muonroi-cli 1.6.2 → 1.6.3

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

@@ -1,2 +1,2 @@
-export declare const PACKAGE_VERSION = "1.6.2";
+export declare const PACKAGE_VERSION = "1.6.3";
 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 = "1.6.2";
+export const PACKAGE_VERSION = "1.6.3";
 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/mcp/__tests__/client-pool.spec.js

@@ -109,6 +109,27 @@ describe("acquireMcpTools — cross-turn client pool", () => {
         expect(results.every((r) => r === "pong")).toBe(true);
         expect(connectOneServer).toHaveBeenCalledTimes(2); // 14 failures → exactly ONE shared reconnect
     });
+    it("waits for a criticalServerId past the normal deadline so it lands THIS turn (session 584ba476c07a)", async () => {
+        // Normal deadline is 500ms (mock). docs connects at ~700ms — past the normal
+        // deadline but within the critical window → must be included when critical.
+        connectOneServer.mockImplementation((s) => new Promise((res) => {
+            if (s.id === "docs")
+                setTimeout(() => res(connected(s.id)), 700);
+            else
+                res(connected(s.id));
+        }));
+        const b = await acquireMcpTools([srv("docs")], { criticalServerIds: ["docs"], criticalDeadlineMs: 3000 });
+        expect(Object.keys(b.tools)).toContain("mcp_docs__ping");
+        expect(b.errors).toHaveLength(0);
+    });
+    it("without criticalServerIds, a slow server is reported still-connecting (available next turn)", async () => {
+        connectOneServer.mockImplementation((s) => new Promise((res) => {
+            setTimeout(() => res(connected(s.id)), 700);
+        }));
+        const b = await acquireMcpTools([srv("docs")]);
+        expect(Object.keys(b.tools)).not.toContain("mcp_docs__ping");
+        expect(b.errors.some((e) => /still connecting/.test(e))).toBe(true);
+    });
     it("keys by cwd/config — a different command reconnects rather than reusing", async () => {
         connectOneServer.mockImplementation(async (s) => connected(s.id));
         await acquireMcpTools([

package/dist/src/mcp/client-pool.js

@@ -179,6 +179,28 @@ export async function acquireMcpTools(servers, opts) {
     await Promise.race([Promise.allSettled(attempts), deadline]);
     if (deadlineTimer)
         clearTimeout(deadlineTimer);
+    // Critical-server extended wait: a turn that MUST have a specific server (e.g.
+    // muonroi-docs on an ecosystem question) waits for just that server's connect
+    // beyond the normal deadline, so a cold first-connect is included THIS turn
+    // rather than reported "still connecting → next turn". Only the named servers
+    // are awaited; everything already settled is untouched (no added latency for
+    // normal turns, which pass no criticalServerIds).
+    const critical = new Set((opts?.criticalServerIds ?? []).filter(Boolean));
+    if (critical.size > 0) {
+        const pendingIdx = enabled.map((s, i) => ({ s, i })).filter(({ s, i }) => critical.has(s.id) && !slots[i].done);
+        if (pendingIdx.length > 0) {
+            const criticalDeadlineMs = Math.max(deadlineMs, opts?.criticalDeadlineMs ?? 8000);
+            const extraMs = Math.max(0, criticalDeadlineMs - deadlineMs);
+            let extraTimer;
+            const extraDeadline = new Promise((resolve) => {
+                extraTimer = setTimeout(resolve, extraMs);
+                extraTimer.unref?.();
+            });
+            await Promise.race([Promise.allSettled(pendingIdx.map(({ i }) => attempts[i])), extraDeadline]);
+            if (extraTimer)
+                clearTimeout(extraTimer);
+        }
+    }
     for (let i = 0; i < slots.length; i++) {
         const slot = slots[i];
         if (slot.done) {

package/dist/src/mcp/runtime.d.ts

@@ -8,6 +8,17 @@ export interface McpToolBundle {
 }
 export interface McpBuildOptions {
     onOAuthRequired?: (serverId: string, url: URL) => void;
+    /**
+     * Server ids the CURRENT turn critically needs (e.g. muonroi-docs on an
+     * ecosystem question). acquireMcpTools waits for these specifically beyond the
+     * normal build deadline — up to `criticalDeadlineMs` — so a cold first-connect
+     * is included THIS turn instead of "ready next turn" (session 584ba476c07a:
+     * first ecosystem question missed muonroi-docs while it was still warming).
+     * Other servers are unaffected — only the named ones get the extended wait.
+     */
+    criticalServerIds?: string[];
+    /** Extended ceiling (ms) for criticalServerIds. Default 8000. */
+    criticalDeadlineMs?: number;
 }
 /**
  * Total wall-clock budget for building the MCP tool set. Servers connect in

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

@@ -60,6 +60,7 @@ import * as phaseTracker from "../ee/phase-tracker.js";
 import { buildScope as buildScopeForVeto } from "../ee/scope.js";
 import { fireTrajectoryEvent } from "../ee/session-trajectory.js";
 import { getTenantId as getTenantIdForVeto } from "../ee/tenant.js";
+import { mentionsEcosystemScope } from "../gsd/directives.js";
 import { acquireMcpTools } from "../mcp/client-pool.js";
 import { dropRedundantFsMcpTools, filterMcpServersByMessage } from "../mcp/smart-filter.js";
 import { getModelInfo } from "../models/registry.js";
@@ -1022,6 +1023,16 @@ export class MessageProcessor {
                         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
@@ -1038,6 +1049,7 @@ export class MessageProcessor {
                                     // command-injection vector the old exec() opener had.
                                     openUrl(url);
                                 },
+                                ...(criticalServerIds && criticalServerIds.length > 0 ? { criticalServerIds } : {}),
                             });
                         }
                         catch (err) {

