@elizaos/app-core 2.0.0-beta.3 → 2.0.11-beta.5

package/services/local-inference/catalog.js

@@ -1,262 +0,0 @@
-/**
- * Eliza-curated local model catalog.
- *
- * Eliza-1 is the only default-eligible model line. There is exactly one
- * default per device tier (`lite-0_6b`, `mobile-1_7b`, `desktop-9b`,
- * `pro-27b`, `server-h200`). The recommendation engine picks one of
- * these tiers based on hardware. See
- * `/Users/shawwalters/eliza-workspace/milady/eliza/packages/inference/AGENTS.md`
- * §2 for the binding tier matrix.
- *
- * HF-search results from outside `elizalabs/eliza-1-*` MUST never be
- * marked default-eligible (handled by `hf-search.ts`, which produces
- * entries that are absent from `DEFAULT_ELIGIBLE_MODEL_IDS`).
- *
- * When upstream naming conventions drift, update `ggufFile` here — we
- * rely on the exact filename for resolved-URL construction in the
- * downloader.
- */
-import { ELIZA_1_TIERS } from "./manifest";
-export const ELIZA_1_TIER_IDS = ELIZA_1_TIERS.map((tier) => `eliza-1-${tier}`);
-/**
- * The model id the engine auto-loads on first run when no preference is
- * set. Resolves to the `mobile-1_7b` tier — the smallest Eliza-1 tier
- * that fits the broadest range of hardware (modern phone or laptop).
- * Hosts that can't fit `mobile-1_7b` get the `lite-0_6b` fallback via
- * the recommendation ladder.
- */
-export const FIRST_RUN_DEFAULT_MODEL_ID = "eliza-1-mobile-1_7b";
-/**
- * The single source of truth for default-eligibility. Only Eliza-1
- * tiers are default-eligible. The recommendation engine MUST refuse to
- * surface anything outside this set as a default; HF-search results
- * MUST never appear here.
- */
-export const DEFAULT_ELIGIBLE_MODEL_IDS = new Set(ELIZA_1_TIER_IDS);
-export function isDefaultEligibleId(id) {
-    return DEFAULT_ELIGIBLE_MODEL_IDS.has(id);
-}
-/** Compatibility export for callers that need the Eliza-1 model id set. */
-export const ELIZA_1_PLACEHOLDER_IDS = new Set(ELIZA_1_TIER_IDS);
-const BASE_REQUIRED_KERNELS = [
-    "dflash",
-    "turbo3",
-    "turbo4",
-    "qjl_full",
-    "polarquant",
-];
-function requiredKernelsForContext(contextLength) {
-    return contextLength > 65536
-        ? [...BASE_REQUIRED_KERNELS, "turbo3_tcq"]
-        : [...BASE_REQUIRED_KERNELS];
-}
-function drafterId(id) {
-    return `${id}-drafter`;
-}
-function runtimeFor(id, contextLength) {
-    return {
-        preferredBackend: "llama-server",
-        optimizations: {
-            parallel: contextLength >= 131072 ? 8 : 4,
-            flashAttention: true,
-            mlock: contextLength >= 131072,
-            requiresKernel: requiredKernelsForContext(contextLength),
-        },
-        dflash: {
-            drafterModelId: drafterId(id),
-            specType: "dflash",
-            contextSize: contextLength,
-            draftContextSize: Math.min(contextLength, 65536),
-            draftMin: 2,
-            draftMax: contextLength >= 131072 ? 8 : 6,
-            gpuLayers: "auto",
-            draftGpuLayers: "auto",
-            disableThinking: true,
-        },
-    };
-}
-function drafterCompanion(args) {
-    return {
-        id: drafterId(args.id),
-        displayName: `${args.displayName} drafter`,
-        hfRepo: `elizalabs/${args.id}`,
-        ggufFile: args.ggufFile,
-        params: args.params,
-        quant: "Eliza-1 drafter companion",
-        sizeGb: args.sizeGb,
-        minRamGb: args.minRamGb,
-        category: "drafter",
-        bucket: args.bucket,
-        hiddenFromCatalog: true,
-        runtimeRole: "dflash-drafter",
-        companionForModelId: args.id,
-        tokenizerFamily: "eliza1",
-        blurb: `${args.displayName} drafter companion.`,
-    };
-}
-export const MODEL_CATALOG = [
-    // ─── Eliza-1 lite (low-RAM phones, CPU fallback) ────────────────────
