muonroi-cli 1.7.2 → 1.8.1

package/dist/src/orchestrator/tool-engine.js

@@ -0,0 +1,3138 @@
+// MessageProcessor — extracted from orchestrator.ts as part of Phase 12.4.
+//
+// Owns the main streaming turn loop that lives in `Agent.processMessage`:
+//   - Abort wiring (external AbortContext + per-turn AbortController)
+//   - Trajectory + phase tracker observations on user_turn / abort
+//   - PIL enrichment pipeline (layers 1/3/6 — fail-open with logging)
+//   - ROUTE-11 per-turn model routing (decide + fallback to non-disabled
+//     provider via CouncilManager)
+//   - Vision proxy (history + current turn)
+//   - Auto-council gate (PIL taskType + heavy tier + role count) — routes
+//     into runCouncilV2 and re-enters processMessage with synthesis
+//   - System prompt assembly (chitchat / playwright gating + PIL suffix +
+//     model constraints)
+//   - SAMR step-aware routing (phase1 reasoning → phase2 execution)
+//   - Tool roundtrip loop:
+//       - Compaction (relax on overflow recovery, B4 top-level prepareStep)
+//       - Tool set assembly: builtin + MCP (smart filter for chitchat /
+//         browser-vocab) + PIL response tools, all wrapped with top-level
+//         cumulative cap (F1), cross-turn dedup (C3), read-path budget
+//       - ProviderOptions composition (buildTurnProviderOptions +
+//         taskTypeToReasoningEffort budget + thinkingType adaptive override
+//         + O1 shape capture)
+//       - streamText({...}) with prepareStep (top-level compactor +
+//         capability sanitizeHistory), onStepStart/Finish, onFinish
+//         (correlation cleanup)
+//       - fullStream consumer (text-delta / reasoning-delta / tool-call
+//         with EE PreToolUse intercept / tool-result with EE PostToolUse
+//         and vision-bridge / tool-error / tool-approval-request /
+//         error / abort)
+//   - Write-ahead persistence (Phase A4 tool_calls, A5 message_seq)
+//   - Context-overflow recovery + transient retry with exponential backoff
+//   - Post-turn compact + Stop / StopFailure hooks
+//   - Debug pipeline trace
+//
+// Zero behavioral changes — every method body mirrors the original
+// `processMessage` (see commit history). The DI surface (`MessageProcessorDeps`)
+// is the minimum proxy onto Agent state needed to reach back into Agent
+// without holding a circular reference. Public `Agent.processMessage`
+// signature is unchanged and continues to be the entrypoint; internally it
+// constructs a `MessageProcessor` per call.
+//
+// Cost-leak code paths preserved here:
+//   - F1 (top-level cumulative cap)         — wrapToolSetWithCap (top-level)
+//   - F1 (openai.promptCacheKey)            — buildTurnProviderOptions
+//   - G1 (OAuth `maxOutputTokens` drop)     — shouldDropParam(runtime, ...)
+//   - B4 (top-level prepareStep compaction) — compactSubAgentMessages
+//   - C3 (cross-turn dedup wrap)            — wrapToolSetWithDedup
+//   - A4 (tool_call write-ahead)            — persistToolCallWriteAhead
+//   - A5 (message_seq write-ahead)          — persistMessageWriteAhead
+//   - O1 (providerOptions shape forensics)  — extractProviderOptionsShape
+//   - siliconflow reasoning-strip           — turnCaps.sanitizeHistory
+import { generateText, stepCountIs, streamText } from "ai";
+import { recordArtifact } from "../ee/artifact-cache.js";
+import { routeFeedback } from "../ee/bridge.js";
+import { getDefaultEEClient } from "../ee/intercept.js";
+import { acquireMcpTools } from "../mcp/client-pool.js";
+import { dropRedundantFsMcpTools, filterMcpServersByMessage } from "../mcp/smart-filter.js";
+import { getModelInfo } from "../models/registry.js";
+import { cheapModelShellLine, injectCheapModelPlaybook, injectCheapModelShellDirective, shouldInjectCheapModelPlaybook, } from "../pil/cheap-model-playbook.js";
+import { injectCheapModelWorkbook, shouldInjectCheapModelWorkbook } from "../pil/cheap-model-workbooks.js";
+import { getResponseTaskType, isResponseTool, normalizeStructuredResponseTaskType, shouldHaltOnResponseTool, } from "../pil/index.js";
+import { isMetaAnalysisPrompt } from "../pil/layer6-output.js";
+import { taskTypeToMaxTokens, taskTypeToReasoningEffort, taskTypeToTier } from "../pil/task-tier-map.js";
+import { mentionsEcosystemScope } from "../playbook/directives.js";
+import { getProviderCapabilities } from "../providers/capabilities.js";
+import { bridgeMcpToolResult, listCachedImages, scrubImagePayloadsInMessages, } from "../providers/mcp-vision-bridge.js";
+import { captureToolSchemas } from "../providers/patch-zod-schema.js";
+import { buildTurnProviderOptions, requireRuntimeProvider, shouldDropParam, } from "../providers/runtime.js";
+import { wireDebug } from "../providers/wire-debug.js";
+import { reportRouteOutcome } from "../router/decide.js";
+import { getNextMessageSequence, logInteraction, markMessageErrored, markToolCallErrored, persistToolCallWriteAhead, } from "../storage/index.js";
+import { persistSessionExperience } from "../storage/session-experience-store.js";
+import { createBuiltinTools } from "../tools/registry.js";
+import { snapshotFromTodoWriteArgs } from "../tools/todo-write-snapshot.js";
+import { visionToolsNeeded } from "../tools/vision-gate.js";
+import { recordTurnTrace } from "../ui/slash/debug.js";
+import { statusBarStore } from "../ui/status-bar/store.js";
+import { appendDecisionLog } from "../usage/decision-log.js";
+import { logger } from "../utils/logger.js";
+import { openUrl } from "../utils/open-url.js";
+import { appendAudit, toolNeedsApproval } from "../utils/permission-mode.js";
+import { getAutoCouncilConfidence, getAutoCouncilMinRoles, getProviderStallRetries, getProviderStallTimeoutMs, getRoleModels, getSteerInjectionEnabled, getTopLevelCompactKeepLast, getTopLevelCompactThresholdChars, getTopLevelToolBudgetChars, isAutoCouncilEnabled, loadMcpServers, } from "../utils/settings.js";
+import { resolveShell } from "../utils/shell.js";
+import { relaxCompactionSettings } from "./compaction.js";
+import { wrapToolSetWithDedup } from "./cross-turn-dedup.js";
+import { humanizeApiError, isAuthenticationError, isContextLimitError, summarizeApiErrorForLog } from "./error-utils.js";
+import { buildGroundingFootnote, findUnverifiedClaims } from "./grounding-check.js";
+import { buildInterruptedTurnNote } from "./interrupted-turn.js";
+import { stableCallId } from "./pending-calls.js";
+import { buildMcpCapabilityBlock, MAX_LLM_CALLS_PER_TURN, } from "./prompts.js";
+import { extractProviderOptionsShape } from "./provider-options-shape.js";
+import { wrapToolSetWithReadBudget } from "./read-path-budget.js";
+import { containsEncryptedReasoning, sanitizeModelMessages } from "./reasoning.js";
+import { repairToolCallHook } from "./repair-tool-call.js";
+import { buildRepetitionReminder, recordAssistantBurst, shouldInjectRepetitionReminder, } from "./repetition-detector.js";
+import { classifyStreamError } from "./retry-classifier.js";
+import { forcedFinalize, incSessionStep, } from "./scope-ceiling.js";
+import { attachReminderToMessages, buildCheckpointReminder, buildScopeReminder, cadenceForSize, shouldInjectCeilingCrossing, shouldInjectReminder, shouldInjectSoftWarn, shouldPreWarnCompaction, } from "./scope-reminder.js";
+import { formatElisionManifest, getSessionExperienceCounts, recordCompaction, recordElision, } from "./session-experience.js";
+import { attemptStallRescue, pushStallToolResult } from "./stall-rescue.js";
+import { createStallWatchdog, STALL_ERROR_MESSAGE, shouldContinueAfterMidLoopStall, shouldRepromptStall, stallRepromptBackoffMs, } from "./stall-watchdog.js";
+import { planSteerInjection } from "./steer-inbox.js";
+import { wrapToolSetWithCap } from "./sub-agent-cap.js";
+import { applyAnthropicPromptCaching, compactSubAgentMessages, cumulativeMessageChars } from "./subagent-compactor.js";
+import { detectTextEmittedToolCall, parseDsmlToolCalls } from "./text-tool-call-detector.js";
+import { createToolLoopCapPredicate } from "./tool-loop-cap.js";
+import { buildToolRepetitionAbortMessage, recordToolError as recordToolRepetitionError, recordToolSuccess as recordToolRepetitionSuccess, } from "./tool-repetition-detector.js";
+/**
+ * F2 — approximate the char cost of the FIXED prompt envelope (system +
+ * tools JSON-Schema) that streamText re-sends on every step. Used to feed
+ * the compactor a realistic total-prompt size so it fires when billed input
+ * is actually large, not when only `messages[]` is.
+ */
+function computeEnvelopeChars(system, tools) {
+    let n = 0;
+    if (typeof system === "string")
+        n += system.length;
+    else if (system && typeof system === "object") {
+        try {
+            n += JSON.stringify(system).length;
+        }
+        catch {
+            /* ignore — best-effort estimate */
+        }
+    }
+    if (tools && typeof tools === "object") {
+        try {
+            n += JSON.stringify(tools).length;
+        }
+        catch {
+            /* ignore */
+        }
+    }
+    return n;
+}
+import { combineAbortSignals, getFinishReason, getStepNumber, getUsage, notifyObserver, toToolCall, toToolResult, } from "./tool-utils.js";
+/**
+ * Session-scoped cache for [EE Session Guidance] dedup. Maps sessionId to the
+ * sha256 prefix of the last-injected guidance content. Prevents the same block
+ * from being re-injected on every turn — once the model has seen a set of
+ * guidance entries it stays informed until new entries arrive.
+ */
+const _injectedGuidanceSha = new Map();
+/**
+ * MessageProcessor — extracted streaming turn loop.
+ *
+ * Lifecycle:
+ *   const processor = new MessageProcessor(deps);
+ *   yield* processor.run(userMessage, observer, images);
+ *
+ * Constructed per call (heap allocation is negligible against the streamText
+ * cost), matching the StreamRunner / CouncilManager pattern.
+ */
+/**
+ * Max response-tool (`respond_*`) calls tolerated within a single turn before
+ * the orchestrator finalizes early with the best answer buffered so far. A
+ * well-behaved turn emits the response tool ONCE; a hedge-then-answer emits 2.
+ * Beyond that is degenerate spam (session 8d8f498268ed: 80× identical
+ * respond_general in one generation). Set to 3 so the legitimate ≤2 patterns
+ * are never cut short.
+ */
+const RESPONSE_TOOL_SPAM_CAP = 3;
+/**
+ * Rewrites tool-result parts in the AI SDK's final response history if the
+ * user manually approved a safety-blocked command. This ensures the model sees
+ * its own retry context accurately (as a success) rather than a repeated block
+ * message, avoiding infinite retry loops or hallucinated failures.
+ */
+export function rewriteSafetyApprovedToolResults(messages) {
+    const _globalSafety = globalThis;
+    const approvedMap = _globalSafety.__muonroiSafetyApproved;
+    if (!approvedMap || approvedMap.size === 0) {
+        return messages;
+    }
+    return messages.map((m) => {
+        if (m.role !== "tool" || !Array.isArray(m.content))
+            return m;
+        let changed = false;
+        const newContent = m.content.map((part) => {
+            if (part.type === "tool-result" && typeof part.toolCallId === "string") {
+                const approved = approvedMap.get(part.toolCallId);
+                if (approved) {
+                    changed = true;
+                    return {
+                        ...part,
+                        isError: false,
+                        result: `Approved (${approved.kind}): blocked command was allowed by user`,
+                    };
+                }
+            }
+            return part;
+        });
+        return changed ? { ...m, content: newContent } : m;
+    });
+}
+import { stripDsmlMarkup } from "./message-processor.js";
+export async function* executeToolEngine(args) {
+    let { deps, stepRouterPhase, phase2Runtime, runtime, modelInfo, _debugSteps, _ceilingHit, userMessage, pilCtx, turnModelId, turnProvider, _stepCeiling, userModelMessage, userEnrichedMessage, signal, observer, taskHash, provider, system, routerStore, attemptedOverflowRecovery, patternLoopForceHalt, userWriteAheadSeq, streamRetryCount, MAX_STREAM_RETRIES, subagents, systemParts, toolTurnSystem, playwrightGuidance, _hasResponseTools, _pilResponseTools, patternLoopInjectCount, agentLoopDecisionCount, MAX_AGENT_LOOP_DECISIONS, _naturalCeiling, _ceilingTaskType, _ceilingSize, textToolReSteerCount, MAX_TEXT_TOOL_RESTEER, turnStartMs, _debugOn, _debugTurnId, _pilEnrichmentDeltaSnapshot, isChitchat, } = args;
+    // Put all extracted code here:
+    let stallTriggered = false;
+    // Time-to-first-byte stall RE-PROMPT: some providers (observed:
+    // xai/grok-build-0.1) accept the request then never send the first byte —
+    // a single wedged socket, not a down backend, so a fresh request usually
+    // goes through. When the watchdog fires with ZERO chunks received this
+    // attempt, we re-issue the SAME request up to `maxStallRetries` times
+    // (loop-persistent counter). Gated on zero-chunks so it can NEVER restart a
+    // turn that already ran tools or emitted text — those go to the partial-
+    // answer rescue path instead. maxStallRetries = 0 restores legacy behaviour.
+    let stallRetryCount = 0;
+    const maxStallRetries = getProviderStallRetries();
+    // Mid-loop dead-socket CONTINUATION counter (loop-persistent). Distinct from
+    // the TTFB re-prompt above: when the watchdog fires AFTER earlier tool steps
+    // ran (chunksThisAttempt > 0) but the in-flight step produced zero bytes
+    // (chunksThisStep === 0 — observed live: xai/grok-build-0.1 wedges the socket
+    // on an inter-step request mid-investigation, session 247a0cea2eac), we
+    // append the completed steps' messages to history and re-issue streamText to
+    // RESUME from the stalled step. Because completed tool-calls + tool-results
+    // are in history, no tool is re-run and no text is duplicated — safe even
+    // for write tools (the TTFB re-prompt is NOT, hence its zero-chunks gate).
+    // Bounded by the same maxStallRetries cap so a persistently-dead provider
+    // still falls through to the partial-answer rescue.
+    let midLoopStallRetryCount = 0;
+    // F3c — per-turn LLM call cap: counts every streamText() invocation
+    // (tool round-trip, stall re-prompt, stream retry) and hard-aborts
+    // the turn when exceeded.  Prevents the session 526a83cf22df pattern
+    // where 3 user messages burnt 82% of 2.44M tokens in 36 LLM calls.
+    let llmCallsThisTurn = 0;
+    // Live-queue steering: messages the user typed mid-turn are drained at a
+    // prepareStep boundary and accumulated here, then re-appended (deduped) to
+    // the messages returned for each subsequent step. Loop-persistent so they
+    // survive a stall-reprompt restart of streamText. NOT pushed into
+    // deps.messages in v1 — model-context only; the assistant response captures
+    // the steering effect and is persisted via appendCompletedTurn.
+    const pendingSteers = [];
+    const steerEnabled = getSteerInjectionEnabled();
+    // Auto-council: route to multi-model debate when EITHER
+    //   (a) PIL classified taskType=plan|analyze with high confidence AND the
+    //       prompt is complex enough to justify the debate cost, OR
+    //   (b) GSD-native tier === "heavy" (wholesale / multi-step / cross-repo work).
+    // After the debate finishes, runCouncilV2 records synthesis on
+    // councilManager.lastSynthesis; we then re-enter processMessage with the synthesis
+    // as the next user turn so the main loop continues with full debate context.
+    // Skip if this is already a council continuation turn (prevent infinite recursion).
+    //
+    // Phase 5 BUG-I (session f1a2a2a547db) — the gate previously fired on
+    // taskType=analyze + conf≥0.85 alone, with no complexity check. Result:
+    // "improve test coverage cho src/X.ts" (single-file, scoreComplexity=low,
+    // score=2) sank 13 minutes into council debate, then halted on pattern-loop
+    // after sprint 1 read 6 files. The complexity gate below bypasses council
+    // for low-complexity analyze prompts — they get the hot-path direct exec
+    // and stay productive. `plan` keeps the old behaviour (architectural
+    // decisions deserve debate regardless of length).
+    const autoCouncilTypes = new Set(["plan", "analyze"]);
+    const councilRoles = getRoleModels();
+    const configuredRoleCount = Object.values(councilRoles).filter(Boolean).length;
+    const heavyTier = pilCtx.complexityTier === "heavy";
+    const autoCouncilConfidence = getAutoCouncilConfidence();
+    const autoCouncilMinRoles = getAutoCouncilMinRoles();
+    const _complexityFromTrace = pilCtx._intentTrace
+        ?.complexity;
+    const _complexityGatePassed = pilCtx.taskType === "plan" || _complexityFromTrace === undefined || _complexityFromTrace !== "low";
+    const taskTypeMatch = pilCtx.taskType &&
+        autoCouncilTypes.has(pilCtx.taskType) &&
+        pilCtx.confidence >= autoCouncilConfidence &&
+        _complexityGatePassed;
+    const shouldAutoCouncil = !deps.councilManager.isContinuation &&
+        isAutoCouncilEnabled() &&
+        configuredRoleCount >= autoCouncilMinRoles &&
+        (taskTypeMatch || heavyTier);
+    // Always log the auto-council decision (taken or skipped) with the gate
+    // values that decided it. Lets reports answer "why did this turn cost
+    // $0.30?" and "is the confidence floor tuned wrong for my prompts?".