package/dist/src/pil/__tests__/layer4-gsd.test.js

@@ -104,6 +104,16 @@ describe("layer4Gsd (gsd-native)", () => {
         const result = await layer4Gsd(makeCtx({ raw, enriched: raw, taskType: "analyze", intentKind: "task", deliverableKind: "answer" }));
         expect(result.enriched).toContain("QUESTION / explanatory");
     });
+    it("deliverableKind='report' is informational (no council/discuss scaffold) — session 666630479c1a", async () => {
+        // "Đọc và tóm tắt kiến trúc…" classifies as deliverableKind 'report'. A
+        // report is human-facing with NO code change, so it must route to the
+        // QUESTION directive, not the heavy implement/discuss/council scaffold that
+        // over-asked with askcards on a read/summarize task.
+        const raw = "đọc và tóm tắt kiến trúc src/orchestrator, src/pil, src/mcp kèm file:line";
+        const result = await layer4Gsd(makeCtx({ raw, enriched: raw, taskType: "analyze", intentKind: "task", deliverableKind: "report" }));
+        expect(result.enriched).toContain("QUESTION / explanatory");
+        expect(result.enriched).not.toContain("MANDATORY");
+    });
     it("Phase 2b: deliverableKind='code' is NOT informational even for a question-shaped prompt", async () => {
         // The raw text reads as a question — the legacy regex would mark it
         // informational. The model's deliverableKind='code' must override that so

package/dist/src/pil/layer4-gsd.js

@@ -84,15 +84,20 @@ export async function layer4Gsd(ctx) {
     // into the human-facing reply as a "2-3 line plan" + process narration
     // (session 829a83888dd2). Route them to the human-facing question directive.
     //
-    // Phase 2b: when the model classified the deliverable, CONSUME it — an
-    // "answer" deliverable IS informational. Only when the model didn't emit one
-    // (deliverableKind null → legacy cascade, or the model omitted the word) do
-    // we fall back to the legacy regex predicates:
+    // Phase 2b: when the model classified the deliverable, CONSUME it. Both an
+    // "answer" AND a "report" deliverable are HUMAN-FACING with no code change, so
+    // both are informational — only "code" routes through the implement/verify (and
+    // heavy discuss/council) scaffold. Treating "report" as non-informational sent
+    // read/summarize/architecture tasks (deliverableKind "report") down the heavy
+    // council + AskUserQuestion path, over-asking on a task that just wanted a
+    // written summary (session 666630479c1a: "Đọc và tóm tắt kiến trúc…" raised 2
+    // askcards + a council loop). Only when the model emitted no deliverable
+    // (deliverableKind null → legacy cascade) do we fall back to regex predicates:
     //   1. isMetaAnalysisPrompt — self/CLI evaluation, prior-turn reflection.
     //   2. taskType "general" classified as a real task by L1.
     //   3. question-shaped prompt that is NOT an implementation request.
     const informational = ctx.deliverableKind
-        ? ctx.deliverableKind === "answer"
+        ? ctx.deliverableKind !== "code"
         : isMetaAnalysisPrompt(ctx.raw) ||
             (ctx.taskType === "general" && ctx.intentKind === "task") ||
             (isQuestionLike(ctx.raw) && !isImplementationIntent(ctx.raw));

package/dist/src/ui/app.js

@@ -2525,7 +2525,7 @@ export function App({ agent, startupConfig, initialMessage, onExit }) {
                         applyLocalAssistantDelta(`\n⚠ [Experience] ${eeChunk.experienceWarning?.message ?? eeChunk.content ?? ""}\nWhy: ${eeChunk.experienceWarning?.why ?? ""}\n`);
                     }
                     else if (eeChunk.type === "experience_injected") {
-                        applyLocalAssistantDelta(`\n💡 [Experience Injected] ${eeChunk.experienceInjected?.pointCount ?? 0} point(s) loaded (score ≥ ${eeChunk.experienceInjected?.scoreFloor ?? 0})\n`);
+                        applyLocalAssistantDelta(formatExperienceInjectedBlock(eeChunk.experienceInjected ?? {}));
                     }
                 });
                 for await (const chunk of agent.processMessage(text.trim(), undefined, images)) {
@@ -3472,10 +3472,7 @@ export function App({ agent, startupConfig, initialMessage, onExit }) {
                                             },
                                         ];
                                     }
-                                    return [
-                                        ...prev,
-                                        buildAssistantEntry(`💡 [Experience Injected] ${chunk.experienceInjected.pointCount} point(s)`),
-                                    ];
+                                    return [...prev, buildAssistantEntry(formatExperienceInjectedBlock(chunk.experienceInjected))];
                                 });
                             }
                             if (chunk.type === "halt" && chunk.haltChunk) {
@@ -3694,10 +3691,7 @@ export function App({ agent, startupConfig, initialMessage, onExit }) {
                                             },
                                         ];
                                     }
-                                    return [
-                                        ...prev,
-                                        buildAssistantEntry(`💡 [Experience Injected] ${chunk.experienceInjected.pointCount} point(s)`),
-                                    ];
+                                    return [...prev, buildAssistantEntry(formatExperienceInjectedBlock(chunk.experienceInjected))];
                                 });
                             }
                             if (chunk.type === "done")

package/package.json

@@ -3,7 +3,7 @@
   "workspaces": [
     "packages/*"
   ],
-  "version": "1.6.2",
+  "version": "1.6.3",
   "description": "BYOK AI coding agent with multi-model council debate, role-based routing, and auto-compact.",
   "repository": {
     "type": "git",