@urun-sh/openai 0.5.4 → 0.5.5

package/dist/responses-turn-OrO4euEN.d.cts

@@ -0,0 +1,513 @@
+import { C as CatalogRow } from './models-NYMZrklp.cjs';
+import { IncomingMessage } from 'node:http';
+import { a as AudioBridge, h as VideoFrameLane, j as VideoOutLane } from './video-out-D20UuJ8G.cjs';
+
+/**
+ * WHAT a proxy serves — the `identity` block on `GET /stats`. Reuse of a
+ * running proxy is allowed iff every field matches the launching invocation
+ * EXACTLY: a partial match (same app, different fn; same everything, older
+ * proxy_version) silently routes the agent at the wrong backend, which is
+ * worse than any error.
+ */
+interface ProxyIdentity {
+    app: string;
+    org: string;
+    fn: string;
+    /**
+     * The control-plane URL the proxy's backhaul session opens against
+     * (URUN_BASE_URL verbatim — same org/app/fn on staging vs prod are
+     * DIFFERENT backends; review finding, #285). Compared exactly: a cosmetic
+     * difference (trailing slash) merely refuses reuse and spawns an ephemeral
+     * proxy — the safe direction.
+     */
+    base_url: string;
+    proxy_version: string;
+}
+
+/**
+ * THE OPENROUTER PROVIDER DOCUMENT — `GET /v1/models?format=openrouter`.
+ *
+ * OpenRouter's provider monitor polls this document (schema 2.4, per
+ * openrouter.ai/docs/guides/community/for-providers) to list a provider's
+ * models on the marketplace. The `id` field is the EXACT model identifier
+ * OpenRouter sends back as `model` on chat completions — which here is the
+ * app slug, the same id the OpenAI surface lists. One id space, two views.
+ *
+ * MODALITY GATING: only rows whose catalog `task` names a chat-completions
+ * shape (chat | code | agent | vl) are declared. A voice/video/image app
+ * behind a chat surface would stream garbage into OpenRouter's baseline
+ * tests; models the catalog cannot vouch for are OMITTED, never guessed.
+ * With no catalog oracle configured the document is `{"data": []}` — honest
+ * emptiness, not invented capability (the chart turns the oracle on).
+ *
+ * PRICING is deliberately omitted entirely: the schema's rule is "a modality
+ * with no pricing array is simply unpriced", OpenRouter configures pricing
+ * during provider onboarding, and the per-model price surface is
+ * urun-infra U3 (model_prices) — when that lands, this module grows a
+ * `prices` input and emits the input/output pricing arrays.
+ *
+ * The full closed-value-domain schema ships as OpenAPI 3.1 at
+ * openrouter.ai/docs/assets/provider-monitor-schema-v2.openapi.json.
+ */
+
+type OpenRouterProviderDoc = {
+    data: OpenRouterProviderModel[];
+};
+interface OpenRouterProviderModel {
+    schema_version: '2.4';
+    /** The EXACT id OpenRouter sends back as `model` — the app slug. */
+    id: string;
+    name: string;
+    created: number;
+    /** Valid enum: int4|int8|fp4|mxfp4|nvfp4|fp6|fp8|mxfp8|fp16|bf16|fp32|null. */
+    quantization: string | null;
+    description: string;
+    hugging_face_id: string;
+    input_modalities: Array<Record<string, unknown>>;
+    output_modalities: Array<Record<string, unknown>>;
+}
+
+/**
+ * Per-model → per-app routing for the compat proxy (owner directive
+ * 2026-08-07): swapping the model in a coding harness routes the request to
+ * the org's DEPLOYED app for that model. v1 is deployed-only — a uRun model
+ * that is not deployed gets a loud 404 naming `urun serve <id>`; a later
+ * phase (explicitly out of scope here; urun-infra#1490 shared-endpoints)
+ * auto-creates from the model catalog on first request.
+ *
+ * MODEL-ID SURFACE (the documented mapping): a model may be named by
+ *   - the app slug itself        ("qwen3-6-27b-bf16"), or
+ *   - the catalog id             ("qwen3.6-27b"), or
+ *   - the catalog id:variant     ("qwen3.6-27b:bf16"),
+ * where slugification mirrors urun-cli `serve.py _default_app_name` exactly:
+ * lowercase, every non-alphanumeric-non-dash character becomes "-", leading/
+ * trailing dashes stripped (catalog id "qwen3.6-27b" + variant "bf16" → app
+ * "qwen3-6-27b-bf16"). COLLISION RULE: an exact slug match always wins over
+ * the catalog-id (prefix) interpretation.
