vidspotai-shared 1.0.55 → 1.0.57

package/lib/services/agent/costPreflight.js

@@ -1,75 +0,0 @@
-"use strict";
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.estimateProjectCost = estimateProjectCost;
-exports.estimatePlanCost = estimatePlanCost;
-/**
- * Cost preflight — sums credits over a Project JSON's clips by their
- * provenance.creditsSpent. For clips not yet generated, falls back to a
- * conservative tier estimate so the user sees an upper bound BEFORE
- * committing spend.
- *
- * Used by:
- *  - the `estimate_cost` agent tool (offered to the planner so it can
- *    stay within tierBudget),
- *  - the /v1/agent/cost preview endpoint (storyboard preview before render).
- */
-const TIER_FALLBACK_PER_SECOND = {
-    T0: 0.5,
-    T1: 2,
-    T2: 6,
-    T3: 14,
-};
-const clipDurationMs = (c) => Math.max(0, c.display.to - c.display.from);
-function estimateProjectCost(project) {
-    const byTier = { T0: 0, T1: 0, T2: 0, T3: 0 };
-    const perClip = [];
-    for (const clip of project.clips) {
-        const tier = clip.provenance?.tier;
-        let credits = clip.provenance?.creditsSpent;
-        let estimated = false;
-        if (credits === undefined) {
-            if (tier) {
-                credits =
-                    (clipDurationMs(clip) / 1000) * TIER_FALLBACK_PER_SECOND[tier];
-                estimated = true;
-            }
-            else {
-                credits = 0;
-            }
-        }
-        if (tier)
-            byTier[tier] += credits;
-        perClip.push({ clipId: clip.id, tier, credits, estimated });
-    }
-    const totalCredits = perClip.reduce((s, x) => s + x.credits, 0);
-    return { totalCredits, byTier, perClip };
-}
-/**
- * Plan-side estimate (pre-execution). The planner already writes
- * `estimatedCredits` per scene + a `totalEstimatedCredits` rollup, so this
- * mostly trusts those values and re-aggregates by tier for the UI chip.
- * Falls back to the same per-second tier curve as `estimateProjectCost`
- * when a scene omits its own estimate.
- */
-function estimatePlanCost(plan) {
-    const byTier = { T0: 0, T1: 0, T2: 0, T3: 0 };
-    const perClip = [];
-    for (const scene of plan.scenes) {
-        const tier = scene.tier;
-        let credits = scene.estimatedCredits;
-        let estimated = false;
-        if (credits === undefined || credits === 0) {
-            credits = (scene.durationMs / 1000) * TIER_FALLBACK_PER_SECOND[tier];
-            estimated = true;
-        }
-        byTier[tier] += credits;
-        perClip.push({
-            clipId: `scene-${scene.sceneIndex}`,
-            tier,
-            credits,
-            estimated,
-        });
-    }
-    const totalCredits = perClip.reduce((s, x) => s + x.credits, 0);
-    return { totalCredits, byTier, perClip };
-}

