muonroi-cli 1.7.2 → 1.8.1

package/dist/src/pil/config.d.ts

@@ -5,6 +5,16 @@
  *   Default OFF during rollout; flip to ON after dual-run validation.
  */
 export declare function isUnifiedPilEnabled(): boolean;
+/**
+ * MUONROI_PIL_UNIFIED_BUDGET_MS: client-side timeout for the unified
+ * /api/pil-context call in Layer 1 Pass 3. The server's uncached path is
+ * classifier-bound (~1.5-3s; its own classifier timeout is 3500ms), so the
+ * legacy 1500ms budget aborted before the server could ever answer — the
+ * unified path then ALWAYS fell back to legacy, making MUONROI_PIL_UNIFIED=1 a
+ * no-op. Default 3500ms clears realistic uncached latency; cache hits return in
+ * ~1ms regardless. Clamped to [1000, 8000].
+ */
+export declare function getUnifiedPilBudgetMs(): number;
 /**
  * MUONROI_LLM_FIRST_CLASSIFY: model-first Layer-1 classification. When enabled
  * (default), the configured model classifies taskType/intentKind/style at the
@@ -14,5 +24,15 @@ export declare function isUnifiedPilEnabled(): boolean;
  * (the orchestrator does this on the main path); without it, the cascade runs.
  */
 export declare function isLlmFirstClassifyEnabled(): boolean;
+/**
+ * G3 (b1): on the model-first classify path, also fetch the unified pil-context
+ * brain so layer3 renders the rich source="unified" injection (server-attributed
+ * t0/t2 + t1_rules + the rateable ledger) instead of falling through to its
+ * legacy dense-only /api/search round-trip. Default ON; set "0" to revert to
+ * the legacy layer3 injection on the default path. Independent of
+ * MUONROI_PIL_UNIFIED (which gates the OFFLINE cascade's unified pass).
+ */
+export declare function isLlmFirstBrainEnabled(): boolean;
 export declare function isDiscoveryEnabled(): boolean;
 export declare function getMaxInterviewQuestions(): number;
+export declare function isPonytailModeEnabled(): boolean;

package/dist/src/pil/config.js

@@ -11,6 +11,21 @@ export function isUnifiedPilEnabled() {
         return true;
     return false;
 }
+/**
+ * MUONROI_PIL_UNIFIED_BUDGET_MS: client-side timeout for the unified
+ * /api/pil-context call in Layer 1 Pass 3. The server's uncached path is
+ * classifier-bound (~1.5-3s; its own classifier timeout is 3500ms), so the
+ * legacy 1500ms budget aborted before the server could ever answer — the
+ * unified path then ALWAYS fell back to legacy, making MUONROI_PIL_UNIFIED=1 a
+ * no-op. Default 3500ms clears realistic uncached latency; cache hits return in
+ * ~1ms regardless. Clamped to [1000, 8000].
+ */
+export function getUnifiedPilBudgetMs() {
+    const raw = Number(process.env.MUONROI_PIL_UNIFIED_BUDGET_MS);
+    if (!Number.isFinite(raw))
+        return 3500;
+    return Math.max(1000, Math.min(8000, Math.trunc(raw)));
+}
 /**
  * MUONROI_LLM_FIRST_CLASSIFY: model-first Layer-1 classification. When enabled
  * (default), the configured model classifies taskType/intentKind/style at the
@@ -22,6 +37,17 @@ export function isUnifiedPilEnabled() {
 export function isLlmFirstClassifyEnabled() {
     return process.env.MUONROI_LLM_FIRST_CLASSIFY !== "0";
 }
+/**
+ * G3 (b1): on the model-first classify path, also fetch the unified pil-context
+ * brain so layer3 renders the rich source="unified" injection (server-attributed
+ * t0/t2 + t1_rules + the rateable ledger) instead of falling through to its
+ * legacy dense-only /api/search round-trip. Default ON; set "0" to revert to
+ * the legacy layer3 injection on the default path. Independent of
+ * MUONROI_PIL_UNIFIED (which gates the OFFLINE cascade's unified pass).
+ */
+export function isLlmFirstBrainEnabled() {
+    return process.env.MUONROI_LLM_FIRST_BRAIN !== "0";
+}
 export function isDiscoveryEnabled() {
     return process.env.MUONROI_PIL_DISCOVERY !== "0";
 }
@@ -33,4 +59,13 @@ export function getMaxInterviewQuestions() {
     const v = Number(process.env.MUONROI_PIL_MAX_QUESTIONS);
     return Number.isFinite(v) && v >= 1 && v <= 5 ? v : 3;
 }
+export function isPonytailModeEnabled() {
+    if (process.env.MUONROI_PONYTAIL_DISABLE === "1")
+        return false;
+    if (process.env.MUONROI_PONYTAIL_DISABLE === "0")
+        return true;
+    if (process.env.VITEST === "true" || process.env.NODE_ENV === "test")
+        return false;
+    return true;
+}
 //# sourceMappingURL=config.js.map

