === event.text ? event : { ...event, text: accepted };
-                                pendingNativeReasoning = appendNativeReasoning(pendingNativeReasoning, forwarded);
-                                provisionalReasoningCharacters += accepted.length;
-                                provisionalReasoningDeltas++;
-                                outputPush(stream, queue, { type: 'reasoning_delta', text: accepted });
-                                continue;
-                            }
-                            if (event.type === 'tool_call_streaming_delta') {
-                                toolArgumentRunaway?.observeToolArgumentProgress(event.argumentCharCount);
-                                generationGuard.observeToolArgumentProgress(event.argumentCharCount);
-                                outputPush(stream, queue, event);
-                                continue;
-                            }
-                            if (event.type === 'context_compaction_started' || event.type === 'context_compaction_completed') {
-                                outputPush(stream, queue, event);
-                                continue;
-                            }
-                            if (event.type === 'tool_call') {
-                                toolArgumentRunaway?.observeToolCallCompleted();
-                                generationGuard.observeToolCall(event.arguments);
-                                // The authoritative completion may introduce a native
-                                // replacement boundary before this call's reasoning.
-                                if (pendingNativeReasoning)
-                                    deferredNativeReasoning.push(pendingNativeReasoning);
-                                pendingNativeReasoning = undefined;
-                                sawToolCall = true;
-                                streamedToolCalls.push(event);
-                            }
-                        }
-                    }
-                    finally {
-                        deadline.dispose();
-                        toolArgumentRunaway?.dispose();
-                        if (iterator && !iteratorFinished)
-                            void iterator.return?.().catch(() => undefined);
+                if (compaction.error)
+                    throw compaction.error;
+            }
+            else if (compaction.notice && compaction.notice !== state.compactionNotice) {
+                state.compactionNotice = compaction.notice;
+                this.#queuePendingSystemNotice(compaction.notice);
+            }
+            // Summarization is an asynchronous part of this same request
+            // boundary. Admit anything that arrived while it was in flight after
+            // applying a replacement, so the steer cannot be overwritten by the
+            // checkpoint or miss a terminal model request.
+            this.#admitPendingSteers(state, stream, queue);
+            this.#evaluateRunBudget(state, stream, queue);
+            if (await this.#waitForBudget(context))
+                return true;
+            ({ state, stream, queue, options } = context);
+        }
+        this.#admitPendingSteers(state, stream, queue);
+        state.turnCount += 1;
+        this.#evaluateRunBudget(state, stream, queue);
+        if (await this.#waitForBudget(context))
+            return true;
+        ({ state, stream, queue, options } = context);
+        // Retain the legacy error only for callers that have not adopted staged
+        // budgets. A configured run budget turns the same count into critical
+        // evidence and lets the decision surface choose what happens next.
+        if (!state.runBudget && state.maxTurns !== undefined && state.turnCount > state.maxTurns) {
+            throw new MaxTurnsExceededError(state.maxTurns);
+        }
+        return false;
+    }
+    /** Own streaming and safe request attempts; never dispatch tools here. */
+    async #askModel(context) {
+        const { state, stream, queue, options, requestAbortController } = context;
+        const model = await this.#deps.resolveModel(state.agent.model);
+        const maxRetries = state.agent.modelSettings?.retry?.maxRetries ?? 0;
+        let attempt = 0;
+        let streamedToolCalls = [];
+        let completion;
+        let sawToolCall = false;
+        let activeRequestId = '';
+        let activeRequest;
+        let pendingNativeReasoning;
+        let deferredNativeReasoning = [];
+        let criticalWrapUp = false;
+        while (true) {
+            streamedToolCalls = [];
+            completion = undefined;
+            sawToolCall = false;
+            pendingNativeReasoning = undefined;
+            deferredNativeReasoning = [];
+            // Stable process-unique request id, allocated immediately before dispatch
+            // so a cost record can be settled exactly once (success or failure).
+            activeRequestId = this.#nextRequestId();
+            const requestId = activeRequestId;
+            const outputLengthBefore = stream.output.length;
+            const newItemsLengthBefore = stream.newItems.length;
+            const inputLengthBefore = state.input.length;
+            const historyLengthBefore = state.history.length;
+            const criticalWrapUpDispatchedBefore = state.criticalWrapUpDispatched;
+            let provisionalTextCharacters = 0;
+            let provisionalReasoningCharacters = 0;
+            let provisionalTextDeltas = 0;
+            let provisionalReasoningDeltas = 0;
+            let lastTextPartId;
+            criticalWrapUp = state.criticalWrapUpPending === true && state.criticalWrapUpDispatched !== true;
+            if (criticalWrapUp)
+                state.criticalWrapUpDispatched = true;
+            const disableChaining = state.disableChainingForAttempt === true;
+            if (disableChaining) {
+                state.disableChainingForAttempt = false;
+                state.responseId = undefined;
+                state.responseProviderId = undefined;
+            }
+            // A per-request signal lets a recoverable guard trip cancel only this
+            // provider request; the segment signal still cancels everything.
+            const requestSignal = new AbortController();
+            const abortRequest = () => requestSignal.abort();
+            if (options.signal?.aborted)
+                requestSignal.abort();
+            else
+                options.signal?.addEventListener('abort', abortRequest, { once: true });
+            const requestInstructions = state.agent.resolveInstructionsForRequest?.() ?? state.agent.instructions;
+            // Same synchronous step as the instructions above, so the snapshot is
+            // exactly the state the model is shown in this request.