package/lib/services/agent/critic.d.ts

@@ -1,103 +0,0 @@
-import { z } from "zod";
-import { AgentPersona } from "../../schemas/agentPersona.schema";
-import { Project } from "../../schemas/project.schema";
-import { VideoBrief } from "../../schemas/brief.schema";
-import { VideoPlan } from "../../schemas/videoPlan.schema";
-import { LlmCaller } from "./llmCaller";
-import { ModelRouter } from "./modelRouter";
-/**
- * Critic — runs after assets land and BEFORE render. Catches off-brief assets
- * so the agent can self-heal (regenerate / replace with stock) without burning
- * a full render cycle.
- *
- * Rubric (V1):
- *   - continuity   : do consecutive scenes match in style / lighting?
- *   - voSync       : does on-screen text align with the VO line?
- *   - brandSafety  : are any blocked claims present?
- *   - consistency  : same product / face across scenes (vision check)
- *   - costVsTier   : did total spend stay within the tier budget?
- *
- * Each rubric item returns pass/warn/fail with a one-line rationale. The
- * agent self-heals up to one pass on the worst-scoring item.
- */
-export declare const CriticSeveritySchema: z.ZodEnum<{
-    pass: "pass";
-    warn: "warn";
-    fail: "fail";
-}>;
-export type CriticSeverity = z.infer<typeof CriticSeveritySchema>;
-export declare const CriticFindingSchema: z.ZodObject<{
-    rubric: z.ZodEnum<{
-        continuity: "continuity";
-        voSync: "voSync";
-        brandSafety: "brandSafety";
-        consistency: "consistency";
-        costVsTier: "costVsTier";
-    }>;
-    severity: z.ZodEnum<{
-        pass: "pass";
-        warn: "warn";
-        fail: "fail";
-    }>;
-    clipId: z.ZodOptional<z.ZodString>;
-    message: z.ZodString;
-    suggestedAction: z.ZodOptional<z.ZodObject<{
-        kind: z.ZodEnum<{
-            trim: "trim";
-            regenerate: "regenerate";
-            replace_with_stock: "replace_with_stock";
-            swap_voice: "swap_voice";
-        }>;
-        detail: z.ZodOptional<z.ZodString>;
-    }, z.core.$strip>>;
-}, z.core.$strip>;
-export type CriticFinding = z.infer<typeof CriticFindingSchema>;
-export declare const CriticReportSchema: z.ZodObject<{
-    overall: z.ZodEnum<{
-        pass: "pass";
-        warn: "warn";
-        fail: "fail";
-    }>;
-    findings: z.ZodArray<z.ZodObject<{
-        rubric: z.ZodEnum<{
-            continuity: "continuity";
-            voSync: "voSync";
-            brandSafety: "brandSafety";
-            consistency: "consistency";
-            costVsTier: "costVsTier";
-        }>;
-        severity: z.ZodEnum<{
-            pass: "pass";
-            warn: "warn";
-            fail: "fail";
-        }>;
-        clipId: z.ZodOptional<z.ZodString>;
-        message: z.ZodString;
-        suggestedAction: z.ZodOptional<z.ZodObject<{
-            kind: z.ZodEnum<{
-                trim: "trim";
-                regenerate: "regenerate";
-                replace_with_stock: "replace_with_stock";
-                swap_voice: "swap_voice";
-            }>;
-            detail: z.ZodOptional<z.ZodString>;
-        }, z.core.$strip>>;
-    }, z.core.$strip>>;
-}, z.core.$strip>;
-export type CriticReport = z.infer<typeof CriticReportSchema>;
-export interface CriticOptions {
-    llm: LlmCaller;
-    router?: ModelRouter;
-}
-export declare class Critic {
-    private readonly opts;
-    private readonly router;
-    constructor(opts: CriticOptions);
-    review(project: Project, brief: VideoBrief, opts?: {
-        persona?: AgentPersona;
-        plan?: VideoPlan;
-    }): Promise<CriticReport>;
-    private formatEmphasis;
-    private formatTierMix;
-}
-//# sourceMappingURL=critic.d.ts.map