-    {
-        id: "eliza-1-lite-0_6b",
-        displayName: "Eliza-1 lite",
-        hfRepo: "elizalabs/eliza-1-lite-0_6b",
-        ggufFile: "text/eliza-1-lite-0_6b-32k.gguf",
-        bundleManifestFile: "eliza-1.manifest.json",
-        params: "1B",
-        quant: "Eliza-1 optimized local runtime",
-        sizeGb: 0.5,
-        minRamGb: 2,
-        category: "chat",
-        bucket: "small",
-        contextLength: 32768,
-        tokenizerFamily: "eliza1",
-        companionModelIds: ["eliza-1-lite-0_6b-drafter"],
-        runtime: runtimeFor("eliza-1-lite-0_6b", 32768),
-        blurb: "Eliza-1 lite — fits low-RAM phones and CPU-only fallback with the optimized local runtime.",
-    },
-    drafterCompanion({
-        id: "eliza-1-lite-0_6b",
-        displayName: "Eliza-1 lite",
-        ggufFile: "dflash/drafter-lite-0_6b.gguf",
-        params: "1B",
-        sizeGb: 0.25,
-        minRamGb: 2,
-        bucket: "small",
-    }),
-    // ─── Eliza-1 mobile (modern phones) ─────────────────────────────────
-    {
-        id: "eliza-1-mobile-1_7b",
-        displayName: "Eliza-1 mobile",
-        hfRepo: "elizalabs/eliza-1-mobile-1_7b",
-        ggufFile: "text/eliza-1-mobile-1_7b-32k.gguf",
-        bundleManifestFile: "eliza-1.manifest.json",
-        params: "1.7B",
-        quant: "Eliza-1 optimized local runtime",
-        sizeGb: 1.2,
-        minRamGb: 4,
-        category: "chat",
-        bucket: "small",
-        contextLength: 32768,
-        tokenizerFamily: "eliza1",
-        companionModelIds: ["eliza-1-mobile-1_7b-drafter"],
-        runtime: runtimeFor("eliza-1-mobile-1_7b", 32768),
-        blurb: "Eliza-1 mobile — modern phone default with text and voice prepared for the optimized local runtime.",
-    },
-    drafterCompanion({
-        id: "eliza-1-mobile-1_7b",
-        displayName: "Eliza-1 mobile",
-        ggufFile: "dflash/drafter-mobile-1_7b.gguf",
-        params: "1.7B",
-        sizeGb: 0.35,
-        minRamGb: 4,
-        bucket: "small",
-    }),
-    // ─── Eliza-1 desktop (laptops, 24GB phones, 48GB Mac) ───────────────
-    {
-        id: "eliza-1-desktop-9b",
-        displayName: "Eliza-1 desktop",
-        hfRepo: "elizalabs/eliza-1-desktop-9b",
-        ggufFile: "text/eliza-1-desktop-9b-64k.gguf",
-        bundleManifestFile: "eliza-1.manifest.json",
-        params: "9B",
-        quant: "Eliza-1 optimized local runtime",
-        sizeGb: 5.4,
-        minRamGb: 12,
-        category: "chat",
-        bucket: "mid",
-        contextLength: 65536,
-        tokenizerFamily: "eliza1",
-        companionModelIds: ["eliza-1-desktop-9b-drafter"],
-        runtime: runtimeFor("eliza-1-desktop-9b", 65536),
-        blurb: "Eliza-1 desktop — laptop / 24 GB phone / 48 GB Mac default with text, voice, and vision in the optimized local runtime.",
-    },
-    drafterCompanion({
-        id: "eliza-1-desktop-9b",
-        displayName: "Eliza-1 desktop",
-        ggufFile: "dflash/drafter-desktop-9b.gguf",
-        params: "9B",
-        sizeGb: 0.8,
-        minRamGb: 12,
-        bucket: "mid",
-    }),
-    // ─── Eliza-1 pro (96GB+ Mac, high-VRAM desktop) ─────────────────────
-    {
-        id: "eliza-1-pro-27b",
-        displayName: "Eliza-1 pro",
-        hfRepo: "elizalabs/eliza-1-pro-27b",
-        ggufFile: "text/eliza-1-pro-27b-128k.gguf",
-        bundleManifestFile: "eliza-1.manifest.json",
-        params: "27B",
-        quant: "Eliza-1 optimized local runtime",
-        sizeGb: 16.8,
-        minRamGb: 32,
-        category: "chat",
-        bucket: "large",
-        contextLength: 131072,
-        tokenizerFamily: "eliza1",
-        companionModelIds: ["eliza-1-pro-27b-drafter"],
-        runtime: runtimeFor("eliza-1-pro-27b", 131072),