+            state.requestSnapshot = state.agent.resolveRequestSnapshot?.();
+            const requestCatalog = getCatalogModel(state.currentProviderId ?? 'openai', state.agent.model);
+            const effectiveOutput = resolveRequestOutput(requestCatalog, state.agent.modelSettings?.maxTokens);
+            const request = {
+                instructions: criticalWrapUp
+                    ? `${requestInstructions}\n\nBudget containment is terminal. Do not call tools. In this one final response, summarize what you completed, the evidence you have, and what remains.`
+                    : requestInstructions,
+                ...(state.supportsConversationChaining &&
+                    !disableChaining &&
+                    state.responseId &&
+                    state.responseProviderId !== undefined &&
+                    state.responseProviderId === state.currentProviderId
+                    ? { previousResponseId: state.responseId }
+                    : {}),
+                ...(disableChaining ? { disableChaining: true } : {}),
+                input: projectModelRequestHistory(state.input),
+                tools: toModelTools(criticalWrapUp ? [] : state.agent.tools),
+                applicationTools: criticalWrapUp ? [] : state.agent.tools,
+                ...(state.agent.modelSettings?.temperature !== undefined
+                    ? { temperature: state.agent.modelSettings.temperature }
+                    : {}),
+                ...(state.agent.modelSettings?.reasoning ? { reasoning: state.agent.modelSettings.reasoning } : {}),
+                ...(effectiveOutput !== undefined ? { maxTokens: effectiveOutput } : {}),
+                ...(state.agent.outputType !== undefined ? { outputType: state.agent.outputType } : {}),
+                ...(state.agent.modelSettings?.codex ? { codex: state.agent.modelSettings.codex } : {}),
+                ...(state.agent.modelSettings?.providerData
+                    ? {
+                        providerOptions: this.#contextCompactionSessionState.disabled
+                            ? Object.fromEntries(Object.entries(state.agent.modelSettings.providerData).filter(([key]) => key !== 'contextCompaction'))
+                            : state.agent.modelSettings.providerData,
                     }
-                };
-                const dispatch = async () => {
-                    state.requestPreparation?.prepare(request);
-                    // A response ID can be established after a self-contained request.
-                    // It does not make this run's full history a delta needing a snapshot.
-                    const needsChainedSnapshot = Boolean(request.previousResponseId && state.usesDeltaHistory);
-                    const fullContext = needsChainedSnapshot && state.compactionHistory
-                        ? mergeChainedCompactionHistory(state.compactionHistory, state.history)
-                        : undefined;
-                    const contextCatalog = getCatalogModel(state.currentProviderId ?? 'openai', state.agent.model);
-                    enforceRequestInputLimit({
-                        contextWindow: contextCatalog?.contextWindow,
-                        maxOutputTokens: request.maxTokens,
-                        limit: state.requestInputLimit
-                            ? state.requestInputLimit()
-                            : state.agent.modelSettings?.maxRequestInputTokens,
-                        history: fullContext?.kind === 'ready'
-                            ? fullContext.history
-                            : needsChainedSnapshot
-                                ? [...(state.compactionHistory ?? []), ...request.input]
-                                : request.input,
-                        instructions: request.instructions ?? '',
-                        tools: request.tools,
-                        lastCompletedInputTokens: state.inputUsageInvalidatedByCompaction
-                            ? undefined
-                            : state.lastCompletedInputTokens,
-                        chainedContextKnown: !needsChainedSnapshot || fullContext?.kind === 'ready',
-                    });
-                    await consume();
-                };
+                    : {}),
+                signal: requestSignal.signal,
+                ...(state.recoveryBudget ? { recoveryBudget: state.recoveryBudget } : {}),
+            };
+            activeRequest = request;
+            const generationGuard = new GenerationGuard({
+                maxOutputCharacters: state.agent.modelSettings?.maxStreamOutputChars,
+                maxTextCharacters: state.agent.modelSettings?.maxStreamOutputChars,
+                maxReasoningCharacters: state.agent.modelSettings?.maxStreamOutputChars,
+                maxToolArgumentCharacters: state.agent.modelSettings?.maxStreamOutputChars,
+                maxCumulativeToolArgumentCharacters: state.agent.modelSettings?.maxStreamOutputChars,
+                requestDeadlineMs: state.agent.modelSettings?.maxModelRequestDurationMs,
+                maxStreamIdleMs: state.agent.modelSettings?.maxModelStreamIdleMs,
+                ...options.generationGuard,
+            });
+            const consume = async () => {
+                const deadline = new GenerationStreamDeadlines({ totalMs: generationGuard.requestDeadlineMs, idleMs: generationGuard.maxStreamIdleMs }, () => requestAbortController.abort(), () => generationGuard.progress);
+                const toolArgumentRunaway = isLunaChainedRequest(state, request)
+                    ? new ToolArgumentRunawayGuard({
+                        ...DEFAULT_TOOL_ARGUMENT_RUNAWAY_SIGNATURE,
+                        timeoutMs: generationGuard.toolArgumentRunawayMs,
+                    }, () => requestAbortController.abort())
+                    : undefined;
+                let iterator;
+                let iteratorFinished = false;
                 try {
-                    if (state.requestPreparation)
-                        await state.requestPreparation.run(dispatch);
-                    else
-                        await dispatch();
-                    break;
-                }
-                catch (error) {
-                    if (error instanceof RequestInputLimitError) {
-                        this.#deps.logDiagnostic?.('Request input admission refused', {
-                            code: error.code,
-                            action: error.action,
-                            reason: error.reason,
-                            guardClass: 'admission',
-                            setting: 'agent.maxRequestInputTokens',
-                            executionScope: state.hookScope ?? 'root',
-                            recovery: 'retained_work',
-                            limit: error.limit,
-                            estimatedTokens: error.estimatedTokens,
-                            observedTokens: error.observedTokens,
-                            capacity: error.capacity,
-                            outputReserve: error.outputReserve,
-                            estimationReserve: error.estimationReserve,
-                        });
-                        throw error;
-                    }
-                    if (isProviderContextOverflow(error)) {
-                        this.#observeTerminalFailure(error, {
-                            requestId: activeRequestId,
-                            provider: state.currentProviderId,
-                            model: state.agent.model,
-                            tier: resolveServiceTier(request),
-                        });
-                        throw new ProviderContextOverflowError(error);
-                    }
-                    const recoverOversizedToolArgument = isOversizedToolArgumentTrip(error) &&
-                        (state.oversizedToolArgumentRecoveries ?? 0) < MAX_OVERSIZED_TOOL_ARGUMENT_RECOVERIES &&
-                        !options.signal?.aborted;
-                    if (recoverOversizedToolArgument)
-                        requestSignal.abort();
-                    else if (error instanceof GenerationGuardError)
-                        requestAbortController.abort();
-                    const decision = classifyInLoopModelRetry(error, attempt, maxRetries, Math.random, {
-                        previousResponseId: activeRequest?.previousResponseId,
-                    });
-                    const rollbackProvisionalAttempt = () => {
-                        stream.output.splice(outputLengthBefore);
-                        if (stream.newItems !== stream.output) {
-                            stream.newItems.splice(newItemsLengthBefore);
+                    iterator = model.stream(request)[Symbol.asyncIterator]();
+                    while (true) {
+                        const nextOperation = iterator.next();
+                        const next = await deadline.wait(toolArgumentRunaway ? toolArgumentRunaway.wait(nextOperation) : nextOperation);
+                        if (next.done) {
+                            iteratorFinished = true;
+                            return;
                         }
-                        if (provisionalTextDeltas > 0 || provisionalReasoningDeltas > 0) {
-                            queue.push({
-                                type: 'model_attempt_rollback',
-                                textCharacters: provisionalTextCharacters,
-                                reasoningCharacters: provisionalReasoningCharacters,
-                                textDeltas: provisionalTextDeltas,
-                                reasoningDeltas: provisionalReasoningDeltas,
-                            });
+                        // Any streamed event is transport activity: re-arm the inactivity
+                        // window so long legitimate reasoning is never mistaken for a stall.