+    const autoCouncilSkipReason = (() => {
+        if (deps.councilManager.isContinuation)
+            return "continuation-turn";
+        if (!isAutoCouncilEnabled())
+            return "feature-disabled";
+        if (configuredRoleCount < autoCouncilMinRoles)
+            return `role-count<${autoCouncilMinRoles} (have ${configuredRoleCount})`;
+        if (!taskTypeMatch && !heavyTier) {
+            if (!pilCtx.taskType || !autoCouncilTypes.has(pilCtx.taskType))
+                return `taskType=${pilCtx.taskType ?? "null"} not in plan|analyze`;
+            if (pilCtx.confidence < autoCouncilConfidence)
+                return `confidence<${autoCouncilConfidence} (got ${pilCtx.confidence.toFixed(2)})`;
+            if (!_complexityGatePassed)
+                return `complexity=low + taskType=${pilCtx.taskType} (analyze needs medium+; plan bypasses gate)`;
+            return "no-trigger";
+        }
+        return "taken";
+    })();
+    appendDecisionLog({
+        ts: Date.now(),
+        sessionId: deps.session?.id ?? null,
+        kind: "auto-council",
+        taken: shouldAutoCouncil,
+        reason: autoCouncilSkipReason,
+        meta: {
+            taskType: pilCtx.taskType ?? null,
+            confidence: pilCtx.confidence,
+            complexityTier: pilCtx.complexityTier ?? null,
+            complexityScore: _complexityFromTrace ?? null,
+            complexityGatePassed: _complexityGatePassed,
+            configuredRoleCount,
+            autoCouncilConfidence,
+            autoCouncilMinRoles,
+            heavyTier,
+            isContinuation: deps.councilManager.isContinuation,
+        },
+    }).catch(() => undefined);
+    if (shouldAutoCouncil) {
+        const reason = heavyTier
+            ? `complexity=heavy${pilCtx.taskType ? ` task=${pilCtx.taskType}` : ""}`
+            : `${pilCtx.taskType} task detected with ${(pilCtx.confidence * 100).toFixed(0)}% confidence`;
+        yield { type: "content", content: `\n[Auto-council triggered: ${reason}]\n` };
+        yield* deps.runCouncilV2(userMessage, { skipClarification: true, observer, userModelMessage });
+        const synthesis = deps.councilManager.lastSynthesis;
+        deps.councilManager.setLastSynthesis(null);
+        if (synthesis) {
+            yield { type: "content", content: "\n[Auto-continuing with council recommendations...]\n" };
+            deps.councilManager.setContinuation(true);
+            try {
+                yield* deps.processMessage(`Council debate completed. Synthesis:\n\n${synthesis}\n\nProceed with the recommended action items.`, observer);
+            }
+            finally {
+                deps.councilManager.setContinuation(false);
+            }
+        }
+        return;
+    }
+    if (deps.batchApi) {
+        try {
+            yield* deps.processMessageBatchTurn({
+                userModelMessage,
+                userEnrichedMessage,
+                observer,
+                provider,
+                subagents,
+                system,
+                runtime,
+                modelInfo,
+                signal,
+            });
+        }
+        finally {
+            if (deps.getAbortController()?.signal === signal) {
+                deps.setAbortController(null);
+            }
+        }
+        return;
+    }
+    try {
+        streamAttempt: while (true) {
+            // SAMR Phase 2: switch to fast model for tool-execution steps
+            if (stepRouterPhase === "phase2" && phase2Runtime) {
+                runtime = phase2Runtime;
+                modelInfo = runtime.modelInfo;
+            }
+            deps.setCompactedThisTurn(false);
+            let assistantText = "";
+            // Count of stream parts received in THIS attempt. Stays 0 only when the
+            // provider never sent a first byte → the safe-to-re-prompt stall case.
+            let chunksThisAttempt = 0;
+            // Count of stream parts received since the last step boundary (reset in
+            // prepareStep). Distinguishes a mid-loop dead socket (a single step's
+            // request got zero bytes while every prior step completed) from a stall
+            // that interrupted text mid-generation. See shouldContinueAfterMidLoopStall.
+            let chunksThisStep = 0;
+            // Decide whether a fired stall watchdog should re-prompt (re-issue the
+            // same request) instead of falling through to rescue/error. Returns the
+            // backoff ms to wait before re-issuing, or null to NOT re-prompt. Reads
+            // the live per-attempt locals; safe to call only when stallTriggered.
+            const planStallReprompt = () => {
+                if (!shouldRepromptStall({
+                    stallTriggered,
+                    stallRetryCount,
+                    maxStallRetries,
+                    chunksThisAttempt,
+                    assistantTextEmpty: assistantText.trim() === "",
+                    aborted: signal.aborted,
+                })) {
+                    return null;
+                }
+                stallRetryCount++;
+                const backoffMs = stallRepromptBackoffMs(stallRetryCount);
+                try {
+                    const _ar = globalThis.__muonroiAgentRuntime;
+                    _ar?.emitEvent({
+                        t: "event",
+                        kind: "stream-retry",
+                        attempt: stallRetryCount,
+                        maxAttempts: maxStallRetries + 1,
+                        errorName: "TimeoutError",
+                        errorMessage: "provider-stall (no first byte) — re-prompting",
+                        nextDelayMs: backoffMs,
+                    });
+                    _ar?.emitEvent({
+                        t: "event",
+                        kind: "toast",
+                        level: "warning",
+                        text: `Model stalled — re-prompting (attempt ${stallRetryCount}/${maxStallRetries})…`,
+                    });
+                }
+                catch (emitErr) {
+                    logger.error("orchestrator", "stall-reprompt telemetry failed", { error: emitErr });
+                }
+                try {
+                    if (deps.session) {
+                        logInteraction(deps.session.id, "stream_retry", {
+                            data: {
+                                attempt: stallRetryCount,
+                                maxAttempts: maxStallRetries + 1,
+                                errorName: "provider-stall",
+                                errorMessage: "no first byte within stall timeout — re-prompted",
+                                nextDelayMs: backoffMs,
+                            },
+                        });
+                    }
+                }
+                catch (logErr) {
+                    logger.error("orchestrator", "stall-reprompt log failed", { error: logErr });
+                }
+                return backoffMs;
+            };
+            // Tracks where `assistantText` was at the previous step boundary so
+            // `onStepFinish` can compute the text emitted within the just-finished
+            // step (input to the self-repetition detector).
+            let _assistantTextAtLastStep = 0;
+            let reasoningPreview = "";
+            let encryptedReasoningHidden = false;
+            let streamOk = false;
+            let closeMcp;
+            let stepNumber = -1;
+            const activeToolCalls = [];
+            // Capped digest of tool outputs gathered this attempt — fuels the
+            // best-effort answer rescue if the stream stalls mid-turn (see
+            // stall-rescue.ts). Reset per attempt; only the most recent results win.
+            const turnToolResults = [];
+            // SAMR: track whether Phase 1 produced tool calls
+            let phase1HadToolCalls = false;
+            let _pendingStructuredResponse = null;
+            let _pendingStructuredResponseLen = -1;
+            try {
+                const { getDatabase } = await import("../storage/db.js");
+                const db = getDatabase();
+                const row = db.prepare("SELECT parent_session_id FROM sessions WHERE id = ?").get(deps.session.id);
+                const isSubSession = !!row?.parent_session_id;
+                let contextWindow = modelInfo?.contextWindow || 0;
+                if (isSubSession && contextWindow > 0) {
+                    contextWindow = Math.min(45000, contextWindow);
+                }
+                const settings = attemptedOverflowRecovery
+                    ? relaxCompactionSettings(deps.getCompactionSettings(contextWindow))
+                    : deps.getCompactionSettings(contextWindow);
+                if (contextWindow) {
+                    await deps.compactForContext(provider, system, contextWindow, signal, settings, attemptedOverflowRecovery);
+                }
+                // Vision-tool gate: for vision-proxy (text-only) models the registry
+                // adds 3 image tools (~500-700 tok) on every turn. Drop them when the
+                // turn has no plausible image involvement. Bias is KEEP — retained on
+                // any image signal, attachment, cached image, or prior-tool turn.
+                const includeVisionTools = visionToolsNeeded({
+                    userMessage,
+                    messages: deps.messages,
+                    cachedImageCount: listCachedImages().length,
+                    priorTurnHadTools: deps.messages.some((m) => m?.role === "tool"),
+                });
+                const baseToolsRaw = createBuiltinTools(deps.bash, deps.mode, {
+                    runTask: (request, abortSignal) => deps.runTask(request, combineAbortSignals(signal, abortSignal)),
+                    runDelegation: (request, abortSignal) => deps.runDelegation(request, combineAbortSignals(signal, abortSignal)),
+                    readDelegation: (id) => deps.readDelegation(id),
+                    listDelegations: () => deps.listDelegations(),
+                    modelId: turnModelId,
+                    includeVisionTools,
+                    consultParentSession: deps.consultParentSession,
+                });
+                // Top-level cumulative cap state. We accumulate the raw tool set
+                // (base + MCP + PIL response tools) across the assembly below,
+                // then apply the cap once. Tier ratios are looser than the
+                // sub-agent cap (50%/80%) so casual single-tool turns are not
+                // trimmed. See sub-agent-cap.ts.
+                // Chitchat: drop builtin tools too (not just MCP). A 1-word greeting
+                // never needs bash/read_file/edit_file/grep — those schemas alone
+                // cost ~1.5K input tokens on this CLI. Falls back to baseTools for
+                // every non-chitchat turn (PIL gates conservatively).
+                //
+                // BUG-A guard — when prior turns already issued tool_calls (their
+                // results still live in the messages history), DROPPING tools on a
+                // continuation chitchat ("tiếp tục" / "continue") causes two
+                // failures: (1) DeepSeek goes into native DSML markup fallback
+                // because it sees tool-call history but no schema (visible in
+                // sessions 002df4014cb4 + fc19b4daee20); (2) the agent has no way
+                // to actually CONTINUE the prior task — the user's clear intent.
+                // Detect prior-tool-context and keep the base tool set in that
+                // case. The 1.5K token saving for true greetings (no prior tool
+                // history) is preserved.
+                const turnCaps = getProviderCapabilities(requireRuntimeProvider(runtime));
+                const _priorTurnHadTools = deps.messages.some((m) => m?.role === "tool");
+                let rawToolSet = !turnCaps.supportsClientTools(runtime.modelInfo)
+                    ? {}
+                    : isChitchat && !_priorTurnHadTools
+                        ? {}
+                        : baseToolsRaw;
+                // MCP skip: chitchat / greeting inputs don't need 7 MCP servers'
+                // worth of tool schemas (~20K input tokens). PIL Layer 1 already
+                // gates this conservatively (≤10 chars + ≤2 words OR brain "none").
+                if (deps.mode === "agent" &&
+                    (!isChitchat || _priorTurnHadTools) &&
+                    turnCaps.supportsClientTools(runtime.modelInfo)) {
+                    // Smart MCP filter: drop OPTIONAL MCP servers whose category the
+                    // current message gives no signal for. Browser/vision servers
+                    // (Playwright/Chrome/Figma/Canva) skip unless the message touches a
+                    // page; docs/web servers (context7/fetch) skip unless the message
+                    // looks like an external lookup. Each MCP contributes 8-15 tools at
+                    // ~150 tok each, so local code work — the majority of turns — saves
+                    // ~13K input tokens it would otherwise pay every turn. Domain
+                    // servers (filesystem/tools/harness) always pass through. Logic is
+                    // a pure helper (src/mcp/smart-filter.ts) so it is unit-tested.
+                    // Override with MUONROI_DISABLE_SMART_MCP=1.
+                    const filteredServers = filterMcpServersByMessage(loadMcpServers(), userMessage, {
+                        disabled: process.env.MUONROI_DISABLE_SMART_MCP === "1",
+                    });
+                    // Ecosystem question → muonroi-docs is the authoritative source the
+                    // agent is nudged to consult FIRST. Wait for it specifically beyond the
+                    // normal deadline so a cold first-connect lands THIS turn instead of
+                    // "ready next turn" (session 584ba476c07a: first ecosystem question
+                    // missed docs while warming → agent guessed from local files).
+                    const criticalServerIds = mentionsEcosystemScope(userMessage)
+                        ? filteredServers.filter((s) => /(^|[-_])docs([-_]|$)/.test(s.id) && /muonroi/i.test(s.id)).map((s) => s.id)
+                        : undefined;
+                    // MCP non-blocking: acquireMcpTools self-bounds — it connects servers
+                    // in parallel and returns PARTIAL results at its internal deadline
+                    // (fast/cached servers included; slow first-connects reported in
+                    // .errors and available next turn). Clients are POOLED across turns
+                    // (client-pool.ts), so a server cold-spawns at most once per session
+                    // instead of every turn. No outer race: the old race discarded the
+                    // WHOLE bundle on timeout (Phase 1c — session f6f7881a5fae).
+                    let mcpBundle = null;
+                    try {
+                        mcpBundle = await acquireMcpTools(filteredServers, {
+                            onOAuthRequired: (_serverId, url) => {
+                                // Server-supplied URL is untrusted — openUrl validates the
+                                // scheme and spawns via execFile (no shell), closing the
+                                // command-injection vector the old exec() opener had.
+                                openUrl(url);
+                            },
+                            ...(criticalServerIds && criticalServerIds.length > 0 ? { criticalServerIds } : {}),
+                        });
+                    }
+                    catch (err) {
+                        logger.error("mcp", "buildMcpToolSet failed, proceeding with builtins only", { error: err });
+                    }
+                    if (mcpBundle) {
+                        closeMcp = mcpBundle.close;
+                        // Drop filesystem-MCP read/write/edit tools that duplicate the
+                        // first-class builtin file tools. Without this, models re-read the
+                        // SAME file via both `read_file` and `mcp_filesystem__read_text_file`
+                        // (live grok session f5dfab0ce0ca: a 772-line file read 6×), wasting
+                        // ~150 tok/schema PLUS re-injecting whole files into context. The
+                        // builtins are strictly better (read-before-write, LSP, CRLF match,
+                        // dedup/read-budget wrappers). Non-duplicate fs tools are untouched.
+                        const _builtinToolNames = new Set(Object.keys(rawToolSet));
+                        const { tools: _dedupedMcpTools, dropped: _droppedFsMcp } = dropRedundantFsMcpTools(mcpBundle.tools, _builtinToolNames);
+                        rawToolSet = { ...rawToolSet, ..._dedupedMcpTools };
+                        // muonroi-tools is THIS CLI: every tool it exposes (ee_query,
+                        // ee_feedback, ee_health, usage_forensics, lsp_query, setup_guide,
+                        // selfverify_*) is now a NATIVE in-process builtin (src/tools/
+                        // native-tools.ts) — strictly better (no subprocess, no cold-start).
+                        // If an external/legacy config still self-spawns muonroi-tools, drop
+                        // any MCP twin whose native equivalent is present so the model never
+                        // sees two interchangeable copies. (The CLI no longer self-spawns it
+                        // by default — see auto-setup.ts.)
+                        for (const key of Object.keys(rawToolSet)) {
+                            const twin = key.match(/^mcp_muonroi-tools__(.+)$/);
+                            if (twin && rawToolSet[twin[1]])
+                                delete rawToolSet[key];
+                        }
+                        if (_droppedFsMcp.length > 0 && deps.session) {
+                            try {
+                                logInteraction(deps.session.id, "routing", {
+                                    model: turnModelId,
+                                    data: { droppedRedundantFsMcp: _droppedFsMcp },
+                                });
+                            }
+                            catch {
+                                /* telemetry best-effort */
+                            }
+                        }
+                        if (mcpBundle.errors.length > 0) {
+                            // A pooled server that is still cold-starting is NOT "unavailable"
+                            // — it's warming up and will be ready next turn. Only surface
+                            // GENUINE failures as "unavailable"; show warming servers as a
+                            // soft, non-alarming note (and only the first time, since the
+                            // pool connects them in the background).
+                            const warming = mcpBundle.errors.filter((e) => /still connecting/.test(e));
+                            const failed = mcpBundle.errors.filter((e) => !/still connecting/.test(e));
+                            if (failed.length > 0) {
+                                yield { type: "content", content: `MCP unavailable: ${failed.join(" | ")}\n\n` };
+                            }
+                            if (warming.length > 0) {
+                                const names = warming.map((e) => e.split(":")[0]).join(", ");
+                                yield { type: "content", content: `MCP warming up (${names}) — ready from the next turn.\n\n` };
+                            }
+                        }
+                    }
+                }
+                // PIL response tools: inject structured output tool when taskType detected
+                if (_hasResponseTools && turnCaps.supportsClientTools(runtime.modelInfo)) {
+                    rawToolSet = { ...rawToolSet, ..._pilResponseTools };
+                    captureToolSchemas(_pilResponseTools);
+                }
+                // Apply the top-level cumulative cap once over the fully-assembled
+                // raw tool set. State is per-turn; each turn gets a fresh budget.
+                const topLevelCap = wrapToolSetWithCap(rawToolSet, {
+                    maxCumulativeChars: getTopLevelToolBudgetChars(deps.maxToolRounds),
+                    midTierRatio: 0.5,
+                    highTierRatio: 0.8,
+                    label: "top-level",
+                });
+                // Phase C3: layer cross-turn dedup on top of the top-level cap.
+                const tools = wrapToolSetWithReadBudget(wrapToolSetWithDedup(topLevelCap.tools, deps.crossTurnDedup), deps.readBudget);
+                captureToolSchemas(tools);
+                let responseToolCalled = false;
+                // A turn must surface exactly ONE final structured answer. Cheap
+                // models sometimes emit the response tool MORE THAN ONCE in a single
+                // step (session 9b1b39bf4dc6: grok emitted respond_general twice —
+                // a 278-char "I must read the code" hedge, then the 3782-char real
+                // answer — both in one step). Yielding each inline appends two
+                // stacked structured_response blocks and shows the hedge first.
+                // Instead we BUFFER the response-tool payloads and yield only the
+                // most complete one (longest serialized data) after the stream
+                // drains — robust to either ordering (hedge-then-answer or
+                // answer-then-summary).
+                _pendingStructuredResponse = null;
+                _pendingStructuredResponseLen = -1;
+                let _responseToolEmitCount = 0;
+                // G3: providerOptions assembly is owned by the capability layer
+                // (src/providers/capabilities.ts). buildTurnProviderOptions feeds
+                // sessionId in so openai.promptCacheKey is derived per turn.
+                // The task-type-driven anthropic.thinking budget override stays
+                // here because it depends on PIL task context, not provider quirks.
+                // biome-ignore lint/suspicious/noExplicitAny: matches RuntimeResult.providerOptions shape (any) used downstream
+                const baseProviderOpts = buildTurnProviderOptions(runtime, { sessionId: deps.session?.id }) ?? {};
+                const providerOpts = runtime.modelInfo?.reasoning && runtime.modelInfo?.provider === "anthropic"
+                    ? {
+                        ...baseProviderOpts,
+                        anthropic: {
+                            ...(baseProviderOpts.anthropic ?? {}),
+                            thinking: {
+                                type: "enabled",
+                                budgetTokens: taskTypeToReasoningEffort(pilCtx.taskType) === "high"
+                                    ? 32_768
+                                    : taskTypeToReasoningEffort(pilCtx.taskType) === "medium"
+                                        ? 8_192
+                                        : 2_048,
+                            },
+                        },
+                    }
+                    : baseProviderOpts;
+                // Use catalog's thinkingType field instead of regex matching.
+                // providerOpts is loosely typed (Record<string, unknown>) after the
+                // g1 capability refactor — narrow with a local typed view.
+                const thinkingModelInfo = getModelInfo(runtime.modelId);
+                const providerOptsAnyView = providerOpts;
+                if (providerOptsAnyView.anthropic?.thinking?.type === "enabled" &&
+                    thinkingModelInfo?.thinkingType === "adaptive") {
+                    providerOptsAnyView.anthropic.thinking = { type: "adaptive" };
+                }
+                // OpenAI api-key path: `store: true` is seeded by OpenAIStrategy
+                // via factory.defaultProviderOptions (Phase 12.2-G4 migration).
+                // OAuth backend (ChatGPT Codex) overrides with `store: false` via
+                // the auth registry. Both flow through resolveModelRuntime →
+                // runtime.providerOptions → buildTurnProviderOptions and arrive
+                // here merged into providerOpts.openai.
+                // Top-level dropParam — shared with sub-agent path via shouldDropParam.
+                // See src/providers/runtime.ts for the central rule.
+                const dropParam = (p) => shouldDropParam(runtime, p);
+                // Tier-aware behavioural suffix. Cheap models (DeepSeek V4 Flash etc.)
+                // ignore well-worded tool descriptions but DO adopt instructions when
+                // surfaced in the system prompt. Smart models don't need this — gated
+                // by `modelInfo.tier === "fast"`. See cheap-model-playbook.ts for
+                // motivation + escape hatch (MUONROI_DISABLE_CHEAP_MODEL_PLAYBOOK=1).
+                // Fast-tier steering, front-loaded for primacy: task convergence
+                // workbook (anti-ramble — cuts tool-call count, the dominant
+                // cheap-model cost) layered UNDER the tool-use playbook so the
+                // CRITICAL tool rules stay at the very front. Both fixed per turn, so
+                // they stay inside the cached prefix.
