@caupulican/pi-adaptative 0.80.86 → 0.80.89

package/dist/core/model-capability.d.ts

@@ -0,0 +1,41 @@
+/**
+ * Model capability auto-detection: derive what the harness may load onto a model FROM the model's
+ * own metadata (`Model.contextWindow`), so small open models (4k/8k/16k windows, sub-1B params)
+ * can still hold a usable chat instead of drowning in tool schemas and background-lane prompts.
+ *
+ * Derivation is metadata-first; defaults apply only when the metadata is missing (unknown/zero
+ * window keeps today's full behavior rather than guessing). Detection can be disabled or forced
+ * per class via the `modelCapability.mode` setting.
+ */
+export type ModelCapabilityClass = "full" | "lean" | "minimal" | "chat";
+export type ModelCapabilityMode = "auto" | "off" | ModelCapabilityClass;
+export interface ModelCapabilityProfile {
+    class: ModelCapabilityClass;
+    contextWindow?: number;
+    reasonCode: string;
+    /** Allow-list; undefined = no allow-list restriction. */
+    allowedToolNames?: readonly string[];
+    /** Block-list applied after the allow-list; undefined = nothing blocked. */
+    blockedToolNames?: readonly string[];
+    /** Whether idle background lanes (goal auto-continue, research) may run on this model. */
+    backgroundLanesEnabled: boolean;
+    /** Output-token cap for lane isolated completions, scaled to the window. */
+    laneMaxOutputTokens: number;
+}
+/** Windows at or above this keep the full harness surface. */
+export declare const MODEL_CAPABILITY_FULL_MIN_CONTEXT = 32768;
+/** Windows at or above this keep core tools but shed background-autonomy extras. */
+export declare const MODEL_CAPABILITY_LEAN_MIN_CONTEXT = 16384;
+/** Windows at or above this get the minimal coding set; below is chat-only. */
+export declare const MODEL_CAPABILITY_MINIMAL_MIN_CONTEXT = 8192;
+export declare const MODEL_CAPABILITY_LEAN_BLOCKED_TOOLS: readonly string[];
+export declare const MODEL_CAPABILITY_MINIMAL_ALLOWED_TOOLS: readonly string[];
+export declare const MODEL_CAPABILITY_CHAT_ALLOWED_TOOLS: readonly string[];
+export declare const DEFAULT_LANE_MAX_OUTPUT_TOKENS = 2048;
+export declare function deriveModelCapabilityProfile(args: {
+    contextWindow?: number;
+    mode?: ModelCapabilityMode;
+}): ModelCapabilityProfile;
+/** Apply the profile's allow/block lists to a requested tool-name list, preserving order. */
+export declare function filterToolNamesForCapability(toolNames: readonly string[], profile: ModelCapabilityProfile): string[];
+//# sourceMappingURL=model-capability.d.ts.map

package/dist/core/model-capability.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-capability.d.ts","sourceRoot":"","sources":["../../src/core/model-capability.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,MAAM,MAAM,oBAAoB,GAAG,MAAM,GAAG,MAAM,GAAG,SAAS,GAAG,MAAM,CAAC;AAExE,MAAM,MAAM,mBAAmB,GAAG,MAAM,GAAG,KAAK,GAAG,oBAAoB,CAAC;AAExE,MAAM,WAAW,sBAAsB;IACtC,KAAK,EAAE,oBAAoB,CAAC;IAC5B,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,UAAU,EAAE,MAAM,CAAC;IACnB,yDAAyD;IACzD,gBAAgB,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IACrC,4EAA4E;IAC5E,gBAAgB,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IACrC,0FAA0F;IAC1F,sBAAsB,EAAE,OAAO,CAAC;IAChC,4EAA4E;IAC5E,mBAAmB,EAAE,MAAM,CAAC;CAC5B;AAED,8DAA8D;AAC9D,eAAO,MAAM,iCAAiC,QAAS,CAAC;AACxD,oFAAoF;AACpF,eAAO,MAAM,iCAAiC,QAAS,CAAC;AACxD,+EAA+E;AAC/E,eAAO,MAAM,oCAAoC,OAAQ,CAAC;AAE1D,eAAO,MAAM,mCAAmC,EAAE,SAAS,MAAM,EAAkC,CAAC;AACpG,eAAO,MAAM,sCAAsC,EAAE,SAAS,MAAM,EAOnE,CAAC;AACF,eAAO,MAAM,mCAAmC,EAAE,SAAS,MAAM,EAAO,CAAC;AAEzE,eAAO,MAAM,8BAA8B,OAAO,CAAC;AA0CnD,wBAAgB,4BAA4B,CAAC,IAAI,EAAE;IAClD,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,IAAI,CAAC,EAAE,mBAAmB,CAAC;CAC3B,GAAG,sBAAsB,CA2BzB;AAED,6FAA6F;AAC7F,wBAAgB,4BAA4B,CAAC,SAAS,EAAE,SAAS,MAAM,EAAE,EAAE,OAAO,EAAE,sBAAsB,GAAG,MAAM,EAAE,CAWpH","sourcesContent":["/**\n * Model capability auto-detection: derive what the harness may load onto a model FROM the model's\n * own metadata (`Model.contextWindow`), so small open models (4k/8k/16k windows, sub-1B params)\n * can still hold a usable chat instead of drowning in tool schemas and background-lane prompts.\n *\n * Derivation is metadata-first; defaults apply only when the metadata is missing (unknown/zero\n * window keeps today's full behavior rather than guessing). Detection can be disabled or forced\n * per class via the `modelCapability.mode` setting.\n */\n\nexport type ModelCapabilityClass = \"full\" | \"lean\" | \"minimal\" | \"chat\";\n\nexport type ModelCapabilityMode = \"auto\" | \"off\" | ModelCapabilityClass;\n\nexport interface ModelCapabilityProfile {\n\tclass: ModelCapabilityClass;\n\tcontextWindow?: number;\n\treasonCode: string;\n\t/** Allow-list; undefined = no allow-list restriction. */\n\tallowedToolNames?: readonly string[];\n\t/** Block-list applied after the allow-list; undefined = nothing blocked. */\n\tblockedToolNames?: readonly string[];\n\t/** Whether idle background lanes (goal auto-continue, research) may run on this model. */\n\tbackgroundLanesEnabled: boolean;\n\t/** Output-token cap for lane isolated completions, scaled to the window. */\n\tlaneMaxOutputTokens: number;\n}\n\n/** Windows at or above this keep the full harness surface. */\nexport const MODEL_CAPABILITY_FULL_MIN_CONTEXT = 32_768;\n/** Windows at or above this keep core tools but shed background-autonomy extras. */\nexport const MODEL_CAPABILITY_LEAN_MIN_CONTEXT = 16_384;\n/** Windows at or above this get the minimal coding set; below is chat-only. */\nexport const MODEL_CAPABILITY_MINIMAL_MIN_CONTEXT = 8_192;\n\nexport const MODEL_CAPABILITY_LEAN_BLOCKED_TOOLS: readonly string[] = [\"delegate\", \"context_audit\"];\nexport const MODEL_CAPABILITY_MINIMAL_ALLOWED_TOOLS: readonly string[] = [\n\t\"read\",\n\t\"bash\",\n\t\"edit\",\n\t\"write\",\n\t// The executor tool: minimal-class models ARE the daily-ops executors, and its schema is tiny.\n\t\"run_toolkit_script\",\n];\nexport const MODEL_CAPABILITY_CHAT_ALLOWED_TOOLS: readonly string[] = [];\n\nexport const DEFAULT_LANE_MAX_OUTPUT_TOKENS = 2048;\nconst MIN_LANE_MAX_OUTPUT_TOKENS = 256;\n\nfunction laneOutputTokensForWindow(contextWindow: number | undefined): number {\n\tif (contextWindow === undefined || contextWindow <= 0) return DEFAULT_LANE_MAX_OUTPUT_TOKENS;\n\t// A lane completion may use at most an eighth of the window for output, floored so tiny\n\t// windows still produce something parseable.\n\treturn Math.min(DEFAULT_LANE_MAX_OUTPUT_TOKENS, Math.max(MIN_LANE_MAX_OUTPUT_TOKENS, Math.floor(contextWindow / 8)));\n}\n\nfunction profileForClass(\n\tcapabilityClass: ModelCapabilityClass,\n\treasonCode: string,\n\tcontextWindow: number | undefined,\n): ModelCapabilityProfile {\n\tconst base = {\n\t\tclass: capabilityClass,\n\t\treasonCode,\n\t\tbackgroundLanesEnabled: true,\n\t\tlaneMaxOutputTokens: laneOutputTokensForWindow(contextWindow),\n\t\t...(contextWindow !== undefined && contextWindow > 0 ? { contextWindow } : {}),\n\t};\n\tswitch (capabilityClass) {\n\t\tcase \"full\":\n\t\t\treturn base;\n\t\tcase \"lean\":\n\t\t\treturn { ...base, blockedToolNames: MODEL_CAPABILITY_LEAN_BLOCKED_TOOLS };\n\t\tcase \"minimal\":\n\t\t\treturn {\n\t\t\t\t...base,\n\t\t\t\tallowedToolNames: MODEL_CAPABILITY_MINIMAL_ALLOWED_TOOLS,\n\t\t\t\tbackgroundLanesEnabled: false,\n\t\t\t};\n\t\tcase \"chat\":\n\t\t\treturn {\n\t\t\t\t...base,\n\t\t\t\tallowedToolNames: MODEL_CAPABILITY_CHAT_ALLOWED_TOOLS,\n\t\t\t\tbackgroundLanesEnabled: false,\n\t\t\t};\n\t}\n}\n\nexport function deriveModelCapabilityProfile(args: {\n\tcontextWindow?: number;\n\tmode?: ModelCapabilityMode;\n}): ModelCapabilityProfile {\n\tconst mode = args.mode ?? \"auto\";\n\tconst contextWindow =\n\t\targs.contextWindow !== undefined && Number.isFinite(args.contextWindow) && args.contextWindow > 0\n\t\t\t? args.contextWindow\n\t\t\t: undefined;\n\tif (mode === \"off\") {\n\t\treturn profileForClass(\"full\", \"detection_disabled\", contextWindow);\n\t}\n\tif (mode !== \"auto\") {\n\t\treturn profileForClass(mode, \"forced_by_setting\", contextWindow);\n\t}\n\n\tif (contextWindow === undefined) {\n\t\t// Metadata missing: defaults, never guesses.\n\t\treturn profileForClass(\"full\", \"unknown_context_window_defaults\", undefined);\n\t}\n\tif (contextWindow >= MODEL_CAPABILITY_FULL_MIN_CONTEXT) {\n\t\treturn profileForClass(\"full\", \"large_context_window\", contextWindow);\n\t}\n\tif (contextWindow >= MODEL_CAPABILITY_LEAN_MIN_CONTEXT) {\n\t\treturn profileForClass(\"lean\", \"lean_context_window\", contextWindow);\n\t}\n\tif (contextWindow >= MODEL_CAPABILITY_MINIMAL_MIN_CONTEXT) {\n\t\treturn profileForClass(\"minimal\", \"minimal_context_window\", contextWindow);\n\t}\n\treturn profileForClass(\"chat\", \"chat_only_context_window\", contextWindow);\n}\n\n/** Apply the profile's allow/block lists to a requested tool-name list, preserving order. */\nexport function filterToolNamesForCapability(toolNames: readonly string[], profile: ModelCapabilityProfile): string[] {\n\tlet filtered = [...toolNames];\n\tif (profile.allowedToolNames !== undefined) {\n\t\tconst allowed = new Set(profile.allowedToolNames);\n\t\tfiltered = filtered.filter((name) => allowed.has(name));\n\t}\n\tif (profile.blockedToolNames !== undefined) {\n\t\tconst blocked = new Set(profile.blockedToolNames);\n\t\tfiltered = filtered.filter((name) => !blocked.has(name));\n\t}\n\treturn filtered;\n}\n"]}