+                        deadline.recordActivity();
+                        const event = next.value;
+                        if (event.type === 'completion') {
+                            toolArgumentRunaway?.observeToolCallCompleted();
+                            generationGuard.observeCompletion(event.output);
+                            completion = event;
+                            continue;
                         }
-                    };
-                    if (recoverOversizedToolArgument) {
-                        // No tool ran: a call is dispatched only after its arguments
-                        // complete. Discard the partial response and ask for smaller calls.
-                        state.oversizedToolArgumentRecoveries = (state.oversizedToolArgumentRecoveries ?? 0) + 1;
-                        state.input.splice(inputLengthBefore);
-                        state.history.splice(historyLengthBefore);
-                        state.criticalWrapUpDispatched = criticalWrapUpDispatchedBefore;
-                        rollbackProvisionalAttempt();
-                        this.#deps.logDiagnostic?.('Recovering oversized tool argument in run loop', {
-                            code: error.code,
-                            recovery: state.oversizedToolArgumentRecoveries,
-                            maxRecoveries: MAX_OVERSIZED_TOOL_ARGUMENT_RECOVERIES,
-                        });
-                        this.#queuePendingSystemNotice(oversizedToolArgumentNotice(error.message));
-                        this.#admitPendingSteers(state, stream, queue);
-                        continue;
-                    }
-                    if (decision.retryable && !options.signal?.aborted) {
-                        attempt++;
-                        state.input.splice(inputLengthBefore);
-                        state.history.splice(historyLengthBefore);
-                        state.criticalWrapUpDispatched = criticalWrapUpDispatchedBefore;
-                        rollbackProvisionalAttempt();
-                        if (decision.kind === 'chain_recovery') {
-                            state.disableChainingForAttempt = true;
-                            state.responseId = undefined;
-                            state.responseProviderId = undefined;
+                        if (event.type === 'text_delta') {
+                            const startsNewPart = event.partId !== undefined &&
+                                lastTextPartId !== undefined &&
+                                event.partId !== lastTextPartId &&
+                                event.text.length > 0;
+                            if (event.partId !== undefined && event.text.length > 0)
+                                lastTextPartId = event.partId;
+                            const text = startsNewPart ? `${ASSISTANT_TEXT_PART_SEPARATOR}${event.text}` : event.text;
+                            toolArgumentRunaway?.observeText();
+                            generationGuard.observeText(text);
+                            provisionalTextCharacters += text.length;
+                            provisionalTextDeltas++;
+                            outputPush(stream, queue, { type: 'text_delta', text });
+                            continue;
+                        }
+                        if (event.type === 'codex_rate_limits') {
+                            outputPush(stream, queue, { type: 'codex_rate_limits', rateLimits: event.rateLimits });
+                            continue;
+                        }
+                        if (event.type === 'reasoning_delta') {
+                            const accepted = generationGuard.observeReasoning(event.text);
+                            if (!accepted)
+                                continue;
+                            const forwarded = accepted === event.text ? event : { ...event, text: accepted };
+                            pendingNativeReasoning = appendNativeReasoning(pendingNativeReasoning, forwarded);
+                            provisionalReasoningCharacters += accepted.length;
+                            provisionalReasoningDeltas++;
+                            outputPush(stream, queue, { type: 'reasoning_delta', text: accepted });
+                            continue;
+                        }
+                        if (event.type === 'tool_call_streaming_delta') {
+                            toolArgumentRunaway?.observeToolArgumentProgress(event.argumentCharCount);
+                            generationGuard.observeToolArgumentProgress(event.argumentCharCount);
+                            outputPush(stream, queue, event);
+                            continue;
+                        }
+                        if (event.type === 'context_compaction_started' || event.type === 'context_compaction_completed') {
+                            outputPush(stream, queue, event);
+                            continue;
+                        }
+                        if (event.type === 'tool_call') {
+                            toolArgumentRunaway?.observeToolCallCompleted();
+                            generationGuard.observeToolCall(event.arguments);
+                            // The authoritative completion may introduce a native
+                            // replacement boundary before this call's reasoning.
+                            if (pendingNativeReasoning)
+                                deferredNativeReasoning.push(pendingNativeReasoning);
+                            pendingNativeReasoning = undefined;
+                            sawToolCall = true;
+                            streamedToolCalls.push(event);
                         }
-                        this.#deps.logDiagnostic?.('Retrying model request in run loop', {
-                            attempt,
-                            maxRetries,
-                            kind: decision.kind,
-                            delayMs: decision.delayMs,
-                            error: describeError(error),
-                        });
-                        await (this.#deps.waitBeforeModelRetry ?? sleepWithAbort)(decision.delayMs, options.signal);
-                        continue;
-                    }
-                    // Session recovery owns full-history rebuild for a chained delta.
-                    // Retract this cycle's provisional tokens so a later chain_recovery
-                    // does not append a second answer after a half-sentence.
-                    if (!decision.retryable && decision.reason === 'chained_delta_not_self_contained') {
-                        rollbackProvisionalAttempt();
                     }
-                    // Dispatch began, but no terminal completion was accepted. Record an
-                    // unpriced marker so the summary stays honest (partial) rather than
-                    // appearing exact. Observational only: the error still propagates.
-                    const cancelled = error instanceof Error && error.name === 'AbortError';
-                    const record = this.#appendCostRecord(state, {
-                        requestId,
-                        provider: state.currentProviderId,
-                        model: state.agent.model,
-                        tier: resolveServiceTier(request),
-                        outcome: cancelled ? 'cancelled' : 'failed',
-                    });
-                    // Only the live queue sees cost records: `output`/`newItems` feed
-                    // persistence and replay, and a cost event must never become a
-                    // restored history item.