+                // F3c — tool-turn: use reduced system prompt (skip
+                // native-capabilities + skills already shown in first call).
+                const activeSystem = llmCallsThisTurn > 0 && toolTurnSystem ? toolTurnSystem : system;
+                const systemWithWorkbook = shouldInjectCheapModelWorkbook(runtime.modelInfo)
+                    ? injectCheapModelWorkbook(activeSystem, pilCtx.taskType)
+                    : activeSystem;
+                const systemWithPlaybook = shouldInjectCheapModelPlaybook(runtime.modelInfo)
+                    ? injectCheapModelPlaybook(systemWithWorkbook)
+                    : systemWithWorkbook;
+                // A2: front-load a one-line shell/env directive for fast-tier models.
+                // The authoritative ENVIRONMENT block already states OS/shell/cwd in
+                // the prompt body, but budget models underweight non-front-loaded
+                // rules — so echo the correct-syntax line at the very front. Derived
+                // from resolveShell({}) (same source as buildEnvironmentBlock) so it
+                // is always accurate to the actual shell the bash tool will spawn.
+                // Gated to fast tier, so the claude branch below still sees `system`.
+                const systemWithShell = shouldInjectCheapModelPlaybook(runtime.modelInfo)
+                    ? injectCheapModelShellDirective(systemWithPlaybook, cheapModelShellLine(resolveShell({}).kind, process.platform))
+                    : systemWithPlaybook;
+                // Append the LIVE MCP tool roster so the agent calls connected MCP
+                // tools by their exact mcp_<server>__<tool> name instead of shelling
+                // out (session f6f7881a5fae). Built from the FINAL toolset for this
+                // iteration (post smart-filter + fs-dedup), so it never names a tool
+                // the model can't actually call. Dynamic per turn → must live OUTSIDE
+                // the cached staticPrefix; for claude it lands in the second
+                // (non-cached) system message via the slice below.
+                const mcpCapabilityBlock = buildMcpCapabilityBlock(Object.keys(tools));
+                const systemWithCaps = mcpCapabilityBlock ? `${systemWithShell}${mcpCapabilityBlock}` : systemWithShell;
+                const systemForModel = runtime.modelId.startsWith("claude")
+                    ? [
+                        {
+                            role: "system",
+                            content: systemParts.staticPrefix,
+                            providerOptions: { anthropic: { cacheControl: { type: "ephemeral" } } },
+                        },
+                        {
+                            role: "system",
+                            content: systemWithCaps.slice(systemParts.staticPrefix.length),
+                        },
+                    ]
+                    : systemWithCaps;
+                // Capture prompt-size breakdown so recordUsage can attach it to the
+                // cost-log entry. Without this, "system prompt is huge" is unfalsifiable.
+                // chars/4 ≈ tokens for English; reported as chars to keep math obvious.
+                const messagesChars = deps.messages.reduce((s, m) => {
+                    const c = m.content;
+                    if (typeof c === "string")
+                        return s + c.length;
+                    if (Array.isArray(c)) {
+                        for (const part of c) {
+                            if (typeof part.text === "string") {
+                                s += part.text.length;
+                            }
+                        }
+                    }
+                    return s;
+                }, 0);
+                let toolsChars = 0;
+                let toolsCount = 0;
+                for (const [name, t] of Object.entries(tools)) {
+                    toolsCount += 1;
+                    toolsChars += name.length;
+                    const desc = t.description;
+                    if (typeof desc === "string")
+                        toolsChars += desc.length;
+                    try {
+                        // Schemas often dominate tool size on non-Anthropic providers
+                        // (Zod-derived JSON schemas can be 2-5K chars per tool).
+                        const params = t.parameters ??
+                            t.inputSchema;
+                        if (params)
+                            toolsChars += JSON.stringify(params).length;
+                    }
+                    catch {
+                        /* best-effort */
+                    }
+                }
+                deps.setLastPromptBreakdown({
+                    systemChars: system.length,
+                    staticPrefixChars: systemParts.staticPrefix.length,
+                    dynamicSuffixChars: systemParts.dynamicSuffix.length,
+                    playwrightGuidanceChars: playwrightGuidance.length,
+                    messagesChars,
+                    messagesCount: deps.messages.length,
+                    toolsChars,
+                    toolsCount,
+                });
+                // Task 2.6a — assign a fresh correlation ID for this top-level streamText call.
+                const _topCallId = crypto.randomUUID();
+                deps.setCurrentCallId(_topCallId);
+                // Capture finishReason so we can surface "round cap hit" as a visible
+                // toast — without this, the agent silently stops mid-flight when
+                // stepCountIs(maxToolRounds) fires and the user sees the TUI freeze
+                // (session 7dcf8fd7d6a4 hit exactly 100 rounds → looked like a crash).
+                let _lastFinishReason = null;
+                // F3b — track hard cap hit for visible toast after stream ends.
+                let _hardCapHit = false;
+                // Phase B4: compact older tool_result parts before each top-level
+                // step once cumulative message chars exceed the configured threshold.
+                // The compactor preserves system + first user verbatim and keeps the
+                // last N tool turns intact; older results are rewritten into short
+                // stubs. Symmetric to the B3 sub-agent path; reuses the same module
+                // with `label: "top-level"` so the stub text reflects which loop
+                // elided the content.
+                const topLevelCompactThreshold = getTopLevelCompactThresholdChars();
+                const topLevelCompactKeepLast = getTopLevelCompactKeepLast();
+                // Phase O1 — capture providerOptions SHAPE (types only) for forensics.
+                deps.setLastProviderOptionsShape(Object.keys(providerOpts).length > 0 ? extractProviderOptionsShape(providerOpts) : null);
+                // Substitute the enriched user message for the current turn so the LLM sees PIL additions,
+                // while leaving the DB-persisted `deps.messages` clean for future turns.
+                const _messagesForCall = deps.messages.map((m) => m === userModelMessage ? userEnrichedMessage : m);
+                if (wireDebug.enabled) {
+                    wireDebug.logRequest({
+                        providerId: runtime.modelInfo?.provider ?? "unknown",
+                        modelId: runtime.modelId,
+                        messages: _messagesForCall,
+                        systemChars: systemForModel?.length ?? 0,
+                        toolNames: tools ? Object.keys(tools) : undefined,
+                        providerOptions: providerOpts,
+                    });
+                }
+                // sanitizeHistory is identity for every provider (kept as a hook
+                // for future provider-specific quirks). Reasoning round-trips
+                // natively via @ai-sdk/openai-compatible — see
+                // src/providers/__tests__/reasoning-roundtrip.test.ts.
+                const _topMessagesForCall = applyAnthropicPromptCaching(turnCaps.sanitizeHistory(_messagesForCall), runtime.modelId);
+                // Closure-mutable cap for the tool-loop askcard rescue.
+                // Phase 1 (SAMR) skips the dynamic cap (it's a single-step path).
+                // Algorithm extracted to ./tool-loop-cap.ts so it can be unit-tested.
+                const _baseDynamicStopWhen = createToolLoopCapPredicate({
+                    initialCap: deps.maxToolRounds,
+                    ask: async (info) => {
+                        if (info.kind === "pattern") {
+                            if (patternLoopInjectCount < 1) {
+                                patternLoopInjectCount++;
+                                patternLoopForceHalt = true;
+                                deps.messages.push({
+                                    role: "user",
+                                    content: `[System Warning: You have called tool '${info.toolName}' ${info.count} times with similar arguments in this turn. If you are stuck in a loop, please re-evaluate your plan, change your approach, or explain the blocker. Do not repeat the same unsuccessful tool call.]`,
+                                });
+                                return "stop";
+                            }
+                        }
+                        // Query the agent itself to make the decision instead of showing a raw user askcard
+                        // Active only in headless (batchApi) mode — TUI users expect the CLI to stop and wait.
+                        if (deps.batchApi && agentLoopDecisionCount < MAX_AGENT_LOOP_DECISIONS) {
+                            try {
+                                const _ar = globalThis.__muonroiAgentRuntime;
+                                _ar?.emitEvent({
+                                    t: "event",
+                                    kind: "toast",
+                                    level: "info",
+                                    text: info.kind === "pattern"
+                                        ? `phát hiện lặp tool liên tục — đang hỏi ý kiến agent có muốn tiếp tục không...`
+                                        : `đạt giới hạn số bước (${info.stepNumber} bước) — đang hỏi ý kiến agent có muốn tiếp tục không...`,
+                                });
+                                const loopContext = info.kind === "pattern"
+                                    ? `You have called the tool '${info.toolName}' repeatedly ${info.count} times in a row with similar inputs. This indicates you might be stuck in a tool loop.`
+                                    : `You have reached the tool execution limit of ${info.stepNumber} steps (cap: ${info.cap}).`;
+                                const systemPrompt = `You are an AI loop-guard assistant. Your task is to evaluate the agent's progress and decide whether they should keep trying to run tools (continue) or stop and present their best answer now (stop).\n` +
+                                    `Choose 'stop' if the agent is stuck in an unproductive loop (e.g. repeated unsuccessful commands, permission errors, file reading loops without making progress), or if they already have enough information to write a final response.\n` +
+                                    `Choose 'continue' only if they are making active progress, are close to a breakthrough, or genuinely need a few more steps to verify their changes.\n\n` +
+                                    `Format your decision EXACTLY as:\n` +
+                                    `<decision>continue</decision> or <decision>stop</decision>\n` +
+                                    `followed by a brief 1-sentence explanation of your decision (in English).`;
+                                const recentMessages = deps.messages.slice(-10);
+                                const modelMessages = [
+                                    ...recentMessages,
+                                    {
+                                        role: "user",
+                                        content: `[System Loop Guard Context: ${loopContext}\nBased on the conversation history and recent tool executions, do you need to continue executing tools or should you stop and summarize now?]`,
+                                    },
+                                ];
+                                const { text: decisionText } = await generateText({
+                                    model: runtime.model,
+                                    system: systemPrompt,
+                                    messages: modelMessages,
+                                    abortSignal: signal,
+                                });
+                                const cleanText = decisionText.trim();
+                                const match = cleanText.match(/<decision>(continue|stop)<\/decision>/i);
+                                const decision = match ? match[1].toLowerCase() : "stop";
+                                let reason = cleanText.replace(/<decision>.*?<\/decision>/gs, "").trim();
+                                // Strip DeepSeek's raw DSML or other XML tool leaks from the toast reason
+                                reason = reason.replace(/<[^>]+>/g, "").trim();
+                                _ar?.emitEvent({
+                                    t: "event",
+                                    kind: "toast",
+                                    level: "info",
+                                    text: `Agent quyết định: ${decision === "continue" ? "TIẾP TỤC" : "DỪNG LẠI"} (${reason || "không có lý do"})`,
+                                });
+                                agentLoopDecisionCount++;
+                                if (decision === "continue") {
+                                    return "continue";
+                                }
+                                else {
+                                    return "stop";
+                                }
+                            }
+                            catch (err) {
+                                logger.error("orchestrator", "loop auto-decision failed", { error: err });
+                            }
+                        }
+                        return deps.askToolLoopContinue ? await deps.askToolLoopContinue(info) : "stop";
+                    },
+                    // Phase 5 BUG-H — thread the resolved natural ceiling down so the
+                    // pattern askcard can pick a context-aware default action (continue
+                    // early in the run, stop once we're past 50% of the natural budget).
+                    naturalCeiling: _naturalCeiling,
+                });
+                // Phase 4 Plan 04 (4B) — compose per-session ceiling alongside the
+                // existing cap + pattern guard. Logical OR: any condition true → halt.
+                // Counter is per-SESSION and increments once per stopWhen invocation
+                // (i.e. once per finished tool step), persisting across user turns
+                // so a wandering 3-turn burst still trips at the matrix limit.
+                const _ceilingSessionId = deps.session?.id ?? "no-session";
+                // Phase 5 Fix 3 — capture the actual step number when the ceiling
+                // trips so the halt toast can report the real value, not the
+                // ceiling/ceiling literal that always showed e.g. "5/5" regardless
+                // of how many steps the turn actually ran.
+                const _ceilingHitAtStep = 0;
+                // Phase 5 Fix 5 — matrix ceiling is now a SOFT BOUNDARY, never a
+                // hard halt. Phase 4's hard halt was a blunt anti-wandering
+                // measure that also blocked legitimate multi-step work: every
+                // long task (improve coverage, optimize startup, refactor) ran
+                // out of budget mid-flight and required the user to manually
+                // type "tiếp tục". Wrong philosophy — "done" must be the agent's
+                // call, not the counter's.
+                //
+                // What replaced the hard halt:
+                //  - Scope-reminders (4A path, prepareStep above) inject
+                //    "[approaching ceiling]" reminder at floor(ceiling*0.7) and
+                //    repeat at K cadence. Past the ceiling, every step gets a
+                //    re-anchor so the model is repeatedly nudged toward closure.
+                //  - The dynamicStopWhen no longer checks the matrix ceiling at
+                //    all. The only halt source is `_baseDynamicStopWhen` which
+                //    enforces `deps.maxToolRounds` as the ULTIMATE runaway safety
+                //    net (default raised; see CLI default).
+                //  - 4R bash repeat detector still catches the dominant wandering
+                //    pattern (identical command twice in a row).
+                //  - F6 synthesis still ensures any natural stream-end without
+                //    text gets a final summary.
+                //
+                // _ceilingHit and _ceilingHitAtStep are kept for telemetry: a
+                // crossing event is logged for forensics, but no action is taken.
+                const dynamicStopWhen = (async (state) => {
+                    // F3b — HARD cap: absolute non-bumpable ceiling per user turn.
+                    // Fires AFTER the soft cap (maxToolRounds) has been bumped by the
+                    // user. Prevents runaway sessions (session 526a83cf22df: 16 LLM
+                    // calls for a single user message, 2.44M total input tokens).
+                    if (state.steps.length > deps.hardMaxToolRounds) {
+                        _hardCapHit = true;
+                        return true;
+                    }
+                    // Terminal response tool: a `respond_*` call IS the model's final
+                    // structured answer (its `execute` is identity — the payload lives
+                    // in the tool-call args). `shouldHaltOnResponseTool` decides if the
+                    // emission is terminal vs a premature "blind" announce:
+                    //  - response tool AFTER real tool work (read/grep/bash) → terminal,
+                    //    halt now (kills d95113d3be09 seq=27: 7 reads → 87× respond loop
+                    //    at call #1, no extra round-trip for the common case).
+                    //  - a single blind response (no prior investigation) → do NOT halt;
+                    //    give the model the step it announced it would use to read code
+                    //    (session e4a9d97a90: lone blind respond_general was force-stopped
+                    //    by the old halt-on-first rule and the agent never investigated).
+                    //  - a 2nd blind response with still no real work → narration loop,
+                    //    halt. In-step spam (80× in one generation) is bounded separately
+                    //    by RESPONSE_TOOL_SPAM_CAP — stopWhen only runs BETWEEN steps.
+                    // Read from `state.steps` (the SDK's own per-step record) rather than
+                    // the for-await consumer's `responseToolCalled` flag — stopWhen runs
+                    // between steps and may evaluate before our consumer processed the
+                    // tool-result part, so the flag would race.
+                    const _steps = state.steps;
+                    if (shouldHaltOnResponseTool(_steps))
+                        return true;
+                    const base = await _baseDynamicStopWhen(state);
+                    if (base)
+                        return true;
+                    const next = incSessionStep(_ceilingSessionId);
+                    // Telemetry-only: record the first time the counter crosses
+                    // the matrix ceiling, so post-hoc queries can correlate the
+                    // ceiling crossing with task completion outcomes. No halt.
+                    if (next === _stepCeiling) {
+                        try {
+                            if (deps.session?.id) {
+                                logInteraction(deps.session.id, "f6_synthesis", {
+                                    data: {
+                                        outcome: "ceiling_crossed_softly",
+                                        stepAtCrossing: next,
+                                        naturalCeiling: _naturalCeiling,
+                                        taskType: _ceilingTaskType,
+                                        size: _ceilingSize,
+                                        hardCapMaxToolRounds: deps.maxToolRounds,
+                                    },
+                                });
+                            }
+                        }
+                        catch {
+                            /* telemetry only */
+                        }
+                    }
+                    return false;
+                });
+                // BUG-A fix — when this turn carries an empty tool set (the
+                // chitchat optimization at line ~1107 drops all schemas), AI SDK
+                // sends `tools:[], tool_choice:undefined` to the provider. DeepSeek
+                // V4 Flash sees prior `tool_call`/`tool_result` parts still in the
+                // messages history (the previous turn used 20 tools) and the
+                // model stays in agent-mode — but with no schema to call, it
+                // falls back to its NATIVE DSML markup syntax and emits that as
+                // plain text. AI SDK does not parse the native format, so the
+                // markup leaks straight to the TUI as garbage and the turn
+                // produces no useful output. Setting `toolChoice:"none"` is the
+                // canonical way to tell the model "you cannot call tools this
+                // turn" so it emits text-only. Verified by stream_start telemetry
+                // on sessions 002df4014cb4 (leak) + fc19b4daee20 (leak): both had
+                // toolCount=0 + toolChoice=undefined on chitchat continuation.
+                const _toolsAreEmpty = Object.keys(tools).length === 0;
+                const _finalToolChoice = _toolsAreEmpty
+                    ? "none"
+                    : _hasResponseTools && turnCaps.supportsClientTools(runtime.modelInfo)
+                        ? "auto"
+                        : undefined;
+                // BUG-C telemetry — record tool availability + toolChoice at the
+                // call site so future regressions show up in telemetry not in TUI.
+                try {
+                    const _toolNamesAtCall = Object.keys(tools);
+                    logInteraction(deps.session?.id ?? "no-session", "stream_start", {
+                        model: turnModelId,
+                        data: {
+                            toolCount: _toolNamesAtCall.length,
+                            hasBash: _toolNamesAtCall.includes("bash"),
+                            toolNames: _toolNamesAtCall.slice(0, 25),
+                            toolChoice: _finalToolChoice ?? "undefined",
+                            hasResponseTools: _hasResponseTools,
+                            supportsClientTools: turnCaps.supportsClientTools(runtime.modelInfo),
+                            priorTurnHadTools: _topMessagesForCall.some((m) => m?.role === "tool"),
+                        },
+                    });
+                }
+                catch {
+                    /* telemetry only */
+                }
+                // Silent-hang guard: abort the stream (and surface a toast in the
+                // catch below) if the provider sends no chunk for too long. Re-armed
+                // on every chunk via stall.pet(), so it never kills an actively
+                // streaming call. Disposed when the stream ends or errors.
+                stallTriggered = false;
+                const stall = createStallWatchdog(getProviderStallTimeoutMs(), () => {
+                    stallTriggered = true;
+                });
+                // F3c — hard-cap LLM calls per turn before this streamText()
+                if (++llmCallsThisTurn > MAX_LLM_CALLS_PER_TURN) {
+                    stall.dispose();
+                    yield {
+                        type: "error",
+                        content: `Turn aborted: reached the limit of ${MAX_LLM_CALLS_PER_TURN} LLM calls for this message. Try a narrower request or break your task into smaller steps.`,
+                        isAuthError: false,
+                    };
+                    yield { type: "done" };
+                    return;
+                }
+                const result = streamText({
+                    model: runtime.model,
+                    system: systemForModel,
+                    messages: _topMessagesForCall,
+                    tools,
+                    toolChoice: _finalToolChoice,
+                    stopWhen: stepRouterPhase === "phase1" ? stepCountIs(1) : dynamicStopWhen,
+                    maxRetries: 0,
+                    abortSignal: combineAbortSignals(signal, stall.signal),
+                    // Repair malformed tool-call JSON args before they bubble up as
+                    // InvalidToolInputError → tool-error → repetition-detector abort.