package/lib/services/agent/critic.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"critic.d.ts","sourceRoot":"","sources":["../../../src/services/agent/critic.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EAAE,YAAY,EAAE,MAAM,mCAAmC,CAAC;AACjE,OAAO,EAAE,OAAO,EAAE,MAAM,8BAA8B,CAAC;AACvD,OAAO,EAAE,UAAU,EAAE,MAAM,4BAA4B,CAAC;AACxD,OAAO,EAAE,SAAS,EAAE,MAAM,gCAAgC,CAAC;AAC3D,OAAO,EAAE,SAAS,EAAE,MAAM,aAAa,CAAC;AACxC,OAAO,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAG5C;;;;;;;;;;;;;;GAcG;AAEH,eAAO,MAAM,oBAAoB;;;;EAAmC,CAAC;AACrE,MAAM,MAAM,cAAc,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,oBAAoB,CAAC,CAAC;AAElE,eAAO,MAAM,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;iBAmB9B,CAAC;AACH,MAAM,MAAM,aAAa,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,mBAAmB,CAAC,CAAC;AAEhE,eAAO,MAAM,kBAAkB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAG7B,CAAC;AACH,MAAM,MAAM,YAAY,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,kBAAkB,CAAC,CAAC;AAE9D,MAAM,WAAW,aAAa;IAC5B,GAAG,EAAE,SAAS,CAAC;IACf,MAAM,CAAC,EAAE,WAAW,CAAC;CACtB;AAED,qBAAa,MAAM;IAGL,OAAO,CAAC,QAAQ,CAAC,IAAI;IAFjC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAc;gBAER,IAAI,EAAE,aAAa;IAI1C,MAAM,CACV,OAAO,EAAE,OAAO,EAChB,KAAK,EAAE,UAAU,EACjB,IAAI,GAAE;QAAE,OAAO,CAAC,EAAE,YAAY,CAAC;QAAC,IAAI,CAAC,EAAE,SAAS,CAAA;KAAO,GACtD,OAAO,CAAC,YAAY,CAAC;IAiExB,OAAO,CAAC,cAAc;IActB,OAAO,CAAC,aAAa;CAKtB"}

package/lib/services/agent/critic.js

@@ -1,132 +0,0 @@
-"use strict";
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.Critic = exports.CriticReportSchema = exports.CriticFindingSchema = exports.CriticSeveritySchema = void 0;
-const zod_1 = require("zod");
-const modelRouter_1 = require("./modelRouter");
-const personas_1 = require("./personas");
-/**
- * Critic — runs after assets land and BEFORE render. Catches off-brief assets
- * so the agent can self-heal (regenerate / replace with stock) without burning
- * a full render cycle.
- *
- * Rubric (V1):
- *   - continuity   : do consecutive scenes match in style / lighting?
- *   - voSync       : does on-screen text align with the VO line?
- *   - brandSafety  : are any blocked claims present?
- *   - consistency  : same product / face across scenes (vision check)
- *   - costVsTier   : did total spend stay within the tier budget?
- *
- * Each rubric item returns pass/warn/fail with a one-line rationale. The
- * agent self-heals up to one pass on the worst-scoring item.
- */
-exports.CriticSeveritySchema = zod_1.z.enum(["pass", "warn", "fail"]);
-exports.CriticFindingSchema = zod_1.z.object({
-    rubric: zod_1.z.enum([
-        "continuity",
-        "voSync",
-        "brandSafety",
-        "consistency",
-        "costVsTier",
-    ]),
-    severity: exports.CriticSeveritySchema,
-    /** Which clip / scene the finding refers to, if any. */
-    clipId: zod_1.z.string().optional(),
-    message: zod_1.z.string(),
-    /** Optional self-heal hint the executor can act on. */
-    suggestedAction: zod_1.z
-        .object({
-        kind: zod_1.z.enum(["regenerate", "replace_with_stock", "trim", "swap_voice"]),
-        detail: zod_1.z.string().optional(),
-    })
-        .optional(),
-});
-exports.CriticReportSchema = zod_1.z.object({
-    overall: exports.CriticSeveritySchema,
-    findings: zod_1.z.array(exports.CriticFindingSchema),
-});
-class Critic {
-    constructor(opts) {
-        this.opts = opts;
-        this.router = opts.router ?? new modelRouter_1.ModelRouter();
-    }
-    async review(project, brief, opts = {}) {
-        const model = this.router.pickFor("critic");
-        // Vision-mode QA: feed clip thumbnails into the critic for continuity check.
-        const thumbnails = project.clips
-            .map((c) => (c.type === "video" || c.type === "image" ? c.src : null))
-            .filter((u) => Boolean(u))
-            .slice(0, 12); // bound vision token spend
-        const personaEmphasis = opts.persona
-            ? this.formatEmphasis(opts.persona)
-            : "";
-        const tierMix = opts.plan
-            ? this.formatTierMix(opts.plan)
-            : "(plan not provided — judge continuity defensively across all scenes)";
-        const result = await this.opts.llm.structured({
-            model,
-            schema: exports.CriticReportSchema,
-            schemaName: "CriticReport",
-            temperature: 0.2,
-            messages: [
-                {
-                    role: "system",
-                    content: [
-                        "You are a quality critic for AI-generated short-form videos.",
-                        "Score the project against the rubric.",
-                        "",
-                        "Continuity rule (tier-aware):",
-                        "- T0/T1 scenes use stock or single AI stills + motion. Do NOT flag",
-                        "  continuity unless lighting / colour palette / subject identity",
-                        "  shifts so hard it breaks comprehension. Mood drift is acceptable.",
-                        "- T2/T3 scenes are AI-generated motion the user paid a premium for.",
-                        "  Flag continuity for any visible character / product identity drift,",
-                        "  lighting jumps, or style mismatches with adjacent scenes.",
-                        "",
-                        "Brand-safety rules (strict):",
-                        "- Any phrase in brief.claims.block → fail immediately.",
-                        "- Any unsubstantiated quantitative claim (\"3x faster\", \"99% of users\",",
-                        "  \"clinically proven\", numeric efficacy without a cited source) → warn",
-                        "  with suggestedAction.kind=\"regenerate\" + a corrective rewrite.",
-                        "- Medical / financial / legal claims default to fail unless brief",
-                        "  explicitly allows them.",
-                        "",
-                        "Per-scene tier reference (use to apply the continuity rule):",
-                        tierMix,
-                        personaEmphasis,
-                    ]
-                        .filter(Boolean)
-                        .join("\n"),
-                },
-                {
-                    role: "user",
-                    content: [
-                        `Brief:\n${JSON.stringify(brief, null, 2)}`,
-                        `\nProject:\n${JSON.stringify(project, null, 2)}`,
-                    ].join("\n"),
-                    imageUrls: thumbnails,
-                },
-            ],
-        });
-        return result.data;
-    }
-    formatEmphasis(persona) {
-        const emphasis = personas_1.PERSONA_PACKS[persona].criticEmphasis;
-        const entries = Object.entries(emphasis);
-        if (!entries.length)
-            return "";
-        const lines = entries
-            .map(([axis, weight]) => `  - ${axis}: extra weight ${weight}`)
-            .join("\n");
-        return [
-            "",
-            `Persona "${persona}" — apply extra scrutiny to:`,
-            lines,
-        ].join("\n");
-    }
-    formatTierMix(plan) {
-        return plan.scenes
-            .map((s) => `  scene ${s.sceneIndex}: tier=${s.tier} (${s.strategy})`)
-            .join("\n");
-    }
-}
-exports.Critic = Critic;