-                    queue.push({ type: 'cost_update', record });
-                    this.#evaluateRunBudget(state, stream, queue);
-                    this.#observeTerminalFailure(error, {
-                        requestId,
-                        provider: state.currentProviderId,
-                        model: state.agent.model,
-                        tier: resolveServiceTier(request),
-                    });
-                    throw error;
                 }
                 finally {
-                    options.signal?.removeEventListener('abort', abortRequest);
+                    deadline.dispose();
+                    toolArgumentRunaway?.dispose();
+                    if (iterator && !iteratorFinished)
+                        void iterator.return?.().catch(() => undefined);
                 }
+            };
+            const dispatch = async () => {
+                state.requestPreparation?.prepare(request);
+                // A response ID can be established after a self-contained request.
+                // It does not make this run's full history a delta needing a snapshot.
+                const needsChainedSnapshot = Boolean(request.previousResponseId && state.usesDeltaHistory);
+                const fullContext = needsChainedSnapshot && state.compactionHistory
+                    ? mergeChainedCompactionHistory(state.compactionHistory, state.history)
+                    : undefined;
+                const contextCatalog = getCatalogModel(state.currentProviderId ?? 'openai', state.agent.model);
+                enforceRequestInputLimit({
+                    contextWindow: contextCatalog?.contextWindow,
+                    maxOutputTokens: request.maxTokens,
+                    limit: state.requestInputLimit ? state.requestInputLimit() : state.agent.modelSettings?.maxRequestInputTokens,
+                    history: fullContext?.kind === 'ready'
+                        ? fullContext.history
+                        : needsChainedSnapshot
+                            ? [...(state.compactionHistory ?? []), ...request.input]
+                            : request.input,
+                    instructions: request.instructions ?? '',
+                    tools: request.tools,
+                    lastCompletedInputTokens: state.inputUsageInvalidatedByCompaction
+                        ? undefined
+                        : state.lastCompletedInputTokens,
+                    chainedContextKnown: !needsChainedSnapshot || fullContext?.kind === 'ready',
+                });
+                await consume();
+            };
+            try {
+                if (state.requestPreparation)
+                    await state.requestPreparation.run(dispatch);
+                else
+                    await dispatch();
+                break;
             }
-            if (!completion) {
-                // A few application-owned model fixtures (and legacy adapters) end a
-                // tool-call stream immediately after the call and use the next request
-                // as the continuation boundary. Preserve that contract while routing
-                // the calls through the same ordered dispatcher. An empty stream still
-                // fails closed as an incomplete model turn.
-                if (streamedToolCalls.length === 0) {
-                    const error = new Error('Application model turn ended without completion');
+            catch (error) {
+                if (error instanceof RequestInputLimitError) {
+                    this.#deps.logDiagnostic?.('Request input admission refused', {
+                        code: error.code,
+                        action: error.action,
+                        reason: error.reason,
+                        guardClass: 'admission',
+                        setting: 'agent.maxRequestInputTokens',
+                        executionScope: state.hookScope ?? 'root',
+                        recovery: 'retained_work',
+                        limit: error.limit,
+                        estimatedTokens: error.estimatedTokens,
+                        observedTokens: error.observedTokens,
+                        capacity: error.capacity,
+                        outputReserve: error.outputReserve,
+                        estimationReserve: error.estimationReserve,
+                    });
+                    throw error;
+                }
+                if (isProviderContextOverflow(error)) {
                     this.#observeTerminalFailure(error, {
                         requestId: activeRequestId,
                         provider: state.currentProviderId,
                         model: state.agent.model,
-                        tier: activeRequest ? resolveServiceTier(activeRequest) : 'standard',
+                        tier: resolveServiceTier(request),
                     });
-                    throw error;
-                }
-                if (criticalWrapUp) {
-                    return finish(stream, state, queue);
+                    throw new ProviderContextOverflowError(error);
                 }
-                for (const reasoning of deferredNativeReasoning)
-                    commitPendingNativeReasoning(state, stream, queue, reasoning);
-                pendingNativeReasoning = commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
-                await this.#dispatchToolCalls(state, stream, queue, streamedToolCalls, toolContext);
-                if (state.toolExecution.terminateAfterExecution)
-                    return finish(stream, state, queue);
-                if (state.toolExecution.pendingApprovals.length > 0) {
-                    stream.interruptions = state.toolExecution.pendingApprovals.map((item) => item.interruption);
-                    return finish(stream, state, queue);
-                }
-                continue;
-            }
-            const finalCompletion = completion;
-            // A terminal response ends recovery, unlike partial tokens or tool
-            // activity. Later failures must not inherit an old episode's deadline.
-            state.recoveryBudget?.noteModelResponseCompleted();
-            // Commit the authoritative terminal state before handling terminal-only
-            // tool calls. Approval may pause immediately after this point, and the
-            // continuation must still carry the response that produced the calls.
-            // Response IDs are provider-chain state, not generic run metadata. Keep
-            // them out of continuation handles for providers that do not support
-            // chaining, while still exposing the completed provider response on the
-            // current stream for diagnostics.
-            if (state.supportsConversationChaining && state.currentProviderId !== undefined) {
-                state.responseId = finalCompletion.responseId;
-                state.responseProviderId = state.currentProviderId;
-            }
-            else {
-                state.responseId = undefined;
-                state.responseProviderId = undefined;
-            }
-            let normalizedCompletionUsage;
-            const upstreamProvider = extractAiSdkUpstreamProvider(finalCompletion.providerMetadata);
-            if (finalCompletion.usage !== undefined) {
-                normalizedCompletionUsage = normalizeModelUsage(finalCompletion.usage);
-                if (upstreamProvider) {
-                    if (normalizedCompletionUsage) {
-                        normalizedCompletionUsage.upstream_provider = upstreamProvider;
+                const recoverOversizedToolArgument = isOversizedToolArgumentTrip(error) &&
+                    (state.oversizedToolArgumentRecoveries ?? 0) < MAX_OVERSIZED_TOOL_ARGUMENT_RECOVERIES &&
+                    !options.signal?.aborted;
+                if (recoverOversizedToolArgument)
+                    requestSignal.abort();
+                else if (error instanceof GenerationGuardError)
+                    requestAbortController.abort();
+                const decision = classifyInLoopModelRetry(error, attempt, maxRetries, Math.random, {
+                    previousResponseId: activeRequest?.previousResponseId,
+                });
+                const rollbackProvisionalAttempt = () => {
+                    stream.output.splice(outputLengthBefore);
+                    if (stream.newItems !== stream.output) {
+                        stream.newItems.splice(newItemsLengthBefore);
                     }
-                    else {
-                        normalizedCompletionUsage = { upstream_provider: upstreamProvider };
+                    if (provisionalTextDeltas > 0 || provisionalReasoningDeltas > 0) {
+                        queue.push({
+                            type: 'model_attempt_rollback',
+                            textCharacters: provisionalTextCharacters,
+                            reasoningCharacters: provisionalReasoningCharacters,
+                            textDeltas: provisionalTextDeltas,
+                            reasoningDeltas: provisionalReasoningDeltas,
+                        });
                     }
+                };
+                if (recoverOversizedToolArgument) {
+                    // No tool ran: a call is dispatched only after its arguments
+                    // complete. Discard the partial response and ask for smaller calls.