+                    // Conservative: only fixes the two observed Qwen-style defects.
+                    // See src/orchestrator/tool-args-repair.ts for the transforms.
+                    experimental_repairToolCall: repairToolCallHook,
+                    prepareStep: ({ stepNumber: sn, messages: stepMessages }) => {
+                        chunksThisStep = 0;
+                        if (deps.isSubSession) {
+                            logger.info("orchestrator", "Sub-session executing tool round", {
+                                stepNumber: sn,
+                                maxToolRounds: deps.maxToolRounds,
+                            });
+                            deps.emitSubagentStatus({
+                                agent: "sub-session",
+                                description: `Running sub-session task...`,
+                                detail: `[Sub-Session] Executing tool round ${sn + 1} of ${deps.maxToolRounds}...`,
+                            });
+                        }
+                        // --- Live-queue steering injection ---------------------------
+                        // Drain the UI steer queue ONCE per prepareStep call (sn >= 1),
+                        // accumulate into pendingSteers, and graft pendingSteers onto the
+                        // messages this step returns. Dedup-by-content makes re-appending
+                        // idempotent even if a stall-reprompt restart re-reads history.
+                        const withSteers = (r) => {
+                            // Guard the drain on !signal.aborted too: planSteerInjection
+                            // already refuses to inject on abort, but draining still CLEARS
+                            // the UI queue — so on a (programmatic) abort we must not drain,
+                            // or a queued-but-uninjected message is lost (spec §143).
+                            const _drained = sn >= 1 && steerEnabled && !signal.aborted ? (deps.drainSteerMessages?.() ?? []) : [];
+                            const _newSteers = planSteerInjection({
+                                drained: _drained,
+                                aborted: signal.aborted,
+                                enabled: steerEnabled,
+                            });
+                            if (_newSteers.length > 0) {
+                                pendingSteers.push(..._newSteers);
+                                deps.appendMidTurnMessages?.(_newSteers);
+                                try {
+                                    const _ar = globalThis.__muonroiAgentRuntime;
+                                    _ar?.emitEvent({
+                                        t: "event",
+                                        kind: "steer-inject",
+                                        count: _newSteers.length,
+                                        atStep: sn,
+                                        runId: deps.getActiveRunId() ?? "",
+                                    });
+                                }
+                                catch (emitErr) {
+                                    logger.error("orchestrator", "steer-inject telemetry failed", { error: emitErr });
+                                }
+                            }
+                            const baseRes = (() => {
+                                if (pendingSteers.length === 0)
+                                    return r;
+                                const _base = r.messages ?? stepMessages;
+                                const _existingContents = new Set(_base
+                                    .filter((m) => m.role === "user")
+                                    .map((m) => (typeof m.content === "string" ? m.content : JSON.stringify(m.content))));
+                                const steersToAdd = pendingSteers.filter((s) => {
+                                    const sContent = typeof s.content === "string" ? s.content : JSON.stringify(s.content);
+                                    return !_existingContents.has(sContent);
+                                });
+                                if (steersToAdd.length === 0)
+                                    return r;
+                                const insertIdx = _base.length;
+                                return {
+                                    ...r,
+                                    messages: [
+                                        ..._base.slice(0, insertIdx),
+                                        ...steersToAdd,
+                                        ..._base.slice(insertIdx),
+                                    ],
+                                };
+                            })();
+                            if (baseRes.messages) {
+                                baseRes.messages = applyAnthropicPromptCaching(baseRes.messages, runtime.modelId);
+                            }
+                            return baseRes;
+                        };
+                        const stripped = turnCaps.sanitizeHistory(stepMessages);
+                        // Agent-controlled veto (PRESERVE) or lighter selective keep (KEEP_TOOL_IDS) for this turn's B4 compaction.
+                        // PRESERVE_FULL_CONTEXT skips the compactor entirely (full history).
+                        // KEEP_TOOL_IDS: id1,id2 (from prior stub " (id=...) ") protects only those specific tool results
+                        // without the cost of a full veto. Parsed from reasoning or assistant note.
+                        let keepToolIds = [];
+                        const hasPreserve = stripped.some((m) => {
+                            const c = m?.content;
+                            const texts = [];
+                            if (typeof c === "string")
+                                texts.push(c);
+                            if (Array.isArray(c)) {
+                                for (const p of c)
+                                    if (typeof p?.text === "string")
+                                        texts.push(p.text);
+                            }
+                            const joined = texts.join(" ");
+                            if (joined.includes("PRESERVE_FULL_CONTEXT"))
+                                return true;
+                            // Idea 3: parse lighter token
+                            const mKeep = joined.match(/KEEP_TOOL_IDS\s*[:=]\s*([a-z0-9_, -]+)/i);
+                            if (mKeep) {
+                                keepToolIds = mKeep[1]
+                                    .split(/[,\s]+/)
+                                    .map((s) => s.trim())
+                                    .filter(Boolean);
+                            }
+                            return false;
+                        });
+                        if (hasPreserve) {
+                            return withSteers({ messages: stripped });
+                        }
+                        // F2 — envelope = system prompt + JSON-Schema of every tool
+                        // re-sent on every step. Without this the threshold check
+                        // ignored 20-50K of fixed prompt overhead and the compactor
+                        // sat dormant just below its limit while billed input climbed.
+                        const envelopeChars = computeEnvelopeChars(systemForModel, tools);
+                        // G1 + G2 — feed the model's context window so the compactor
+                        // can pick a token-aware threshold and shrink keepLastTurns
+                        // when the window is approaching its ceiling.
+                        const contextWindowTokens = runtime.modelInfo?.contextWindow ?? 0;
+                        // Idea 4: fire-and-forget persist of elided tool outputs to EE (source=tool-artifact)
+                        // so later layer3/ee.query "tool-artifact id=xxx" or "full tool result id=..." can re-hydrate.
+                        // Use process-level fallbacks (prepareStep closure does not directly expose outer cwd/session in this scope).
+                        const _cwd = process.cwd();
+                        const _sess = undefined; // best-effort; EE artifact still indexable by content + meta.toolCallId
+                        const persistArtifact = (toolCallId, toolName, fullContent, reason) => {
+                            // Local-first: record the FULL output in-process so ee_query can
+                            // rehydrate it even if EE is down (the EE extract below caps at 8k
+                            // and needs the network; the cache keeps up to 200k, no network).
+                            recordArtifact(toolCallId, toolName, fullContent);
+                            // Lived-experience telemetry: count this elision so a later
+                            // "cảm nhận trong CLI" question answers from data, and so the
+                            // post-compaction note can list what it just stubbed.
+                            recordElision(toolCallId, toolName, fullContent.length, sn);
+                            try {
+                                getDefaultEEClient()
+                                    .extract({
+                                    transcript: fullContent.slice(0, 8000),
+                                    projectPath: _cwd,
+                                    meta: {
+                                        source: "tool-artifact",
+                                        toolCallId,
+                                        toolName,
+                                        reason,
+                                        sessionId: _sess,
+                                        elidedAtStep: sn,
+                                    },
+                                }, AbortSignal.timeout(700))
+                                    .catch(() => { });
+                            }
+                            catch {
+                                /* fail-open, no silent swallow of the decision */
+                            }
+                        };
+                        // T1.1 + T1.2 — reasoning models (DeepSeek V4 Flash, R1) emit 2K-5K
+                        // CoT tokens per turn that accumulate across the multi-step loop.
+                        // Strip old reasoning and compact earlier (ratio 0.3 vs 0.5) to
+                        // cut ~40-60% of cumulative input tokens.
+                        const isReasoningModel = runtime.modelInfo?.reasoning === true;
+                        const compacted = compactSubAgentMessages(stripped, {
+                            thresholdChars: topLevelCompactThreshold,
+                            // Rec #1 (cheap part): on meta/self-eval turns keep a couple more
+                            // trailing tool turns verbatim — those carry the reasoning the
+                            // agent is being asked to reflect on, and over-eliding them is
+                            // exactly what starves a self-evaluation. One boolean, no new
+                            // detection logic (isMetaAnalysisPrompt already gates layer3/5).
+                            keepLastTurns: topLevelCompactKeepLast + (isMetaAnalysisPrompt(userMessage) ? 2 : 0),
+                            label: "top-level",
+                            envelopeChars,
+                            contextWindowTokens,
+                            contextFillRatio: isReasoningModel ? 0.3 : undefined,
+                            keepToolIds: keepToolIds.length ? keepToolIds : undefined,
+                            persistArtifact,
+                            stripOldReasoning: isReasoningModel,
+                        });
+                        if (compacted !== stripped)
+                            recordCompaction(sn);
+                        // Pre-compaction visibility: give the agent one step of notice
+                        // before B4 actually rewrites history into stubs. This is the
+                        // advance warning that was missing — agent can now decide to
+                        // summarize, finish, or request preservation. Fires when we did
+                        // NOT compact this step (compacted === stripped, restored by the
+                        // compactSubAgentMessages no-op ref contract) AND the prompt is
+                        // approaching the threshold. Must compare CHARS (messages +
+                        // envelope), not stripped.length (a message count that never
+                        // exceeds a char-scaled threshold) — session 2b7a10219499.
+                        const _preWarnChars = cumulativeMessageChars(stripped) + envelopeChars;
+                        if (compacted === stripped && shouldPreWarnCompaction(_preWarnChars, topLevelCompactThreshold)) {
+                            const _cp = buildCheckpointReminder(sn, true);
+                            const _pre = `[pre-compaction warning at step ${sn} — next step(s) will likely rewrite older tool results to stubs (threshold ${topLevelCompactThreshold}, keepLast=${topLevelCompactKeepLast}). ${_cp} Summarize or finish if possible, or warn the user they can run the "/compact" command if they want a clean compressed history.]`;
+                            return withSteers({ messages: attachReminderToMessages(stripped, _pre) });
+                        }
+                        // Phase 4A — scope reminder injection (REQ-005).
+                        // Cadence K = 3/5/8 for small/medium/large. Soft-warn fires
+                        // ONCE per session at floor(ceiling*0.7). Reminder lives in
+                        // the tool_result/system channel so B3/B4 compaction cannot
+                        // strip it (system-prompt path is unsafe at high step counts).
+                        // Ceiling reuses the 4B (task_type × size) matrix result
+                        // resolved above (`_stepCeiling`, `_ceilingTaskType`,
+                        // `_ceilingSize`, `_ceilingSessionId`) so the reminder and the
+                        // halt boundary agree on the same number.
+                        const _scopeSize = _ceilingSize;
+                        const _scopeK = cadenceForSize(_scopeSize);
+                        const _scopeCeiling = Math.max(1, _stepCeiling ?? deps.maxToolRounds ?? 30);
+                        const _scopeStep = sn;
+                        const _shouldRemind = shouldInjectReminder(_scopeStep, _scopeK);
+                        const _shouldWarn = shouldInjectSoftWarn(_scopeStep, _scopeCeiling, _ceilingSessionId);
+                        // Phase 5 Fix 5 (revised) — past the natural matrix ceiling the
+                        // orchestrator emits a STRONG re-anchor reminder, but only when
+                        //   (a) crossing the ceiling for the first time (one-shot), OR
+                        //   (b) hitting a normal cadence step (multiple of K).
+                        // Original Phase 5 Fix 5 fired on EVERY step past ceiling, which
+                        // on long-running sessions (e.g. step 77 / ceiling 6 in session
+                        // 1f29e238a816) produced 70+ redundant reminders that bloated
+                        // the tool_result channel and forced the model into a "YES still
+                        // on scope" loop on every tool call.
+                        const _pastNaturalCeiling = _scopeStep > _naturalCeiling;
+                        const _justCrossedCeiling = shouldInjectCeilingCrossing(_scopeStep, _naturalCeiling, _ceilingSessionId);
+                        const _pastCeilingAtCadence = _pastNaturalCeiling && _shouldRemind;
+                        // Fix #8 — self-repetition one-shot. Fires when the assistant
+                        // has opened the last 3 streamText steps with the same 4-word
+                        // phrase (e.g. "YES still on scope" — session 1f29e238a816
+                        // emitted 15 such bursts past ceiling). Reminder is attached
+                        // alongside (and before) any scope reminder so the model sees
+                        // the behavioural correction first.
+                        const _shouldRepeatReminder = shouldInjectRepetitionReminder(_ceilingSessionId);
+                        if (_shouldRemind || _shouldWarn || _justCrossedCeiling || _shouldRepeatReminder) {
+                            const _baseReminder = buildScopeReminder({
+                                step: _scopeStep,
+                                ceiling: _scopeCeiling,
+                                taskType: _ceilingTaskType,
+                                size: _scopeSize,
+                                originalPrompt: userMessage,
+                            });
+                            // Strong "past natural budget" prefix only applies when we
+                            // ACTUALLY want the model to consider wrapping up — i.e. on
+                            // the crossing event or at a cadence step past ceiling, not
+                            // on every silent step in between.
+                            const _useStrong = _justCrossedCeiling || _pastCeilingAtCadence;
+                            const _scopePart = _shouldRemind || _shouldWarn || _justCrossedCeiling
+                                ? _useStrong
+                                    ? `[past natural budget — step ${_scopeStep}/${_naturalCeiling}] If task is COMPLETE, emit final answer NOW. If you need to keep working in this long session, suggest that the user run the "/compact" slash command to compress the conversation history before continuing. Otherwise, simplify the next step. ${_baseReminder}`
+                                    : _shouldWarn
+                                        ? `[approaching ceiling] ${_baseReminder}`
+                                        : _baseReminder
+                                : null;
+                            const _reminder = _shouldRepeatReminder
+                                ? _scopePart
+                                    ? `${buildRepetitionReminder(_ceilingSessionId)}\n${_scopePart}`
+                                    : buildRepetitionReminder(_ceilingSessionId)
+                                : _scopePart;
+                            const withReminder = attachReminderToMessages(compacted, _reminder);
+                            return withSteers({ messages: withReminder });
+                        }
+                        if (compacted === stripped && stripped === stepMessages)
+                            return withSteers({});
+                        // Self-awareness note: tell the model compaction happened so it
+                        // knows earlier context was elided and can adjust its behavior.
+                        // Enhanced per EE anti-mù plan (docs/ee-anti-mu-compaction-plan.md Phase 2): include proactive
+                        // "task finished?", "compacted yet?", "EE checkpoint" so agent can self-assess and avoid mù
+                        // even when the top-level summary is not in its immediate focus (sub-agents, long loops).
+                        const _compactNote = compacted !== stripped
+                            ? (() => {
+                                // Rec #2: turn the generic "high-value elided? use ee_query"
+                                // prose into a concrete, actionable manifest of what was just
+                                // stubbed (id/tool/size) — sourced from the elisions recorded
+                                // by persistArtifact above — so the rehydrate round-trip is
+                                // informed, not blind.
+                                const _m = formatElisionManifest();
+                                return `[context compacted at step ${sn} — older or low-value tool results rewritten to stubs to fit budget. High-value evidence (file reads, bash, your previous responses) is kept verbatim. ${buildCheckpointReminder(sn, true)}${_m ? ` ${_m}` : ""}]`;
+                            })()
+                            : null;
+                        if (_compactNote) {
+                            return withSteers({ messages: attachReminderToMessages(compacted, _compactNote) });
+                        }
+                        return withSteers({ messages: compacted });
+                    },
+                    ...(dropParam("temperature") ? {} : { temperature: 0.7 }),
+                    ...(dropParam("maxOutputTokens") ? {} : { maxOutputTokens: taskTypeToMaxTokens(pilCtx.taskType) }),
+                    ...(Object.keys(providerOpts).length > 0 ? { providerOptions: providerOpts } : {}),
+                    experimental_onStepStart: (event) => {
+                        stepNumber = getStepNumber(event, stepNumber + 1);
+                        notifyObserver(observer?.onStepStart, {
+                            stepNumber,
+                            timestamp: Date.now(),
+                        });
+                    },
+                    onStepFinish: (event) => {
+                        const currentStep = getStepNumber(event, Math.max(stepNumber, 0));
+                        stepNumber = Math.max(stepNumber, currentStep);
+                        const stepUsage = getUsage(event);
+                        notifyObserver(observer?.onStepFinish, {
+                            stepNumber: currentStep,
+                            timestamp: Date.now(),
+                            finishReason: getFinishReason(event),
+                            usage: stepUsage,
+                        });
+                        // Realtime status bar update per step
+                        if (stepUsage.inputTokens || stepUsage.outputTokens) {
+                            // O1 — thread THIS turn's providerOptions shape per step so every
+                            // step event records it (not just step 1) and an interleaved task
+                            // can't overwrite it. Mirrors the gate used for the call itself.
+                            deps.recordUsage(stepUsage, "message", runtime.modelId, Object.keys(providerOpts).length > 0 ? extractProviderOptionsShape(providerOpts) : null);
+                        }
+                        // Fix #8 — feed the assistant text emitted in this step into
+                        // the self-repetition detector. The slice covers everything
+                        // appended to `assistantText` since the previous step boundary;
+                        // a step with no text (pure tool call) records as empty, which
+                        // recordAssistantBurst treats as a no-op so the current run is
+                        // preserved across tool interludes.
+                        const _stepText = assistantText.slice(_assistantTextAtLastStep);
+                        _assistantTextAtLastStep = assistantText.length;
+                        recordAssistantBurst(_ceilingSessionId, _stepText);
+                    },
+                    onFinish: ({ finishReason }) => {
+                        _lastFinishReason = finishReason ?? null;
+                        // Task 2.6b — emit llm-done (agent-mode only).
+                        try {
+                            const _ar = globalThis.__muonroiAgentRuntime;
+                            _ar?.emitEvent({
+                                t: "event",
+                                kind: "llm-done",
+                                correlationId: _topCallId,
+                                totalChars: assistantText.length,
+                                finishReason: finishReason ?? "stop",
+                            });
+                        }
+                        catch (err) {
+                            logger.error("orchestrator", "failed to emit llm-done", { error: err });
+                        }
+                        deps.setCurrentCallId("");
+                        // Rec #1 persisted forensics: onFinish fires once per top-level turn,
+                        // so flush this session's cumulative experience counts here. Readers
+                        // take the latest row per session, so the last turn's row is the
+                        // session total. No-ops on missing id / all-zero. Fail-open.
+                        try {
+                            persistSessionExperience(deps.session?.id ?? null, getSessionExperienceCounts());
+                        }
+                        catch (err) {
+                            logger.error("orchestrator", "persistSessionExperience failed", { error: err });
+                        }
+                    },
+                });
+                let _topTokenIndex = 0;
+                const _wireProviderIdTop = runtime.modelInfo?.provider ?? "unknown";
+                for await (const part of result.fullStream) {
+                    stall.pet(); // chunk arrived — reset the stall watchdog
+                    // Count only real content parts. The watchdog abort itself surfaces
+                    // as an "abort" part — counting it would defeat the TTFB-stall gate
+                    // (a frozen-before-first-byte stall yields ONLY the abort part).