+ *
+ * RESOLUTION ORDER (one canonical path, documented end to end):
+ *   1. model absent / "urun" / an alias of the startup app  → DEFAULT app.
+ *   2. exact slug match on a deployed serve app             → that app.
+ *   3. catalog-id form matching exactly one deployed app    → that app
+ *      (two or more candidates → loud ambiguity error naming them).
+ *   4. the name maps to an org app that is NOT an active app exposing the
+ *      proxy's serve function                               → loud 404
+ *      naming `urun serve <model>` and the available models.
+ *   5. the name matches a catalog row but no deployed app   → loud 404
+ *      naming `urun serve <id>` (deployed-only v1).
+ *   6. anything else — a model name outside the uRun namespace entirely
+ *      (e.g. the harness's own upstream default, "claude-*"/"gpt-*") →
+ *      DEFAULT app. This IS today's single-app contract, kept deliberately
+ *      so `urun compat <agent>` with the agent's stock model keeps working
+ *      with zero new env; the per-model /stats table records every such
+ *      mapping so it is visible, never silent. Models the proxy ADVERTISES
+ *      on /v1/models can never land here — they resolve (2/3) or fail loud
+ *      (4/5) above.
+ *
+ * NO-DEFAULT MODE (`defaultApp: null`) — the HOSTED multi-tenant lane
+ * (`src/hosted/`): a shared endpoint serving every org has no "the app this
+ * proxy was started for", so rules 1 and 6 have nothing to fall back TO.
+ * Rather than inventing one (picking "some" app for a caller would be the
+ * worst kind of silent divergence), both rules become the SAME loud
+ * {@link UnknownModelError} that rules 4/5 already raise: name a deployed
+ * model, here is the list. Rules 2–5 are byte-for-byte the local behavior —
+ * one router, one resolution order, two configurations.
+ */
+
+/** One org app row from `GET {orgApi}/apps` (urun-cli `ApiClient.list_apps`). */
+interface DeployedApp {
+    app_slug: string;
+    function_name?: string | null;
+    deployment_status?: string | null;
+    [k: string]: unknown;
+}
+/**
+ * A model that does not resolve to a deployed app the proxy may serve.
+ * Rendered by the server in each lane's NATIVE error format as a 404 — never
+ * silently served by the default app.
+ */
+declare class UnknownModelError extends Error {
+}
+/**
+ * A session handle names a pooled session that no longer exists (closed,
+ * evicted after its pod died, replaced by a re-home, or the proxy restarted).
+ * The caller asked to REATTACH that exact session — opening a fresh one and
+ * calling it "resumed" would be a silent lie, so this is always loud.
+ */
+declare class SessionGoneError extends Error {
+}
+/**
+ * OpenAI-shaped model list (same shape as models.ts listModels). When the
+ * catalog oracle is configured, each entry ALSO carries the catalog
+ * enrichment fields (additive JSON — OpenAI clients ignore unknown fields;
+ * the OpenRouter + Vercel AI Gateway provider listings need them). Entries
+ * whose slug matches no catalog row stay at the bare four fields.
+ */
+interface RouterModelEntry {
+    id: string;
+    object: 'model';
+    created: number;
+    owned_by: string;
+    /** Display name — the canonical `<model_id>:<variant>` catalog ref. */
+    name?: string;
+    /** User-facing description from the catalog (console Endpoints copy). */
+    description?: string;
+    /** Catalog modality (chat | code | agent | vl | audio | image | ...). */
+    task?: string;
+    /** Serving engine (vllm | sglang | llamacpp | ...). */
+    engine?: string;
+    /** The catalog lane this app deploys, e.g. 'rtx6000:1'. */
+    gpu_spec?: string;
+    /** Context window in tokens (chat rows carry it in engine_args). */
+    context_length?: number;
+}
+interface RouterModelList {
+    object: 'list';
+    data: RouterModelEntry[];
+}
+interface ModelRouterOptions<S> {
+    /**
+     * The startup app slug (URUN_APP) — the DEFAULT model — or `null` for the
+     * hosted multi-tenant lane, which has no per-proxy default app: there,
+     * every request must NAME a deployed model and an unnamed/unknown one
+     * fails loud instead of silently landing somewhere (see the module header,
+     * "NO-DEFAULT MODE").