package/lib/services/agent/eval/index.d.ts

@@ -1,5 +0,0 @@
-export * from "./types";
-export * from "./recorder";
-export * from "./judge";
-export * from "./seedBriefs";
-//# sourceMappingURL=index.d.ts.map

package/lib/services/agent/eval/index.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../src/services/agent/eval/index.ts"],"names":[],"mappings":"AAAA,cAAc,SAAS,CAAC;AACxB,cAAc,YAAY,CAAC;AAC3B,cAAc,SAAS,CAAC;AACxB,cAAc,cAAc,CAAC"}

package/lib/services/agent/eval/index.js

@@ -1,20 +0,0 @@
-"use strict";
-var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
-    if (k2 === undefined) k2 = k;
-    var desc = Object.getOwnPropertyDescriptor(m, k);
-    if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
-      desc = { enumerable: true, get: function() { return m[k]; } };
-    }
-    Object.defineProperty(o, k2, desc);
-}) : (function(o, m, k, k2) {
-    if (k2 === undefined) k2 = k;
-    o[k2] = m[k];
-}));
-var __exportStar = (this && this.__exportStar) || function(m, exports) {
-    for (var p in m) if (p !== "default" && !Object.prototype.hasOwnProperty.call(exports, p)) __createBinding(exports, m, p);
-};
-Object.defineProperty(exports, "__esModule", { value: true });
-__exportStar(require("./types"), exports);
-__exportStar(require("./recorder"), exports);
-__exportStar(require("./judge"), exports);
-__exportStar(require("./seedBriefs"), exports);