-        blurb: "Eliza-1 pro — 96 GB+ Mac and high-VRAM desktop default. Fused text + voice + vision; longest-context Eliza-1 tier on workstation hardware.",
-    },
-    drafterCompanion({
-        id: "eliza-1-pro-27b",
-        displayName: "Eliza-1 pro",
-        ggufFile: "dflash/drafter-pro-27b.gguf",
-        params: "9B",
-        sizeGb: 1.2,
-        minRamGb: 32,
-        bucket: "large",
-    }),
-    // ─── Eliza-1 server (workstation / server) ──────────────────────────
-    {
-        id: "eliza-1-server-h200",
-        displayName: "Eliza-1 server",
-        hfRepo: "elizalabs/eliza-1-server-h200",
-        ggufFile: "text/eliza-1-server-h200-256k.gguf",
-        bundleManifestFile: "eliza-1.manifest.json",
-        params: "27B",
-        quant: "Eliza-1 optimized local runtime",
-        sizeGb: 16.8,
-        minRamGb: 96,
-        category: "chat",
-        bucket: "large",
-        contextLength: 262144,
-        tokenizerFamily: "eliza1",
-        companionModelIds: ["eliza-1-server-h200-drafter"],
-        runtime: runtimeFor("eliza-1-server-h200", 262144),
-        blurb: "Eliza-1 server — H200-class workstation / server tier with the largest context window in the line.",
-    },
-    drafterCompanion({
-        id: "eliza-1-server-h200",
-        displayName: "Eliza-1 server",
-        ggufFile: "dflash/drafter-server-h200.gguf",
-        params: "9B",
-        sizeGb: 1.2,
-        minRamGb: 96,
-        bucket: "large",
-    }),
-];
-export function findCatalogModel(id) {
-    return MODEL_CATALOG.find((m) => m.id === id);
-}
-/**
- * Construct the HuggingFace resolve URL for a given catalog entry.
- *
- * Respects `ELIZA_HF_BASE_URL` when set so self-hosted HF mirrors and the
- * downloader e2e test suite can redirect all downloads without touching
- * the catalog.
- */
-export function buildHuggingFaceResolveUrlForPath(model, filePath) {
-    const base = process.env.ELIZA_HF_BASE_URL?.trim().replace(/\/+$/, "") ||
-        "https://huggingface.co";
-    // Encode each path segment separately so nested bundle layouts like
-    // `text/eliza-1-mobile-1_7b-32k.gguf` keep their slashes.
-    const encodedPath = filePath
-        .split("/")
-        .map((segment) => encodeURIComponent(segment))
-        .join("/");
-    return `${base}/${model.hfRepo}/resolve/main/${encodedPath}?download=true`;
-}
-export function buildHuggingFaceResolveUrl(model) {
-    return buildHuggingFaceResolveUrlForPath(model, model.ggufFile);
-}

package/services/local-inference/conversation-registry.d.ts

@@ -1,122 +0,0 @@
-/**
- * Conversation registry for the local-inference path.
- *
- * Today's slot allocation is purely a hash function: `deriveSlotId` maps a
- * `promptCacheKey` (or any stable string) to `slot_id` in `[0, parallel)`.
- * That works for one-shot calls but breaks for long agentic loops:
- *
- *   - Two distinct conversations whose cache keys hash to the same slot
- *     evict each other's KV every turn (slot thrashing).
- *   - The current high-water mark of concurrent conversations is invisible,
- *     so `--parallel N` cannot be tuned to fit.
- *   - There is no notion of an explicit "I am still using this slot" lease,
- *     so eviction is purely best-effort.
- *
- * This registry keeps a per-conversation reservation. `openConversation`
- * picks the lowest-loaded slot and pins the conversation to it; subsequent
- * `generateInConversation` calls always land on the same slot. When the
- * pool is full, slot reuse falls back to the same-as-before hash policy
- * (two leases on the same slot still serialise correctly via the dispatcher's
- * generation queue).
- *
- * The registry tracks the high-water mark of concurrently-open conversations
- * so the engine can warn (or, in the future, restart llama-server with a
- * higher --parallel) when the load outgrows the configured slot count.