+     */
+    defaultApp: string | null;
+    /** The serve function name every routed app must expose (URUN_FUNCTION). */
+    fnName: string;
+    /** Open a backhaul session for an app slug (called at most once per app). */
+    openSession: (appSlug: string) => S | Promise<S>;
+    /** Terminal release for one pool entry (Session.end() underneath). */
+    closeSession: (entry: S) => Promise<void>;
+    /**
+     * List the org's deployed apps, or null when the credentials cannot
+     * (URUN_JWT lane: the pre-vended token is scoped to the default app, so
+     * there is no org listing AND no cross-app session — routing degrades to
+     * the default-app-only contract, which is exactly today's behavior).
+     */
+    listApps: (() => Promise<DeployedApp[]>) | null;
+    /**
+     * Catalog rows (models.ts fetchCatalogRows) as the uRun-namespace oracle
+     * for rule 5 AND the /v1/models enrichment + OpenRouter provider-doc
+     * source, or null when catalog access is not configured. Optional fields
+     * beyond model_id/variant are tolerated (thin rows still typecheck).
+     */
+    listCatalog: (() => Promise<CatalogRow[]>) | null;
+    /** Deployed-apps cache TTL (the list changes on deploys, not per request). */
+    appsTtlMs?: number;
+    /**
+     * The stable NATIVE identity of one pooled entry (the uRun session id in
+     * the proxy wiring — the same identity the serve-side session-affinity tag
+     * rides, urun-python#1556). Powers the session-identity seam
+     * ({@link ModelRouter.handleFor} / {@link ModelRouter.sessionForHandle});
+     * a router without it fails LOUD on those calls, never approximates.
+     */
+    sessionKey?: (entry: S) => string;
+}
+/**
+ * The image-capability modes the Images lane gates on (images.ts's
+ * ImageCapabilityMode — declared here because the router is the LOWER layer:
+ * the lane's gate type stays structurally satisfied by
+ * {@link ModelRouter.supportsImages} with no router→lane import).
+ */
+type ImageCapabilityMode = 'generation' | 'edit';
+/**
+ * The session pool: one backhaul session per deployed app, keyed by app slug,
+ * opened lazily on the first request that routes to it and reused for every
+ * subsequent one. The startup app is seeded eagerly by the CLI. Sessions
+ * close on proxy shutdown via {@link closeAll}; there is NO idle-close policy
+ * (deliberate v1 simplification — noted as a follow-up in the PR).
+ */
+declare class ModelRouter<S> {
+    private readonly opts;
+    private readonly pool;
+    private appsCache;
+    private rowsCache;
+    constructor(opts: ModelRouterOptions<S>);
+    /** Seed an already-open session (the CLI's eagerly-opened startup app). */
+    seed(appSlug: string, entry: S): void;
+    private deployedApps;
+    /**
+     * Catalog rows with the same TTL discipline as {@link deployedApps} (the
+     * catalog changes on reseed migrations, not per request). Null when no
+     * oracle is configured — callers degrade to the bare four-field entries.
+     */
+    private catalogRows;
+    /** Apps this proxy may serve: active AND exposing the serve function. */
+    private servable;
+    private availableIds;
+    /**
+     * NO-DEFAULT MODE's terminal for rules 1 and 6: there is no app to fall
+     * back to, so say so loudly and list what the CALLER'S org actually has.
+     * Never returns.
+     */
+    private noDefaultApp;
+    /**
+     * Resolve a request's `model` to an app slug — the documented resolution
+     * order from the module header. Throws {@link UnknownModelError} for a uRun
+     * model that is not deployed (rules 4/5).
+     */
+    resolveApp(model: string | undefined): Promise<string>;
+    /** The pooled session for a model — opened lazily, reused afterwards. */
+    sessionFor(model: string | undefined): Promise<{
+        app: string;
+        entry: S;
+    }>;
+    private keyOf;
+    /**
+     * SESSION-IDENTITY SEAM (a): the opaque stable handle for the pooled
+     * session currently serving `model`'s turns. Rides the SAME acquisition
+     * path as every request ({@link sessionFor}) — the session opens lazily if
+     * this model has none yet — and derives the handle from native identity
+     * (app slug + uRun session id), zero bespoke bookkeeping.