package/dist/src/pil/discovery-types.d.ts

@@ -23,13 +23,32 @@ export interface RelevantModule {
     relevance: string;
     exists: boolean;
 }
-export type ClarityDimension = "outcome" | "scope" | "constraint";
-export interface ClarityGap {
-    dimension: ClarityDimension;
-    description: string;
-    suggestedQuestion: string;
-    options: string[];
-    defaultIndex: number;
+/**
+ * A card designed entirely by the model — the model controls the question,
+ * context, options (labels, kinds, which is cancel/adjust), and default.
+ * The CLI only assigns questionId and renders via CouncilQuestionData.
+ */
+export interface ModelCard {
+    /** The question text shown to the user */
+    question: string;
+    /** Optional context/explanation shown below the question */
+    context?: string;
+    /** User-selectable options. Model controls labels, kinds, cancel/adjust markers */
+    options: ModelCardOption[];
+    /** Which option is pre-selected (0 = first). Defaults to 0. */
+    defaultIndex?: number;
+}
+export interface ModelCardOption {
+    /** Button label or freetext field label */
+    label: string;
+    /** Optional tooltip / description */
+    description?: string;
+    /** 'choice' = clickable button, 'freetext' = free-text input field */
+    kind: "choice" | "freetext";
+    /** If true, picking this option cancels the entire interaction */
+    isCancel?: boolean;
+    /** If true, picking this indicates the user wants to clarify further (triggers re-interview) */
+    isAdjust?: boolean;
 }
 export interface ClarifiedIntent {
     intentStatement?: string;
@@ -59,7 +78,7 @@ export type ModelClarificationProposer = (input: {
         domain?: string | null;
     };
     additionalContext?: string;
-}) => Promise<string[]>;
+}) => Promise<ModelCard[]>;
 export interface FeasibilityResult {
     viable: boolean;
     warnings: string[];
@@ -81,6 +100,11 @@ export interface DiscoveryResult {
     domain: string | null;
     outputStyle: OutputStyle | null;
     discoveryMs: number;
+    /** Raw Q&A pairs from the interview, visible to the model in enrichment */
+    interviewTranscript: Array<{
+        question: string;
+        answer: string;
+    }>;
 }
 export interface AcceptanceCardData {
     intentStatement: string;
@@ -90,5 +114,4 @@ export interface AcceptanceCardData {
 }
 export interface DiscoveryInteractionHandler {
     askQuestion(question: CouncilQuestionData): Promise<CouncilQuestionAnswer>;
-    showAcceptance(card: AcceptanceCardData): Promise<"accept" | "adjust" | "cancel">;
 }

package/dist/src/pil/discovery.d.ts

@@ -11,8 +11,8 @@ export interface L1Result {
 export declare function runDiscovery(raw: string, l1: L1Result, cwd: string, handler: DiscoveryInteractionHandler | null, sessionId?: string | null, clarificationProposer?: ModelClarificationProposer | null, recentTurnsSummary?: string | null): Promise<DiscoveryResult>;
 /**
  * Create a ModelClarificationProposer backed by the actual task model.
- * The model receives the user raw + CLI enrichment (l1, project modules, etc.)
- * and outputs the specific questions *it* needs the user to answer.
- * This is the effective way to let the model interview based on what is still missing.
+ * The model receives the user raw + CLI enrichment and outputs the
+ * exact ModelCard[] it wants shown — full control over questions, options,
+ * option kinds, and cancel/adjust markers.
  */
 export declare function createModelClarificationProposer(providerFactory: any, modelId: string): ModelClarificationProposer;