muonroi-cli 1.6.3 → 1.6.5

package/dist/packages/agent-harness-core/src/event-filter.js

@@ -28,6 +28,7 @@ export const LIFECYCLE_PRESET = new Set([
     "sprint-stage",
     "sprint-halt",
     "route-decision",
+    "steer-inject",
     "usage",
 ]);
 /**

package/dist/packages/agent-harness-core/src/event-redact.js

@@ -85,7 +85,12 @@ const ALLOWED_FIELDS = {
         forceCouncil: "pass",
         runId: "pass",
     },
-    "toast": {
+    "steer-inject": {
+        count: "pass",
+        atStep: "pass",
+        runId: "pass",
+    },
+    toast: {
         level: "pass",
         text: 500, // cap to 500 chars, then scrub
         ttlMs: "pass",
@@ -94,7 +99,7 @@ const ALLOWED_FIELDS = {
         target: "pass",
         text: 500, // cap to 500 chars
     },
-    "usage": {
+    usage: {
         source: "pass",
         model: "pass",
         inputTokens: "pass",

package/dist/packages/agent-harness-core/src/protocol.d.ts

@@ -177,6 +177,14 @@ export type LiveEvent = {
     /** Total number of unverified claims in this turn. */
     count: number;
     ts: number;
+} | {
+    t: "event";
+    kind: "steer-inject";
+    /** How many queued messages were injected at this boundary. */
+    count: number;
+    /** The prepareStep step number at which injection occurred (>= 1). */
+    atStep: number;
+    runId: string;
 } | {
     t: "idle";
 };

package/dist/src/generated/version.d.ts

@@ -1,2 +1,2 @@
-export declare const PACKAGE_VERSION = "1.6.3";
+export declare const PACKAGE_VERSION = "1.6.5";
 export declare const PACKAGE_DESCRIPTION = "BYOK AI coding agent with multi-model council debate, role-based routing, and auto-compact.";

package/dist/src/generated/version.js

@@ -1,5 +1,5 @@
 // AUTO-GENERATED by scripts/sync-version.cjs. DO NOT EDIT BY HAND.
 // Sourced from package.json at build time so it survives bun --compile bundling.
-export const PACKAGE_VERSION = "1.6.3";
+export const PACKAGE_VERSION = "1.6.5";
 export const PACKAGE_DESCRIPTION = "BYOK AI coding agent with multi-model council debate, role-based routing, and auto-compact.";
 //# sourceMappingURL=version.js.map

package/dist/src/gsd/__tests__/directives.test.js

@@ -98,5 +98,42 @@ describe("buildDirective", () => {
             expect(out.text).toMatch(/\[recommended\]/);
         }
     });
+    // Language nudge — re-anchors the "reply in user's language" rule INSIDE the
+    // directive so layered brevity / FIX-FIRST directives can't drown it (live
+    // miss: storyflow_ui session 22661c8de9f2).
+    describe("language nudge", () => {
+        it("appends the nudge when replyLanguage is set", () => {
+            const out = buildDirective({
+                complexity: scoreComplexity("fix CI fail"),
+                phase: "debug",
+                grayAreas: [],
+                replyLanguage: "Vietnamese",
+            });
+            expect(out.text).toMatch(/LANGUAGE — the user wrote in Vietnamese/);
+            expect(out.text).toMatch(/Reply in Vietnamese/);
+            expect(out.text).toMatch(/OVERRIDES any brevity/);
+        });
+        it("omits the nudge when replyLanguage is undefined", () => {
+            const out = buildDirective({
+                complexity: scoreComplexity("fix CI fail"),
+                phase: "debug",
+                grayAreas: [],
+            });
+            expect(out.text).not.toMatch(/LANGUAGE —/);
+        });
+        it("stacks with the ecosystem nudge when both apply", () => {
+            const out = buildDirective({
+                complexity: scoreComplexity("how does the muonroi ecosystem work"),
+                phase: null,
+                grayAreas: [],
+                ecosystem: true,
+                replyLanguage: "Vietnamese",
+            });
+            expect(out.text).toMatch(/ECOSYSTEM SCOPE/);
+            expect(out.text).toMatch(/LANGUAGE —/);
+            // ecosystem nudge precedes language nudge (deterministic order)
+            expect(out.text.indexOf("ECOSYSTEM SCOPE")).toBeLessThan(out.text.indexOf("LANGUAGE —"));
+        });
+    });
 });
 //# sourceMappingURL=directives.test.js.map