+                    if (part.type !== "abort") {
+                        chunksThisAttempt++;
+                        chunksThisStep++;
+                    }
+                    if (signal.aborted) {
+                        yield { type: "content", content: "\n\n[Cancelled]" };
+                        break;
+                    }
+                    if (wireDebug.enabled) {
+                        wireDebug.logChunk(_wireProviderIdTop, String(part.type ?? "unknown"), {
+                            hasText: typeof part.text === "string"
+                                ? part.text.length
+                                : undefined,
+                            hasReasoning: typeof part.reasoning === "string"
+                                ? part.reasoning.length
+                                : undefined,
+                        });
+                        if (part.type === "error") {
+                            wireDebug.logError(_wireProviderIdTop, part.error);
+                        }
+                    }
+                    switch (part.type) {
+                        case "text-delta":
+                            assistantText += part.text;
+                            // Task 2.6b — emit llm-token (agent-mode only; high-volume, default-off per Phase 4).
+                            try {
+                                const _ar = globalThis.__muonroiAgentRuntime;
+                                _ar?.emitEvent({
+                                    t: "event",
+                                    kind: "llm-token",
+                                    correlationId: _topCallId,
+                                    delta: part.text,
+                                    tokenIndex: _topTokenIndex++,
+                                });
+                            }
+                            catch {
+                                /* best-effort */
+                            }
+                            yield { type: "content", content: part.text };
+                            break;
+                        case "reasoning-delta":
+                            reasoningPreview = `${reasoningPreview}${part.text}`.slice(-256);
+                            if (containsEncryptedReasoning(reasoningPreview)) {
+                                if (!encryptedReasoningHidden) {
+                                    encryptedReasoningHidden = true;
+                                    yield { type: "reasoning", content: "[Encrypted reasoning hidden]" };
+                                }
+                                break;
+                            }
+                            // P0 native observation: accumulate reasoning for intent context.
+                            deps.appendTurnAssistantReasoning(part.text);
+                            yield { type: "reasoning", content: part.text };
+                            break;
+                        case "tool-call": {
+                            const tc = toToolCall(part);
+                            activeToolCalls.push(tc);
+                            // SAMR: track that Phase 1 produced tool calls → transition to Phase 2
+                            if (stepRouterPhase === "phase1")
+                                phase1HadToolCalls = true;
+                            // Response tool = the terminal final answer (identity execute;
+                            // the payload lives in the call args). Buffer it (longest-wins)
+                            // straight from the args and gate UI/DB/exec spam: cheap models
+                            // sometimes emit the response tool MANY times in ONE generation
+                            // (session 8d8f498268ed: 80× identical respond_general hedge in
+                            // one step). stopWhen only halts BETWEEN steps, so it can't stop
+                            // an in-step spam — this does. Surface only the first indicator;
+                            // if the model spams past the cap, finalize NOW with the
+                            // buffered answer instead of streaming out the degenerate step.
+                            if (isResponseTool(tc.function.name)) {
+                                _responseToolEmitCount += 1;
+                                try {
+                                    const _payload = JSON.parse(tc.function.arguments || "{}");
+                                    const _len = JSON.stringify(_payload).length;
+                                    if (_len > _pendingStructuredResponseLen) {
+                                        _pendingStructuredResponseLen = _len;
+                                        _pendingStructuredResponse = {
+                                            taskType: getResponseTaskType(tc.function.name) ?? tc.function.name,
+                                            data: _payload,
+                                        };
+                                    }
+                                }
+                                catch {
+                                    /* keep the prior buffered payload */
+                                }
+                                responseToolCalled = true;
+                                // Only the first response-tool call gets a UI indicator.
+                                if (_responseToolEmitCount === 1) {
+                                    yield { type: "tool_calls", toolCalls: [tc] };
+                                }
+                                if (_responseToolEmitCount >= RESPONSE_TOOL_SPAM_CAP && _pendingStructuredResponse) {
+                                    if (deps.session) {
+                                        try {
+                                            logInteraction(deps.session.id, "f6_synthesis", {
+                                                eventSubtype: "response_tool_spam_abort",
+                                                data: { emitted: _responseToolEmitCount, keptChars: _pendingStructuredResponseLen },
+                                            });
+                                        }
+                                        catch {
+                                            /* telemetry best-effort */
+                                        }
+                                    }
+                                    // Persist a clean turn (user + the single buffered answer)
+                                    // so history stays usable; the spam is dropped. Mirrors the
+                                    // tool-repetition abort: yield + done + return (do NOT await
+                                    // result.response — the stream is still spewing calls).
+                                    const _data = _pendingStructuredResponse.data;
+                                    const _answerText = typeof _data.response === "string"
+                                        ? _data.response
+                                        : JSON.stringify(_pendingStructuredResponse.data);
+                                    try {
+                                        deps.appendCompletedTurn(userModelMessage, [
+                                            { role: "assistant", content: _answerText },
+                                        ]);
+                                    }
+                                    catch (persistErr) {
+                                        console.error(`[message-processor] response-tool-spam persist failed: ${persistErr?.message}`);
+                                    }
+                                    yield {
+                                        type: "structured_response",
+                                        structuredResponse: _pendingStructuredResponse,
+                                    };
+                                    yield { type: "done" };
+                                    return;
+                                }
+                                break; // response tools skip write-ahead/hooks/normal tool_calls yield
+                            }
+                            // EE PreToolUse hook: fire intercept before tool execution.
+                            {
+                                const turnAssistantReasoning = deps.getTurnAssistantReasoning();
+                                const intentContext = {
+                                    ...(turnAssistantReasoning ? { assistantReasoningExcerpt: turnAssistantReasoning.slice(-200) } : {}),
+                                    ...(deps.priorWarningIdsInSession?.size > 0
+                                        ? {
+                                            priorWarningIdsInSession: Array.from(deps.priorWarningIdsInSession).slice(-20),
+                                        }
+                                        : {}),
+                                    ...(pilCtx.gsdPhase ? { gsdPhase: pilCtx.gsdPhase } : {}),
+                                    ...(userMessage.slice(0, 200) ? { userGoalExcerpt: userMessage.slice(0, 200) } : {}),
+                                };
+                                const preInput = {
+                                    hook_event_name: "PreToolUse",
+                                    tool_name: tc.function.name,
+                                    tool_input: JSON.parse(tc.function.arguments || "{}"),
+                                    session_id: deps.session?.id,
+                                    cwd: deps.bash.getCwd(),
+                                    ...(Object.keys(intentContext).length > 0 ? { intent_context: intentContext } : {}),
+                                };
+                                const preResult = await deps.fireHook(preInput, signal).catch(() => ({
+                                    blocked: false,
+                                    blockingErrors: [],
+                                    preventContinuation: false,
+                                    additionalContexts: [],
+                                    results: [],
+                                    eeMatches: [],
+                                }));
+                                for (const ctx of preResult.additionalContexts ?? []) {
+                                    yield { type: "content", content: `${ctx}\n` };
+                                }
+                                // Store structured EE matches for session guidance injection on next turn.
+                                for (const m of preResult.eeMatches ?? []) {
+                                    deps.sessionEEGuidance.set(m.id, {
+                                        toolName: m.toolName,
+                                        message: m.message,
+                                        why: m.why,
+                                        confidence: m.confidence,
+                                    });
+                                    // Cap at 30 entries — oldest first, trim when exceeded.
+                                    if (deps.sessionEEGuidance.size > 30) {
+                                        const firstKey = deps.sessionEEGuidance.keys().next().value;
+                                        if (firstKey !== undefined)
+                                            deps.sessionEEGuidance.delete(firstKey);
+                                    }
+                                }
+                                // P0 native observation: track which principle IDs surfaced
+                                // this turn so the next intercept can dedup server-side.
+                                try {
+                                    const { getLastSurfacedState } = await import("../ee/intercept.js");
+                                    const { surfacedIds } = getLastSurfacedState();
+                                    for (const id of surfacedIds)
+                                        deps.priorWarningIdsInSession.add(id);
+                                    // Cap memory: keep only most-recent 100 IDs.
+                                    if (deps.priorWarningIdsInSession.size > 100) {
+                                        const arr = Array.from(deps.priorWarningIdsInSession);
+                                        deps.setPriorWarningIdsInSession(new Set(arr.slice(-100)));
+                                    }
+                                }
+                                catch {
+                                    /* fail-open */
+                                }
+                            }
+                            // Pitfall 9: log the pending call so reconcile() can recover any
+                            // staged .tmp files if the process is killed before tool-result.
+                            if (deps.pendingCalls) {
+                                const turnId = deps.session?.id ?? "anon";
+                                const callId = stableCallId(turnId, tc.function.name, tc.function.arguments);
+                                // Phase 0: predictStagedPaths = [] for all tools (refined in Phase 1).
+                                void deps.pendingCalls.begin({ call_id: callId, tool_name: tc.function.name }).catch(() => { });
+                                // Attach callId to the ToolCall so tool-result can end it.
+                                tc._pendingCallId = callId;
+                            }
+                            // Phase A4: write-ahead persistence — insert a pending row into
+                            // tool_calls BEFORE executing the tool. If the stream throws
+                            // mid-call (e.g. provider 5xx, abort, network drop), this row
+                            // remains as `pending` so `usage forensics` can show the args
+                            // the model passed. The post-stream appendMessages() path
+                            // (INSERT OR IGNORE + UPDATE) will finalize this row to
+                            // `completed` once the turn settles normally.
+                            if (deps.sessionStore && deps.session) {
+                                // Predicted assistant seq: user message + assistant message
+                                // are appended atomically by appendCompletedTurn().
+                                // getNextMessageSequence() returns the seq the user message
+                                // will get; the assistant message is the next one after.
+                                let predictedSeq = -1;
+                                try {
+                                    predictedSeq = getNextMessageSequence(deps.session.id) + 1;
+                                }
+                                catch {
+                                    /* fail-open — leave predictedSeq=-1; post-stream UPDATE corrects it */
+                                }
+                                persistToolCallWriteAhead(deps.session.id, predictedSeq, tc.id, tc.function.name, tc.function.arguments || "{}");
+                            }
+                            notifyObserver(observer?.onToolStart, {
+                                toolCall: tc,
+                                timestamp: Date.now(),
+                            });
+                            // Interaction log: tool call start
+                            try {
+                                if (deps.session) {
+                                    logInteraction(deps.session.id, "tool_call", {
+                                        eventSubtype: tc.function.name,
+                                        data: {
+                                            toolCallId: tc.id,
+                                            argsPreview: tc.function.arguments.slice(0, 200),
+                                        },
+                                    });
+                                }
+                            }
+                            catch {
+                                /* fail-open */
+                            }
+                            yield { type: "tool_calls", toolCalls: [tc] };
+                            break;
+                        }
+                        case "tool-result": {
+                            const tc = {
+                                id: part.toolCallId,
+                                type: "function",
+                                function: { name: part.toolName, arguments: JSON.stringify(part.input ?? {}) },
+                            };
+                            let tr = toToolResult(part.output);
+                            // Vision Bridge: proxy image-bearing tool results for text-only models (any tool, not just MCP)
+                            try {
+                                const bridgeResult = await bridgeMcpToolResult(part.toolName, tr.output, turnModelId, signal, part.toolCallId);
+                                if (bridgeResult.proxied) {
+                                    tr = {
+                                        ...tr,
+                                        output: typeof bridgeResult.output === "string"
+                                            ? bridgeResult.output
+                                            : JSON.stringify(bridgeResult.output),
+                                    };
+                                    yield { type: "content", content: `[Vision Bridge: image → text for ${turnModelId}]\n` };
+                                }
+                            }
+                            catch (err) {
+                                console.error("[Agent:visionBridge] failed to process image for tool result", err);
+                            }
+                            // Safety-block intercept: when bash.execute returns a
+                            // "BLOCKED (...):" error, surface an askcard to the user via
+                            // deps.askSafetyOverride. If allowed, rewrite the output as a
+                            // success so the model knows the command was approved (it may
+                            // retry on the next turn). The approved command is stored in
+                            // the global __muonroiSafetyApproved map so registry.ts's
+                            // bash.execute bypasses the block on the retry.
+                            const _outputText = [tr.output, tr.error].filter((x) => typeof x === "string").join("\n");
+                            const _blockMatch = _outputText.match(/^BLOCKED \(([^)]+)\):\s*(.*)/);
+                            if (_blockMatch && part.toolName === "bash") {
+                                const _blockKind = _blockMatch[1];
+                                const _blockReason = _blockMatch[2];
+                                const _command = typeof part.input === "object" && part.input != null
+                                    ? String(part.input.command ?? "")
+                                    : "";
+                                // empty-bash: auto-block without askcard.
+                                // The BLOCKED error from registry.ts already steers the agent
+                                // to self-correct. Showing an askcard adds user friction for
+                                // no benefit (empty bash calls are never intentional) and
+                                // wastes a turn of user interaction per strike.
+                                if (_blockKind === "empty-bash") {
+                                    tr = { ...tr, success: false, error: _outputText, output: _outputText };
+                                }
+                                else {
+                                    const _verdict = deps.askSafetyOverride
+                                        ? await deps.askSafetyOverride({
+                                            kind: _blockKind,
+                                            toolName: part.toolName,
+                                            blockedItem: _command,
+                                            reason: _blockReason,
+                                            source: "bash.execute",
+                                        })
+                                        : { action: "block" };
+                                    if (_verdict.action === "allow-once" || _verdict.action === "allow-session") {
+                                        // Store approval so registry.ts can bypass the block on retry.
+                                        const _globalSafety = globalThis;
+                                        if (!_globalSafety.__muonroiSafetyApproved) {
+                                            _globalSafety.__muonroiSafetyApproved = new Map();
+                                        }
+                                        _globalSafety.__muonroiSafetyApproved.set(part.toolCallId, {
+                                            kind: _verdict.action === "allow-session" ? "session" : "once",
+                                            command: _command,
+                                        });
+                                        // Rewrite tool result as success so the stream continues
+                                        // without an error. The model will see "Approved: ..." and
+                                        // may retry the tool call on the next turn, at which point
+                                        // registry.ts will see the approval and actually run it.
+                                        tr = { ...tr, success: true, output: `Approved (${_verdict.action}): ${_blockReason}` };
+                                        yield {
+                                            type: "content",
+                                            content: `[User approved blocked command: ${_blockKind} — ${_verdict.action}]\n`,
+                                        };
+                                    }
+                                    if (_verdict.action === "block") {
+                                        tr = { ...tr, success: false, error: _outputText, output: _outputText };
+                                    }
+                                }
+                            }
+                            // Capture into the stall-rescue digest before any further
+                            // processing — if the stream stalls after this, these outputs
+                            // are all we have to synthesize a final answer from.
+                            pushStallToolResult(turnToolResults, part.toolName, typeof tr.output === "string" ? tr.output : JSON.stringify(tr.output));
+                            // Pitfall 9: settle the pending call log entry.
+                            if (deps.pendingCalls) {
+                                const pending = activeToolCalls.find((t) => t.id === part.toolCallId);
+                                const callId = pending?._pendingCallId;
+                                if (callId) {
+                                    const endStatus = signal.aborted ? "aborted" : "settled";
+                                    void deps.pendingCalls.end(callId, endStatus).catch(() => { });
+                                }
+                            }
+                            // EE PostToolUse hook: fire-and-forget after tool execution.
+                            {
+                                const postInput = {
+                                    hook_event_name: "PostToolUse",
+                                    tool_name: part.toolName,
+                                    tool_input: part.input ?? {},
+                                    tool_output: typeof tr.output === "string"
+                                        ? { text: tr.output }
+                                        : (tr.output ?? {}),
+                                    session_id: deps.session?.id,
+                                    cwd: deps.bash.getCwd(),
+                                };
+                                await deps.fireHook(postInput, signal).catch((err) => {
+                                    console.error("[Agent:PostToolUse hook] failed", err);
+                                });
+                            }
+                            // Response tool: yield as structured_response instead of tool_result.
+                            // AI SDK v5 wraps tool outputs as `{type:"json", value:{...}}`; unwrap
+                            // to expose the schema-shaped payload to the UI renderer.
+                            if (isResponseTool(part.toolName)) {
+                                responseToolCalled = true;
+                                // Payload was already buffered (longest-wins) from the
+                                // tool-CALL args above; re-buffer from the executed result as
+                                // a fallback (unwraps the AI-SDK `{type:"json",value}` shape).
+                                // Counting + the spam cap live in the tool-call branch.
+                                const taskType = getResponseTaskType(part.toolName);
+                                const rawOutput = part.output;
+                                const unwrapped = rawOutput && typeof rawOutput === "object" && rawOutput.type === "json"
+                                    ? (rawOutput.value ?? {})
+                                    : (rawOutput ?? {});
+                                const _len = JSON.stringify(unwrapped ?? {}).length;
+                                if (_len > _pendingStructuredResponseLen) {
+                                    _pendingStructuredResponseLen = _len;
+                                    _pendingStructuredResponse = {
+                                        taskType: taskType ?? part.toolName,
+                                        data: unwrapped,
+                                    };
+                                }
+                                notifyObserver(observer?.onToolFinish, { toolCall: tc, toolResult: tr, timestamp: Date.now() });
+                                break;
+                            }
+                            notifyObserver(observer?.onToolFinish, {
+                                toolCall: tc,
+                                toolResult: tr,
+                                timestamp: Date.now(),
+                            });
+                            // Interaction log: tool result.
+                            // Phase 5 BUG-J — for edit/write/update tools, persist the
+                            // structured diff (file_path, +N/-M counts, isNew flag, and
+                            // a bounded patch preview) so forensics queries can audit
+                            // what actually changed in each turn without re-reading
+                            // git history. Earlier the log only had the summary string
+                            // ("Edited X (+1 -1)") — the patch text was lost.
+                            try {
+                                if (deps.session) {
+                                    const outputPreview = typeof tr.output === "string" ? tr.output.slice(0, 200) : JSON.stringify(tr.output).slice(0, 200);
+                                    const _trWithDiff = tr;
+                                    const diffMeta = _trWithDiff.diff &&
+                                        (tc.function.name === "edit_file" ||
+                                            tc.function.name === "write_file" ||
+                                            tc.function.name === "update_file")
+                                        ? {
+                                            filePath: _trWithDiff.diff.filePath,
+                                            additions: _trWithDiff.diff.additions,
+                                            removals: _trWithDiff.diff.removals,
+                                            isNew: _trWithDiff.diff.isNew,
+                                            // Cap at 4000 chars — enough to inspect small/medium
+                                            // edits without ballooning the SQLite row. Large
+                                            // refactors get truncated with a tail marker so
+                                            // readers know the patch is partial.
+                                            patchPreview: _trWithDiff.diff.patch.length > 4000
+                                                ? `${_trWithDiff.diff.patch.slice(0, 4000)}\n…[truncated]`
+                                                : _trWithDiff.diff.patch,
+                                        }
+                                        : undefined;
+                                    logInteraction(deps.session.id, "tool_result", {
+                                        eventSubtype: tc.function.name,
+                                        data: { success: tr.success, outputPreview, ...(diffMeta ? { diff: diffMeta } : {}) },
+                                    });
+                                }
+                            }
+                            catch {
+                                /* fail-open */
+                            }
+                            yield { type: "tool_result", toolCall: tc, toolResult: tr };
+                            // Reset tool-repetition counter on any non-error result. A
+                            // successful call between two failures of the same shape is
+                            // progress and should not accumulate toward the abort gate.
+                            if (tr.success) {
+                                recordToolRepetitionSuccess(deps.session?.id ?? null);
+                            }
+                            // todo_write side-effect: surface the task list to the UI via a
+                            // dedicated chunk so the sticky checklist panel can re-render
+                            // without parsing tool args itself. Skipped when the snapshot
+                            // doesn't parse (malformed args) so the UI is never poisoned.