+                    state.oversizedToolArgumentRecoveries = (state.oversizedToolArgumentRecoveries ?? 0) + 1;
+                    state.input.splice(inputLengthBefore);
+                    state.history.splice(historyLengthBefore);
+                    state.criticalWrapUpDispatched = criticalWrapUpDispatchedBefore;
+                    rollbackProvisionalAttempt();
+                    this.#deps.logDiagnostic?.('Recovering oversized tool argument in run loop', {
+                        code: error.code,
+                        recovery: state.oversizedToolArgumentRecoveries,
+                        maxRecoveries: MAX_OVERSIZED_TOOL_ARGUMENT_RECOVERIES,
+                    });
+                    this.#queuePendingSystemNotice(oversizedToolArgumentNotice(error.message));
+                    this.#admitPendingSteers(state, stream, queue);
+                    continue;
                 }
-                if (normalizedCompletionUsage) {
-                    if (normalizedCompletionUsage.prompt_tokens !== undefined) {
-                        state.lastCompletedInputTokens = normalizedCompletionUsage.prompt_tokens;
-                        state.inputUsageInvalidatedByCompaction = false;
+                if (decision.retryable && !options.signal?.aborted) {
+                    attempt++;
+                    state.input.splice(inputLengthBefore);
+                    state.history.splice(historyLengthBefore);
+                    state.criticalWrapUpDispatched = criticalWrapUpDispatchedBefore;
+                    rollbackProvisionalAttempt();
+                    if (decision.kind === 'chain_recovery') {
+                        state.disableChainingForAttempt = true;
+                        state.responseId = undefined;
+                        state.responseProviderId = undefined;
                     }
-                    const accumulated = addTokenUsage(normalizeModelUsage(state.usage), normalizedCompletionUsage);
-                    state.usage = {
-                        ...(accumulated.prompt_tokens !== undefined ? { inputTokens: accumulated.prompt_tokens } : {}),
-                        ...(accumulated.completion_tokens !== undefined ? { outputTokens: accumulated.completion_tokens } : {}),
-                        // `total_tokens` is the billable normalized total, including cache
-                        // creation tokens. Keep it alongside the application aliases so the
-                        // authoritative run accumulator never reconstructs a smaller total.
-                        ...(accumulated.total_tokens !== undefined ? { totalTokens: accumulated.total_tokens } : {}),
-                        ...(accumulated.cache_read_tokens !== undefined
-                            ? { cachedInputTokens: accumulated.cache_read_tokens }
-                            : {}),
-                        ...(accumulated.cache_creation_tokens !== undefined
-                            ? { cacheWriteTokens: accumulated.cache_creation_tokens }
-                            : {}),
-                    };
+                    this.#deps.logDiagnostic?.('Retrying model request in run loop', {
+                        attempt,
+                        maxRetries,
+                        kind: decision.kind,
+                        delayMs: decision.delayMs,
+                        error: describeError(error),
+                    });
+                    await (this.#deps.waitBeforeModelRetry ?? sleepWithAbort)(decision.delayMs, options.signal);
+                    continue;
                 }
-                else {
-                    // Providers may carry a future/opaque usage shape. Preserve the old
-                    // pass-through behavior when normalization cannot recognize it.
-                    state.usage = finalCompletion.usage;
+                // Session recovery owns full-history rebuild for a chained delta.
+                // Retract this cycle's provisional tokens so a later chain_recovery
+                // does not append a second answer after a half-sentence.
+                if (!decision.retryable && decision.reason === 'chained_delta_not_self_contained') {
+                    rollbackProvisionalAttempt();
                 }
+                // Dispatch began, but no terminal completion was accepted. Record an
+                // unpriced marker so the summary stays honest (partial) rather than
+                // appearing exact. Observational only: the error still propagates.
+                const cancelled = error instanceof Error && error.name === 'AbortError';
+                const record = this.#appendCostRecord(state, {
+                    requestId,
+                    provider: state.currentProviderId,
+                    model: state.agent.model,
+                    tier: resolveServiceTier(request),
+                    outcome: cancelled ? 'cancelled' : 'failed',
+                });
+                // Only the live queue sees cost records: `output`/`newItems` feed
+                // persistence and replay, and a cost event must never become a
+                // restored history item.
+                queue.push({ type: 'cost_update', record });
+                this.#evaluateRunBudget(state, stream, queue);
+                this.#observeTerminalFailure(error, {
+                    requestId,
+                    provider: state.currentProviderId,
+                    model: state.agent.model,
+                    tier: resolveServiceTier(request),
+                });
+                throw error;
             }
-            else if (upstreamProvider) {
-                normalizedCompletionUsage = { upstream_provider: upstreamProvider };
-            }
-            else {
-                normalizedCompletionUsage = undefined;
+            finally {
+                options.signal?.removeEventListener('abort', abortRequest);
             }
-            // Attribute cost while the dispatched request's identity is still known.
-            // Emit it immediately so the UI can show this request's cost before the
-            // run ends; the terminal result re-delivers the cumulative record list
-            // and the session accumulator ignores the duplicates by request id.
-            // Only the live queue sees the event: `output`/`newItems` feed
-            // persistence and replay, and a cost event must never become a restored
-            // history item.