package/lib/services/agent/eval/judge.d.ts

@@ -1,14 +0,0 @@
-import { AgentPersona } from "../../../schemas/agentPersona.schema";
-import { Project } from "../../../schemas/project.schema";
-import { LlmCaller } from "../llmCaller";
-import { ModelRouter } from "../modelRouter";
-import { EvalBrief, EvalScore } from "./types";
-export declare class EvalJudge {
-    private readonly llm;
-    private readonly router;
-    constructor(llm: LlmCaller, router?: ModelRouter);
-    judge(brief: EvalBrief, project: Project, agentRunId: string, opts?: {
-        persona?: AgentPersona;
-    }): Promise<EvalScore>;
-}
-//# sourceMappingURL=judge.d.ts.map

package/lib/services/agent/eval/judge.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"judge.d.ts","sourceRoot":"","sources":["../../../../src/services/agent/eval/judge.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,YAAY,EAAE,MAAM,sCAAsC,CAAC;AACpE,OAAO,EAAE,OAAO,EAAE,MAAM,iCAAiC,CAAC;AAC1D,OAAO,EAAE,SAAS,EAAE,MAAM,cAAc,CAAC;AACzC,OAAO,EAAE,WAAW,EAAE,MAAM,gBAAgB,CAAC;AAE7C,OAAO,EAAE,SAAS,EAAgC,SAAS,EAAE,MAAM,SAAS,CAAC;AAgD7E,qBAAa,SAAS;IAGR,OAAO,CAAC,QAAQ,CAAC,GAAG;IAFhC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAc;gBAER,GAAG,EAAE,SAAS,EAAE,MAAM,CAAC,EAAE,WAAW;IAI3D,KAAK,CACT,KAAK,EAAE,SAAS,EAChB,OAAO,EAAE,OAAO,EAChB,UAAU,EAAE,MAAM,EAClB,IAAI,GAAE;QAAE,OAAO,CAAC,EAAE,YAAY,CAAA;KAAO,GACpC,OAAO,CAAC,SAAS,CAAC;CA8CtB"}