package/dist/core/model-capability.js

@@ -0,0 +1,101 @@
+/**
+ * Model capability auto-detection: derive what the harness may load onto a model FROM the model's
+ * own metadata (`Model.contextWindow`), so small open models (4k/8k/16k windows, sub-1B params)
+ * can still hold a usable chat instead of drowning in tool schemas and background-lane prompts.
+ *
+ * Derivation is metadata-first; defaults apply only when the metadata is missing (unknown/zero
+ * window keeps today's full behavior rather than guessing). Detection can be disabled or forced
+ * per class via the `modelCapability.mode` setting.
+ */
+/** Windows at or above this keep the full harness surface. */
+export const MODEL_CAPABILITY_FULL_MIN_CONTEXT = 32_768;
+/** Windows at or above this keep core tools but shed background-autonomy extras. */
+export const MODEL_CAPABILITY_LEAN_MIN_CONTEXT = 16_384;
+/** Windows at or above this get the minimal coding set; below is chat-only. */
+export const MODEL_CAPABILITY_MINIMAL_MIN_CONTEXT = 8_192;
+export const MODEL_CAPABILITY_LEAN_BLOCKED_TOOLS = ["delegate", "context_audit"];
+export const MODEL_CAPABILITY_MINIMAL_ALLOWED_TOOLS = [
+    "read",
+    "bash",
+    "edit",
+    "write",
+    // The executor tool: minimal-class models ARE the daily-ops executors, and its schema is tiny.
+    "run_toolkit_script",
+];
+export const MODEL_CAPABILITY_CHAT_ALLOWED_TOOLS = [];
+export const DEFAULT_LANE_MAX_OUTPUT_TOKENS = 2048;
+const MIN_LANE_MAX_OUTPUT_TOKENS = 256;
+function laneOutputTokensForWindow(contextWindow) {
+    if (contextWindow === undefined || contextWindow <= 0)
+        return DEFAULT_LANE_MAX_OUTPUT_TOKENS;
+    // A lane completion may use at most an eighth of the window for output, floored so tiny
+    // windows still produce something parseable.
+    return Math.min(DEFAULT_LANE_MAX_OUTPUT_TOKENS, Math.max(MIN_LANE_MAX_OUTPUT_TOKENS, Math.floor(contextWindow / 8)));
+}
+function profileForClass(capabilityClass, reasonCode, contextWindow) {
+    const base = {
+        class: capabilityClass,
+        reasonCode,
+        backgroundLanesEnabled: true,
+        laneMaxOutputTokens: laneOutputTokensForWindow(contextWindow),
+        ...(contextWindow !== undefined && contextWindow > 0 ? { contextWindow } : {}),
+    };
+    switch (capabilityClass) {
+        case "full":
+            return base;
+        case "lean":
+            return { ...base, blockedToolNames: MODEL_CAPABILITY_LEAN_BLOCKED_TOOLS };
+        case "minimal":
+            return {
+                ...base,
+                allowedToolNames: MODEL_CAPABILITY_MINIMAL_ALLOWED_TOOLS,
+                backgroundLanesEnabled: false,
+            };
+        case "chat":
+            return {
+                ...base,
+                allowedToolNames: MODEL_CAPABILITY_CHAT_ALLOWED_TOOLS,
+                backgroundLanesEnabled: false,
+            };
+    }
+}
+export function deriveModelCapabilityProfile(args) {
+    const mode = args.mode ?? "auto";
+    const contextWindow = args.contextWindow !== undefined && Number.isFinite(args.contextWindow) && args.contextWindow > 0
+        ? args.contextWindow
+        : undefined;
+    if (mode === "off") {
+        return profileForClass("full", "detection_disabled", contextWindow);
+    }
+    if (mode !== "auto") {
+        return profileForClass(mode, "forced_by_setting", contextWindow);
+    }
+    if (contextWindow === undefined) {
+        // Metadata missing: defaults, never guesses.
+        return profileForClass("full", "unknown_context_window_defaults", undefined);
+    }
+    if (contextWindow >= MODEL_CAPABILITY_FULL_MIN_CONTEXT) {
+        return profileForClass("full", "large_context_window", contextWindow);
+    }
+    if (contextWindow >= MODEL_CAPABILITY_LEAN_MIN_CONTEXT) {
+        return profileForClass("lean", "lean_context_window", contextWindow);
+    }
+    if (contextWindow >= MODEL_CAPABILITY_MINIMAL_MIN_CONTEXT) {
+        return profileForClass("minimal", "minimal_context_window", contextWindow);
+    }
+    return profileForClass("chat", "chat_only_context_window", contextWindow);
+}
+/** Apply the profile's allow/block lists to a requested tool-name list, preserving order. */
+export function filterToolNamesForCapability(toolNames, profile) {
+    let filtered = [...toolNames];
+    if (profile.allowedToolNames !== undefined) {
+        const allowed = new Set(profile.allowedToolNames);
+        filtered = filtered.filter((name) => allowed.has(name));
+    }
+    if (profile.blockedToolNames !== undefined) {
+        const blocked = new Set(profile.blockedToolNames);
+        filtered = filtered.filter((name) => !blocked.has(name));
+    }
+    return filtered;
+}
+//# sourceMappingURL=model-capability.js.map

package/dist/core/model-capability.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-capability.js","sourceRoot":"","sources":["../../src/core/model-capability.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAoBH,8DAA8D;AAC9D,MAAM,CAAC,MAAM,iCAAiC,GAAG,MAAM,CAAC;AACxD,oFAAoF;AACpF,MAAM,CAAC,MAAM,iCAAiC,GAAG,MAAM,CAAC;AACxD,+EAA+E;AAC/E,MAAM,CAAC,MAAM,oCAAoC,GAAG,KAAK,CAAC;AAE1D,MAAM,CAAC,MAAM,mCAAmC,GAAsB,CAAC,UAAU,EAAE,eAAe,CAAC,CAAC;AACpG,MAAM,CAAC,MAAM,sCAAsC,GAAsB;IACxE,MAAM;IACN,MAAM;IACN,MAAM;IACN,OAAO;IACP,+FAA+F;IAC/F,oBAAoB;CACpB,CAAC;AACF,MAAM,CAAC,MAAM,mCAAmC,GAAsB,EAAE,CAAC;AAEzE,MAAM,CAAC,MAAM,8BAA8B,GAAG,IAAI,CAAC;AACnD,MAAM,0BAA0B,GAAG,GAAG,CAAC;AAEvC,SAAS,yBAAyB,CAAC,aAAiC,EAAU;IAC7E,IAAI,aAAa,KAAK,SAAS,IAAI,aAAa,IAAI,CAAC;QAAE,OAAO,8BAA8B,CAAC;IAC7F,wFAAwF;IACxF,6CAA6C;IAC7C,OAAO,IAAI,CAAC,GAAG,CAAC,8BAA8B,EAAE,IAAI,CAAC,GAAG,CAAC,0BAA0B,EAAE,IAAI,CAAC,KAAK,CAAC,aAAa,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;AAAA,CACrH;AAED,SAAS,eAAe,CACvB,eAAqC,EACrC,UAAkB,EAClB,aAAiC,EACR;IACzB,MAAM,IAAI,GAAG;QACZ,KAAK,EAAE,eAAe;QACtB,UAAU;QACV,sBAAsB,EAAE,IAAI;QAC5B,mBAAmB,EAAE,yBAAyB,CAAC,aAAa,CAAC;QAC7D,GAAG,CAAC,aAAa,KAAK,SAAS,IAAI,aAAa,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,aAAa,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC9E,CAAC;IACF,QAAQ,eAAe,EAAE,CAAC;QACzB,KAAK,MAAM;YACV,OAAO,IAAI,CAAC;QACb,KAAK,MAAM;YACV,OAAO,EAAE,GAAG,IAAI,EAAE,gBAAgB,EAAE,mCAAmC,EAAE,CAAC;QAC3E,KAAK,SAAS;YACb,OAAO;gBACN,GAAG,IAAI;gBACP,gBAAgB,EAAE,sCAAsC;gBACxD,sBAAsB,EAAE,KAAK;aAC7B,CAAC;QACH,KAAK,MAAM;YACV,OAAO;gBACN,GAAG,IAAI;gBACP,gBAAgB,EAAE,mCAAmC;gBACrD,sBAAsB,EAAE,KAAK;aAC7B,CAAC;IACJ,CAAC;AAAA,CACD;AAED,MAAM,UAAU,4BAA4B,CAAC,IAG5C,EAA0B;IAC1B,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,IAAI,MAAM,CAAC;IACjC,MAAM,aAAa,GAClB,IAAI,CAAC,aAAa,KAAK,SAAS,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,aAAa,CAAC,IAAI,IAAI,CAAC,aAAa,GAAG,CAAC;QAChG,CAAC,CAAC,IAAI,CAAC,aAAa;QACpB,CAAC,CAAC,SAAS,CAAC;IACd,IAAI,IAAI,KAAK,KAAK,EAAE,CAAC;QACpB,OAAO,eAAe,CAAC,MAAM,EAAE,oBAAoB,EAAE,aAAa,CAAC,CAAC;IACrE,CAAC;IACD,IAAI,IAAI,KAAK,MAAM,EAAE,CAAC;QACrB,OAAO,eAAe,CAAC,IAAI,EAAE,mBAAmB,EAAE,aAAa,CAAC,CAAC;IAClE,CAAC;IAED,IAAI,aAAa,KAAK,SAAS,EAAE,CAAC;QACjC,6CAA6C;QAC7C,OAAO,eAAe,CAAC,MAAM,EAAE,iCAAiC,EAAE,SAAS,CAAC,CAAC;IAC9E,CAAC;IACD,IAAI,aAAa,IAAI,iCAAiC,EAAE,CAAC;QACxD,OAAO,eAAe,CAAC,MAAM,EAAE,sBAAsB,EAAE,aAAa,CAAC,CAAC;IACvE,CAAC;IACD,IAAI,aAAa,IAAI,iCAAiC,EAAE,CAAC;QACxD,OAAO,eAAe,CAAC,MAAM,EAAE,qBAAqB,EAAE,aAAa,CAAC,CAAC;IACtE,CAAC;IACD,IAAI,aAAa,IAAI,oCAAoC,EAAE,CAAC;QAC3D,OAAO,eAAe,CAAC,SAAS,EAAE,wBAAwB,EAAE,aAAa,CAAC,CAAC;IAC5E,CAAC;IACD,OAAO,eAAe,CAAC,MAAM,EAAE,0BAA0B,EAAE,aAAa,CAAC,CAAC;AAAA,CAC1E;AAED,6FAA6F;AAC7F,MAAM,UAAU,4BAA4B,CAAC,SAA4B,EAAE,OAA+B,EAAY;IACrH,IAAI,QAAQ,GAAG,CAAC,GAAG,SAAS,CAAC,CAAC;IAC9B,IAAI,OAAO,CAAC,gBAAgB,KAAK,SAAS,EAAE,CAAC;QAC5C,MAAM,OAAO,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,gBAAgB,CAAC,CAAC;QAClD,QAAQ,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;IACzD,CAAC;IACD,IAAI,OAAO,CAAC,gBAAgB,KAAK,SAAS,EAAE,CAAC;QAC5C,MAAM,OAAO,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,gBAAgB,CAAC,CAAC;QAClD,QAAQ,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;IAC1D,CAAC;IACD,OAAO,QAAQ,CAAC;AAAA,CAChB","sourcesContent":["/**\n * Model capability auto-detection: derive what the harness may load onto a model FROM the model's\n * own metadata (`Model.contextWindow`), so small open models (4k/8k/16k windows, sub-1B params)\n * can still hold a usable chat instead of drowning in tool schemas and background-lane prompts.\n *\n * Derivation is metadata-first; defaults apply only when the metadata is missing (unknown/zero\n * window keeps today's full behavior rather than guessing). Detection can be disabled or forced\n * per class via the `modelCapability.mode` setting.\n */\n\nexport type ModelCapabilityClass = \"full\" | \"lean\" | \"minimal\" | \"chat\";\n\nexport type ModelCapabilityMode = \"auto\" | \"off\" | ModelCapabilityClass;\n\nexport interface ModelCapabilityProfile {\n\tclass: ModelCapabilityClass;\n\tcontextWindow?: number;\n\treasonCode: string;\n\t/** Allow-list; undefined = no allow-list restriction. */\n\tallowedToolNames?: readonly string[];\n\t/** Block-list applied after the allow-list; undefined = nothing blocked. */\n\tblockedToolNames?: readonly string[];\n\t/** Whether idle background lanes (goal auto-continue, research) may run on this model. */\n\tbackgroundLanesEnabled: boolean;\n\t/** Output-token cap for lane isolated completions, scaled to the window. */\n\tlaneMaxOutputTokens: number;\n}\n\n/** Windows at or above this keep the full harness surface. */\nexport const MODEL_CAPABILITY_FULL_MIN_CONTEXT = 32_768;\n/** Windows at or above this keep core tools but shed background-autonomy extras. */\nexport const MODEL_CAPABILITY_LEAN_MIN_CONTEXT = 16_384;\n/** Windows at or above this get the minimal coding set; below is chat-only. */\nexport const MODEL_CAPABILITY_MINIMAL_MIN_CONTEXT = 8_192;\n\nexport const MODEL_CAPABILITY_LEAN_BLOCKED_TOOLS: readonly string[] = [\"delegate\", \"context_audit\"];\nexport const MODEL_CAPABILITY_MINIMAL_ALLOWED_TOOLS: readonly string[] = [\n\t\"read\",\n\t\"bash\",\n\t\"edit\",\n\t\"write\",\n\t// The executor tool: minimal-class models ARE the daily-ops executors, and its schema is tiny.\n\t\"run_toolkit_script\",\n];\nexport const MODEL_CAPABILITY_CHAT_ALLOWED_TOOLS: readonly string[] = [];\n\nexport const DEFAULT_LANE_MAX_OUTPUT_TOKENS = 2048;\nconst MIN_LANE_MAX_OUTPUT_TOKENS = 256;\n\nfunction laneOutputTokensForWindow(contextWindow: number | undefined): number {\n\tif (contextWindow === undefined || contextWindow <= 0) return DEFAULT_LANE_MAX_OUTPUT_TOKENS;\n\t// A lane completion may use at most an eighth of the window for output, floored so tiny\n\t// windows still produce something parseable.\n\treturn Math.min(DEFAULT_LANE_MAX_OUTPUT_TOKENS, Math.max(MIN_LANE_MAX_OUTPUT_TOKENS, Math.floor(contextWindow / 8)));\n}\n\nfunction profileForClass(\n\tcapabilityClass: ModelCapabilityClass,\n\treasonCode: string,\n\tcontextWindow: number | undefined,\n): ModelCapabilityProfile {\n\tconst base = {\n\t\tclass: capabilityClass,\n\t\treasonCode,\n\t\tbackgroundLanesEnabled: true,\n\t\tlaneMaxOutputTokens: laneOutputTokensForWindow(contextWindow),\n\t\t...(contextWindow !== undefined && contextWindow > 0 ? { contextWindow } : {}),\n\t};\n\tswitch (capabilityClass) {\n\t\tcase \"full\":\n\t\t\treturn base;\n\t\tcase \"lean\":\n\t\t\treturn { ...base, blockedToolNames: MODEL_CAPABILITY_LEAN_BLOCKED_TOOLS };\n\t\tcase \"minimal\":\n\t\t\treturn {\n\t\t\t\t...base,\n\t\t\t\tallowedToolNames: MODEL_CAPABILITY_MINIMAL_ALLOWED_TOOLS,\n\t\t\t\tbackgroundLanesEnabled: false,\n\t\t\t};\n\t\tcase \"chat\":\n\t\t\treturn {\n\t\t\t\t...base,\n\t\t\t\tallowedToolNames: MODEL_CAPABILITY_CHAT_ALLOWED_TOOLS,\n\t\t\t\tbackgroundLanesEnabled: false,\n\t\t\t};\n\t}\n}\n\nexport function deriveModelCapabilityProfile(args: {\n\tcontextWindow?: number;\n\tmode?: ModelCapabilityMode;\n}): ModelCapabilityProfile {\n\tconst mode = args.mode ?? \"auto\";\n\tconst contextWindow =\n\t\targs.contextWindow !== undefined && Number.isFinite(args.contextWindow) && args.contextWindow > 0\n\t\t\t? args.contextWindow\n\t\t\t: undefined;\n\tif (mode === \"off\") {\n\t\treturn profileForClass(\"full\", \"detection_disabled\", contextWindow);\n\t}\n\tif (mode !== \"auto\") {\n\t\treturn profileForClass(mode, \"forced_by_setting\", contextWindow);\n\t}\n\n\tif (contextWindow === undefined) {\n\t\t// Metadata missing: defaults, never guesses.\n\t\treturn profileForClass(\"full\", \"unknown_context_window_defaults\", undefined);\n\t}\n\tif (contextWindow >= MODEL_CAPABILITY_FULL_MIN_CONTEXT) {\n\t\treturn profileForClass(\"full\", \"large_context_window\", contextWindow);\n\t}\n\tif (contextWindow >= MODEL_CAPABILITY_LEAN_MIN_CONTEXT) {\n\t\treturn profileForClass(\"lean\", \"lean_context_window\", contextWindow);\n\t}\n\tif (contextWindow >= MODEL_CAPABILITY_MINIMAL_MIN_CONTEXT) {\n\t\treturn profileForClass(\"minimal\", \"minimal_context_window\", contextWindow);\n\t}\n\treturn profileForClass(\"chat\", \"chat_only_context_window\", contextWindow);\n}\n\n/** Apply the profile's allow/block lists to a requested tool-name list, preserving order. */\nexport function filterToolNamesForCapability(toolNames: readonly string[], profile: ModelCapabilityProfile): string[] {\n\tlet filtered = [...toolNames];\n\tif (profile.allowedToolNames !== undefined) {\n\t\tconst allowed = new Set(profile.allowedToolNames);\n\t\tfiltered = filtered.filter((name) => allowed.has(name));\n\t}\n\tif (profile.blockedToolNames !== undefined) {\n\t\tconst blocked = new Set(profile.blockedToolNames);\n\t\tfiltered = filtered.filter((name) => !blocked.has(name));\n\t}\n\treturn filtered;\n}\n"]}