- */
-/**
- * Opaque handle returned by `openConversation`. Callers MUST treat this as
- * opaque — the registry owns the slot id and lifetime.
- */
-export interface ConversationHandle {
-    readonly conversationId: string;
-    readonly modelId: string;
-    /**
-     * Pinned slot id in `[0, parallel)`, or `-1` when slot pinning is disabled
-     * (parallel <= 0). Used by both backends as the cache key:
-     *   - llama-server: forwarded as `slot_id` in the request payload.
-     *   - node-llama-cpp: combined with the conversation id to derive the
-     *     session-pool key so identical conversations share a session.
-     */
-    readonly slotId: number;
-    /** Wall-clock ms when the handle was opened. */
-    readonly openedAtMs: number;
-    /** Wall-clock ms when the handle was last touched (open or generate). */
-    lastUsedMs: number;
-    /** TTL after which the registry MAY auto-close on the next sweep. */
-    readonly ttlMs: number;
-    /** True when `closeConversation` has been called; further use is rejected. */
-    closed: boolean;
-}
-export interface OpenConversationArgs {
-    conversationId: string;
-    modelId: string;
-    /** Slot count from the running server (`--parallel N`). Defaults to 1. */
-    parallel?: number;
-    /**
-     * TTL after which the handle is considered idle and may be auto-closed
-     * by `evictIdle`. Defaults to 60 minutes — long enough for an LLM call
-     * to finish even on a slow drafter, short enough to recover from forgotten
-     * close calls within the long-cache window.
-     */
-    ttlMs?: number;
-}
-/**
- * In-memory registry of open conversation handles. A single instance is
- * shared by the engine; each backend reads from it on every generate to
- * decide which slot to pin to.
- */
-export declare class ConversationRegistry {
-    private readonly handles;
-    /** Per-slot reference count; lowest-loaded slot wins on next open. */
-    private readonly slotLoad;
-    /** Largest concurrent open count seen; the engine reads this for parallel auto-tune. */
-    private highWaterMark;
-    /**
-     * Lookup / open a conversation handle. Idempotent for the same
-     * conversation id + model id; callers can call this on every turn
-     * without leaking handles. When the call is reusing an existing handle,
-     * `lastUsedMs` is bumped for LRU-style eviction tracking.
-     */
-    open(args: OpenConversationArgs): ConversationHandle;
-    /**
-     * Lookup an open handle by conversation+model. Returns null when the
-     * conversation has not been opened or has already been closed. Bumps
-     * `lastUsedMs` so an LRU sweep treats reads as activity.
-     */
-    get(conversationId: string, modelId: string): ConversationHandle | null;
-    /**
-     * Close + drop a handle. Idempotent — closing an unknown / already-closed
-     * handle is a no-op so callers can call this from cleanup paths
-     * unconditionally.
-     */
-    close(conversationId: string, modelId: string): void;
-    /**
-     * Sweep handles whose `lastUsedMs` is older than their TTL. Returns the
-     * conversation ids dropped so callers can persist final KV state to
-     * disk, etc. Safe to call on a timer.
-     */
-    evictIdle(now?: number): string[];
-    /**
-     * Snapshot every currently-open handle. Used by the shutdown path to
-     * emit a save-state request per slot.
-     */
-    snapshot(): readonly ConversationHandle[];
-    /** Largest concurrent open count seen since the registry was created. */
-    highWater(): number;
-    /** Number of currently-open handles. */
-    size(): number;
-    /**
-     * Pick the slot with the fewest in-flight handles. Ties are broken by a
-     * deterministic hash of the conversation id, which avoids consistently
-     * loading slot 0 when N concurrent opens race.
-     */
-    private pickLowestLoadedSlot;
-    private compositeKey;
-}
-/**
- * Module-singleton registry. The engine reads this on every generate; the
- * conversation lifecycle API (`openConversation`, `closeConversation`)
- * mutates it.