package/lib/services/agent/eval/judge.js

@@ -1,95 +0,0 @@
-"use strict";
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.EvalJudge = void 0;
-const zod_1 = require("zod");
-const modelRouter_1 = require("../modelRouter");
-const personas_1 = require("../personas");
-const types_1 = require("./types");
-/**
- * LLM-as-judge — scores a finalized Project against an EvalBrief's rubric.
- *
- * Uses the critic-tier model (vision-capable Sonnet) with a strict structured
- * output. Independent from the live Critic that runs during a real agent run;
- * the judge is allowed to be slower / more expensive because it runs offline.
- */
-const RUBRIC_WEIGHTS = {
-    hookStrength: 0.2,
-    voVisualSync: 0.2,
-    pacingFit: 0.15,
-    brandSafety: 0.15,
-    costVsTier: 0.1,
-    intentMatch: 0.2,
-};
-/**
- * Apply persona criticEmphasis as additive deltas to the base weights, then
- * renormalise so the weight set still sums to 1. Keeps scores comparable
- * across personas while letting each persona's "what matters most" lean
- * the aggregate score.
- */
-function weightsForPersona(persona) {
-    if (!persona)
-        return RUBRIC_WEIGHTS;
-    const emphasis = personas_1.PERSONA_PACKS[persona].criticEmphasis;
-    const adjusted = { ...RUBRIC_WEIGHTS };
-    for (const [axis, delta] of Object.entries(emphasis)) {
-        const key = axis;
-        adjusted[key] = Math.max(0, adjusted[key] + (delta ?? 0));
-    }
-    const total = Object.values(adjusted).reduce((s, v) => s + v, 0);
-    if (total <= 0)
-        return RUBRIC_WEIGHTS;
-    for (const k of Object.keys(adjusted)) {
-        adjusted[k] = adjusted[k] / total;
-    }
-    return adjusted;
-}
-const JudgeOutputSchema = zod_1.z.object({
-    rubric: types_1.EvalRubricSchema,
-    notes: zod_1.z.string(),
-});
-class EvalJudge {
-    constructor(llm, router) {
-        this.llm = llm;
-        this.router = router ?? new modelRouter_1.ModelRouter();
-    }
-    async judge(brief, project, agentRunId, opts = {}) {
-        const model = this.router.pickFor("critic");
-        const weights = weightsForPersona(opts.persona);
-        const messages = [
-            {
-                role: "system",
-                content: [
-                    "You are an offline quality judge for AI-generated short videos.",
-                    "Score the project against the rubric on each axis (0-1, two decimals).",
-                    "Be harsh: 1.0 means publish-without-edit; 0.7 is good but with one fixable issue;",
-                    "0.4 means a re-roll is needed. Cite concrete clip ids in `notes`.",
-                ].join(" "),
-            },
-            {
-                role: "user",
-                content: [
-                    `Brief id: ${brief.id} (niche: ${brief.niche})`,
-                    `What "good" looks like:\n${brief.goodOutputNotes}`,
-                    `\nProject JSON:\n${JSON.stringify(project, null, 2)}`,
-                ].join("\n\n"),
-            },
-        ];
-        const result = await this.llm.structured({
-            model,
-            messages,
-            schema: JudgeOutputSchema,
-            schemaName: "EvalJudgeResult",
-            temperature: 0,
-        });
-        const aggregate = Object.keys(weights).reduce((sum, key) => sum + result.data.rubric[key] * weights[key], 0);
-        return {
-            briefId: brief.id,
-            agentRunId,
-            rubric: result.data.rubric,
-            score: Number(aggregate.toFixed(3)),
-            judgeNotes: result.data.notes,
-            scoredAt: new Date().toISOString(),
-        };
-    }
-}
-exports.EvalJudge = EvalJudge;

package/lib/services/agent/eval/recorder.d.ts

@@ -1,17 +0,0 @@
-import { ToolContext, ToolOutcome } from "../toolRegistry";
-import { AgentRunTrace } from "./types";
-export declare function setTracePersistence(fn: (trace: AgentRunTrace) => Promise<void>): void;
-export declare function startTrace(opts: {
-    agentRunId: string;
-    briefId?: string;
-    modelRouting: Record<string, string>;
-    rngSeed: string;
-}): AgentRunTrace;
-export declare function getTrace(agentRunId: string): AgentRunTrace | undefined;
-export declare function endTrace(agentRunId: string): Promise<AgentRunTrace | undefined>;
-/**
- * Drop-in replacement for runTool that records the call in the active trace.
- * Use this in eval mode; production code can keep calling runTool directly.
- */
-export declare function runToolRecorded<O = unknown>(name: string, input: unknown, ctx: ToolContext): Promise<ToolOutcome<O>>;
-//# sourceMappingURL=recorder.d.ts.map

package/lib/services/agent/eval/recorder.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"recorder.d.ts","sourceRoot":"","sources":["../../../../src/services/agent/eval/recorder.ts"],"names":[],"mappings":"AAAA,OAAO,EAA0B,WAAW,EAAE,WAAW,EAAE,MAAM,iBAAiB,CAAC;AACnF,OAAO,EAAE,aAAa,EAAiB,MAAM,SAAS,CAAC;AAcvD,wBAAgB,mBAAmB,CAAC,EAAE,EAAE,CAAC,KAAK,EAAE,aAAa,KAAK,OAAO,CAAC,IAAI,CAAC,GAAG,IAAI,CAErF;AAED,wBAAgB,UAAU,CAAC,IAAI,EAAE;IAC/B,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,YAAY,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACrC,OAAO,EAAE,MAAM,CAAC;CACjB,GAAG,aAAa,CAWhB;AAED,wBAAgB,QAAQ,CAAC,UAAU,EAAE,MAAM,GAAG,aAAa,GAAG,SAAS,CAEtE;AAED,wBAAsB,QAAQ,CAAC,UAAU,EAAE,MAAM,GAAG,OAAO,CAAC,aAAa,GAAG,SAAS,CAAC,CAYrF;AAED;;;GAGG;AACH,wBAAsB,eAAe,CAAC,CAAC,GAAG,OAAO,EAC/C,IAAI,EAAE,MAAM,EACZ,KAAK,EAAE,OAAO,EACd,GAAG,EAAE,WAAW,GACf,OAAO,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CASzB"}