package/dist/core/model-router/config-diagnostics.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"config-diagnostics.d.ts","sourceRoot":"","sources":["../../../src/core/model-router/config-diagnostics.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,sBAAsB,CAAC;AAE1D,OAAO,KAAK,EAAE,yBAAyB,EAAE,MAAM,aAAa,CAAC;AAyB7D,wBAAgB,mCAAmC,CAClD,QAAQ,EAAE,yBAAyB,EACnC,aAAa,EAAE,aAAa,GAC1B,MAAM,EAAE,CAgBV","sourcesContent":["import type { Model } from \"@caupulican/pi-ai\";\nimport type { ModelRegistry } from \"../model-registry.ts\";\nimport { resolveCliModel } from \"../model-resolver.ts\";\nimport type { ModelRouterStatusSettings } from \"./status.ts\";\n\nfunction formatModel(model: Model<any>): string {\n\treturn `${model.provider}/${model.id}`;\n}\n\nfunction collectModelRouterModelDiagnostics(\n\tlabel: \"cheap model\" | \"expensive model\",\n\tsettingKey: \"modelRouter.cheapModel\" | \"modelRouter.expensiveModel\",\n\tmodelPattern: string | undefined,\n\tmodelRegistry: ModelRegistry,\n): string[] {\n\tif (!modelPattern) {\n\t\treturn [`Model router ${label} is unset; configure ${settingKey} or disable modelRouter.enabled.`];\n\t}\n\tconst resolved = resolveCliModel({ cliModel: modelPattern, modelRegistry });\n\tif (!resolved.model) {\n\t\treturn [`Model router ${label} is unresolved: ${modelPattern}.`];\n\t}\n\tif (!modelRegistry.hasConfiguredAuth(resolved.model)) {\n\t\treturn [`Model router ${label} is missing auth: ${formatModel(resolved.model)}.`];\n\t}\n\treturn [];\n}\n\nexport function collectModelRouterConfigDiagnostics(\n\tsettings: ModelRouterStatusSettings,\n\tmodelRegistry: ModelRegistry,\n): string[] {\n\tif (!settings.enabled) return [];\n\treturn [\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"cheap model\",\n\t\t\t\"modelRouter.cheapModel\",\n\t\t\tsettings.cheapModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"expensive model\",\n\t\t\t\"modelRouter.expensiveModel\",\n\t\t\tsettings.expensiveModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t];\n}\n"]}
+{"version":3,"file":"config-diagnostics.d.ts","sourceRoot":"","sources":["../../../src/core/model-router/config-diagnostics.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,sBAAsB,CAAC;AAE1D,OAAO,KAAK,EAAE,yBAAyB,EAAE,MAAM,aAAa,CAAC;AAyB7D,wBAAgB,mCAAmC,CAClD,QAAQ,EAAE,yBAAyB,EACnC,aAAa,EAAE,aAAa,GAC1B,MAAM,EAAE,CAsBV","sourcesContent":["import type { Model } from \"@caupulican/pi-ai\";\nimport type { ModelRegistry } from \"../model-registry.ts\";\nimport { resolveCliModel } from \"../model-resolver.ts\";\nimport type { ModelRouterStatusSettings } from \"./status.ts\";\n\nfunction formatModel(model: Model<any>): string {\n\treturn `${model.provider}/${model.id}`;\n}\n\nfunction collectModelRouterModelDiagnostics(\n\tlabel: \"cheap model\" | \"medium model\" | \"expensive model\",\n\tsettingKey: \"modelRouter.cheapModel\" | \"modelRouter.mediumModel\" | \"modelRouter.expensiveModel\",\n\tmodelPattern: string | undefined,\n\tmodelRegistry: ModelRegistry,\n): string[] {\n\tif (!modelPattern) {\n\t\treturn [`Model router ${label} is unset; configure ${settingKey} or disable modelRouter.enabled.`];\n\t}\n\tconst resolved = resolveCliModel({ cliModel: modelPattern, modelRegistry });\n\tif (!resolved.model) {\n\t\treturn [`Model router ${label} is unresolved: ${modelPattern}.`];\n\t}\n\tif (!modelRegistry.hasConfiguredAuth(resolved.model)) {\n\t\treturn [`Model router ${label} is missing auth: ${formatModel(resolved.model)}.`];\n\t}\n\treturn [];\n}\n\nexport function collectModelRouterConfigDiagnostics(\n\tsettings: ModelRouterStatusSettings,\n\tmodelRegistry: ModelRegistry,\n): string[] {\n\tif (!settings.enabled) return [];\n\treturn [\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"cheap model\",\n\t\t\t\"modelRouter.cheapModel\",\n\t\t\tsettings.cheapModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"medium model\",\n\t\t\t\"modelRouter.mediumModel\",\n\t\t\tsettings.mediumModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"expensive model\",\n\t\t\t\"modelRouter.expensiveModel\",\n\t\t\tsettings.expensiveModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t];\n}\n"]}

package/dist/core/model-router/config-diagnostics.js

@@ -20,6 +20,7 @@ export function collectModelRouterConfigDiagnostics(settings, modelRegistry) {
         return [];
     return [
         ...collectModelRouterModelDiagnostics("cheap model", "modelRouter.cheapModel", settings.cheapModel, modelRegistry),
+        ...collectModelRouterModelDiagnostics("medium model", "modelRouter.mediumModel", settings.mediumModel, modelRegistry),
         ...collectModelRouterModelDiagnostics("expensive model", "modelRouter.expensiveModel", settings.expensiveModel, modelRegistry),
     ];
 }