package/dist/src/gsd/directives.d.ts

@@ -46,6 +46,15 @@ export interface DirectiveInput {
      * the shipped authoritative source.
      */
     ecosystem?: boolean;
+    /**
+     * User's reply language (heuristic — Vietnamese|undefined). When set, the
+     * directive appends an explicit language nudge so the rule survives the
+     * personality/GSD instructions stacking on top of it (storyflow_ui session
+     * 22661c8de9f2: user wrote Vietnamese, layered directives + a stalled
+     * forced-finalize drowned out the base "reply in user's language" rule and
+     * the agent answered in English).
+     */
+    replyLanguage?: string;
 }
 export interface DirectiveOutput {
     text: string;
@@ -60,4 +69,13 @@ export declare function mentionsEcosystemScope(message: string): boolean;
  * configured — the model simply finds no such tool and falls back to local files.
  */
 export declare const ECOSYSTEM_DOCS_NUDGE: string;
+/**
+ * Appended to any directive when the user's reply language is non-English.
+ * The base system prompt's "reply in user's language" rule normally suffices,
+ * but `concise` / `FIX-FIRST` / GSD-debug directive bodies stack on top of it
+ * with strong "be terse / code over prose" language that crowds the rule out
+ * — observed live (storyflow_ui 22661c8de9f2). This NUDGE re-anchors the rule
+ * inside the directive itself so brevity preferences cannot override it.
+ */
+export declare function buildLanguageNudge(lang: string): string;
 export declare function buildDirective(input: DirectiveInput): DirectiveOutput;

package/dist/src/gsd/directives.js

@@ -36,6 +36,20 @@ export const ECOSYSTEM_DOCS_NUDGE = [
     `${HEADER} ECOSYSTEM SCOPE — this turn concerns the Muonroi ecosystem (platform overview, BB/.NET packages, building-block, open-core boundary, setup).`,
     "If the muonroi-docs MCP is available, it is the AUTHORITATIVE source — call it FIRST (docs_search / setup_guide / bb_recipe_list / bb_package_describe), THEN ground with local files. Do NOT characterize the ecosystem from local repo files alone.",
 ].join("\n");
+/**
+ * Appended to any directive when the user's reply language is non-English.
+ * The base system prompt's "reply in user's language" rule normally suffices,
+ * but `concise` / `FIX-FIRST` / GSD-debug directive bodies stack on top of it
+ * with strong "be terse / code over prose" language that crowds the rule out
+ * — observed live (storyflow_ui 22661c8de9f2). This NUDGE re-anchors the rule
+ * inside the directive itself so brevity preferences cannot override it.
+ */
+export function buildLanguageNudge(lang) {
+    return [
+        `${HEADER} LANGUAGE — the user wrote in ${lang}. Reply in ${lang}.`,
+        "This rule OVERRIDES any brevity / concise / code-over-prose directive: terseness is fine, but the response language stays the user's.",
+    ].join("\n");
+}
 function renderGrayAreas(qs) {
     if (qs.length === 0)
         return "  (no gray areas detected — confirm the request is fully specified before proceeding)";
@@ -125,9 +139,16 @@ export function buildDirective(input) {
     // Ecosystem-scoped turns get a docs-first nudge regardless of tier (question
     // OR task): muonroi-docs is the authoritative source and must not be skipped
     // in favour of guessing from local files (session 41ccfeb2ceee turn 1).
+    let text = base.text;
     if (input.ecosystem) {
-        return { ...base, text: `${base.text}\n${ECOSYSTEM_DOCS_NUDGE}` };
+        text = `${text}\n${ECOSYSTEM_DOCS_NUDGE}`;
+    }
+    // Language nudge: re-anchor the "reply in user's language" rule INSIDE the
+    // directive when the user wrote in a non-English language, so layered
+    // brevity/concise directives can't drown it (storyflow_ui 22661c8de9f2).
+    if (input.replyLanguage) {
+        text = `${text}\n${buildLanguageNudge(input.replyLanguage)}`;
     }
-    return base;
+    return { ...base, text };
 }
 //# sourceMappingURL=directives.js.map