-            const record = this.#appendCostRecord(state, {
-                requestId: activeRequestId,
-                provider: state.currentProviderId,
-                model: state.agent.model,
-                tier: activeRequest ? resolveServiceTier(activeRequest) : 'standard',
-                outcome: 'completed',
-                usage: normalizedCompletionUsage,
-                providerUsd: finalCompletion.costUsd,
-            });
-            queue.push({ type: 'cost_update', record });
-            // Cost and usage arrive in the same completion metadata, so surface both
-            // live. The footer is a per-request indicator: emit this request's own
-            // usage, never the run accumulator in `state.usage`.
-            if (normalizedCompletionUsage)
-                queue.push({ type: 'usage_update', usage: normalizedCompletionUsage });
-            this.#evaluateRunBudget(state, stream, queue);
-            stream.lastResponseId = finalCompletion.responseId;
-            stream.rawResponses?.push(finalCompletion);
-            // Some provider adapters report function calls only in the terminal
-            // completion rather than as separate stream events. Their reasoning may
-            // likewise be terminal-only, so associate it before replaying calls.
-            const toolCalls = [...streamedToolCalls];
-            const hasTerminalReasoning = finalCompletion.output.some((item) => item.type === 'reasoning' && appendNativeReasoning(undefined, item) !== undefined);
-            if (hasTerminalReasoning) {
-                // Terminal output supplies native ordering and authoritative metadata.
-                // Streamed fragments have already been displayed, but are not history.
-                pendingNativeReasoning = undefined;
+        }
+        return {
+            streamedToolCalls,
+            completion,
+            sawToolCall,
+            activeRequestId,
+            activeRequest,
+            pendingNativeReasoning,
+            deferredNativeReasoning,
+            criticalWrapUp,
+        };
+    }
+    /** Accept response facts before any tool from that response can execute. */
+    #recordResponse(context, response) {
+        const { state, stream, queue } = context;
+        const { streamedToolCalls, completion, sawToolCall, activeRequestId, activeRequest, deferredNativeReasoning, criticalWrapUp, } = response;
+        let { pendingNativeReasoning } = response;
+        if (!completion) {
+            if (streamedToolCalls.length === 0) {
+                const error = new Error('Application model turn ended without completion');
+                this.#observeTerminalFailure(error, {
+                    requestId: activeRequestId,
+                    provider: state.currentProviderId,
+                    model: state.agent.model,
+                    tier: activeRequest ? resolveServiceTier(activeRequest) : 'standard',
+                });
+                throw error;
             }
-            else {
+            if (!criticalWrapUp) {
                 for (const reasoning of deferredNativeReasoning)
                     commitPendingNativeReasoning(state, stream, queue, reasoning);
+                commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
             }
-            for (const item of finalCompletion.output) {
-                if (item.type === 'reasoning') {
-                    // Each terminal native item is independently replayable; encrypted
-                    // reasoning cannot be concatenated or overwritten by its sibling.
-                    pendingNativeReasoning = commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
-                    pendingNativeReasoning = appendNativeReasoning(undefined, item);
-                    continue;
-                }
-                if (item.type === 'tool_call') {
-                    pendingNativeReasoning = commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
-                    if (!sawToolCall)
-                        toolCalls.push(item);
-                    continue;
+            return criticalWrapUp ? [] : streamedToolCalls;
+        }
+        const finalCompletion = completion;
+        // A terminal response ends recovery, unlike partial tokens or tool
+        // activity. Later failures must not inherit an old episode's deadline.
+        state.recoveryBudget?.noteModelResponseCompleted();
+        // Commit the authoritative terminal state before handling terminal-only
+        // tool calls. Approval may pause immediately after this point, and the
+        // continuation must still carry the response that produced the calls.
+        // Response IDs are provider-chain state, not generic run metadata. Keep
+        // them out of continuation handles for providers that do not support
+        // chaining, while still exposing the completed provider response on the
+        // current stream for diagnostics.
+        if (state.supportsConversationChaining && state.currentProviderId !== undefined) {
+            state.responseId = finalCompletion.responseId;
+            state.responseProviderId = state.currentProviderId;
+        }
+        else {
+            state.responseId = undefined;
+            state.responseProviderId = undefined;
+        }
+        let normalizedCompletionUsage;
+        const upstreamProvider = extractAiSdkUpstreamProvider(finalCompletion.providerMetadata);
+        if (finalCompletion.usage !== undefined) {
+            normalizedCompletionUsage = normalizeModelUsage(finalCompletion.usage);
+            if (upstreamProvider) {
+                if (normalizedCompletionUsage) {
+                    normalizedCompletionUsage.upstream_provider = upstreamProvider;
                 }
-                if (item.type !== 'provider_opaque')
-                    continue;
-                pendingNativeReasoning = commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
-                const historyItem = {
-                    ...item.item,
-                    providerOpaque: {
-                        provider: item.provider,
-                        ...(item.sourceProvider ? { sourceProvider: item.sourceProvider } : {}),
-                    },
-                };
-                state.history.push(historyItem);
-                state.input.push(item);
-                if (isOpenAICompaction(item)) {
-                    // Completion usage describes the request before native replacement.
-                    // Retain billed usage/costs; invalidate only the next admission hint.
-                    state.lastCompletedInputTokens = undefined;
-                    state.inputUsageInvalidatedByCompaction = true;
+                else {
+                    normalizedCompletionUsage = { upstream_provider: upstreamProvider };
                 }
-                outputPush(stream, queue, { type: 'item', item });
             }
-            sawToolCall = toolCalls.length > 0;
-            pendingNativeReasoning = commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
-            const assistantText = finalCompletion.output
-                .filter((item) => item.type === 'message')
-                .flatMap((item) => item.content)
-                // Whitespace-only output (e.g. a background-notification acknowledgment)
-                // must not enter the canonical history: the truthy check used to let
-                // "\n\n" through, accumulating junk in model context.
-                .map((part) => part.text)
-                .filter((text) => text.length > 0)
-                .join(ASSISTANT_TEXT_PART_SEPARATOR)
-                .trim();
-            if (assistantText) {
-                stream.finalOutput = assistantText;
-                const item = {
-                    type: 'message',
-                    role: 'assistant',
-                    content: [{ type: 'output_text', text: assistantText }],
+            if (normalizedCompletionUsage) {
+                if (normalizedCompletionUsage.prompt_tokens !== undefined) {
+                    state.lastCompletedInputTokens = normalizedCompletionUsage.prompt_tokens;
+                    state.inputUsageInvalidatedByCompaction = false;
+                }
+                const accumulated = addTokenUsage(normalizeModelUsage(state.usage), normalizedCompletionUsage);
+                state.usage = {
+                    ...(accumulated.prompt_tokens !== undefined ? { inputTokens: accumulated.prompt_tokens } : {}),
+                    ...(accumulated.completion_tokens !== undefined ? { outputTokens: accumulated.completion_tokens } : {}),
+                    // `total_tokens` is the billable normalized total, including cache
+                    // creation tokens. Keep it alongside the application aliases so the
+                    // authoritative run accumulator never reconstructs a smaller total.