package/dist/core/model-router/config-diagnostics.js.map

@@ -1 +1 @@
-{"version":3,"file":"config-diagnostics.js","sourceRoot":"","sources":["../../../src/core/model-router/config-diagnostics.ts"],"names":[],"mappings":"AAEA,OAAO,EAAE,eAAe,EAAE,MAAM,sBAAsB,CAAC;AAGvD,SAAS,WAAW,CAAC,KAAiB,EAAU;IAC/C,OAAO,GAAG,KAAK,CAAC,QAAQ,IAAI,KAAK,CAAC,EAAE,EAAE,CAAC;AAAA,CACvC;AAED,SAAS,kCAAkC,CAC1C,KAAwC,EACxC,UAAmE,EACnE,YAAgC,EAChC,aAA4B,EACjB;IACX,IAAI,CAAC,YAAY,EAAE,CAAC;QACnB,OAAO,CAAC,gBAAgB,KAAK,wBAAwB,UAAU,kCAAkC,CAAC,CAAC;IACpG,CAAC;IACD,MAAM,QAAQ,GAAG,eAAe,CAAC,EAAE,QAAQ,EAAE,YAAY,EAAE,aAAa,EAAE,CAAC,CAAC;IAC5E,IAAI,CAAC,QAAQ,CAAC,KAAK,EAAE,CAAC;QACrB,OAAO,CAAC,gBAAgB,KAAK,mBAAmB,YAAY,GAAG,CAAC,CAAC;IAClE,CAAC;IACD,IAAI,CAAC,aAAa,CAAC,iBAAiB,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC;QACtD,OAAO,CAAC,gBAAgB,KAAK,qBAAqB,WAAW,CAAC,QAAQ,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IACnF,CAAC;IACD,OAAO,EAAE,CAAC;AAAA,CACV;AAED,MAAM,UAAU,mCAAmC,CAClD,QAAmC,EACnC,aAA4B,EACjB;IACX,IAAI,CAAC,QAAQ,CAAC,OAAO;QAAE,OAAO,EAAE,CAAC;IACjC,OAAO;QACN,GAAG,kCAAkC,CACpC,aAAa,EACb,wBAAwB,EACxB,QAAQ,CAAC,UAAU,EACnB,aAAa,CACb;QACD,GAAG,kCAAkC,CACpC,iBAAiB,EACjB,4BAA4B,EAC5B,QAAQ,CAAC,cAAc,EACvB,aAAa,CACb;KACD,CAAC;AAAA,CACF","sourcesContent":["import type { Model } from \"@caupulican/pi-ai\";\nimport type { ModelRegistry } from \"../model-registry.ts\";\nimport { resolveCliModel } from \"../model-resolver.ts\";\nimport type { ModelRouterStatusSettings } from \"./status.ts\";\n\nfunction formatModel(model: Model<any>): string {\n\treturn `${model.provider}/${model.id}`;\n}\n\nfunction collectModelRouterModelDiagnostics(\n\tlabel: \"cheap model\" | \"expensive model\",\n\tsettingKey: \"modelRouter.cheapModel\" | \"modelRouter.expensiveModel\",\n\tmodelPattern: string | undefined,\n\tmodelRegistry: ModelRegistry,\n): string[] {\n\tif (!modelPattern) {\n\t\treturn [`Model router ${label} is unset; configure ${settingKey} or disable modelRouter.enabled.`];\n\t}\n\tconst resolved = resolveCliModel({ cliModel: modelPattern, modelRegistry });\n\tif (!resolved.model) {\n\t\treturn [`Model router ${label} is unresolved: ${modelPattern}.`];\n\t}\n\tif (!modelRegistry.hasConfiguredAuth(resolved.model)) {\n\t\treturn [`Model router ${label} is missing auth: ${formatModel(resolved.model)}.`];\n\t}\n\treturn [];\n}\n\nexport function collectModelRouterConfigDiagnostics(\n\tsettings: ModelRouterStatusSettings,\n\tmodelRegistry: ModelRegistry,\n): string[] {\n\tif (!settings.enabled) return [];\n\treturn [\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"cheap model\",\n\t\t\t\"modelRouter.cheapModel\",\n\t\t\tsettings.cheapModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"expensive model\",\n\t\t\t\"modelRouter.expensiveModel\",\n\t\t\tsettings.expensiveModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t];\n}\n"]}
+{"version":3,"file":"config-diagnostics.js","sourceRoot":"","sources":["../../../src/core/model-router/config-diagnostics.ts"],"names":[],"mappings":"AAEA,OAAO,EAAE,eAAe,EAAE,MAAM,sBAAsB,CAAC;AAGvD,SAAS,WAAW,CAAC,KAAiB,EAAU;IAC/C,OAAO,GAAG,KAAK,CAAC,QAAQ,IAAI,KAAK,CAAC,EAAE,EAAE,CAAC;AAAA,CACvC;AAED,SAAS,kCAAkC,CAC1C,KAAyD,EACzD,UAA+F,EAC/F,YAAgC,EAChC,aAA4B,EACjB;IACX,IAAI,CAAC,YAAY,EAAE,CAAC;QACnB,OAAO,CAAC,gBAAgB,KAAK,wBAAwB,UAAU,kCAAkC,CAAC,CAAC;IACpG,CAAC;IACD,MAAM,QAAQ,GAAG,eAAe,CAAC,EAAE,QAAQ,EAAE,YAAY,EAAE,aAAa,EAAE,CAAC,CAAC;IAC5E,IAAI,CAAC,QAAQ,CAAC,KAAK,EAAE,CAAC;QACrB,OAAO,CAAC,gBAAgB,KAAK,mBAAmB,YAAY,GAAG,CAAC,CAAC;IAClE,CAAC;IACD,IAAI,CAAC,aAAa,CAAC,iBAAiB,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC;QACtD,OAAO,CAAC,gBAAgB,KAAK,qBAAqB,WAAW,CAAC,QAAQ,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IACnF,CAAC;IACD,OAAO,EAAE,CAAC;AAAA,CACV;AAED,MAAM,UAAU,mCAAmC,CAClD,QAAmC,EACnC,aAA4B,EACjB;IACX,IAAI,CAAC,QAAQ,CAAC,OAAO;QAAE,OAAO,EAAE,CAAC;IACjC,OAAO;QACN,GAAG,kCAAkC,CACpC,aAAa,EACb,wBAAwB,EACxB,QAAQ,CAAC,UAAU,EACnB,aAAa,CACb;QACD,GAAG,kCAAkC,CACpC,cAAc,EACd,yBAAyB,EACzB,QAAQ,CAAC,WAAW,EACpB,aAAa,CACb;QACD,GAAG,kCAAkC,CACpC,iBAAiB,EACjB,4BAA4B,EAC5B,QAAQ,CAAC,cAAc,EACvB,aAAa,CACb;KACD,CAAC;AAAA,CACF","sourcesContent":["import type { Model } from \"@caupulican/pi-ai\";\nimport type { ModelRegistry } from \"../model-registry.ts\";\nimport { resolveCliModel } from \"../model-resolver.ts\";\nimport type { ModelRouterStatusSettings } from \"./status.ts\";\n\nfunction formatModel(model: Model<any>): string {\n\treturn `${model.provider}/${model.id}`;\n}\n\nfunction collectModelRouterModelDiagnostics(\n\tlabel: \"cheap model\" | \"medium model\" | \"expensive model\",\n\tsettingKey: \"modelRouter.cheapModel\" | \"modelRouter.mediumModel\" | \"modelRouter.expensiveModel\",\n\tmodelPattern: string | undefined,\n\tmodelRegistry: ModelRegistry,\n): string[] {\n\tif (!modelPattern) {\n\t\treturn [`Model router ${label} is unset; configure ${settingKey} or disable modelRouter.enabled.`];\n\t}\n\tconst resolved = resolveCliModel({ cliModel: modelPattern, modelRegistry });\n\tif (!resolved.model) {\n\t\treturn [`Model router ${label} is unresolved: ${modelPattern}.`];\n\t}\n\tif (!modelRegistry.hasConfiguredAuth(resolved.model)) {\n\t\treturn [`Model router ${label} is missing auth: ${formatModel(resolved.model)}.`];\n\t}\n\treturn [];\n}\n\nexport function collectModelRouterConfigDiagnostics(\n\tsettings: ModelRouterStatusSettings,\n\tmodelRegistry: ModelRegistry,\n): string[] {\n\tif (!settings.enabled) return [];\n\treturn [\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"cheap model\",\n\t\t\t\"modelRouter.cheapModel\",\n\t\t\tsettings.cheapModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"medium model\",\n\t\t\t\"modelRouter.mediumModel\",\n\t\t\tsettings.mediumModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t\t...collectModelRouterModelDiagnostics(\n\t\t\t\"expensive model\",\n\t\t\t\"modelRouter.expensiveModel\",\n\t\t\tsettings.expensiveModel,\n\t\t\tmodelRegistry,\n\t\t),\n\t];\n}\n"]}

package/dist/core/model-router/intent-classifier.d.ts

@@ -1,3 +1,5 @@
+import type { RouteDecision } from "../autonomy/contracts.ts";
 export type ModelRouterIntent = "research" | "modify";
+export declare function classifyModelRouterRoute(prompt: string): RouteDecision;
 export declare function classifyModelRouterIntent(prompt: string): ModelRouterIntent;
 //# sourceMappingURL=intent-classifier.d.ts.map