package/dist/src/orchestrator/message-processor.d.ts

@@ -100,6 +100,14 @@ export interface MessageProcessorDeps extends TurnRunnerDepsBase {
      * — preserves backward compat for batch / headless paths that have no UI to
      * surface the askcard.
      */
+    /**
+     * Live-queue steering drain (UI-provided). Returns and CLEARS any messages
+     * the user typed while this turn is streaming, so prepareStep can inject them
+     * mid-turn. Undefined / returns [] → no steering (legacy deferred queue).
+     */
+    drainSteerMessages?: () => {
+        text: string;
+    }[];
     askToolLoopContinue?: ToolLoopCapAsk;
     runCouncilV2(userMessage: string, opts: {
         skipClarification: boolean;

package/dist/src/orchestrator/message-processor.js

@@ -89,7 +89,7 @@ import { statusBarStore } from "../ui/status-bar/store.js";
 import { appendDecisionLog } from "../usage/decision-log.js";
 import { openUrl } from "../utils/open-url.js";
 import { appendAudit, toolNeedsApproval } from "../utils/permission-mode.js";
-import { getAutoCouncilConfidence, getAutoCouncilMinRoles, getProviderStallTimeoutMs, getRoleModels, getTopLevelCompactKeepLast, getTopLevelCompactThresholdChars, getTopLevelToolBudgetChars, isAutoCouncilEnabled, isProviderDisabled, loadMcpServers, loadValidSubAgents, } from "../utils/settings.js";
+import { getAutoCouncilConfidence, getAutoCouncilMinRoles, getProviderStallRetries, getProviderStallTimeoutMs, getRoleModels, getSteerInjectionEnabled, getTopLevelCompactKeepLast, getTopLevelCompactThresholdChars, getTopLevelToolBudgetChars, isAutoCouncilEnabled, isProviderDisabled, loadMcpServers, loadValidSubAgents, } from "../utils/settings.js";
 import { resolveShell } from "../utils/shell.js";
 import { relaxCompactionSettings } from "./compaction.js";
 import { wrapToolSetWithDedup } from "./cross-turn-dedup.js";
@@ -108,7 +108,8 @@ import { forcedFinalize, getSessionLastTask, incSessionStep, parseBudgetOverride
 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 } from "./stall-watchdog.js";
+import { createStallWatchdog, STALL_ERROR_MESSAGE, shouldRepromptStall, stallRepromptBackoffMs, } from "./stall-watchdog.js";
+import { planSteerInjection } from "./steer-inbox.js";
 import { wrapToolSetWithCap } from "./sub-agent-cap.js";
 import { compactSubAgentMessages, cumulativeMessageChars } from "./subagent-compactor.js";
 import { detectTextEmittedToolCall, parseDsmlToolCalls } from "./text-tool-call-detector.js";
@@ -813,6 +814,24 @@ export class MessageProcessor {
         // clear toast and SKIP the transient-retry (a stalled provider just stalls
         // again, wasting another full timeout of silence).
         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();
+        // 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
@@ -928,7 +947,7 @@ export class MessageProcessor {
             return;
         }
         try {
-            while (true) {
+            streamAttempt: while (true) {
                 // SAMR Phase 2: switch to fast model for tool-execution steps
                 if (stepRouterPhase === "phase2" && phase2Runtime) {
                     runtime = phase2Runtime;
@@ -936,6 +955,65 @@ export class MessageProcessor {
                 }
                 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;
+                // 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) {
+                        console.error(`[message-processor] stall-reprompt telemetry failed: ${emitErr?.message}`);
+                    }
+                    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) {
+                        console.error(`[message-processor] stall-reprompt log failed: ${logErr?.message}`);
+                    }
+                    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).
@@ -1469,6 +1547,46 @@ export class MessageProcessor {
                         prepareStep: ({ stepNumber: sn, messages: stepMessages }) => {
                             if (sn < 1)
                                 return {};
+                            // --- 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 = steerEnabled && !signal.aborted ? (deps.drainSteerMessages?.() ?? []) : [];
+                                const _newSteers = planSteerInjection({
+                                    drained: _drained,
+                                    aborted: signal.aborted,
+                                    enabled: steerEnabled,
+                                });
+                                if (_newSteers.length > 0) {
+                                    pendingSteers.push(..._newSteers);
+                                    try {
+                                        const _ar = globalThis.__muonroiAgentRuntime;