+                    ...(accumulated.total_tokens !== undefined ? { totalTokens: accumulated.total_tokens } : {}),
+                    ...(accumulated.cache_read_tokens !== undefined ? { cachedInputTokens: accumulated.cache_read_tokens } : {}),
+                    ...(accumulated.cache_creation_tokens !== undefined
+                        ? { cacheWriteTokens: accumulated.cache_creation_tokens }
+                        : {}),
                 };
-                state.history.push(item);
-                outputPush(stream, queue, { type: 'item', item });
-                state.input.push({ type: 'message', role: 'assistant', content: [{ type: 'text', text: assistantText }] });
             }
-            // Tool calls are dispatched only after the reasoning, provider-opaque,
-            // and assistant-text items from the same completion are committed. A
-            // response that carries both prose and a tool call must serialize as
-            // text -> function_call -> function_call_result; dispatching first left
-            // the text stranded after its own tool result, where the assistant
-            // message merger glued it onto the *next* turn's tool call and every
-            // request ended with a bare assistant message. Models read that as a
-            // turn truncated mid-sentence and restart instead of progressing.
-            if (!state.criticalWrapUpPending && toolCalls.length > 0) {
-                await this.#dispatchToolCalls(state, stream, queue, toolCalls, toolContext);
+            else {
+                // Providers may carry a future/opaque usage shape. Preserve the old
+                // pass-through behavior when normalization cannot recognize it.
+                state.usage = finalCompletion.usage;
             }
-            if (state.toolExecution.terminateAfterExecution)
-                return finish(stream, state, queue);
-            // A critical subagent gets exactly this final tool-free model call.
-            if (criticalWrapUp)
-                return finish(stream, state, queue);
-            if (state.pendingRunBudgetInteraction && toolCalls.length > 0) {
-                return this.#pauseForRunBudgetInteraction(state, stream, queue);
+        }
+        else if (upstreamProvider) {
+            normalizedCompletionUsage = { upstream_provider: upstreamProvider };
+        }
+        else {
+            normalizedCompletionUsage = undefined;
+        }
+        // Attribute cost while the dispatched request's identity is still known.
+        // Emit it immediately so the UI can show this request's cost before the
+        // run ends; the terminal result re-delivers the cumulative record list
+        // and the session accumulator ignores the duplicates by request id.
+        // Only the live queue sees the event: `output`/`newItems` feed
+        // persistence and replay, and a cost event must never become a restored
+        // history item.
+        const record = this.#appendCostRecord(state, {
+            requestId: activeRequestId,
+            provider: state.currentProviderId,
+            model: state.agent.model,
+            tier: activeRequest ? resolveServiceTier(activeRequest) : 'standard',
+            outcome: 'completed',
+            usage: normalizedCompletionUsage,
+            providerUsd: finalCompletion.costUsd,
+        });
+        queue.push({ type: 'cost_update', record });
+        // Cost and usage arrive in the same completion metadata, so surface both
+        // live. The footer is a per-request indicator: emit this request's own
+        // usage, never the run accumulator in `state.usage`.
+        if (normalizedCompletionUsage)
+            queue.push({ type: 'usage_update', usage: normalizedCompletionUsage });
+        this.#evaluateRunBudget(state, stream, queue);
+        stream.lastResponseId = finalCompletion.responseId;
+        stream.rawResponses?.push(finalCompletion);
+        // Some provider adapters report function calls only in the terminal
+        // completion rather than as separate stream events. Their reasoning may
+        // likewise be terminal-only, so associate it before replaying calls.
+        const toolCalls = [...streamedToolCalls];
+        const hasTerminalReasoning = finalCompletion.output.some((item) => item.type === 'reasoning' && appendNativeReasoning(undefined, item) !== undefined);
+        if (hasTerminalReasoning) {
+            // Terminal output supplies native ordering and authoritative metadata.
+            // Streamed fragments have already been displayed, but are not history.
+            pendingNativeReasoning = undefined;
+        }
+        else {
+            for (const reasoning of deferredNativeReasoning)
+                commitPendingNativeReasoning(state, stream, queue, reasoning);
+        }
+        for (const item of finalCompletion.output) {
+            if (item.type === 'reasoning') {
+                // Each terminal native item is independently replayable; encrypted
+                // reasoning cannot be concatenated or overwritten by its sibling.
+                pendingNativeReasoning = commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
+                pendingNativeReasoning = appendNativeReasoning(undefined, item);
+                continue;
             }
+            if (item.type === 'tool_call') {
+                pendingNativeReasoning = commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
+                if (!sawToolCall)
+                    toolCalls.push(item);
+                continue;
+            }
+            if (item.type !== 'provider_opaque')
+                continue;
+            pendingNativeReasoning = commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
+            const historyItem = {
+                ...item.item,
+                providerOpaque: {
+                    provider: item.provider,
+                    ...(item.sourceProvider ? { sourceProvider: item.sourceProvider } : {}),
+                },
+            };
+            state.history.push(historyItem);
+            state.input.push(item);
+            if (isOpenAICompaction(item)) {
+                // Completion usage describes the request before native replacement.
+                // Retain billed usage/costs; invalidate only the next admission hint.
+                state.lastCompletedInputTokens = undefined;
+                state.inputUsageInvalidatedByCompaction = true;
+            }
+            outputPush(stream, queue, { type: 'item', item });
+        }
+        pendingNativeReasoning = commitPendingNativeReasoning(state, stream, queue, pendingNativeReasoning);
+        const assistantText = finalCompletion.output
+            .filter((item) => item.type === 'message')
+            .flatMap((item) => item.content)
+            // Whitespace-only output (e.g. a background-notification acknowledgment)
+            // must not enter the canonical history: the truthy check used to let
+            // "\n\n" through, accumulating junk in model context.