package/dist/core/model-router/intent-classifier.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"intent-classifier.d.ts","sourceRoot":"","sources":["../../../src/core/model-router/intent-classifier.ts"],"names":[],"mappings":"AAAA,MAAM,MAAM,iBAAiB,GAAG,UAAU,GAAG,QAAQ,CAAC;AAUtD,wBAAgB,yBAAyB,CAAC,MAAM,EAAE,MAAM,GAAG,iBAAiB,CAK3E","sourcesContent":["export type ModelRouterIntent = \"research\" | \"modify\";\n\nconst MODIFY_INTENT_RE =\n\t/\\b(add|apply|build|change|commit|create|delete|edit|fix|generate|implement|install|modify|patch|refactor|remove|rename|replace|run|test|update|write)\\b|\\b(npm|pnpm|yarn|bun|git|gh|pytest|vitest|tsc|biome)\\b|[;&|]\\s*\\w+/i;\n\nconst EXPLICIT_MODIFY_REQUEST_RE =\n\t/^(?:can you|could you|please|pls|go ahead and|let'?s|i need you to|we need to|you should)\\s+.*\\b(add|apply|build|change|commit|create|delete|edit|fix|generate|implement|install|modify|patch|refactor|remove|rename|replace|run|test|update|write)\\b/i;\n\nconst READ_ONLY_QUESTION_RE = /^(?:how|what|why|when|where|which|who|explain|summarize|compare|describe|list|show)\\b/i;\n\nexport function classifyModelRouterIntent(prompt: string): ModelRouterIntent {\n\tconst text = prompt.trim();\n\tif (EXPLICIT_MODIFY_REQUEST_RE.test(text)) return \"modify\";\n\tif (READ_ONLY_QUESTION_RE.test(text)) return \"research\";\n\treturn MODIFY_INTENT_RE.test(text) ? \"modify\" : \"research\";\n}\n"]}
+{"version":3,"file":"intent-classifier.d.ts","sourceRoot":"","sources":["../../../src/core/model-router/intent-classifier.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,0BAA0B,CAAC;AAE9D,MAAM,MAAM,iBAAiB,GAAG,UAAU,GAAG,QAAQ,CAAC;AAiCtD,wBAAgB,wBAAwB,CAAC,MAAM,EAAE,MAAM,GAAG,aAAa,CA0ItE;AAED,wBAAgB,yBAAyB,CAAC,MAAM,EAAE,MAAM,GAAG,iBAAiB,CAG3E","sourcesContent":["import type { RouteDecision } from \"../autonomy/contracts.ts\";\n\nexport type ModelRouterIntent = \"research\" | \"modify\";\n\nconst EXPLICIT_MODIFY_REQUEST_RE =\n\t/^(?:can you|could you|please|pls|go ahead and|let'?s|i need you to|we need to|you should)\\s+.*\\b(add|apply|build|change|commit|create|delete|edit|fix|generate|implement|install|modify|patch|refactor|remove|rename|replace|run|test|update|write|publish|release|push|deploy|tag|reset|clean|rewrite)\\b/i;\n\nconst READ_ONLY_QUESTION_RE =\n\t/^(?:(?:can you|could you|please|pls|go ahead and|let'?s|i need you to|we need to|you should)\\s+)?(?:how|what|why|when|where|which|who|explain|summarize|compare|describe|list|show|search|find|view|read|locate)\\b/i;\n\nconst RELEASE_PUBLISH_RE = /\\b(publish|release|push|deploy|tag)\\b/i;\nconst SECURITY_AUTH_RE = /\\b(auth|token|credential|credentials|secret|api[-_]key)\\b/i;\nconst DESTRUCTIVE_RE = /\\b(delete|reset|rm\\s+-rf|clean)\\b/i;\n\nconst SELF_MOD_MUTATE_RE =\n\t/\\b(modify|change|write|update|edit|delete|add|remove)\\s+.*\\b(skills|prompts|settings|tools|behavior)\\b|self[-_]modification/i;\nconst ARCHITECTURE_MUTATE_RE = /\\b(rewrite|redesign|change|modify|rearchitect)\\s+.*\\b(architecture|architect)\\b/i;\n\n// Planning floor: plans steer all downstream work, so planning never routes cheap by default.\n// Only the route judge may downgrade a planning prompt back to cheap (explicit trivial verdict).\n// Core terms are always planning; design/architecture words count only with prospective phrasing,\n// so lookups like \"show me the architecture\" stay cheap.\nconst PLANNING_CORE_RE = /\\b(plan|planning|roadmap|strategy)\\b/i;\nconst PLANNING_DESIGN_WORD_RE = /\\b(design|architect\\w*|structure|approach)\\b/i;\nconst PLANNING_PROSPECTIVE_RE =\n\t/\\b(how (?:should|would|do we|can we)|what(?:'s| is) the best|propose|draft|come up with|figure out|decide (?:on|how))\\b/i;\n\nfunction isPlanningPrompt(text: string): boolean {\n\treturn PLANNING_CORE_RE.test(text) || (PLANNING_DESIGN_WORD_RE.test(text) && PLANNING_PROSPECTIVE_RE.test(text));\n}\n\nconst REFACTOR_RE = /\\b(refactor|refactoring)\\b/i;\nconst TEST_VALIDATION_RE = /\\b(test|testing|validation|lint|vitest|jest|run)\\b/i;\nconst IMPLEMENT_RE = /\\b(implement|fix|apply|change|update|create|write|generate|modify|edit|patch|add)\\b/i;\n\nexport function classifyModelRouterRoute(prompt: string): RouteDecision {\n\tconst text = prompt.trim();\n\n\tif (text.length === 0) {\n\t\treturn {\n\t\t\ttier: \"cheap\",\n\t\t\trisk: \"read-only\",\n\t\t\tconfidence: 0.1,\n\t\t\treasonCode: \"empty_prompt\",\n\t\t\treasons: [\"Empty or whitespace prompt\"],\n\t\t};\n\t}\n\n\t// 1. Explicit read-only questions/lookups dominate (unless prefixed by explicit mutation verb).\n\t// Planning-shaped questions are the exception: a plan steers expensive downstream work, so the\n\t// floor is medium even when phrased as a question.\n\tif (READ_ONLY_QUESTION_RE.test(text) && !EXPLICIT_MODIFY_REQUEST_RE.test(text)) {\n\t\tif (isPlanningPrompt(text)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"read-only\",\n\t\t\t\tconfidence: 0.75,\n\t\t\t\treasonCode: \"planning_min_medium\",\n\t\t\t\treasons: [\"Planning/design prompts never route cheap by default; a judge may deem them trivial\"],\n\t\t\t};\n\t\t}\n\t\treturn {\n\t\t\ttier: \"cheap\",\n\t\t\trisk: \"read-only\",\n\t\t\tconfidence: 0.9,\n\t\t\treasonCode: \"read_only_question\",\n\t\t\treasons: [\"Prompt asks a question or requests an explanation, search, or lookup\"],\n\t\t};\n\t}\n\n\t// Helper function to match patterns and return appropriate decision\n\tfunction matchKeywords(input: string): RouteDecision | null {\n\t\t// A. High-risk / approval-required/expensive signals\n\t\tif (RELEASE_PUBLISH_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"approval-required\",\n\t\t\t\tconfidence: 0.9,\n\t\t\t\treasonCode: \"release_or_publish\",\n\t\t\t\treasons: [\"Prompt mentions publishing, releasing, pushing, or deploying\"],\n\t\t\t};\n\t\t}\n\t\tif (SECURITY_AUTH_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"high-impact\",\n\t\t\t\tconfidence: 0.95,\n\t\t\t\treasonCode: \"security_or_auth\",\n\t\t\t\treasons: [\"Prompt mentions credentials, authentication, tokens, or secrets\"],\n\t\t\t};\n\t\t}\n\t\tif (DESTRUCTIVE_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"approval-required\",\n\t\t\t\tconfidence: 0.85,\n\t\t\t\treasonCode: \"destructive_or_git_history\",\n\t\t\t\treasons: [\"Prompt mentions deleting, resetting, cleaning, or destructive operations\"],\n\t\t\t};\n\t\t}\n\t\tif (SELF_MOD_MUTATE_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"approval-required\",\n\t\t\t\tconfidence: 0.9,\n\t\t\t\treasonCode: \"settings_or_self_modification\",\n\t\t\t\treasons: [\"Prompt mentions modifying skills, prompts, settings, tools, or self-modification\"],\n\t\t\t};\n\t\t}\n\t\tif (ARCHITECTURE_MUTATE_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"high-impact\",\n\t\t\t\tconfidence: 0.9,\n\t\t\t\treasonCode: \"architecture_or_ambiguous\",\n\t\t\t\treasons: [\"Prompt mentions core architecture or rewrite\"],\n\t\t\t};\n\t\t}\n\n\t\t// B. Explicit implementation/scoped-write signals route medium\n\t\tif (isPlanningPrompt(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"read-only\",\n\t\t\t\tconfidence: 0.75,\n\t\t\t\treasonCode: \"planning_min_medium\",\n\t\t\t\treasons: [\"Planning/design prompts never route cheap by default; a judge may deem them trivial\"],\n\t\t\t};\n\t\t}\n\t\tif (REFACTOR_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"scoped-write\",\n\t\t\t\tconfidence: 0.8,\n\t\t\t\treasonCode: \"mechanical_refactor\",\n\t\t\t\treasons: [\"Prompt mentions refactoring code structure\"],\n\t\t\t};\n\t\t}\n\t\tif (TEST_VALIDATION_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"scoped-write\",\n\t\t\t\tconfidence: 0.8,\n\t\t\t\treasonCode: \"test_or_validation\",\n\t\t\t\treasons: [\"Prompt mentions testing, validation, or linting\"],\n\t\t\t};\n\t\t}\n\t\tif (IMPLEMENT_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"scoped-write\",\n\t\t\t\tconfidence: 0.85,\n\t\t\t\treasonCode: \"normal_implementation\",\n\t\t\t\treasons: [\"Prompt mentions implementing, updating, creating, or modifying code\"],\n\t\t\t};\n\t\t}\n\n\t\treturn null;\n\t}\n\n\tconst match = matchKeywords(text);\n\tif (match) {\n\t\treturn match;\n\t}\n\n\t// 4. Default fallbacks\n\treturn {\n\t\ttier: \"cheap\",\n\t\trisk: \"read-only\",\n\t\tconfidence: 0.5,\n\t\treasonCode: \"default_read_only\",\n\t\treasons: [\"No explicit implementation, destructive, or release patterns detected\"],\n\t};\n}\n\nexport function classifyModelRouterIntent(prompt: string): ModelRouterIntent {\n\tconst decision = classifyModelRouterRoute(prompt);\n\treturn decision.tier === \"cheap\" ? \"research\" : \"modify\";\n}\n"]}

package/dist/core/model-router/intent-classifier.js

@@ -1,12 +1,157 @@
-const MODIFY_INTENT_RE = /\b(add|apply|build|change|commit|create|delete|edit|fix|generate|implement|install|modify|patch|refactor|remove|rename|replace|run|test|update|write)\b|\b(npm|pnpm|yarn|bun|git|gh|pytest|vitest|tsc|biome)\b|[;&|]\s*\w+/i;
-const EXPLICIT_MODIFY_REQUEST_RE = /^(?:can you|could you|please|pls|go ahead and|let'?s|i need you to|we need to|you should)\s+.*\b(add|apply|build|change|commit|create|delete|edit|fix|generate|implement|install|modify|patch|refactor|remove|rename|replace|run|test|update|write)\b/i;
-const READ_ONLY_QUESTION_RE = /^(?:how|what|why|when|where|which|who|explain|summarize|compare|describe|list|show)\b/i;
-export function classifyModelRouterIntent(prompt) {
+const EXPLICIT_MODIFY_REQUEST_RE = /^(?:can you|could you|please|pls|go ahead and|let'?s|i need you to|we need to|you should)\s+.*\b(add|apply|build|change|commit|create|delete|edit|fix|generate|implement|install|modify|patch|refactor|remove|rename|replace|run|test|update|write|publish|release|push|deploy|tag|reset|clean|rewrite)\b/i;
+const READ_ONLY_QUESTION_RE = /^(?:(?:can you|could you|please|pls|go ahead and|let'?s|i need you to|we need to|you should)\s+)?(?:how|what|why|when|where|which|who|explain|summarize|compare|describe|list|show|search|find|view|read|locate)\b/i;
+const RELEASE_PUBLISH_RE = /\b(publish|release|push|deploy|tag)\b/i;
+const SECURITY_AUTH_RE = /\b(auth|token|credential|credentials|secret|api[-_]key)\b/i;
+const DESTRUCTIVE_RE = /\b(delete|reset|rm\s+-rf|clean)\b/i;
+const SELF_MOD_MUTATE_RE = /\b(modify|change|write|update|edit|delete|add|remove)\s+.*\b(skills|prompts|settings|tools|behavior)\b|self[-_]modification/i;
+const ARCHITECTURE_MUTATE_RE = /\b(rewrite|redesign|change|modify|rearchitect)\s+.*\b(architecture|architect)\b/i;
+// Planning floor: plans steer all downstream work, so planning never routes cheap by default.
+// Only the route judge may downgrade a planning prompt back to cheap (explicit trivial verdict).
+// Core terms are always planning; design/architecture words count only with prospective phrasing,
+// so lookups like "show me the architecture" stay cheap.
+const PLANNING_CORE_RE = /\b(plan|planning|roadmap|strategy)\b/i;
+const PLANNING_DESIGN_WORD_RE = /\b(design|architect\w*|structure|approach)\b/i;
+const PLANNING_PROSPECTIVE_RE = /\b(how (?:should|would|do we|can we)|what(?:'s| is) the best|propose|draft|come up with|figure out|decide (?:on|how))\b/i;
+function isPlanningPrompt(text) {
+    return PLANNING_CORE_RE.test(text) || (PLANNING_DESIGN_WORD_RE.test(text) && PLANNING_PROSPECTIVE_RE.test(text));
+}
+const REFACTOR_RE = /\b(refactor|refactoring)\b/i;
+const TEST_VALIDATION_RE = /\b(test|testing|validation|lint|vitest|jest|run)\b/i;
+const IMPLEMENT_RE = /\b(implement|fix|apply|change|update|create|write|generate|modify|edit|patch|add)\b/i;
+export function classifyModelRouterRoute(prompt) {
     const text = prompt.trim();
-    if (EXPLICIT_MODIFY_REQUEST_RE.test(text))
-        return "modify";
-    if (READ_ONLY_QUESTION_RE.test(text))
-        return "research";
-    return MODIFY_INTENT_RE.test(text) ? "modify" : "research";
+    if (text.length === 0) {
+        return {
+            tier: "cheap",
+            risk: "read-only",
+            confidence: 0.1,
+            reasonCode: "empty_prompt",
+            reasons: ["Empty or whitespace prompt"],
+        };
+    }
+    // 1. Explicit read-only questions/lookups dominate (unless prefixed by explicit mutation verb).
+    // Planning-shaped questions are the exception: a plan steers expensive downstream work, so the
+    // floor is medium even when phrased as a question.
+    if (READ_ONLY_QUESTION_RE.test(text) && !EXPLICIT_MODIFY_REQUEST_RE.test(text)) {
+        if (isPlanningPrompt(text)) {
+            return {
+                tier: "medium",
+                risk: "read-only",
+                confidence: 0.75,
+                reasonCode: "planning_min_medium",
+                reasons: ["Planning/design prompts never route cheap by default; a judge may deem them trivial"],
+            };
+        }
+        return {
+            tier: "cheap",
+            risk: "read-only",
+            confidence: 0.9,
+            reasonCode: "read_only_question",
+            reasons: ["Prompt asks a question or requests an explanation, search, or lookup"],
+        };
+    }
+    // Helper function to match patterns and return appropriate decision
+    function matchKeywords(input) {
+        // A. High-risk / approval-required/expensive signals
+        if (RELEASE_PUBLISH_RE.test(input)) {
+            return {
+                tier: "expensive",
+                risk: "approval-required",
+                confidence: 0.9,
+                reasonCode: "release_or_publish",
+                reasons: ["Prompt mentions publishing, releasing, pushing, or deploying"],
+            };
+        }
+        if (SECURITY_AUTH_RE.test(input)) {
+            return {
+                tier: "expensive",
+                risk: "high-impact",
+                confidence: 0.95,
+                reasonCode: "security_or_auth",
+                reasons: ["Prompt mentions credentials, authentication, tokens, or secrets"],
+            };
+        }
+        if (DESTRUCTIVE_RE.test(input)) {
+            return {
+                tier: "expensive",
+                risk: "approval-required",
+                confidence: 0.85,
+                reasonCode: "destructive_or_git_history",
+                reasons: ["Prompt mentions deleting, resetting, cleaning, or destructive operations"],
+            };
+        }
+        if (SELF_MOD_MUTATE_RE.test(input)) {
+            return {
+                tier: "expensive",
+                risk: "approval-required",
+                confidence: 0.9,
+                reasonCode: "settings_or_self_modification",
+                reasons: ["Prompt mentions modifying skills, prompts, settings, tools, or self-modification"],
+            };
+        }
+        if (ARCHITECTURE_MUTATE_RE.test(input)) {
+            return {
+                tier: "expensive",
+                risk: "high-impact",
+                confidence: 0.9,
+                reasonCode: "architecture_or_ambiguous",
+                reasons: ["Prompt mentions core architecture or rewrite"],
+            };
+        }
+        // B. Explicit implementation/scoped-write signals route medium
+        if (isPlanningPrompt(input)) {
+            return {
+                tier: "medium",
+                risk: "read-only",
+                confidence: 0.75,
+                reasonCode: "planning_min_medium",
+                reasons: ["Planning/design prompts never route cheap by default; a judge may deem them trivial"],
+            };
+        }
+        if (REFACTOR_RE.test(input)) {
+            return {
+                tier: "medium",
+                risk: "scoped-write",
+                confidence: 0.8,
+                reasonCode: "mechanical_refactor",
+                reasons: ["Prompt mentions refactoring code structure"],
+            };
+        }
+        if (TEST_VALIDATION_RE.test(input)) {
+            return {
+                tier: "medium",
+                risk: "scoped-write",
+                confidence: 0.8,
+                reasonCode: "test_or_validation",
+                reasons: ["Prompt mentions testing, validation, or linting"],
+            };
+        }
+        if (IMPLEMENT_RE.test(input)) {
+            return {
+                tier: "medium",
+                risk: "scoped-write",
+                confidence: 0.85,
+                reasonCode: "normal_implementation",
+                reasons: ["Prompt mentions implementing, updating, creating, or modifying code"],
+            };
+        }
+        return null;
+    }
+    const match = matchKeywords(text);
+    if (match) {
+        return match;
+    }
+    // 4. Default fallbacks
+    return {
+        tier: "cheap",
+        risk: "read-only",
+        confidence: 0.5,
+        reasonCode: "default_read_only",
+        reasons: ["No explicit implementation, destructive, or release patterns detected"],
+    };
+}
+export function classifyModelRouterIntent(prompt) {
+    const decision = classifyModelRouterRoute(prompt);
+    return decision.tier === "cheap" ? "research" : "modify";
 }
 //# sourceMappingURL=intent-classifier.js.map