+                                        _ar?.emitEvent({
+                                            t: "event",
+                                            kind: "steer-inject",
+                                            count: _newSteers.length,
+                                            atStep: sn,
+                                            runId: deps.getActiveRunId() ?? "",
+                                        });
+                                    }
+                                    catch (emitErr) {
+                                        console.error(`[message-processor] steer-inject telemetry failed: ${emitErr?.message}`);
+                                    }
+                                }
+                                if (pendingSteers.length === 0)
+                                    return r;
+                                const _base = r.messages ?? stepMessages;
+                                const _steerContents = new Set(pendingSteers.map((s) => (typeof s.content === "string" ? s.content : JSON.stringify(s.content))));
+                                const _deduped = _base.filter((m) => !(m.role === "user" &&
+                                    _steerContents.has(typeof m.content === "string" ? m.content : JSON.stringify(m.content))));
+                                return { ...r, messages: [..._deduped, ...pendingSteers] };
+                            };
                             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).
@@ -1499,7 +1617,7 @@ export class MessageProcessor {
                                 return false;
                             });
                             if (hasPreserve) {
-                                return { messages: stripped };
+                                return withSteers({ messages: stripped });
                             }
                             // F2 — envelope = system prompt + JSON-Schema of every tool
                             // re-sent on every step. Without this the threshold check
@@ -1573,7 +1691,7 @@ export class MessageProcessor {
                             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.]`;
-                                return { messages: attachReminderToMessages(stripped, _pre) };
+                                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
@@ -1635,10 +1753,10 @@ export class MessageProcessor {
                                         : buildRepetitionReminder(_ceilingSessionId)
                                     : _scopePart;
                                 const withReminder = attachReminderToMessages(compacted, _reminder);
-                                return { messages: withReminder };
+                                return withSteers({ messages: withReminder });
                             }
                             if (compacted === stripped && stripped === stepMessages)
-                                return {};
+                                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
@@ -1656,9 +1774,9 @@ export class MessageProcessor {
                                 })()
                                 : null;
                             if (_compactNote) {
-                                return { messages: attachReminderToMessages(compacted, _compactNote) };
+                                return withSteers({ messages: attachReminderToMessages(compacted, _compactNote) });
                             }
-                            return { messages: compacted };
+                            return withSteers({ messages: compacted });
                         },
                         ...(dropParam("temperature") ? {} : { temperature: 0.7 }),
                         ...(dropParam("maxOutputTokens") ? {} : { maxOutputTokens: taskTypeToMaxTokens(pilCtx.taskType) }),
@@ -1730,6 +1848,11 @@ export class MessageProcessor {
                     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++;
                         if (signal.aborted) {
                             yield { type: "content", content: "\n\n[Cancelled]" };
                             break;
@@ -2283,6 +2406,19 @@ export class MessageProcessor {
                                 // 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;
+                                        }
+                                    }
                                     stall.dispose();
                                     // A response tool already produced the terminal structured
                                     // answer (buffered from its call args) before the provider
@@ -2913,6 +3049,20 @@ export class MessageProcessor {
                         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.