- */
-export declare const conversationRegistry: ConversationRegistry;
-//# sourceMappingURL=conversation-registry.d.ts.map

package/services/local-inference/conversation-registry.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"conversation-registry.d.ts","sourceRoot":"","sources":["../../../../../../src/services/local-inference/conversation-registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAIH;;;GAGG;AACH,MAAM,WAAW,kBAAkB;IACjC,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB;;;;;;OAMG;IACH,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,gDAAgD;IAChD,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,yEAAyE;IACzE,UAAU,EAAE,MAAM,CAAC;IACnB,qEAAqE;IACrE,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,8EAA8E;IAC9E,MAAM,EAAE,OAAO,CAAC;CACjB;AAED,MAAM,WAAW,oBAAoB;IACnC,cAAc,EAAE,MAAM,CAAC;IACvB,OAAO,EAAE,MAAM,CAAC;IAChB,0EAA0E;IAC1E,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;;;OAKG;IACH,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAID;;;;GAIG;AACH,qBAAa,oBAAoB;IAC/B,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAyC;IACjE,sEAAsE;IACtE,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAA6B;IACtD,wFAAwF;IACxF,OAAO,CAAC,aAAa,CAAK;IAE1B;;;;;OAKG;IACH,IAAI,CAAC,IAAI,EAAE,oBAAoB,GAAG,kBAAkB;IAqCpD;;;;OAIG;IACH,GAAG,CAAC,cAAc,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,kBAAkB,GAAG,IAAI;IAOvE;;;;OAIG;IACH,KAAK,CAAC,cAAc,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,IAAI;IAcpD;;;;OAIG;IACH,SAAS,CAAC,GAAG,GAAE,MAAmB,GAAG,MAAM,EAAE;IAkB7C;;;OAGG;IACH,QAAQ,IAAI,SAAS,kBAAkB,EAAE;IAIzC,yEAAyE;IACzE,SAAS,IAAI,MAAM;IAInB,wCAAwC;IACxC,IAAI,IAAI,MAAM;IAId;;;;OAIG;IACH,OAAO,CAAC,oBAAoB;IAqB5B,OAAO,CAAC,YAAY;CAGrB;AAED;;;;GAIG;AACH,eAAO,MAAM,oBAAoB,sBAA6B,CAAC"}

package/services/local-inference/conversation-registry.js

@@ -1,182 +0,0 @@
-/**
- * Conversation registry for the local-inference path.
- *
- * Today's slot allocation is purely a hash function: `deriveSlotId` maps a
- * `promptCacheKey` (or any stable string) to `slot_id` in `[0, parallel)`.
- * That works for one-shot calls but breaks for long agentic loops:
- *
- *   - Two distinct conversations whose cache keys hash to the same slot
- *     evict each other's KV every turn (slot thrashing).
- *   - The current high-water mark of concurrent conversations is invisible,
- *     so `--parallel N` cannot be tuned to fit.
- *   - There is no notion of an explicit "I am still using this slot" lease,
- *     so eviction is purely best-effort.
- *
- * This registry keeps a per-conversation reservation. `openConversation`
- * picks the lowest-loaded slot and pins the conversation to it; subsequent
- * `generateInConversation` calls always land on the same slot. When the
- * pool is full, slot reuse falls back to the same-as-before hash policy
- * (two leases on the same slot still serialise correctly via the dispatcher's
- * generation queue).
- *
- * The registry tracks the high-water mark of concurrently-open conversations
- * so the engine can warn (or, in the future, restart llama-server with a
- * higher --parallel) when the load outgrows the configured slot count.
- */
-import { createHash } from "node:crypto";
-const DEFAULT_HANDLE_TTL_MS = 60 * 60 * 1000;
-/**
- * In-memory registry of open conversation handles. A single instance is
- * shared by the engine; each backend reads from it on every generate to
- * decide which slot to pin to.
- */
-export class ConversationRegistry {
-    handles = new Map();
-    /** Per-slot reference count; lowest-loaded slot wins on next open. */
-    slotLoad = new Map();
-    /** Largest concurrent open count seen; the engine reads this for parallel auto-tune. */
-    highWaterMark = 0;
-    /**
-     * Lookup / open a conversation handle. Idempotent for the same
-     * conversation id + model id; callers can call this on every turn
-     * without leaking handles. When the call is reusing an existing handle,
-     * `lastUsedMs` is bumped for LRU-style eviction tracking.