package/dist/core/model-router/intent-classifier.js.map

@@ -1 +1 @@
-{"version":3,"file":"intent-classifier.js","sourceRoot":"","sources":["../../../src/core/model-router/intent-classifier.ts"],"names":[],"mappings":"AAEA,MAAM,gBAAgB,GACrB,6NAA6N,CAAC;AAE/N,MAAM,0BAA0B,GAC/B,wPAAwP,CAAC;AAE1P,MAAM,qBAAqB,GAAG,wFAAwF,CAAC;AAEvH,MAAM,UAAU,yBAAyB,CAAC,MAAc,EAAqB;IAC5E,MAAM,IAAI,GAAG,MAAM,CAAC,IAAI,EAAE,CAAC;IAC3B,IAAI,0BAA0B,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,QAAQ,CAAC;IAC3D,IAAI,qBAAqB,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,UAAU,CAAC;IACxD,OAAO,gBAAgB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,UAAU,CAAC;AAAA,CAC3D","sourcesContent":["export type ModelRouterIntent = \"research\" | \"modify\";\n\nconst MODIFY_INTENT_RE =\n\t/\\b(add|apply|build|change|commit|create|delete|edit|fix|generate|implement|install|modify|patch|refactor|remove|rename|replace|run|test|update|write)\\b|\\b(npm|pnpm|yarn|bun|git|gh|pytest|vitest|tsc|biome)\\b|[;&|]\\s*\\w+/i;\n\nconst EXPLICIT_MODIFY_REQUEST_RE =\n\t/^(?:can you|could you|please|pls|go ahead and|let'?s|i need you to|we need to|you should)\\s+.*\\b(add|apply|build|change|commit|create|delete|edit|fix|generate|implement|install|modify|patch|refactor|remove|rename|replace|run|test|update|write)\\b/i;\n\nconst READ_ONLY_QUESTION_RE = /^(?:how|what|why|when|where|which|who|explain|summarize|compare|describe|list|show)\\b/i;\n\nexport function classifyModelRouterIntent(prompt: string): ModelRouterIntent {\n\tconst text = prompt.trim();\n\tif (EXPLICIT_MODIFY_REQUEST_RE.test(text)) return \"modify\";\n\tif (READ_ONLY_QUESTION_RE.test(text)) return \"research\";\n\treturn MODIFY_INTENT_RE.test(text) ? \"modify\" : \"research\";\n}\n"]}
+{"version":3,"file":"intent-classifier.js","sourceRoot":"","sources":["../../../src/core/model-router/intent-classifier.ts"],"names":[],"mappings":"AAIA,MAAM,0BAA0B,GAC/B,4SAA4S,CAAC;AAE9S,MAAM,qBAAqB,GAC1B,qNAAqN,CAAC;AAEvN,MAAM,kBAAkB,GAAG,wCAAwC,CAAC;AACpE,MAAM,gBAAgB,GAAG,4DAA4D,CAAC;AACtF,MAAM,cAAc,GAAG,oCAAoC,CAAC;AAE5D,MAAM,kBAAkB,GACvB,8HAA8H,CAAC;AAChI,MAAM,sBAAsB,GAAG,kFAAkF,CAAC;AAElH,8FAA8F;AAC9F,iGAAiG;AACjG,kGAAkG;AAClG,yDAAyD;AACzD,MAAM,gBAAgB,GAAG,uCAAuC,CAAC;AACjE,MAAM,uBAAuB,GAAG,+CAA+C,CAAC;AAChF,MAAM,uBAAuB,GAC5B,0HAA0H,CAAC;AAE5H,SAAS,gBAAgB,CAAC,IAAY,EAAW;IAChD,OAAO,gBAAgB,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,uBAAuB,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,uBAAuB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC;AAAA,CACjH;AAED,MAAM,WAAW,GAAG,6BAA6B,CAAC;AAClD,MAAM,kBAAkB,GAAG,qDAAqD,CAAC;AACjF,MAAM,YAAY,GAAG,sFAAsF,CAAC;AAE5G,MAAM,UAAU,wBAAwB,CAAC,MAAc,EAAiB;IACvE,MAAM,IAAI,GAAG,MAAM,CAAC,IAAI,EAAE,CAAC;IAE3B,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACvB,OAAO;YACN,IAAI,EAAE,OAAO;YACb,IAAI,EAAE,WAAW;YACjB,UAAU,EAAE,GAAG;YACf,UAAU,EAAE,cAAc;YAC1B,OAAO,EAAE,CAAC,4BAA4B,CAAC;SACvC,CAAC;IACH,CAAC;IAED,gGAAgG;IAChG,+FAA+F;IAC/F,mDAAmD;IACnD,IAAI,qBAAqB,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,0BAA0B,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;QAChF,IAAI,gBAAgB,CAAC,IAAI,CAAC,EAAE,CAAC;YAC5B,OAAO;gBACN,IAAI,EAAE,QAAQ;gBACd,IAAI,EAAE,WAAW;gBACjB,UAAU,EAAE,IAAI;gBAChB,UAAU,EAAE,qBAAqB;gBACjC,OAAO,EAAE,CAAC,qFAAqF,CAAC;aAChG,CAAC;QACH,CAAC;QACD,OAAO;YACN,IAAI,EAAE,OAAO;YACb,IAAI,EAAE,WAAW;YACjB,UAAU,EAAE,GAAG;YACf,UAAU,EAAE,oBAAoB;YAChC,OAAO,EAAE,CAAC,sEAAsE,CAAC;SACjF,CAAC;IACH,CAAC;IAED,oEAAoE;IACpE,SAAS,aAAa,CAAC,KAAa,EAAwB;QAC3D,qDAAqD;QACrD,IAAI,kBAAkB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YACpC,OAAO;gBACN,IAAI,EAAE,WAAW;gBACjB,IAAI,EAAE,mBAAmB;gBACzB,UAAU,EAAE,GAAG;gBACf,UAAU,EAAE,oBAAoB;gBAChC,OAAO,EAAE,CAAC,8DAA8D,CAAC;aACzE,CAAC;QACH,CAAC;QACD,IAAI,gBAAgB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YAClC,OAAO;gBACN,IAAI,EAAE,WAAW;gBACjB,IAAI,EAAE,aAAa;gBACnB,UAAU,EAAE,IAAI;gBAChB,UAAU,EAAE,kBAAkB;gBAC9B,OAAO,EAAE,CAAC,iEAAiE,CAAC;aAC5E,CAAC;QACH,CAAC;QACD,IAAI,cAAc,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YAChC,OAAO;gBACN,IAAI,EAAE,WAAW;gBACjB,IAAI,EAAE,mBAAmB;gBACzB,UAAU,EAAE,IAAI;gBAChB,UAAU,EAAE,4BAA4B;gBACxC,OAAO,EAAE,CAAC,0EAA0E,CAAC;aACrF,CAAC;QACH,CAAC;QACD,IAAI,kBAAkB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YACpC,OAAO;gBACN,IAAI,EAAE,WAAW;gBACjB,IAAI,EAAE,mBAAmB;gBACzB,UAAU,EAAE,GAAG;gBACf,UAAU,EAAE,+BAA+B;gBAC3C,OAAO,EAAE,CAAC,kFAAkF,CAAC;aAC7F,CAAC;QACH,CAAC;QACD,IAAI,sBAAsB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YACxC,OAAO;gBACN,IAAI,EAAE,WAAW;gBACjB,IAAI,EAAE,aAAa;gBACnB,UAAU,EAAE,GAAG;gBACf,UAAU,EAAE,2BAA2B;gBACvC,OAAO,EAAE,CAAC,8CAA8C,CAAC;aACzD,CAAC;QACH,CAAC;QAED,+DAA+D;QAC/D,IAAI,gBAAgB,CAAC,KAAK,CAAC,EAAE,CAAC;YAC7B,OAAO;gBACN,IAAI,EAAE,QAAQ;gBACd,IAAI,EAAE,WAAW;gBACjB,UAAU,EAAE,IAAI;gBAChB,UAAU,EAAE,qBAAqB;gBACjC,OAAO,EAAE,CAAC,qFAAqF,CAAC;aAChG,CAAC;QACH,CAAC;QACD,IAAI,WAAW,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YAC7B,OAAO;gBACN,IAAI,EAAE,QAAQ;gBACd,IAAI,EAAE,cAAc;gBACpB,UAAU,EAAE,GAAG;gBACf,UAAU,EAAE,qBAAqB;gBACjC,OAAO,EAAE,CAAC,4CAA4C,CAAC;aACvD,CAAC;QACH,CAAC;QACD,IAAI,kBAAkB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YACpC,OAAO;gBACN,IAAI,EAAE,QAAQ;gBACd,IAAI,EAAE,cAAc;gBACpB,UAAU,EAAE,GAAG;gBACf,UAAU,EAAE,oBAAoB;gBAChC,OAAO,EAAE,CAAC,iDAAiD,CAAC;aAC5D,CAAC;QACH,CAAC;QACD,IAAI,YAAY,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YAC9B,OAAO;gBACN,IAAI,EAAE,QAAQ;gBACd,IAAI,EAAE,cAAc;gBACpB,UAAU,EAAE,IAAI;gBAChB,UAAU,EAAE,uBAAuB;gBACnC,OAAO,EAAE,CAAC,qEAAqE,CAAC;aAChF,CAAC;QACH,CAAC;QAED,OAAO,IAAI,CAAC;IAAA,CACZ;IAED,MAAM,KAAK,GAAG,aAAa,CAAC,IAAI,CAAC,CAAC;IAClC,IAAI,KAAK,EAAE,CAAC;QACX,OAAO,KAAK,CAAC;IACd,CAAC;IAED,uBAAuB;IACvB,OAAO;QACN,IAAI,EAAE,OAAO;QACb,IAAI,EAAE,WAAW;QACjB,UAAU,EAAE,GAAG;QACf,UAAU,EAAE,mBAAmB;QAC/B,OAAO,EAAE,CAAC,uEAAuE,CAAC;KAClF,CAAC;AAAA,CACF;AAED,MAAM,UAAU,yBAAyB,CAAC,MAAc,EAAqB;IAC5E,MAAM,QAAQ,GAAG,wBAAwB,CAAC,MAAM,CAAC,CAAC;IAClD,OAAO,QAAQ,CAAC,IAAI,KAAK,OAAO,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,QAAQ,CAAC;AAAA,CACzD","sourcesContent":["import type { RouteDecision } from \"../autonomy/contracts.ts\";\n\nexport type ModelRouterIntent = \"research\" | \"modify\";\n\nconst EXPLICIT_MODIFY_REQUEST_RE =\n\t/^(?:can you|could you|please|pls|go ahead and|let'?s|i need you to|we need to|you should)\\s+.*\\b(add|apply|build|change|commit|create|delete|edit|fix|generate|implement|install|modify|patch|refactor|remove|rename|replace|run|test|update|write|publish|release|push|deploy|tag|reset|clean|rewrite)\\b/i;\n\nconst READ_ONLY_QUESTION_RE =\n\t/^(?:(?:can you|could you|please|pls|go ahead and|let'?s|i need you to|we need to|you should)\\s+)?(?:how|what|why|when|where|which|who|explain|summarize|compare|describe|list|show|search|find|view|read|locate)\\b/i;\n\nconst RELEASE_PUBLISH_RE = /\\b(publish|release|push|deploy|tag)\\b/i;\nconst SECURITY_AUTH_RE = /\\b(auth|token|credential|credentials|secret|api[-_]key)\\b/i;\nconst DESTRUCTIVE_RE = /\\b(delete|reset|rm\\s+-rf|clean)\\b/i;\n\nconst SELF_MOD_MUTATE_RE =\n\t/\\b(modify|change|write|update|edit|delete|add|remove)\\s+.*\\b(skills|prompts|settings|tools|behavior)\\b|self[-_]modification/i;\nconst ARCHITECTURE_MUTATE_RE = /\\b(rewrite|redesign|change|modify|rearchitect)\\s+.*\\b(architecture|architect)\\b/i;\n\n// Planning floor: plans steer all downstream work, so planning never routes cheap by default.\n// Only the route judge may downgrade a planning prompt back to cheap (explicit trivial verdict).\n// Core terms are always planning; design/architecture words count only with prospective phrasing,\n// so lookups like \"show me the architecture\" stay cheap.