package/dist/src/orchestrator/orchestrator.d.ts

@@ -20,6 +20,8 @@ export declare class Agent {
     private messages;
     private messageSeqs;
     private abortController;
+    /** UI-registered live-queue steer drain; see Agent.setSteerDrain. */
+    private steerDrain;
     private maxToolRounds;
     private mode;
     private modelId;
@@ -97,6 +99,14 @@ export declare class Agent {
      */
     private _initFlow;
     getModel(): string;
+    /**
+     * Register (or clear with null) the UI callback that drains messages typed
+     * while a turn is streaming, for mid-turn steering injection. Called from the
+     * TUI when MUONROI_STEER_INJECTION is enabled.
+     */
+    setSteerDrain(fn: (() => {
+        text: string;
+    }[]) | null): void;
     getActiveRunId(): string | null;
     setModel(model: string): void;
     getMode(): AgentMode;

package/dist/src/orchestrator/orchestrator.js

@@ -130,6 +130,8 @@ export class Agent {
     messages = [];
     messageSeqs = [];
     abortController = null;
+    /** UI-registered live-queue steer drain; see Agent.setSteerDrain. */
+    steerDrain = null;
     maxToolRounds;
     mode = "agent";
     modelId;
@@ -309,6 +311,14 @@ export class Agent {
     getModel() {
         return this.modelId;
     }
+    /**
+     * Register (or clear with null) the UI callback that drains messages typed
+     * while a turn is streaming, for mid-turn steering injection. Called from the
+     * TUI when MUONROI_STEER_INJECTION is enabled.
+     */
+    setSteerDrain(fn) {
+        this.steerDrain = fn;
+    }
     getActiveRunId() {
         return this._activeRunId;
     }
@@ -2123,6 +2133,7 @@ export class Agent {
             runDelegation: (request, signal) => self.runDelegation(request, signal),
             readDelegation: (id) => self.readDelegation(id),
             listDelegations: () => self.listDelegations(),
+            drainSteerMessages: () => self.steerDrain?.() ?? [],
             appendCompletedTurn: (user, asst) => self.appendCompletedTurn(user, asst),
             discardAbortedTurn: (user) => self.discardAbortedTurn(user),
             recordUsage: (usage, source, model, shape) => self.recordUsage(usage, source, model, shape),

package/dist/src/orchestrator/stall-rescue.d.ts

@@ -31,6 +31,7 @@ export interface StallToolResult {
 export declare const STALL_RESCUE_MAX_RESULTS = 8;
 /** Max chars kept per tool output in the digest. */
 export declare const STALL_RESCUE_MAX_CHARS_PER_RESULT = 1500;
+export declare function detectReplyLanguageHint(userText: string): string | null;
 /**
  * Capture a tool result into a capped ring buffer (mutates `buffer`). Keeps the
  * buffer bounded in BOTH count and per-entry size so a long turn can't blow

package/dist/src/orchestrator/stall-rescue.js

@@ -26,6 +26,22 @@
 export const STALL_RESCUE_MAX_RESULTS = 8;
 /** Max chars kept per tool output in the digest. */
 export const STALL_RESCUE_MAX_CHARS_PER_RESULT = 1500;
+/**
+ * Heuristic: detect the user's reply language from the original request so the
+ * synthesis prompt can carry an EXPLICIT language directive. The base system
+ * prompt's "reply in user's language" rule is normally enough, but after a
+ * stall the rescue path issues a NEW synthesis call whose extra directives
+ * ("use ONLY tool outputs", "no more tools") can crowd out the language rule
+ * — observed live (storyflow_ui session 22661c8de9f2): user wrote Vietnamese,
+ * the rescued answer came back in English. Returns a language name or null
+ * (skip the hint = treat as English-default).
+ */
+const VI_DIACRITIC_RE = /[à-ỹÀ-Ỹ]/;
+export function detectReplyLanguageHint(userText) {
+    if (VI_DIACRITIC_RE.test(userText))
+        return "Vietnamese";
+    return null;
+}
 /**
  * Capture a tool result into a capped ring buffer (mutates `buffer`). Keeps the
  * buffer bounded in BOTH count and per-entry size so a long turn can't blow
@@ -44,7 +60,10 @@ export function pushStallToolResult(buffer, tool, rawText) {
  */
 export function buildStallSynthesisMessages(baseMessages, userText, toolResults) {
     const digest = toolResults.map((r, i) => `[${i + 1}] ${r.tool}:\n${r.text}`).join("\n\n");
-    const content = "The connection to the model stalled before it could finish its answer. " +
+    const lang = detectReplyLanguageHint(userText);
+    const langLine = lang ? `Reply in ${lang} (the user wrote in ${lang}).\n\n` : "";
+    const content = langLine +
+        "The connection to the model stalled before it could finish its answer. " +
         "You already ran the tools below this turn — use ONLY their outputs to give " +
         "your best final answer now. Do NOT call any more tools.\n\n" +
         `Original request:\n${userText}\n\n` +