+                            if (tc.function.name === "todo_write" && tr.success) {
+                                const snap = snapshotFromTodoWriteArgs(tc.function.arguments);
+                                if (snap)
+                                    yield { type: "task_list_update", taskListSnapshot: snap };
+                            }
+                            break;
+                        }
+                        case "tool-error": {
+                            // AI SDK emits this when tool execution throws/aborts before
+                            // producing a tool-result. Without this branch, the tool_call
+                            // log row has no matching tool_result and the EE judge never
+                            // sees the failure → silent ~1.6% pairing leak in prod DB.
+                            const errPart = part;
+                            const tc = {
+                                id: errPart.toolCallId,
+                                type: "function",
+                                function: { name: errPart.toolName, arguments: JSON.stringify(errPart.input ?? {}) },
+                            };
+                            const errMsg = errPart.error instanceof Error
+                                ? errPart.error.message
+                                : typeof errPart.error === "string"
+                                    ? errPart.error
+                                    : JSON.stringify(errPart.error);
+                            const tr = { success: false, output: `[tool-error] ${errMsg}` };
+                            // A respond_* (response) tool that ERRORED still carries the
+                            // model's terminal answer in its call args — for response tools
+                            // execution is identity, so an execution failure is usually a
+                            // post-processing issue and the payload is intact. Recover it so
+                            // the turn is not left empty (textLength:0) and the answer is not
+                            // swallowed on turn-finalize. Mirrors the tool-call / tool-result
+                            // buffering above.
+                            if (isResponseTool(errPart.toolName)) {
+                                if (!_pendingStructuredResponse && errPart.input && typeof errPart.input === "object") {
+                                    const data = errPart.input;
+                                    const _len = JSON.stringify(data).length;
+                                    if (_len > _pendingStructuredResponseLen) {
+                                        _pendingStructuredResponseLen = _len;
+                                        _pendingStructuredResponse = {
+                                            taskType: getResponseTaskType(errPart.toolName) ?? errPart.toolName,
+                                            data,
+                                        };
+                                    }
+                                }
+                                if (!_pendingStructuredResponse) {
+                                    // Nothing recoverable (args never parsed). Do NOT suppress F6
+                                    // synthesis — let it produce a visible prose answer instead
+                                    // of an empty turn.
+                                    responseToolCalled = false;
+                                }
+                            }
+                            // Settle pending-call ledger so we don't leak stale .tmp files.
+                            if (deps.pendingCalls) {
+                                const pending = activeToolCalls.find((t) => t.id === errPart.toolCallId);
+                                const callId = pending?._pendingCallId;
+                                if (callId)
+                                    void deps.pendingCalls.end(callId, "settled").catch(() => { });
+                            }
+                            // Phase A4: mark the write-ahead tool_calls row as `errored`.
+                            // The post-stream appendMessages() path does NOT see tool-error
+                            // parts in the assistant message content (the SDK doesn't emit
+                            // them there), so without this explicit update the row would
+                            // remain `pending` after a clean tool failure.
+                            if (deps.session) {
+                                markToolCallErrored(deps.session.id, errPart.toolCallId, errMsg);
+                            }
+                            // Fire PostToolUseFailure so EE judge can record IGNORED outcome.
+                            {
+                                const failInput = {
+                                    hook_event_name: "PostToolUseFailure",
+                                    tool_name: errPart.toolName,
+                                    tool_input: errPart.input ?? {},
+                                    error: errMsg,
+                                    session_id: deps.session?.id,
+                                    cwd: deps.bash.getCwd(),
+                                };
+                                await deps.fireHook(failInput, signal).catch(() => { });
+                            }
+                            try {
+                                if (deps.session) {
+                                    logInteraction(deps.session.id, "tool_result", {
+                                        eventSubtype: errPart.toolName,
+                                        data: { success: false, error: errMsg.slice(0, 500), reason: "tool-error" },
+                                    });
+                                }
+                            }
+                            catch (logErr) {
+                                console.error(`[message-processor] interaction-log tool_result failed: ${logErr?.message}`);
+                            }
+                            notifyObserver(observer?.onToolFinish, { toolCall: tc, toolResult: tr, timestamp: Date.now() });
+                            yield { type: "tool_result", toolCall: tc, toolResult: tr };
+                            // Tool-call perseveration guard. After N consecutive identical
+                            // (toolName, args, error) triples, abort the streaming loop
+                            // before TPM rate limits do (session 080fe2fcbf24).
+                            const repetition = recordToolRepetitionError(deps.session?.id ?? null, errPart.toolName, errPart.input, errMsg);
+                            if (repetition.shouldAbort) {
+                                const abortMsg = buildToolRepetitionAbortMessage(errPart.toolName, repetition.runLength, errMsg);
+                                try {
+                                    if (deps.session) {
+                                        logInteraction(deps.session.id, "error", {
+                                            eventSubtype: "tool_repetition_abort",
+                                            data: {
+                                                toolName: errPart.toolName,
+                                                runLength: repetition.runLength,
+                                                errorPreview: errMsg.slice(0, 200),
+                                            },
+                                        });
+                                    }
+                                }
+                                catch (logErr) {
+                                    console.error(`[message-processor] interaction-log tool_repetition_abort failed: ${logErr?.message}`);
+                                }
+                                notifyObserver(observer?.onError, { message: abortMsg, timestamp: Date.now() });
+                                yield { type: "error", content: abortMsg, isAuthError: false };
+                                yield { type: "done" };
+                                return;
+                            }
+                            break;
+                        }
+                        case "tool-approval-request": {
+                            const approvalPart = part;
+                            const toolCallId = approvalPart.toolCall?.toolCallId ?? "";
+                            const pendingTc = activeToolCalls.find((tc) => tc.id === toolCallId);
+                            const tcForChunk = pendingTc ?? {
+                                id: toolCallId,
+                                type: "function",
+                                function: {
+                                    name: approvalPart.toolCall?.toolName ?? "paid_request",
+                                    arguments: JSON.stringify(approvalPart.toolCall?.input ?? {}),
+                                },
+                            };
+                            // Payment pre-check disabled — Stripe billing pending.
+                            const paymentPrecheck = undefined;
+                            // Plan 03-01: check permission mode before yielding approval request to UI.
+                            // auto-edit auto-approves file ops; yolo auto-approves everything.
+                            const toolName = approvalPart.toolCall?.toolName ?? "";
+                            const input = approvalPart.toolCall?.input ?? {};
+                            const context = toolName === "bash"
+                                ? { command: String(input.command ?? "") }
+                                : toolName === "write_file" ||
+                                    toolName === "edit_file" ||
+                                    toolName === "read_file" ||
+                                    toolName === "grep"
+                                    ? { path: String(input.path ?? input.file_path ?? "") }
+                                    : undefined;
+                            if (!toolNeedsApproval(toolName, deps.permissionMode, context)) {
+                                // Auto-approve: respond directly without surfacing to UI.
+                                deps.respondToToolApproval(approvalPart.approvalId, true);
+                                appendAudit({
+                                    kind: deps.permissionMode === "yolo" ? "yolo-override" : "permission-override",
+                                    tool: toolName,
+                                    mode: deps.permissionMode,
+                                    context,
+                                    ts: Date.now(),
+                                });
+                                break;
+                            }
+                            yield {
+                                type: "tool_approval_request",
+                                approvalId: approvalPart.approvalId,
+                                toolCall: tcForChunk,
+                                paymentPrecheck,
+                            };
+                            break;
+                        }
+                        case "error": {
+                            // F6b: a TRANSIENT error delivered as a stream PART (e.g. a provider
+                            // 5xx / dropped socket surfaced mid-loop after a tool step) otherwise
+                            // BYPASSES the thrown-error transient-retry path below — the turn would
+                            // surface the error and stop instead of retrying a flaky transient
+                            // (observed live: SiliconFlow 500 error-part ended a multi-step turn).
+                            // Route it through the SAME bounded budget (shared streamRetryCount /
+                            // MAX_STREAM_RETRIES) and the same no-content guard, so an error-part
+                            // and a thrown error behave identically. (A 2026-06-20 council debate
+                            // independently endorsed bounded auto-retry for exactly this case.)
+                            {
+                                const { transient: _partTransient } = classifyStreamError(part.error);
+                                if (_partTransient &&
+                                    !assistantText.trim() &&
+                                    streamRetryCount < MAX_STREAM_RETRIES &&
+                                    !signal.aborted &&
+                                    !stallTriggered) {
+                                    streamRetryCount++;
+                                    // Exponential backoff: 500 → 8000 ms with ±25% jitter (mirrors the
+                                    // thrown-error path).
+                                    const baseMs = 500;
+                                    const expMs = Math.min(baseMs * 4 ** (streamRetryCount - 1), 8_000);
+                                    const spread = expMs * 0.25;
+                                    const nextDelayMs = Math.round(expMs + (Math.random() * 2 - 1) * spread);
+                                    const errName = part.error instanceof Error ? part.error.name : "Error";
+                                    const errMsg = part.error instanceof Error ? part.error.message : String(part.error);
+                                    try {
+                                        const _ar = globalThis.__muonroiAgentRuntime;
+                                        _ar?.emitEvent({
+                                            t: "event",
+                                            kind: "stream-retry",
+                                            attempt: streamRetryCount,
+                                            maxAttempts: MAX_STREAM_RETRIES + 1,
+                                            errorName: errName,
+                                            errorMessage: errMsg,
+                                            nextDelayMs,
+                                        });
+                                    }
+                                    catch {
+                                        /* best-effort telemetry */
+                                    }
+                                    try {
+                                        if (deps.session) {
+                                            logInteraction(deps.session.id, "stream_retry", {
+                                                data: {
+                                                    attempt: streamRetryCount,
+                                                    maxAttempts: MAX_STREAM_RETRIES + 1,
+                                                    errorName: errName,
+                                                    errorMessage: errMsg.slice(0, 200),
+                                                    nextDelayMs,
+                                                    source: "error-part",
+                                                },
+                                            });
+                                        }
+                                    }
+                                    catch (logErr) {
+                                        console.error(`[message-processor] error-part retry log failed: ${logErr?.message}`);
+                                    }
+                                    await new Promise((resolve) => setTimeout(resolve, nextDelayMs));
+                                    if (!signal.aborted) {
+                                        continue streamAttempt;
+                                    }
+                                }
+                            }
+                            // F6b-2: a TRANSIENT error-part that arrives AFTER content/tool steps
+                            // (F6b's no-content guard above skipped it — assistantText/tools already
+                            // flowed). Try to graft the COMPLETED steps onto history and re-issue, so
+                            // a flaky 5xx after a tool step RESUMES instead of ending the turn (the
+                            // observed Cycle-1 "fail liên tục" case: read big file → finish-step → 500).
+                            // SAFE by construction: if result.response can't yield the completed steps
+                            // (it rejects on a hard stream error), _appended stays 0 and we fall through
+                            // to surface the error — identical to prior behavior. No tool re-run: the
+                            // completed steps are grafted onto history, not replayed. Shares the
+                            // mid-loop recovery budget (midLoopStallRetryCount / maxStallRetries).
+                            {
+                                const { transient: _contTransient } = classifyStreamError(part.error);
+                                if (_contTransient && midLoopStallRetryCount < maxStallRetries && !signal.aborted && !stallTriggered) {
+                                    let _appended = 0;
+                                    try {
+                                        const _resp = (await Promise.race([
+                                            result.response,
+                                            new Promise((_r, rej) => setTimeout(() => rej(new Error("response-timeout")), 3_000)),
+                                        ]));
+                                        const _gen = sanitizeModelMessages(scrubImagePayloadsInMessages(_resp.messages));
+                                        for (const _m of _gen) {
+                                            deps.messages.push(_m);
+                                            _appended++;
+                                        }
+                                    }
+                                    catch (respErr) {
+                                        console.error(`[message-processor] error-part continuation: completed steps unavailable: ${respErr?.message}`);
+                                    }
+                                    if (_appended > 0) {
+                                        midLoopStallRetryCount++;
+                                        const _contBackoff = stallRepromptBackoffMs(midLoopStallRetryCount);
+                                        const _errName = part.error instanceof Error ? part.error.name : "Error";
+                                        const _errMsg = part.error instanceof Error ? part.error.message : String(part.error);
+                                        try {
+                                            const _ar = globalThis.__muonroiAgentRuntime;
+                                            _ar?.emitEvent({
+                                                t: "event",
+                                                kind: "stream-retry",
+                                                attempt: midLoopStallRetryCount,
+                                                maxAttempts: maxStallRetries + 1,
+                                                errorName: _errName,
+                                                errorMessage: "transient error-part after tool step — resuming from completed steps",
+                                                nextDelayMs: _contBackoff,
+                                            });
+                                            _ar?.emitEvent({
+                                                t: "event",
+                                                kind: "toast",
+                                                level: "warning",
+                                                text: `Transient error mid-task — resuming (attempt ${midLoopStallRetryCount}/${maxStallRetries})…`,
+                                            });
+                                        }
+                                        catch {
+                                            /* best-effort telemetry */
+                                        }
+                                        try {
+                                            if (deps.session) {
+                                                logInteraction(deps.session.id, "stream_retry", {
+                                                    data: {
+                                                        attempt: midLoopStallRetryCount,
+                                                        maxAttempts: maxStallRetries + 1,
+                                                        errorName: _errName,
+                                                        errorMessage: `${_errMsg.slice(0, 160)} — resumed from completed steps`,
+                                                        appendedMessages: _appended,
+                                                        nextDelayMs: _contBackoff,
+                                                        source: "error-part-continuation",
+                                                    },
+                                                });
+                                            }
+                                        }
+                                        catch (logErr) {
+                                            console.error(`[message-processor] error-part continuation log failed: ${logErr?.message}`);
+                                        }
+                                        await new Promise((resolve) => setTimeout(resolve, _contBackoff));
+                                        if (!signal.aborted) {
+                                            continue streamAttempt;
+                                        }
+                                    }
+                                }
+                            }
+                            const authError = isAuthenticationError(part.error);
+                            const friendly = humanizeApiError(part.error, {
+                                modelId: runtime.modelId,
+                                providerId: runtime.modelInfo?.provider,
+                            });
+                            const forensics = summarizeApiErrorForLog(part.error);
+                            notifyObserver(observer?.onError, {
+                                message: friendly,
+                                timestamp: Date.now(),
+                            });
+                            // Interaction log: error + forensics envelope so opaque
+                            // provider 4xx ("parameter is invalid" / unknown 400s) leave
+                            // an actionable wire-level trace without needing a repro.
+                            try {
+                                if (deps.session) {
+                                    logInteraction(deps.session.id, "error", {
+                                        eventSubtype: authError ? "auth" : "api",
+                                        data: {
+                                            message: friendly.slice(0, 200),
+                                            ...(forensics ? { forensics } : {}),
+                                        },
+                                    });
+                                }
+                            }
+                            catch (logErr) {
+                                console.error(`[message-processor] interaction-log error failed: ${logErr?.message}`);
+                            }
+                            yield {
+                                type: "error",
+                                content: friendly,
+                                isAuthError: authError,
+                            };
+                            break;
+                        }
+                        case "abort":
+                            // A stall-watchdog abort arrives here as an "abort" stream part
+                            // (the SDK surfaces it as a part, not a throw). Distinguish it
+                            // from a genuine user cancel — which is caught at the top of the
+                            // loop via `signal.aborted` — and surface it as a visible error
+                            // instead of a benign "[Cancelled]" so a hung provider no longer
+                            // looks like a silent freeze.
+                            if (stallTriggered) {
+                                // Time-to-first-byte stall (no real chunk this attempt): the
+                                // socket wedged before any output — re-issue the SAME request
+                                // rather than giving up. Bounded by maxStallRetries; never
+                                // fires once tools ran or text flowed (planStallReprompt gate).
+                                const _stallBackoff = planStallReprompt();
+                                if (_stallBackoff != null) {
+                                    stall.dispose();
+                                    await new Promise((r) => setTimeout(r, _stallBackoff));
+                                    if (!signal.aborted) {
+                                        stallTriggered = false;
+                                        continue streamAttempt;
+                                    }
+                                }
+                                // Mid-loop dead-socket CONTINUATION (distinct from the TTFB
+                                // re-prompt above): the watchdog fired AFTER earlier steps ran
+                                // (chunksThisAttempt > 0) but the in-flight step produced zero
+                                // bytes (chunksThisStep === 0) — a single wedged inter-step
+                                // socket (xai/grok-build-0.1 mid-investigation, session
+                                // 247a0cea2eac), not a down backend. The TTFB re-prompt can't
+                                // recover this (its zero-chunks gate refuses, since restarting
+                                // the whole request would re-run the tools that already ran).
+                                // Instead, append the COMPLETED steps' generated messages
+                                // (assistant tool-calls + their tool-results) to history and
+                                // re-issue streamText: with the results already in context, no
+                                // tool re-runs and no text duplicates, so this is safe even for
+                                // write/commit tools. Falls through to the rescue path if the
+                                // completed steps can't be recovered.
+                                if (shouldContinueAfterMidLoopStall({
+                                    stallTriggered,
+                                    chunksThisAttempt,
+                                    chunksThisStep,
+                                    retryCount: midLoopStallRetryCount,
+                                    maxRetries: maxStallRetries,
+                                    aborted: signal.aborted,
+                                })) {
+                                    let _appended = 0;
+                                    try {
+                                        // result.response settles fast here (the stream was already
+                                        // aborted via stall.signal). Race a short timeout so a
+                                        // doubly-wedged provider can't re-hang the recovery itself.
+                                        const _resp = (await Promise.race([
+                                            result.response,
+                                            new Promise((_r, rej) => setTimeout(() => rej(new Error("response-timeout")), 3_000)),
+                                        ]));
+                                        const _gen = sanitizeModelMessages(scrubImagePayloadsInMessages(_resp.messages));
+                                        for (const _m of _gen) {
+                                            deps.messages.push(_m);
+                                            _appended++;
+                                        }
+                                    }
+                                    catch (respErr) {
+                                        console.error(`[message-processor] mid-loop stall continuation: completed steps unavailable: ${respErr?.message}`);
+                                    }
+                                    // A re-issue with zero preserved steps would restart from the
+                                    // original prompt and re-run every tool — only continue when
+                                    // the completed steps were actually grafted onto history.
+                                    if (_appended > 0) {
+                                        midLoopStallRetryCount++;
+                                        const _midBackoff = stallRepromptBackoffMs(midLoopStallRetryCount);
+                                        try {
+                                            const _ar = globalThis.__muonroiAgentRuntime;
+                                            _ar?.emitEvent({
+                                                t: "event",
+                                                kind: "stream-retry",
+                                                attempt: midLoopStallRetryCount,
+                                                maxAttempts: maxStallRetries + 1,
+                                                errorName: "TimeoutError",
+                                                errorMessage: "provider-stall (mid-loop) — resuming from completed steps",
+                                                nextDelayMs: _midBackoff,
+                                            });
+                                            _ar?.emitEvent({
+                                                t: "event",
+                                                kind: "toast",
+                                                level: "warning",
+                                                text: `Model stalled mid-task — resuming (attempt ${midLoopStallRetryCount}/${maxStallRetries})…`,
+                                            });
+                                        }
+                                        catch (emitErr) {
+                                            console.error(`[message-processor] mid-loop continuation telemetry failed: ${emitErr?.message}`);
+                                        }
+                                        try {
+                                            if (deps.session) {
+                                                logInteraction(deps.session.id, "stream_retry", {
+                                                    data: {
+                                                        attempt: midLoopStallRetryCount,
+                                                        maxAttempts: maxStallRetries + 1,
+                                                        errorName: "provider-stall-midloop",
+                                                        errorMessage: "no byte on inter-step request — resumed from completed steps",
+                                                        appendedMessages: _appended,
+                                                        nextDelayMs: _midBackoff,
+                                                    },
+                                                });
+                                            }
+                                        }
+                                        catch (logErr) {
+                                            console.error(`[message-processor] mid-loop continuation log failed: ${logErr?.message}`);
+                                        }
+                                        stall.dispose();
+                                        await new Promise((r) => setTimeout(r, _midBackoff));
+                                        if (!signal.aborted) {
+                                            stallTriggered = false;
+                                            continue streamAttempt;
+                                        }
+                                    }
+                                }
+                                stall.dispose();
+                                // A response tool already produced the terminal structured
+                                // answer (buffered from its call args) before the provider
+                                // stalled on a LATER step. Surface it and finish cleanly —
+                                // never bury the model's actual answer behind a "not
+                                // responding" error. Root cause of the "respond_* indicator
+                                // shows but no answer block renders" report: this stall-abort
+                                // path returned before the post-loop structured_response yield,
+                                // dropping the captured answer. A response tool is terminal,
+                                // so there is nothing to rescue — just emit what we have.