package/lib/services/agent/eval/recorder.js

@@ -1,65 +0,0 @@
-"use strict";
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.setTracePersistence = setTracePersistence;
-exports.startTrace = startTrace;
-exports.getTrace = getTrace;
-exports.endTrace = endTrace;
-exports.runToolRecorded = runToolRecorded;
-const toolRegistry_1 = require("../toolRegistry");
-/**
- * Recording wrapper around runTool — every tool call is appended to a trace
- * buffer keyed by agentRunId. Replay reads this buffer back; offline eval
- * compares two recordings against the same brief.
- *
- * Storage is in-memory by default; the host swaps in a Firestore-backed
- * implementation by setting the persistence callback.
- */
-const traces = new Map();
-let persistFn = null;
-function setTracePersistence(fn) {
-    persistFn = fn;
-}
-function startTrace(opts) {
-    const trace = {
-        agentRunId: opts.agentRunId,
-        briefId: opts.briefId,
-        startedAt: new Date().toISOString(),
-        modelRouting: opts.modelRouting,
-        rngSeed: opts.rngSeed,
-        toolCalls: [],
-    };
-    traces.set(opts.agentRunId, trace);
-    return trace;
-}
-function getTrace(agentRunId) {
-    return traces.get(agentRunId);
-}
-async function endTrace(agentRunId) {
-    const trace = traces.get(agentRunId);
-    if (!trace)
-        return undefined;
-    trace.endedAt = new Date().toISOString();
-    if (persistFn) {
-        try {
-            await persistFn(trace);
-        }
-        catch {
-            // Persistence failure must not poison the agent run.
-        }
-    }
-    return trace;
-}
-/**
- * Drop-in replacement for runTool that records the call in the active trace.
- * Use this in eval mode; production code can keep calling runTool directly.
- */
-async function runToolRecorded(name, input, ctx) {
-    const ts = Date.now();
-    const outcome = await (0, toolRegistry_1.runTool)(name, input, ctx);
-    const trace = traces.get(ctx.agentRunId);
-    if (trace) {
-        const entry = { ts, tool: name, input, outcome };
-        trace.toolCalls.push(entry);
-    }
-    return outcome;
-}

package/lib/services/agent/eval/seedBriefs.d.ts

@@ -1,16 +0,0 @@
-import { EvalBrief } from "./types";
-/**
- * Eval set — 50 hand-curated briefs covering the 5 launch personas plus
- * adjacent niches that should route to the closest persona (e.g. recipe →
- * faceless-yt, real-estate → product-demo).
- *
- * Use these to:
- *   1. Catch routing regressions (router persona !== expected niche).
- *   2. Catch planner regressions (tier mix, schema validity, scene count).
- *   3. Score quality drift after prompt / model / rubric changes.
- *
- * Add new briefs alongside the change that needs them; do not bulk-grow
- * without a reason — every brief costs 4 LLM calls per run.
- */
-export declare const SEED_EVAL_BRIEFS: EvalBrief[];
-//# sourceMappingURL=seedBriefs.d.ts.map

package/lib/services/agent/eval/seedBriefs.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"seedBriefs.d.ts","sourceRoot":"","sources":["../../../../src/services/agent/eval/seedBriefs.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AAEpC;;;;;;;;;;;;GAYG;AACH,eAAO,MAAM,gBAAgB,EAAE,SAAS,EA4pCvC,CAAC"}