package/dist/src/orchestrator/stall-rescue.test.js

@@ -1,5 +1,5 @@
 import { describe, expect, it } from "vitest";
-import { attemptStallRescue, buildStallSynthesisMessages, pushStallToolResult, STALL_RESCUE_MAX_CHARS_PER_RESULT, STALL_RESCUE_MAX_RESULTS, } from "./stall-rescue.js";
+import { attemptStallRescue, buildStallSynthesisMessages, detectReplyLanguageHint, pushStallToolResult, STALL_RESCUE_MAX_CHARS_PER_RESULT, STALL_RESCUE_MAX_RESULTS, } from "./stall-rescue.js";
 describe("pushStallToolResult", () => {
     it("caps the buffer count, keeping the most recent results", () => {
         const buf = [];
@@ -34,6 +34,35 @@ describe("buildStallSynthesisMessages", () => {
         // does not mutate the base array
         expect(base.length).toBe(1);
     });
+    it("prepends a Vietnamese reply directive when user wrote in Vietnamese", () => {
+        const out = buildStallSynthesisMessages([], "sửa thêm popup reload không xoá login", [
+            { tool: "grep", text: "found" },
+        ]);
+        const last = out[out.length - 1];
+        expect(last.content).toMatch(/^Reply in Vietnamese/);
+        expect(last.content).toContain("the user wrote in Vietnamese");
+    });
+    it("omits the language directive for plain English input", () => {
+        const out = buildStallSynthesisMessages([], "find a bug please", [{ tool: "grep", text: "found" }]);
+        const last = out[out.length - 1];
+        expect(last.content).not.toMatch(/^Reply in /);
+    });
+});
+describe("detectReplyLanguageHint", () => {
+    it("detects Vietnamese from any diacritic", () => {
+        expect(detectReplyLanguageHint("tìm")).toBe("Vietnamese");
+        expect(detectReplyLanguageHint("sửa thêm cho tôi")).toBe("Vietnamese");
+        expect(detectReplyLanguageHint("Cập nhật")).toBe("Vietnamese");
+    });
+    it("returns null for ASCII / English text", () => {
+        expect(detectReplyLanguageHint("commit and push please")).toBeNull();
+        expect(detectReplyLanguageHint("")).toBeNull();
+    });
+    it("returns null for a VN-only word with NO diacritics (the heuristic is conservative)", () => {
+        // 'commit push de auto ci/cd len prod nhe' — VN sans diacritics is indistinguishable
+        // from English under this heuristic. Acceptable: false negative, never false positive.
+        expect(detectReplyLanguageHint("commit push de auto ci/cd len prod nhe")).toBeNull();
+    });
 });
 describe("attemptStallRescue", () => {
     it("returns null when there are no tool results (nothing to synthesize)", async () => {

package/dist/src/orchestrator/stall-watchdog.d.ts

@@ -31,4 +31,35 @@ export interface StallWatchdog {
 export declare const STALL_ABORT_REASON = "provider-stall";
 /** User-facing message surfaced when the stall watchdog fires. */
 export declare const STALL_ERROR_MESSAGE: string;
+/** Inputs to the stall re-prompt decision — see {@link shouldRepromptStall}. */
+export interface StallRepromptState {
+    /** The watchdog fired for this attempt. */
+    stallTriggered: boolean;
+    /** How many stall re-prompts have already happened this turn. */
+    stallRetryCount: number;
+    /** Configured cap (getProviderStallRetries); 0 disables re-prompt. */
+    maxStallRetries: number;
+    /** Real content parts received this attempt (the abort part is NOT counted). */
+    chunksThisAttempt: number;
+    /** True when no assistant text has flowed this attempt. */
+    assistantTextEmpty: boolean;
+    /** True on genuine user cancel (never re-prompt over a cancel). */
+    aborted: boolean;
+}
+/**
+ * Decide whether a fired stall watchdog should trigger a re-prompt (re-issue
+ * the same request) instead of surfacing the stall.
+ *
+ * ONLY a time-to-first-byte stall qualifies: zero real chunks AND no assistant
+ * text this attempt, under the retry cap, and not a user cancel. Re-issuing
+ * after tools ran or text flowed would corrupt/duplicate output — those cases
+ * fall through to the partial-answer rescue path instead. Pure (no side
+ * effects) so it is unit-testable in isolation from the orchestrator loop.
+ */
+export declare function shouldRepromptStall(s: StallRepromptState): boolean;
+/**
+ * Exponential backoff (ms, capped at 4s) before the Nth stall re-prompt
+ * (1-based): 500 → 1000 → 2000 → 4000 → 4000.
+ */
+export declare function stallRepromptBackoffMs(attempt: number): number;
 export declare function createStallWatchdog(timeoutMs: number, onFire?: () => void): StallWatchdog;