+                                if (_pendingStructuredResponse) {
+                                    if (!streamOk) {
+                                        try {
+                                            const _d = _pendingStructuredResponse.data;
+                                            const _ans = typeof _d.response === "string" ? _d.response : JSON.stringify(_pendingStructuredResponse.data);
+                                            deps.appendCompletedTurn(userModelMessage, [
+                                                { role: "assistant", content: _ans },
+                                            ]);
+                                            streamOk = true;
+                                        }
+                                        catch (persistErr) {
+                                            console.error(`[message-processor] stall+response-tool persist failed: ${persistErr?.message}`);
+                                        }
+                                    }
+                                    yield {
+                                        type: "structured_response",
+                                        structuredResponse: _pendingStructuredResponse,
+                                    };
+                                    yield { type: "done" };
+                                    return;
+                                }
+                                // Best-effort answer rescue: a turn that already ran tools but
+                                // stalled before the final synthesis would otherwise return
+                                // ONLY "Model not responding", discarding all that work (live
+                                // obs 2026-06-04, deepseek session 734e65cffdf6: 67 tool calls
+                                // → user got nothing). Make ONE guarded forced-finalize call
+                                // over the gathered tool outputs. forcedFinalize has its own
+                                // stall timeout, so a still-dead provider just falls through.
+                                let _rescued = null;
+                                if (turnToolResults.length > 0) {
+                                    try {
+                                        const _userText = typeof userModelMessage?.content === "string"
+                                            ? userModelMessage.content
+                                            : JSON.stringify(userModelMessage?.content ?? "");
+                                        _rescued = await attemptStallRescue({
+                                            baseMessages: _topMessagesForCall,
+                                            userText: _userText.slice(0, 4000),
+                                            toolResults: turnToolResults,
+                                            system: typeof systemForModel === "string" ? systemForModel : undefined,
+                                            finalize: (a) => forcedFinalize({ model: runtime.model, messages: a.messages, system: a.system }),
+                                        });
+                                    }
+                                    catch {
+                                        _rescued = null;
+                                    }
+                                    try {
+                                        if (deps.session) {
+                                            logInteraction(deps.session.id, "stall_rescue", {
+                                                data: {
+                                                    outcome: _rescued ? "rescued" : "no_text",
+                                                    toolResultCount: turnToolResults.length,
+                                                    chars: _rescued?.length ?? 0,
+                                                },
+                                            });
+                                        }
+                                    }
+                                    catch {
+                                        /* telemetry is best-effort */
+                                    }
+                                }
+                                if (_rescued) {
+                                    assistantText += (assistantText ? "\n\n" : "") + _rescued;
+                                    yield { type: "content", content: _rescued };
+                                }
+                                // Persist a record of the interrupted turn BEFORE returning so
+                                // the next turn is not amnesiac. Previously this returned with
+                                // nothing persisted → the next turn saw "no previous turn" and
+                                // redid the work, orphaning any edits the stalled turn applied
+                                // (live obs 2026-06-04, deepseek-v4-flash). When rescued, the
+                                // note now carries the synthesized answer too (assistantText).
+                                // Best-effort: never let persistence failure block surfacing.
+                                if (!streamOk) {
+                                    try {
+                                        const _stallNote = buildInterruptedTurnNote(assistantText, activeToolCalls.map((c) => c.function.name));
+                                        deps.appendCompletedTurn(userModelMessage, [
+                                            { role: "assistant", content: _stallNote },
+                                        ]);
+                                        streamOk = true;
+                                    }
+                                    catch {
+                                        /* best-effort — surface the stall regardless */
+                                    }
+                                }
+                                if (_rescued) {
+                                    // Recovered a best-effort answer from partial data — surface
+                                    // a soft notice instead of the scary "not responding" error.
+                                    yield {
+                                        type: "content",
+                                        content: "\n\n[Note: the model connection stalled; the answer above is a best-effort synthesis " +
+                                            "from the tool results gathered before the stall and may be incomplete.]",
+                                    };
+                                    yield { type: "done" };
+                                    return;
+                                }
+                                notifyObserver(observer?.onError, { message: STALL_ERROR_MESSAGE, timestamp: Date.now() });
+                                yield { type: "error", content: STALL_ERROR_MESSAGE, isAuthError: false };
+                                yield { type: "done" };
+                                return;
+                            }
+                            yield { type: "content", content: "\n\n[Cancelled]" };
+                            break;
+                    }
+                }
+                stall.dispose(); // stream drained normally — stop the stall watchdog
+                // ─── SAMR Phase 1 → Phase 2 transition ─────────────────────────
+                // Phase 1 (premium model) produced tool calls but the SDK stopped
+                // before executing them (stopWhen: stepCountIs(1)). Append the
+                // assistant message to deps.messages and restart the loop with
+                // the fast execution model. Phase 2's streamText call will see
+                // the pending tool calls and execute them automatically.
+                //
+                // EXCEPT when Phase 1 emitted a response tool: a `respond_*` call IS
+                // the terminal structured answer (identity execute), not work to hand
+                // to Phase 2. Transitioning here would (a) skip the structured_response
+                // yield below — the answer never reaches the TUI — and (b) append a
+                // dangling assistant tool-call WITHOUT its tool-result (only assistant
+                // msgs are pushed), corrupting Phase 2's history. Fall through instead
+                // so the buffered answer is yielded + persisted on this turn.
+                if (stepRouterPhase === "phase1" && phase1HadToolCalls && !responseToolCalled) {
+                    try {
+                        const phase1Response = await result.response;
+                        // Append only new messages (assistant message with tool calls)
+                        const newMsgs = phase1Response.messages.slice(deps.messages.length);
+                        for (const msg of newMsgs) {
+                            if (msg.role === "assistant") {
+                                deps.messages.push(msg);
+                            }
+                        }
+                    }
+                    catch {
+                        // If response extraction fails, fall through to normal completion
+                    }
+                    stepRouterPhase = "phase2";
+                    continue; // Re-enter while loop with Phase 2 (fast) model
+                }
+                // Surface the single most-complete response-tool answer buffered
+                // during the stream (see _pendingStructuredResponse). Yielding here —
+                // once, after the stream drained and after the Phase 1 transition —
+                // collapses any duplicate response-tool emissions in the turn into a
+                // single structured_response block for the UI.
+                if (_pendingStructuredResponse) {
+                    // Schema-mismatch normalization: when the model calls a typed
+                    // respond_<task> (e.g. respond_analyze) but sends a payload shaped
+                    // like respond_general ({ response: "..." } without schema-specific
+                    // fields like `findings`), the TUI typed renderer renders an empty
+                    // box. Normalize taskType to 'general' so the plain-markdown
+                    // renderer is used. See normalizeStructuredResponseTaskType() docs.
+                    // Session 48d22fe436f6: respond_analyze({ response: "..." }) →
+                    // analyze renderer showed empty findings list → answer swallowed.
+                    const _normalizedType = normalizeStructuredResponseTaskType(_pendingStructuredResponse.taskType, _pendingStructuredResponse.data);
+                    if (_normalizedType !== _pendingStructuredResponse.taskType) {
+                        _pendingStructuredResponse = { taskType: _normalizedType, data: _pendingStructuredResponse.data };
+                    }
+                    yield {
+                        type: "structured_response",
+                        structuredResponse: _pendingStructuredResponse,
+                    };
+                    // Ensure DB has the answer as assistant row even if later _finalMessages
+                    // or buildChatEntries path drops the tool-call-only assistant (sanitize,
+                    // response-tool terminal, vision bridge, sub-delegate). Prevents TUI
+                    // "flashed then vanished" on finalizeActiveTurn + getChatEntries().
+                    if (deps.session) {
+                        try {
+                            const _d = _pendingStructuredResponse.data;
+                            const _ans = typeof _d.response === "string" ? _d.response : JSON.stringify(_pendingStructuredResponse.data);
+                            deps.appendCompletedTurn(userModelMessage, [{ role: "assistant", content: _ans }]);
+                        }
+                        catch { }
+                    }
+                    if (_responseToolEmitCount > 1 && deps.session) {
+                        try {
+                            logInteraction(deps.session.id, "f6_synthesis", {
+                                eventSubtype: "response_tool_deduped",
+                                data: { emitted: _responseToolEmitCount, keptChars: _pendingStructuredResponseLen },
+                            });
+                        }
+                        catch {
+                            /* telemetry best-effort */
+                        }
+                    }
+                }
+                if (signal.aborted) {
+                    if (_pendingStructuredResponse) {
+                        try {
+                            const _d = _pendingStructuredResponse.data;
+                            const _ans = typeof _d.response === "string" ? _d.response : JSON.stringify(_pendingStructuredResponse.data);
+                            deps.appendCompletedTurn(userModelMessage, [{ role: "assistant", content: _ans }]);
+                        }
+                        catch { }
+                    }
+                    deps.discardAbortedTurn(userModelMessage);
+                    yield { type: "done" };
+                    return;
+                }
+                try {
+                    const response = await result.response;
+                    if (!signal.aborted) {
+                        // Scrub oversized base64 image payloads from tool-result parts
+                        // BEFORE persisting. The vision bridge above only modified the
+                        // transient `tr` shown to the user — `response.messages` from
+                        // the AI SDK still carries the full base64 (e.g. Playwright
+                        // screenshot, ~1.5MB). Persisting that lets it accumulate and
+                        // overflow the model's context on subsequent turns.
+                        const scrubbed = rewriteSafetyApprovedToolResults(scrubImagePayloadsInMessages(response.messages));
+                        // Phase 5 F6 — synthesis step when stream ended without a final
+                        // text response. Cheap models (DeepSeek V4 Flash) frequently
+                        // emit only tool-calls in their last step and stop, leaving the
+                        // user staring at "Here's the summary:..." truncation that
+                        // required a manual "tiếp tục" turn-2 to coax out. Detect that
+                        // shape and inject ONE forcedFinalize call (same path as 4B
+                        // ceiling-hit) so the answer arrives on turn 1.
+                        //
+                        // Skip when 4B ceiling already triggered its own forcedFinalize
+                        // below — running both would double-bill and duplicate text.
+                        let _f6SynthesisText = null;
+                        const _f6LastMsg = scrubbed[scrubbed.length - 1];
+                        const _f6LastRole = _f6LastMsg?.role ?? "none";
+                        let _f6Outcome = "skip_ceiling";
+                        let _f6Elapsed = 0;
+                        let _f6ChunkChars = 0;
+                        let _f6Error = null;
+                        // A response tool already produced the final structured answer —
+                        // F6 synthesis would duplicate it as prose. Skip entirely. With
+                        // the stopWhen terminal-halt above, the turn now ends right after
+                        // the response tool (last scrubbed message is the response
+                        // tool-result, role "tool"), which would otherwise trip the
+                        // _needsSynthesis "ended on a tool" branch and double-respond.
+                        if (responseToolCalled) {
+                            _f6Outcome = "skip_response_tool";
+                        }
+                        else if (!_ceilingHit) {
+                            const _needsSynthesis = (() => {
+                                if (!_f6LastMsg)
+                                    return false;
+                                if (_f6LastMsg.role === "tool")
+                                    return true;
+                                if (_f6LastMsg.role !== "assistant")
+                                    return false;
+                                const _c = _f6LastMsg.content;
+                                if (typeof _c === "string")
+                                    return !_c.trim();
+                                if (!Array.isArray(_c))
+                                    return false;
+                                return !_c.some((p) => p && p.type === "text" && typeof p.text === "string" && p.text.trim().length > 0);
+                            })();
+                            if (!_needsSynthesis) {
+                                _f6Outcome = "skip_has_text";
+                            }
+                            else {
+                                const _f6Start = Date.now();
+                                try {
+                                    const _ff = await forcedFinalize({
+                                        model: runtime.model,
+                                        messages: _topMessagesForCall,
+                                        system: typeof systemForModel === "string" ? systemForModel : undefined,
+                                    });
+                                    _f6Elapsed = Date.now() - _f6Start;
+                                    _f6ChunkChars = (_ff.text ?? "").length;
+                                    if (_ff.text.trim()) {
+                                        _f6SynthesisText = _ff.text;
+                                        assistantText += _ff.text;
+                                        yield { type: "content", content: _ff.text };
+                                        _f6Outcome = "fired_text";
+                                    }
+                                    else {
+                                        _f6Outcome = "fired_empty";
+                                    }
+                                }
+                                catch (_err) {
+                                    _f6Elapsed = Date.now() - _f6Start;
+                                    _f6Outcome = "error";
+                                    _f6Error = _err?.message?.slice(0, 200) ?? String(_err).slice(0, 200);
+                                }
+                            }
+                        }
+                        try {
+                            if (deps.session) {
+                                logInteraction(deps.session.id, "f6_synthesis", {
+                                    data: {
+                                        outcome: _f6Outcome,
+                                        lastMsgRole: _f6LastRole,
+                                        elapsedMs: _f6Elapsed,
+                                        chars: _f6ChunkChars,
+                                        error: _f6Error,
+                                        ceilingHit: _ceilingHit,
+                                        scrubbedLen: scrubbed.length,
+                                    },
+                                });
+                            }
+                        }
+                        catch {
+                            /* telemetry is best-effort */
+                        }
+                        // Summary-phase grounding check (Agent Operating Contract, runtime
+                        // half). Soft-flag counts / file:line refs in the final synthesis
+                        // that don't appear in this turn's tool outputs — possible
+                        // hallucination. Never blocks: emits a grounding-flag event + a
+                        // warn toast + an inline advisory footnote. Only runs when the
+                        // turn actually produced tool output (a corpus to ground against)
+                        // and is not chitchat. See grounding-check.ts.
+                        if (process.env.MUONROI_DISABLE_GROUNDING_CHECK !== "1" && !isChitchat && assistantText.trim().length > 0) {
+                            try {
+                                const _gParts = [];
+                                let _gHadTool = false;
+                                for (const _gm of scrubbed) {
+                                    if (!_gm || _gm.role === "assistant")
+                                        continue;
+                                    if (_gm.role === "tool")
+                                        _gHadTool = true;
+                                    const _gc = _gm.content;
+                                    _gParts.push(typeof _gc === "string" ? _gc : JSON.stringify(_gc));
+                                }
+                                if (_gHadTool) {
+                                    const _claims = findUnverifiedClaims(assistantText, _gParts.join("\n"));
+                                    if (_claims.length > 0) {
+                                        const _footnote = buildGroundingFootnote(_claims);
+                                        assistantText += _footnote;
+                                        yield { type: "content", content: _footnote };
+                                        const _gar = globalThis.__muonroiAgentRuntime;
+                                        const _claimTexts = _claims.map((c) => c.text);
+                                        _gar?.emitEvent({
+                                            t: "event",
+                                            kind: "grounding-flag",
+                                            claims: _claimTexts,
+                                            count: _claims.length,
+                                            ts: Date.now(),
+                                        });
+                                        _gar?.emitEvent({
+                                            t: "event",
+                                            kind: "toast",
+                                            level: "warn",
+                                            text: `grounding: ${_claims.length} unverified claim(s) — ${_claimTexts.join(", ")}`,
+                                        });
+                                        if (deps.session) {
+                                            try {
+                                                logInteraction(deps.session.id, "grounding_flag", {
+                                                    data: { claims: _claimTexts, count: _claims.length },
+                                                });
+                                            }
+                                            catch {
+                                                /* telemetry is best-effort */
+                                            }
+                                        }
+                                    }
+                                }
+                            }
+                            catch {
+                                /* grounding check is best-effort — never break finalize */
+                            }
+                        }
+                        const _finalMessages = sanitizeModelMessages(scrubbed);
+                        if (_f6SynthesisText !== null) {
+                            _finalMessages.push({
+                                role: "assistant",
+                                content: _f6SynthesisText,
+                            });
+                        }
+                        deps.appendCompletedTurn(userModelMessage, _finalMessages);
+                        streamOk = true;
+                    }
+                }
+                catch (responseError) {
+                    if (!attemptedOverflowRecovery && !assistantText.trim() && modelInfo && isContextLimitError(responseError)) {
+                        attemptedOverflowRecovery = true;
+                        continue;
+                    }
+                }
+                if (signal.aborted) {
+                    deps.discardAbortedTurn(userModelMessage);
+                    yield { type: "done" };
+                    return;
+                }
+                if (patternLoopForceHalt) {
+                    patternLoopForceHalt = false;
+                    try {
+                        const _ar = globalThis.__muonroiAgentRuntime;
+                        _ar?.emitEvent({
+                            t: "event",
+                            kind: "toast",
+                            level: "info",
+                            text: "vòng lặp tool được phát hiện — đang tự động tiêm prompt nhắc nhở để agent tự sửa đổi",
+                        });
+                    }
+                    catch {
+                        /* best-effort */
+                    }
+                    stall.dispose();
+                    await closeMcp?.().catch(() => { });
+                    continue;
+                }
+                // Phase 5 Fix 5 — the Phase 4 4B forced-finalize-on-ceiling-hit
+                // block lived here. With the matrix ceiling no longer halting the
+                // stream (it's pure telemetry now), _ceilingHit can never be true
+                // and this branch is dead. F6 synthesis above already covers the
+                // "stream ended with no final text" case for both natural model
+                // termination AND maxToolRounds halt. Keeping the comment as a
+                // breadcrumb for future archaeology.
+                if (!streamOk && assistantText.trim()) {
+                    deps.appendCompletedTurn(userModelMessage, [{ role: "assistant", content: stripDsmlMarkup(assistantText) }]);
+                }
+                // Fallback: model responded in text despite tool_choice=required
+                // Attempt JSON extraction from assistant text → yield as structured_response
+                if (_hasResponseTools && !responseToolCalled && pilCtx.taskType && assistantText.trim()) {
+                    try {
+                        const jsonMatch = assistantText.match(/\{[\s\S]*\}/);
+                        if (jsonMatch) {
+                            const parsed = JSON.parse(jsonMatch[0]);
+                            if (Object.keys(parsed).length > 0) {
+                                responseToolCalled = true;
+                                yield {
+                                    type: "structured_response",
+                                    structuredResponse: {
+                                        taskType: pilCtx.taskType,
+                                        data: parsed,
+                                    },
+                                };
+                            }
+                        }
+                    }
+                    catch {
+                        // JSON parse failed — leave as text-fallback
+                    }
+                }
+                // Track PIL output mode for /optimize metrics
+                {
+                    const { setLastOutputMode } = await import("../pil/store.js");
+                    if (!_hasResponseTools)
+                        setLastOutputMode("conversational");
+                    else if (responseToolCalled)
+                        setLastOutputMode("structured");
+                    else
+                        setLastOutputMode("text-fallback");
+                }
+                // ROUTE-11: Fire routeFeedback after turn completes (success path).