\nconst PLANNING_CORE_RE = /\\b(plan|planning|roadmap|strategy)\\b/i;\nconst PLANNING_DESIGN_WORD_RE = /\\b(design|architect\\w*|structure|approach)\\b/i;\nconst PLANNING_PROSPECTIVE_RE =\n\t/\\b(how (?:should|would|do we|can we)|what(?:'s| is) the best|propose|draft|come up with|figure out|decide (?:on|how))\\b/i;\n\nfunction isPlanningPrompt(text: string): boolean {\n\treturn PLANNING_CORE_RE.test(text) || (PLANNING_DESIGN_WORD_RE.test(text) && PLANNING_PROSPECTIVE_RE.test(text));\n}\n\nconst REFACTOR_RE = /\\b(refactor|refactoring)\\b/i;\nconst TEST_VALIDATION_RE = /\\b(test|testing|validation|lint|vitest|jest|run)\\b/i;\nconst IMPLEMENT_RE = /\\b(implement|fix|apply|change|update|create|write|generate|modify|edit|patch|add)\\b/i;\n\nexport function classifyModelRouterRoute(prompt: string): RouteDecision {\n\tconst text = prompt.trim();\n\n\tif (text.length === 0) {\n\t\treturn {\n\t\t\ttier: \"cheap\",\n\t\t\trisk: \"read-only\",\n\t\t\tconfidence: 0.1,\n\t\t\treasonCode: \"empty_prompt\",\n\t\t\treasons: [\"Empty or whitespace prompt\"],\n\t\t};\n\t}\n\n\t// 1. Explicit read-only questions/lookups dominate (unless prefixed by explicit mutation verb).\n\t// Planning-shaped questions are the exception: a plan steers expensive downstream work, so the\n\t// floor is medium even when phrased as a question.\n\tif (READ_ONLY_QUESTION_RE.test(text) && !EXPLICIT_MODIFY_REQUEST_RE.test(text)) {\n\t\tif (isPlanningPrompt(text)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"read-only\",\n\t\t\t\tconfidence: 0.75,\n\t\t\t\treasonCode: \"planning_min_medium\",\n\t\t\t\treasons: [\"Planning/design prompts never route cheap by default; a judge may deem them trivial\"],\n\t\t\t};\n\t\t}\n\t\treturn {\n\t\t\ttier: \"cheap\",\n\t\t\trisk: \"read-only\",\n\t\t\tconfidence: 0.9,\n\t\t\treasonCode: \"read_only_question\",\n\t\t\treasons: [\"Prompt asks a question or requests an explanation, search, or lookup\"],\n\t\t};\n\t}\n\n\t// Helper function to match patterns and return appropriate decision\n\tfunction matchKeywords(input: string): RouteDecision | null {\n\t\t// A. High-risk / approval-required/expensive signals\n\t\tif (RELEASE_PUBLISH_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"approval-required\",\n\t\t\t\tconfidence: 0.9,\n\t\t\t\treasonCode: \"release_or_publish\",\n\t\t\t\treasons: [\"Prompt mentions publishing, releasing, pushing, or deploying\"],\n\t\t\t};\n\t\t}\n\t\tif (SECURITY_AUTH_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"high-impact\",\n\t\t\t\tconfidence: 0.95,\n\t\t\t\treasonCode: \"security_or_auth\",\n\t\t\t\treasons: [\"Prompt mentions credentials, authentication, tokens, or secrets\"],\n\t\t\t};\n\t\t}\n\t\tif (DESTRUCTIVE_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"approval-required\",\n\t\t\t\tconfidence: 0.85,\n\t\t\t\treasonCode: \"destructive_or_git_history\",\n\t\t\t\treasons: [\"Prompt mentions deleting, resetting, cleaning, or destructive operations\"],\n\t\t\t};\n\t\t}\n\t\tif (SELF_MOD_MUTATE_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"approval-required\",\n\t\t\t\tconfidence: 0.9,\n\t\t\t\treasonCode: \"settings_or_self_modification\",\n\t\t\t\treasons: [\"Prompt mentions modifying skills, prompts, settings, tools, or self-modification\"],\n\t\t\t};\n\t\t}\n\t\tif (ARCHITECTURE_MUTATE_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"expensive\",\n\t\t\t\trisk: \"high-impact\",\n\t\t\t\tconfidence: 0.9,\n\t\t\t\treasonCode: \"architecture_or_ambiguous\",\n\t\t\t\treasons: [\"Prompt mentions core architecture or rewrite\"],\n\t\t\t};\n\t\t}\n\n\t\t// B. Explicit implementation/scoped-write signals route medium\n\t\tif (isPlanningPrompt(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"read-only\",\n\t\t\t\tconfidence: 0.75,\n\t\t\t\treasonCode: \"planning_min_medium\",\n\t\t\t\treasons: [\"Planning/design prompts never route cheap by default; a judge may deem them trivial\"],\n\t\t\t};\n\t\t}\n\t\tif (REFACTOR_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"scoped-write\",\n\t\t\t\tconfidence: 0.8,\n\t\t\t\treasonCode: \"mechanical_refactor\",\n\t\t\t\treasons: [\"Prompt mentions refactoring code structure\"],\n\t\t\t};\n\t\t}\n\t\tif (TEST_VALIDATION_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"scoped-write\",\n\t\t\t\tconfidence: 0.8,\n\t\t\t\treasonCode: \"test_or_validation\",\n\t\t\t\treasons: [\"Prompt mentions testing, validation, or linting\"],\n\t\t\t};\n\t\t}\n\t\tif (IMPLEMENT_RE.test(input)) {\n\t\t\treturn {\n\t\t\t\ttier: \"medium\",\n\t\t\t\trisk: \"scoped-write\",\n\t\t\t\tconfidence: 0.85,\n\t\t\t\treasonCode: \"normal_implementation\",\n\t\t\t\treasons: [\"Prompt mentions implementing, updating, creating, or modifying code\"],\n\t\t\t};\n\t\t}\n\n\t\treturn null;\n\t}\n\n\tconst match = matchKeywords(text);\n\tif (match) {\n\t\treturn match;\n\t}\n\n\t// 4. Default fallbacks\n\treturn {\n\t\ttier: \"cheap\",\n\t\trisk: \"read-only\",\n\t\tconfidence: 0.5,\n\t\treasonCode: \"default_read_only\",\n\t\treasons: [\"No explicit implementation, destructive, or release patterns detected\"],\n\t};\n}\n\nexport function classifyModelRouterIntent(prompt: string): ModelRouterIntent {\n\tconst decision = classifyModelRouterRoute(prompt);\n\treturn decision.tier === \"cheap\" ? \"research\" : \"modify\";\n}\n"]}

package/dist/core/model-router/route-judge.d.ts

@@ -0,0 +1,54 @@
+import type { RouteDecision } from "../autonomy/contracts.ts";
+/**
+ * Routing-only judge: a bounded, tool-less completion (default: the medium model) that decides the
+ * final cheap/medium/expensive tier for a user prompt, refining the regex classifier's baseline.
+ * Core rule: planning is never cheap unless the judge explicitly deems the task trivial. The judge
+ * proposes; the existing pipeline (model resolution, auth, escalation, gates) still decides.
+ * Failure is honest: unparseable/timeout/unavailable falls back to the baseline with a visible
+ * reasonCode, never silently.
+ */
+/** Static across calls — callers pass cacheRetention "short" so only the variable tail is billed. */
+export declare const ROUTE_JUDGE_SYSTEM_PROMPT: string;
+export declare const ROUTE_JUDGE_MAX_OUTPUT_TOKENS = 128;
+export declare const ROUTE_JUDGE_MAX_WALL_CLOCK_MS = 10000;
+export interface RouteJudgeVerdict {
+    tier: "cheap" | "medium" | "expensive";
+    risk: RouteDecision["risk"];
+    trivial: boolean;
+    reason: string;
+}
+export declare function buildRouteJudgeUserPrompt(args: {
+    prompt: string;
+    baseline: RouteDecision;
+}): string;
+export declare function parseRouteJudgeVerdict(text: string): RouteJudgeVerdict | undefined;
+/** Merge a judge verdict into the baseline decision (pure; never returns learning). */
+export declare function applyRouteJudgeVerdict(baseline: RouteDecision, verdict: RouteJudgeVerdict): RouteDecision;
+export interface RouteJudgeRunResult {
+    decision: RouteDecision;
+    verdict?: RouteJudgeVerdict;
+    /** Set when the judge could not decide and the baseline was kept. */
+    fallbackReason?: string;
+    costUsd: number;
+}
+/**
+ * Run the judge over a baseline decision. The completion executor is injected (production:
+ * AgentSession.runIsolatedCompletion on the judge model). Never throws; every failure keeps the
+ * baseline with a visible fallbackReason.
+ */
+export declare function runRouteJudge(args: {
+    prompt: string;
+    baseline: RouteDecision;
+    complete: (input: {
+        systemPrompt: string;
+        userPrompt: string;
+        signal?: AbortSignal;
+    }) => Promise<{
+        text: string;
+        costUsd: number;
+        stopReason: string;
+    }>;
+    signal?: AbortSignal;
+    maxWallClockMs?: number;
+}): Promise<RouteJudgeRunResult>;
+//# sourceMappingURL=route-judge.d.ts.map