+            .map((part) => part.text)
+            .filter((text) => text.length > 0)
+            .join(ASSISTANT_TEXT_PART_SEPARATOR)
+            .trim();
+        if (assistantText) {
+            stream.finalOutput = assistantText;
+            const item = {
+                type: 'message',
+                role: 'assistant',
+                content: [{ type: 'output_text', text: assistantText }],
+            };
+            state.history.push(item);
+            outputPush(stream, queue, { type: 'item', item });
+            state.input.push({ type: 'message', role: 'assistant', content: [{ type: 'text', text: assistantText }] });
+        }
+        return toolCalls;
+    }
+    /** Execute and settle one tool plan; approvals suspend this operation in place. */
+    async #executeTools(context, response, calls) {
+        let { state, stream, queue, options, toolContext } = context;
+        if (response.criticalWrapUp)
+            return true;
+        if (!state.criticalWrapUpPending && calls.length > 0) {
+            await this.#dispatchToolCalls(state, stream, queue, calls, toolContext);
+        }
+        if (state.toolExecution.terminateAfterExecution)
+            return true;
+        if (calls.length === 0)
+            return false;
+        while (state.pendingRunBudgetInteraction || state.toolExecution.pendingApprovals.length > 0) {
+            if (await this.#waitForBudget(context))
+                return true;
+            ({ state, stream, queue, options, toolContext } = context);
+            let answeredApproval = false;
             if (state.toolExecution.pendingApprovals.length > 0) {
                 stream.interruptions = state.toolExecution.pendingApprovals.map((item) => item.interruption);
-                return finish(stream, state, queue);
-            }
-            if (!sawToolCall) {
-                // The normal stop seam. A stop policy may keep the same run going; the
-                // reminder is admitted at the next request boundary, where pending
-                // budget evidence pauses first and turn accounting applies as usual.
-                const stopDecision = options.onNormalStop && !options.signal?.aborted && !state.criticalWrapUpPending
-                    ? options.onNormalStop(state.history, state.agent)
-                    : undefined;
-                if (stopDecision) {
-                    this.#deps.logDiagnostic?.('Normal stop policy decided', { action: stopDecision.action, turnCount: state.turnCount, ...stopDecision.diagnostics }, { severity: 'debug', eventType: 'run_loop.normal_stop_policy' });
+                await this.#waitForDecision(context);
+                ({ state, stream, queue, options, toolContext } = context);
+                if (state.approvalDecision) {
+                    state.toolExecution.resolveApproval({ callId: state.approvalDecisionCallId, interruption: state.approvalDecisionInterruption }, state.approvalDecision, state.approvalMessage, state.approvals);
+                    state.approvalDecision = undefined;
+                    state.approvalDecisionCallId = undefined;
+                    state.approvalDecisionInterruption = undefined;
+                    state.approvalMessage = undefined;
+                    answeredApproval = true;
                 }
-                if (stopDecision?.action === 'continue') {
-                    this.#queuePendingSystemNotice(stopDecision.reminder);
-                    continue;
-                }
-                if (stopDecision?.terminalCause)
-                    state.terminalCause = stopDecision.terminalCause;
-                // Evidence discovered when a response has already completed needs no
-                // human boundary: there is no later request or tool to block.
-                state.pendingRunBudgetInteraction = undefined;
-                return finish(stream, state, queue);
             }
+            this.#assertLive(context);
+            if (state.toolExecution.hasPlan) {
+                await state.toolExecution.settle(this.#toolCallContext(state, stream, queue, toolContext));
+            }
+            stream.interruptions = state.toolExecution.pendingApprovals.map((item) => item.interruption);
+            if (answeredApproval && options.stopAfterApprovalResolution && stream.interruptions.length === 0)
+                return true;
+            if (state.toolExecution.terminateAfterExecution)
+                return true;
+        }
+        return false;
+    }
+    /** Normal no-tool completion; only an explicit stop-policy continuation repeats. */
+    async #finishTurn(context) {
+        const { state, options } = context;
+        // The normal stop seam. A stop policy may keep the same run going; the
+        // reminder is admitted at the next request boundary, where pending
+        // budget evidence pauses first and turn accounting applies as usual.
+        const stopDecision = options.onNormalStop && !options.signal?.aborted && !state.criticalWrapUpPending
+            ? await options.onNormalStop(state.history, state.agent, context.requestAbortController.signal)
+            : undefined;
+        if (stopDecision) {
+            this.#deps.logDiagnostic?.('Normal stop policy decided', { action: stopDecision.action, turnCount: state.turnCount, ...stopDecision.diagnostics }, { severity: 'debug', eventType: 'run_loop.normal_stop_policy' });
+        }
+        if (stopDecision?.action === 'continue') {
+            if (stopDecision.repairDeadline !== undefined)
+                state.completionRepairDeadline = stopDecision.repairDeadline;
+            this.#queuePendingSystemNotice(stopDecision.reminder);
+            return false;
         }
+        if (stopDecision?.terminalCause)
+            state.terminalCause = stopDecision.terminalCause;
+        if (stopDecision?.action === 'stop' && stopDecision.notice) {
+            this.#appendStopNotice(context, stopDecision.notice);
+        }
+        // Evidence discovered when a response has already completed needs no
+        // human boundary: there is no later request or tool to block.
+        state.pendingRunBudgetInteraction = undefined;
+        return true;
     }
     async #dispatchToolCalls(state, stream, queue, events, toolContext) {
         const context = this.#toolCallContext(state, stream, queue, toolContext);
@@ -1288,6 +1463,7 @@ export class ApplicationRunLoop {
                 });
                 outputPush(stream, queue, { type: 'item', item: callItem });
             },
+            onExecutionStarted: (callId) => state.live?.session?.acceptToolDispatch(callId),
             onDispatch: (entry) => {
                 outputPush(stream, queue, {
                     type: 'tool_call_dispatched',
@@ -1388,13 +1564,6 @@ export class ApplicationRunLoop {
     #requiresHumanBudgetDecision(event) {
         return event.type === 'tool_stall' || (event.type === 'budget_stage' && event.stage !== 'soft');
     }
-    #pauseForRunBudgetInteraction(state, stream, queue) {
-        const interaction = state.pendingRunBudgetInteraction;
-        if (!interaction)
-            return finish(stream, state, queue);
-        stream.interruptions = [interaction];
-        return finish(stream, state, queue);
-    }
 }
 function isLunaChainedRequest(state, request) {
     return (state.currentProviderId === 'codex' &&