-     */
-    open(args) {
-        if (!args.conversationId) {
-            throw new Error("[conversation-registry] conversationId is required");
-        }
-        if (!args.modelId) {
-            throw new Error("[conversation-registry] modelId is required");
-        }
-        const compositeKey = this.compositeKey(args.conversationId, args.modelId);
-        const existing = this.handles.get(compositeKey);
-        if (existing && !existing.closed) {
-            existing.lastUsedMs = Date.now();
-            return existing;
-        }
-        const parallel = typeof args.parallel === "number" && args.parallel > 0
-            ? Math.floor(args.parallel)
-            : 1;
-        const slotId = this.pickLowestLoadedSlot(parallel, args.conversationId);
-        const now = Date.now();
-        const handle = {
-            conversationId: args.conversationId,
-            modelId: args.modelId,
-            slotId,
-            openedAtMs: now,
-            lastUsedMs: now,
-            ttlMs: args.ttlMs ?? DEFAULT_HANDLE_TTL_MS,
-            closed: false,
-        };
-        this.handles.set(compositeKey, handle);
-        this.slotLoad.set(slotId, (this.slotLoad.get(slotId) ?? 0) + 1);
-        if (this.handles.size > this.highWaterMark) {
-            this.highWaterMark = this.handles.size;
-        }
-        return handle;
-    }
-    /**
-     * Lookup an open handle by conversation+model. Returns null when the
-     * conversation has not been opened or has already been closed. Bumps
-     * `lastUsedMs` so an LRU sweep treats reads as activity.
-     */
-    get(conversationId, modelId) {
-        const handle = this.handles.get(this.compositeKey(conversationId, modelId));
-        if (!handle || handle.closed)
-            return null;
-        handle.lastUsedMs = Date.now();
-        return handle;
-    }
-    /**
-     * Close + drop a handle. Idempotent — closing an unknown / already-closed
-     * handle is a no-op so callers can call this from cleanup paths
-     * unconditionally.
-     */
-    close(conversationId, modelId) {
-        const compositeKey = this.compositeKey(conversationId, modelId);
-        const handle = this.handles.get(compositeKey);
-        if (!handle)
-            return;
-        handle.closed = true;
-        this.handles.delete(compositeKey);
-        const remaining = (this.slotLoad.get(handle.slotId) ?? 0) - 1;
-        if (remaining <= 0) {
-            this.slotLoad.delete(handle.slotId);
-        }
-        else {
-            this.slotLoad.set(handle.slotId, remaining);
-        }
-    }
-    /**
-     * Sweep handles whose `lastUsedMs` is older than their TTL. Returns the
-     * conversation ids dropped so callers can persist final KV state to
-     * disk, etc. Safe to call on a timer.
-     */
-    evictIdle(now = Date.now()) {
-        const dropped = [];
-        for (const [compositeKey, handle] of this.handles) {
-            if (now - handle.lastUsedMs > handle.ttlMs) {
-                handle.closed = true;
-                this.handles.delete(compositeKey);
-                const remaining = (this.slotLoad.get(handle.slotId) ?? 0) - 1;
-                if (remaining <= 0) {
-                    this.slotLoad.delete(handle.slotId);
-                }
-                else {
-                    this.slotLoad.set(handle.slotId, remaining);
-                }
-                dropped.push(handle.conversationId);
-            }
-        }
-        return dropped;
-    }
-    /**
-     * Snapshot every currently-open handle. Used by the shutdown path to
-     * emit a save-state request per slot.
-     */
-    snapshot() {
-        return [...this.handles.values()];
-    }
-    /** Largest concurrent open count seen since the registry was created. */
-    highWater() {
-        return this.highWaterMark;
-    }
-    /** Number of currently-open handles. */
-    size() {
-        return this.handles.size;
-    }
-    /**
-     * Pick the slot with the fewest in-flight handles. Ties are broken by a
-     * deterministic hash of the conversation id, which avoids consistently
-     * loading slot 0 when N concurrent opens race.
-     */
-    pickLowestLoadedSlot(parallel, conversationId) {
-        if (parallel <= 1)
-            return 0;
-        let bestSlot = 0;
-        let bestLoad = Number.POSITIVE_INFINITY;
-        for (let slot = 0; slot < parallel; slot += 1) {
-            const load = this.slotLoad.get(slot) ?? 0;
-            if (load < bestLoad) {
-                bestLoad = load;
-                bestSlot = slot;
-            }
-        }
-        if (bestLoad === 0)
-            return bestSlot;
-        // All slots are loaded equally — use the conversation hash for a
-        // deterministic tie-break. Same conversation, same slot when reopened.
-        const digest = createHash("sha256").update(conversationId).digest();
-        return digest.readUInt32BE(0) % parallel;
-    }
-    compositeKey(conversationId, modelId) {
-        return `${modelId}::${conversationId}`;
-    }
-}
-/**
- * Module-singleton registry. The engine reads this on every generate; the
- * conversation lifecycle API (`openConversation`, `closeConversation`)
- * mutates it.
- */
-export const conversationRegistry = new ConversationRegistry();