package/dist/core/model-router/route-judge.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"route-judge.d.ts","sourceRoot":"","sources":["../../../src/core/model-router/route-judge.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,0BAA0B,CAAC;AAE9D;;;;;;;GAOG;AAEH,uGAAqG;AACrG,eAAO,MAAM,yBAAyB,QAS1B,CAAC;AAEb,eAAO,MAAM,6BAA6B,MAAM,CAAC;AACjD,eAAO,MAAM,6BAA6B,QAAS,CAAC;AAEpD,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,OAAO,GAAG,QAAQ,GAAG,WAAW,CAAC;IACvC,IAAI,EAAE,aAAa,CAAC,MAAM,CAAC,CAAC;IAC5B,OAAO,EAAE,OAAO,CAAC;IACjB,MAAM,EAAE,MAAM,CAAC;CACf;AAED,wBAAgB,yBAAyB,CAAC,IAAI,EAAE;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,aAAa,CAAA;CAAE,GAAG,MAAM,CAMnG;AAKD,wBAAgB,sBAAsB,CAAC,IAAI,EAAE,MAAM,GAAG,iBAAiB,GAAG,SAAS,CAiClF;AAED,uFAAuF;AACvF,wBAAgB,sBAAsB,CAAC,QAAQ,EAAE,aAAa,EAAE,OAAO,EAAE,iBAAiB,GAAG,aAAa,CAczG;AAED,MAAM,WAAW,mBAAmB;IACnC,QAAQ,EAAE,aAAa,CAAC;IACxB,OAAO,CAAC,EAAE,iBAAiB,CAAC;IAC5B,qEAAqE;IACrE,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;CAChB;AAED;;;;GAIG;AACH,wBAAsB,aAAa,CAAC,IAAI,EAAE;IACzC,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE,aAAa,CAAC;IACxB,QAAQ,EAAE,CAAC,KAAK,EAAE;QAAE,YAAY,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,WAAW,CAAA;KAAE,KAAK,OAAO,CAAC;QAChG,IAAI,EAAE,MAAM,CAAC;QACb,OAAO,EAAE,MAAM,CAAC;QAChB,UAAU,EAAE,MAAM,CAAC;KACnB,CAAC,CAAC;IACH,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,cAAc,CAAC,EAAE,MAAM,CAAC;CACxB,GAAG,OAAO,CAAC,mBAAmB,CAAC,CAyC/B","sourcesContent":["import { runBoundedCompletion } from \"../autonomy/bounded-completion.ts\";\nimport type { RouteDecision } from \"../autonomy/contracts.ts\";\n\n/**\n * Routing-only judge: a bounded, tool-less completion (default: the medium model) that decides the\n * final cheap/medium/expensive tier for a user prompt, refining the regex classifier's baseline.\n * Core rule: planning is never cheap unless the judge explicitly deems the task trivial. The judge\n * proposes; the existing pipeline (model resolution, auth, escalation, gates) still decides.\n * Failure is honest: unparseable/timeout/unavailable falls back to the baseline with a visible\n * reasonCode, never silently.\n */\n\n/** Static across calls — callers pass cacheRetention \"short\" so only the variable tail is billed. */\nexport const ROUTE_JUDGE_SYSTEM_PROMPT = [\n\t\"You are a routing judge for a coding agent. You only route; you never answer the task.\",\n\t\"Pick which model tier should handle the user prompt:\",\n\t'- \"cheap\": trivial, mechanical, read-only lookups only.',\n\t'- \"medium\": normal implementation, scoped edits, tests, and NON-trivial planning/design.',\n\t'- \"expensive\": architecture, ambiguity, security/auth, destructive or release operations, high-impact changes.',\n\t\"Planning, design, and strategy prompts are NEVER cheap unless the task is genuinely trivial.\",\n\t\"Respond with STRICT JSON only - no prose:\",\n\t'{\"tier\":\"cheap\"|\"medium\"|\"expensive\",\"risk\":\"read-only\"|\"scoped-write\"|\"high-impact\"|\"approval-required\",\"trivial\":true|false,\"reason\":\"<short reason>\"}',\n].join(\"\\n\");\n\nexport const ROUTE_JUDGE_MAX_OUTPUT_TOKENS = 128;\nexport const ROUTE_JUDGE_MAX_WALL_CLOCK_MS = 10_000;\n\nexport interface RouteJudgeVerdict {\n\ttier: \"cheap\" | \"medium\" | \"expensive\";\n\trisk: RouteDecision[\"risk\"];\n\ttrivial: boolean;\n\treason: string;\n}\n\nexport function buildRouteJudgeUserPrompt(args: { prompt: string; baseline: RouteDecision }): string {\n\treturn [\n\t\t`Baseline (regex) verdict: tier=${args.baseline.tier}, risk=${args.baseline.risk}, reason=${args.baseline.reasonCode}.`,\n\t\t\"User prompt:\",\n\t\targs.prompt.slice(0, 4000),\n\t].join(\"\\n\");\n}\n\nconst JUDGE_TIERS: readonly string[] = [\"cheap\", \"medium\", \"expensive\"];\nconst JUDGE_RISKS: readonly string[] = [\"read-only\", \"scoped-write\", \"high-impact\", \"approval-required\"];\n\nexport function parseRouteJudgeVerdict(text: string): RouteJudgeVerdict | undefined {\n\tconst trimmed = text.trim();\n\tconst candidates: string[] = [trimmed];\n\tconst fenced = /```(?:json)?\\s*([\\s\\S]*?)```/.exec(trimmed);\n\tif (fenced?.[1]) candidates.push(fenced[1].trim());\n\tconst start = trimmed.indexOf(\"{\");\n\tconst end = trimmed.lastIndexOf(\"}\");\n\tif (start >= 0 && end > start) candidates.push(trimmed.slice(start, end + 1));\n\n\tfor (const candidate of candidates) {\n\t\tlet parsed: unknown;\n\t\ttry {\n\t\t\tparsed = JSON.parse(candidate);\n\t\t} catch {\n\t\t\tcontinue;\n\t\t}\n\t\tif (!parsed || typeof parsed !== \"object\" || Array.isArray(parsed)) continue;\n\t\tconst record = parsed as Record<string, unknown>;\n\t\t// The judge may never select the learning tier or anything outside the three foreground tiers.\n\t\tif (typeof record.tier !== \"string\" || !JUDGE_TIERS.includes(record.tier)) continue;\n\t\tconst risk =\n\t\t\ttypeof record.risk === \"string\" && JUDGE_RISKS.includes(record.risk)\n\t\t\t\t? (record.risk as RouteDecision[\"risk\"])\n\t\t\t\t: undefined;\n\t\tif (!risk) continue;\n\t\treturn {\n\t\t\ttier: record.tier as RouteJudgeVerdict[\"tier\"],\n\t\t\trisk,\n\t\t\ttrivial: record.trivial === true,\n\t\t\treason: typeof record.reason === \"string\" ? record.reason.slice(0, 200) : \"\",\n\t\t};\n\t}\n\treturn undefined;\n}\n\n/** Merge a judge verdict into the baseline decision (pure; never returns learning). */\nexport function applyRouteJudgeVerdict(baseline: RouteDecision, verdict: RouteJudgeVerdict): RouteDecision {\n\t// Enforce the core rule in code, not just in the judge prompt: downgrading a non-cheap\n\t// baseline to cheap requires an EXPLICIT trivial verdict. An untrusted judge saying\n\t// {tier:\"cheap\", trivial:false} for an elevated prompt keeps the baseline tier.\n\tconst tier =\n\t\tverdict.tier === \"cheap\" && baseline.tier !== \"cheap\" && !verdict.trivial ? baseline.tier : verdict.tier;\n\treturn {\n\t\t...baseline,\n\t\ttier,\n\t\trisk: verdict.risk,\n\t\tconfidence: Math.max(baseline.confidence, 0.75),\n\t\treasonCode: `judge_${tier}${verdict.trivial ? \"_trivial\" : \"\"}`,\n\t\treasons: [...baseline.reasons, `Route judge: ${verdict.reason || \"no reason given\"}`],\n\t};\n}\n\nexport interface RouteJudgeRunResult {\n\tdecision: RouteDecision;\n\tverdict?: RouteJudgeVerdict;\n\t/** Set when the judge could not decide and the baseline was kept. */\n\tfallbackReason?: string;\n\tcostUsd: number;\n}\n\n/**\n * Run the judge over a baseline decision. The completion executor is injected (production:\n * AgentSession.runIsolatedCompletion on the judge model). Never throws; every failure keeps the\n * baseline with a visible fallbackReason.\n */\nexport async function runRouteJudge(args: {\n\tprompt: string;\n\tbaseline: RouteDecision;\n\tcomplete: (input: { systemPrompt: string; userPrompt: string; signal?: AbortSignal }) => Promise<{\n\t\ttext: string;\n\t\tcostUsd: number;\n\t\tstopReason: string;\n\t}>;\n\tsignal?: AbortSignal;\n\tmaxWallClockMs?: number;\n}): Promise<RouteJudgeRunResult> {\n\tconst bounded = await runBoundedCompletion({\n\t\tmaxWallClockMs: args.maxWallClockMs ?? ROUTE_JUDGE_MAX_WALL_CLOCK_MS,\n\t\tsignal: args.signal,\n\t\texecute: (signal) =>\n\t\t\targs.complete({\n\t\t\t\tsystemPrompt: ROUTE_JUDGE_SYSTEM_PROMPT,\n\t\t\t\tuserPrompt: buildRouteJudgeUserPrompt({ prompt: args.prompt, baseline: args.baseline }),\n\t\t\t\tsignal,\n\t\t\t}),\n\t});\n\tconst costUsd = bounded.completion?.costUsd ?? 0;\n\n\tif (bounded.failure || !bounded.completion) {\n\t\treturn {\n\t\t\tdecision: {\n\t\t\t\t...args.baseline,\n\t\t\t\treasons: [...args.baseline.reasons, \"Route judge unavailable; baseline kept\"],\n\t\t\t},\n\t\t\tfallbackReason: bounded.failure ? `judge_${bounded.failure.reasonCode}` : \"judge_unavailable_fallback\",\n\t\t\tcostUsd,\n\t\t};\n\t}\n\tif (bounded.completion.stopReason === \"error\" || bounded.completion.stopReason === \"aborted\") {\n\t\treturn {\n\t\t\tdecision: { ...args.baseline, reasons: [...args.baseline.reasons, \"Route judge errored; baseline kept\"] },\n\t\t\tfallbackReason: \"judge_model_error\",\n\t\t\tcostUsd,\n\t\t};\n\t}\n\n\tconst verdict = parseRouteJudgeVerdict(bounded.completion.text);\n\tif (!verdict) {\n\t\treturn {\n\t\t\tdecision: { ...args.baseline, reasons: [...args.baseline.reasons, \"Route judge unparseable; baseline kept\"] },\n\t\t\tfallbackReason: \"judge_unparseable_fallback\",\n\t\t\tcostUsd,\n\t\t};\n\t}\n\n\treturn { decision: applyRouteJudgeVerdict(args.baseline, verdict), verdict, costUsd };\n}\n"]}