+                // Must come AFTER posttool calls (posttool fires during tool-result processing above).
+                // Fire-and-forget — no await. Skipped when taskHash is null (bridge absent).
+                {
+                    const turnDuration = Date.now() - turnStartMs;
+                    if (taskHash) {
+                        const tier = taskTypeToTier(pilCtx.taskType);
+                        void routeFeedback(taskHash, tier, runtime.modelId, "success", // Phase 6: all normal completions = 'success'
+                        0, // retryCount: 0 for first attempt
+                        turnDuration);
+                    }
+                    // HTTP path: also report via router store taskHash (covers warm/cold EE routes)
+                    const storeHash = routerStore.getState().taskHash;
+                    if (storeHash) {
+                        reportRouteOutcome(storeHash, "success", turnDuration);
+                    }
+                }
+                // Detect a tool call emitted as plain TEXT (wrong dialect) in the final
+                // assistant answer — the action never ran, so the turn would otherwise
+                // end silently with broken/half-done work (live: deepseek session
+                // 905d564dbde4 emitted `<read_file>` as text after a destructive edit).
+                // Detect regardless of how many real tool calls already succeeded: the
+                // common failure is the model doing a few real tools, then emitting the
+                // NEXT call as text and stopping (live deepseek-native, full-fix CLI: 2
+                // real read_file calls, then `<read_file><path>` as text → silent stop).
+                // An earlier `activeToolCalls.length === 0` guard suppressed exactly
+                // that case. Detector precision (structural invocation shape, not a bare
+                // mention) guards against false-firing on a normal final answer.
+                const _textToolCall = detectTextEmittedToolCall(assistantText);
+                // Interaction log: agent response complete
+                try {
+                    if (deps.session) {
+                        const sb = statusBarStore.getState();
+                        const turnDurationMs = Date.now() - turnStartMs;
+                        // BUG-A telemetry — detect raw DeepSeek native tool-call markup
+                        // leaking into assistant text. Signature is `<｜｜DSML｜｜` (the
+                        // fullwidth vertical bars are NOT pipes, they're U+FF5C).
+                        const _dsmlSig = "｜｜DSML｜｜";
+                        const _dsmlMatches = assistantText.includes(_dsmlSig);
+                        const _codeBlockBash = /```\s*bash\b/i.test(assistantText);
+                        logInteraction(deps.session.id, "agent_response", {
+                            model: turnModelId,
+                            inputTokens: sb.in_tokens,
+                            outputTokens: sb.out_tokens,
+                            durationMs: turnDurationMs,
+                            data: {
+                                textLength: assistantText.length,
+                                toolCallCount: activeToolCalls.length,
+                                compacted: deps.getCompactedThisTurn(),
+                                dsmlLeak: _dsmlMatches,
+                                bashCodeBlock: _codeBlockBash,
+                                textToolXmlLeak: _textToolCall.detected,
+                                textToolXmlTool: _textToolCall.tool,
+                            },
+                        });
+                    }
+                }
+                catch {
+                    /* fail-open */
+                }
+                // F3b — surface hard-cap stop (absolute ceiling, cannot be bumped).
+                if (_hardCapHit) {
+                    yield {
+                        type: "content",
+                        content: `\n\n[Hard limit reached: Agent hit the absolute step ceiling of ${deps.hardMaxToolRounds} steps. ` +
+                            `This turn cannot continue — too many LLM round-trips. ` +
+                            `Start a new turn or break your task into smaller steps.]\n`,
+                    };
+                }
+                // Surface the round-cap stop so the user knows why the agent halted
+                // (session 7dcf8fd7d6a4 hit stepCountIs(100) silently, looked like a
+                // crash). AI SDK reports finishReason='tool-calls' when the step cap
+                // fires with tool calls still pending — distinct from 'stop' (model
+                // chose to end). We only warn when stepNumber ≥ cap so a model that
+                // legitimately terminates mid-tool-call (rare) doesn't get a false
+                // warning.
+                if (_lastFinishReason === "tool-calls" && stepNumber >= deps.maxToolRounds - 1) {
+                    yield {
+                        type: "content",
+                        content: `\n\n[Stopped: Agent paused execution to prevent runaway loops (hit step cap of ${deps.maxToolRounds}). ` +
+                            `If you want to continue the current task, simply reply with "continue" or "tiếp tục". ` +
+                            `For long tasks, we strongly recommend running the \`/compact\` command first to free up context memory before continuing.]\n`,
+                    };
+                }
+                // Tool-call-as-text leak: the model wrote a tool invocation as plain
+                // text (wrong dialect) and made NO real tool call, so the action never
+                // ran. Auto-recover ONCE: append a corrective message and re-run the
+                // turn so the model can invoke the tool properly. The just-finished
+                // (text-only) turn is already persisted above — the model sees its own
+                // mistake plus the correction. Mirrors the proven phase-switch re-entry
+                // (it also pushes to deps.messages then `continue`s); bounded by
+                // MAX_TEXT_TOOL_RESTEER so a persistently-degrading model can't loop.
+                if (_textToolCall.detected && streamOk && textToolReSteerCount < MAX_TEXT_TOOL_RESTEER) {
+                    textToolReSteerCount++;
+                    // Recover the model's INTENT from the leaked markup (DeepSeek-native
+                    // DSML carries the tool + args) so the corrective restates the exact
+                    // call — far more effective than a generic "use the tool" nudge.
+                    const _parsedCalls = parseDsmlToolCalls(assistantText);
+                    const _intent = _parsedCalls.length > 0
+                        ? ` You appear to have intended: ${_parsedCalls
+                            .map((c) => `${c.name}(${Object.entries(c.args)
+                            .map(([k, v]) => `${k}=${JSON.stringify(v)}`)
+                            .join(", ")})`)
+                            .join("; ")}. Make those exact call(s) via the tool interface now.`
+                        : "";
+                    deps.messages.push({
+                        role: "user",
+                        content: `Your previous reply wrote a \`${_textToolCall.tool}\` tool call as XML/text. That is NOT how tools are invoked here — ` +
+                            "writing tool calls as text does nothing, so the action did not run. " +
+                            "Use the actual tool-calling interface (function/tool calls) to perform the action now. " +
+                            "Do NOT output XML tags like <read_file>, <write_to_file>, <execute_command>, or <tool_call> (or DSML markup) as text." +
+                            _intent,
+                    });
+                    if (deps.session) {
+                        try {
+                            logInteraction(deps.session.id, "text_tool_resteer", {
+                                model: turnModelId,
+                                data: { tool: _textToolCall.tool, attempt: textToolReSteerCount },
+                            });
+                        }
+                        catch {
+                            /* telemetry best-effort */
+                        }
+                    }
+                    {
+                        const _gar = globalThis.__muonroiAgentRuntime;
+                        _gar?.emitEvent({
+                            t: "event",
+                            kind: "toast",
+                            level: "info",
+                            text: `model wrote a ${_textToolCall.tool} tool call as text — re-steering to use the tool interface`,
+                        });
+                    }
+                    await closeMcp?.().catch(() => { });
+                    continue;
+                }
+                // Re-steer exhausted: inject DSML intent as a user message so the
+                // model re-issues the same calls via the real tool interface.
+                if (_textToolCall.detected) {
+                    const _parsedDsml = assistantText ? parseDsmlToolCalls(assistantText) : [];
+                    if (_parsedDsml.length > 0 && textToolReSteerCount < MAX_TEXT_TOOL_RESTEER) {
+                        textToolReSteerCount++;
+                        const _intentStr = _parsedDsml
+                            .map((c) => `${c.name}(${Object.entries(c.args)
+                            .map(([k, v]) => `${k}=${JSON.stringify(v)}`)
+                            .join(", ")})`)
+                            .join("; ");
+                        deps.messages.push({
+                            role: "user",
+                            content: `[DSML intent: ${_intentStr} — re-issue via the tool interface now.]`,
+                        });
+                        await closeMcp?.().catch(() => { });
+                        continue;
+                    }
+                    // No parseable DSML — surface the dead-end warning.
+                    yield {
+                        type: "content",
+                        content: `\n\n[⚠ The model wrote a \`${_textToolCall.tool}\` tool call as TEXT instead of invoking the tool, ` +
+                            "so that action did NOT run and this turn made no real progress. " +
+                            "Re-run the request (optionally with a more capable model) — the tool interface was not used.]\n",
+                    };
+                    const _gar2 = globalThis.__muonroiAgentRuntime;
+                    _gar2?.emitEvent({
+                        t: "event",
+                        kind: "toast",
+                        level: "warn",
+                        text: `model emitted a ${_textToolCall.tool} tool call as text — action not executed`,
+                    });
+                }
+                const stopInput = {
+                    hook_event_name: "Stop",
+                    session_id: deps.session?.id,
+                    cwd: deps.bash.getCwd(),
+                };
+                await deps.fireHook(stopInput, signal).catch(() => { });
+                // Debug trace: emit pipeline summary
+                if (_debugOn) {
+                    const sb = statusBarStore.getState();
+                    const defaultInfo = getModelInfo(deps.modelId);
+                    const usedInfo = getModelInfo(turnModelId);
+                    const routerSaved = defaultInfo && usedInfo && defaultInfo.outputPrice > usedInfo.outputPrice
+                        ? (sb.out_tokens * (defaultInfo.outputPrice - usedInfo.outputPrice)) / 1_000_000
+                        : 0;
+                    const cacheSaved = sb.cache_read_tokens > 0 && defaultInfo
+                        ? (sb.cache_read_tokens *
+                            (defaultInfo.inputPrice - (defaultInfo.cachedInputPrice ?? defaultInfo.inputPrice * 0.1))) /
+                            1_000_000
+                        : 0;
+                    const trace = {
+                        turn_id: _debugTurnId,
+                        timestamp: turnStartMs,
+                        raw_prompt: userMessage,
+                        steps: _debugSteps,
+                        model_requested: deps.modelId,
+                        model_used: turnModelId,
+                        routed: turnModelId !== deps.modelId,
+                        input_tokens: sb.in_tokens,
+                        output_tokens: sb.out_tokens,
+                        cache_read_tokens: sb.cache_read_tokens,
+                        cost_usd: sb.session_usd,
+                        estimated_savings: {
+                            pil_tokens_saved: _pilEnrichmentDeltaSnapshot > 0 ? _pilEnrichmentDeltaSnapshot : 0,
+                            cache_tokens_saved: sb.cache_read_tokens,
+                            router_cost_saved_usd: routerSaved,
+                            total_tokens_saved: (_pilEnrichmentDeltaSnapshot > 0 ? _pilEnrichmentDeltaSnapshot : 0) + sb.cache_read_tokens,
+                            total_cost_saved_usd: routerSaved + cacheSaved,
+                        },
+                    };
+                    recordTurnTrace(trace);
+                    const traceLines = [];
+                    traceLines.push("\n┌─ Pipeline Trace ─────────────────────────");
+                    for (const step of _debugSteps) {
+                        const dur = step.duration_ms < 1 ? "<1ms" : `${step.duration_ms}ms`;
+                        const saved = step.tokens_saved ? ` (saved ~${step.tokens_saved} tok)` : "";
+                        traceLines.push(`│ ▸ ${step.name} [${dur}]${saved}`);
+                        traceLines.push(`│   ${step.output_summary}`);
+                    }
+                    const routeLabel = trace.routed ? `${trace.model_requested}→${trace.model_used}` : trace.model_used;
+                    traceLines.push(`│ Model: ${routeLabel} | ↑${sb.in_tokens} ↓${sb.out_tokens} | $${sb.session_usd.toFixed(4)}`);
+                    if (trace.estimated_savings.total_cost_saved_usd > 0) {
+                        traceLines.push(`│ Savings: ~${trace.estimated_savings.total_tokens_saved} tok, ~$${trace.estimated_savings.total_cost_saved_usd.toFixed(4)}`);
+                    }
+                    traceLines.push("└──────────────────────────────────────────\n");
+                    yield { type: "content", content: traceLines.join("\n") };
+                }
+                // Yield done FIRST so the UI releases isProcessing immediately.
+                // postTurnCompact is a background optimization that must not block
+                // the composer — without this, DeepSeek's 5-30s compaction call
+                // keeps the input in loading state and queued messages hang.
+                //
+                // CRITICAL: postTurnCompact MUST be fire-and-forget (void, NOT await).
+                // Even though yield-done happens first, the for-await loop in
+                // use-app-logic.tsx does NOT exit until the generator returns.
+                // An awaited postTurnCompact blocks the generator's return, which
+                // blocks the for-await loop, which blocks finalizeActiveTurn.
+                if (modelInfo?.contextWindow) {
+                    void deps.postTurnCompact(provider, system, modelInfo.contextWindow, signal).catch(() => { });
+                }
+                yield { type: "done" };
+                return;
+            }
+            catch (err) {
+                if (signal.aborted) {
+                    deps.discardAbortedTurn(userModelMessage);
+                    // ROUTE-11: Fire routeFeedback for cancelled turns (abort path).
+                    // Fire-and-forget — no await. Skipped when taskHash is null.
+                    {
+                        const turnDuration = Date.now() - turnStartMs;
+                        if (taskHash) {
+                            const tier = taskTypeToTier(pilCtx.taskType);
+                            void routeFeedback(taskHash, tier, runtime.modelId, "cancelled", 0, turnDuration);
+                        }
+                        const storeHash = routerStore.getState().taskHash;
+                        if (storeHash) {
+                            reportRouteOutcome(storeHash, "cancelled", turnDuration);
+                        }
+                    }
+                    yield { type: "content", content: "\n\n[Cancelled]" };
+                    yield { type: "done" };
+                    return;
+                }
+                if (!attemptedOverflowRecovery && !assistantText.trim() && modelInfo && isContextLimitError(err)) {
+                    attemptedOverflowRecovery = true;
+                    continue;
+                }
+                // Stall surfaced as a throw (rather than an "abort" stream part):
+                // apply the SAME time-to-first-byte re-prompt as the abort-part path.
+                // The watchdog already fired (stallTriggered) so its timer is spent —
+                // no dispose needed; the next attempt arms a fresh watchdog.
+                if (stallTriggered) {
+                    const _stallBackoff = planStallReprompt();
+                    if (_stallBackoff != null) {
+                        await new Promise((r) => setTimeout(r, _stallBackoff));
+                        if (!signal.aborted) {
+                            stallTriggered = false;
+                            continue;
+                        }
+                    }
+                }
+                // Transient network/server error retry — up to MAX_STREAM_RETRIES extra attempts.
+                // Only retry when no content has flowed yet (assistantText empty) to avoid
+                // partial-output corruption. Honour the abort signal between retries.
+                // Skip retry on a stall abort: the provider is unresponsive, so a retry
+                // just burns another full stall timeout of silence — surface it instead.
+                if (!assistantText.trim() && streamRetryCount < MAX_STREAM_RETRIES && !signal.aborted && !stallTriggered) {
+                    const { transient } = classifyStreamError(err);
+                    if (transient) {
+                        streamRetryCount++;
+                        // Exponential backoff: 500 → 2000 ms with ±25% jitter
+                        const baseMs = 500;
+                        const expMs = Math.min(baseMs * 4 ** (streamRetryCount - 1), 8_000);
+                        const spread = expMs * 0.25;
+                        const nextDelayMs = Math.round(expMs + (Math.random() * 2 - 1) * spread);
+                        const errorName = err instanceof Error ? err.name : "Error";
+                        const errorMessage = err instanceof Error ? err.message : String(err);
+                        // Emit harness telemetry event
+                        try {
+                            const _ar = globalThis.__muonroiAgentRuntime;
+                            _ar?.emitEvent({
+                                t: "event",
+                                kind: "stream-retry",
+                                attempt: streamRetryCount,
+                                maxAttempts: MAX_STREAM_RETRIES + 1,
+                                errorName,
+                                errorMessage,
+                                nextDelayMs,
+                            });
+                        }
+                        catch {
+                            /* best-effort */
+                        }
+                        try {
+                            if (deps.session) {
+                                logInteraction(deps.session.id, "stream_retry", {
+                                    data: {
+                                        attempt: streamRetryCount,
+                                        maxAttempts: MAX_STREAM_RETRIES + 1,
+                                        errorName,
+                                        errorMessage: errorMessage.slice(0, 200),
+                                        nextDelayMs,
+                                    },
+                                });
+                            }
+                        }
+                        catch {
+                            /* fail-open */
+                        }
+                        await new Promise((resolve) => setTimeout(resolve, nextDelayMs));
+                        if (!signal.aborted) {
+                            continue;
+                        }
+                    }
+                }
+                const authError = isAuthenticationError(err);
+                // Stall aborts carry an opaque DOMException; show the clear stall
+                // message instead of the raw abort reason.
+                const friendly = stallTriggered
+                    ? STALL_ERROR_MESSAGE
+                    : humanizeApiError(err, { modelId: runtime.modelId, providerId: runtime.modelInfo?.provider });
+                notifyObserver(observer?.onError, {
+                    message: friendly,
+                    timestamp: Date.now(),
+                });
+                yield {
+                    type: "error",
+                    content: friendly,
+                    isAuthError: authError,
+                };
+                if (_pendingStructuredResponse) {
+                    const _d = _pendingStructuredResponse.data;
+                    const _ans = typeof _d.response === "string" ? _d.response : JSON.stringify(_pendingStructuredResponse.data);
+                    deps.appendCompletedTurn(userModelMessage, [{ role: "assistant", content: _ans }]);
+                }
+                else if (assistantText.trim()) {
+                    deps.appendCompletedTurn(userModelMessage, [{ role: "assistant", content: stripDsmlMarkup(assistantText) }]);
+                }
+                else if (deps.session && userWriteAheadSeq != null) {
+                    markMessageErrored(deps.session.id, userWriteAheadSeq);
+                }
+                // ROUTE-11: Fire routeFeedback for failed turns (error path).
+                // Must come AFTER posttool calls. Fire-and-forget — no await.
+                {
+                    const turnDuration = Date.now() - turnStartMs;
+                    if (taskHash) {
+                        const tier = taskTypeToTier(pilCtx.taskType);
+                        void routeFeedback(taskHash, tier, runtime.modelId, "fail", 0, turnDuration);
+                    }
+                    const storeHash = routerStore.getState().taskHash;
+                    if (storeHash) {
+                        reportRouteOutcome(storeHash, "fail", turnDuration);
+                    }
+                }
+                const stopFailureInput = {
+                    hook_event_name: "StopFailure",
+                    error: friendly,
+                    session_id: deps.session?.id,
+                    cwd: deps.bash.getCwd(),
+                };
+                await deps.fireHook(stopFailureInput, signal).catch(() => { });
+                // Yield done FIRST — same rationale as the success path above.
+                // postTurnCompact fire-and-forget so generator returns immediately.
+                if (modelInfo?.contextWindow) {
+                    void deps.postTurnCompact(provider, system, modelInfo.contextWindow, signal).catch(() => { });
+                }
+                yield { type: "done" };
+                return;
+            }
+            finally {
+                await closeMcp?.().catch(() => { });
+            }
+        }
+    }
+    catch (err) {
+        throw err;
+    }
+    finally {
+        if (deps.isSubSession) {
+            deps.emitSubagentStatus(null);
+        }
+    }
+}
+//# sourceMappingURL=tool-engine.js.map