+     */
+    handleFor(model: string | undefined): Promise<{
+        app: string;
+        handle: string;
+    }>;
+    /**
+     * SESSION-IDENTITY SEAM (b): the exact pooled session a handle names.
+     * NEVER opens a fresh session — a handle whose session is gone (closed,
+     * evicted, re-homed to a replacement, proxy restarted) or malformed throws
+     * {@link SessionGoneError} loudly. Resume is reattach-or-fail, not
+     * reattach-or-quietly-restart.
+     */
+    sessionForHandle(handle: string): Promise<{
+        app: string;
+        entry: S;
+    }>;
+    /**
+     * Drop ONE pooled session whose backhaul died (its pod was restarted /
+     * drained / deleted) and release it — the next {@link sessionFor} opens a
+     * fresh one, i.e. asks the control plane for a new assignment. Used by the
+     * one-shot re-home (rehome.ts, urun-sh/urun-python#1592).
+     *
+     * IDENTITY-GUARDED (the same rule the pi lane's SessionPool follows): a
+     * concurrent request that already re-homed this app has put a NEWER entry
+     * under the key, and evicting that would close a healthy session out from
+     * under it.
+     */
+    evict(app: string, entry: S): Promise<void>;
+    /**
+     * `GET /v1/models`: the org's deployed serve apps as model entries, the
+     * default app FIRST. On the JWT lane (no org listing) this is the default
+     * app plus any app already in the pool — the gap is called out loudly in
+     * the PR, not papered over here.
+     */
+    modelList(): Promise<RouterModelList>;
+    /**
+     * The OpenRouter PROVIDER document (`GET /v1/models?format=openrouter`):
+     * schema 2.4 per openrouter.ai/docs/guides/community/for-providers. Only
+     * chat-completions-shaped models are declared (task chat | code | agent |
+     * vl — a voice or video app behind a chat surface would stream garbage);
+     * models the catalog cannot vouch for are OMITTED, never guessed. Pricing
+     * is deliberately omitted entirely (schema rule: "A modality with no
+     * pricing array is simply unpriced" — OpenRouter configures pricing during
+     * provider onboarding; the per-model price surface is urun-infra U3).
+     */
+    openRouterModels(): Promise<OpenRouterProviderDoc>;
+    /**
+     * IMAGE CAPABILITY GATE (the Images lane's `imageModels` oracle, images.ts):
+     * is `model`'s deployed app a native destination for `mode`? Authorized
+     * resolution ({@link resolveApp} — undeployed uRun models throw
+     * {@link UnknownModelError}, which the lane renders as a 404) then the
+     * SAME catalog join as the /v1/models enrichment (PR421's rowForSlug):
+     * exact `<model_id>-<variant>` slug across ALL its GPU-placement rows, a
+     * bare model_id only when one variant owns it.
+     *
+     * Capability is read ONLY from catalog modality metadata, never from the
+     * serve function name: every row must carry task 'image' INVARIANTLY across
+     * the placements, and the mode comes from engine_args.expects_image —
+     * false ⇒ pure text-to-image ('generation'; diffusers v0.38.0
+     * pipeline_qwenimage_edit_plus raises torch.cat on an EMPTY image list, so
+     * an edit destination is never generation-capable); true or ABSENT (the
+     * native default) ⇒ 'edit'. The consensus rule is per-placement:
+     * capability must agree across every placement — a CONFLICTING flag (mixed
+     * true/false/absent) or a non-boolean one is a refused capability, never
+     * silently defaulted to edit.
+     *
+     * A CONFIGURED catalog oracle that FAILS propagates its rejection — the
+     * gate never fabricates a `false` (which would render as a misleading 404)
+     * out of an infrastructure outage. No oracle configured, no matching row,
+     * or a slug the catalog cannot vouch for ⇒ false (loud 404 upstream).
+     */
+    supportsImages(model: string | undefined, mode: ImageCapabilityMode): Promise<boolean>;
+    /** Close every pooled session (Session.end() underneath) — proxy shutdown. */
+    closeAll(): Promise<void>;
+}
+
+/**
+ * The shared Responses execution/store seam — the ONE execution + event-
+ * envelope path and the authorized tenant store, imported by BOTH the HTTP
+ * SSE lane (server.ts) and the `/v1/responses` WebSocket lane
+ * (responses-ws.ts). This module deliberately imports NO transport acceptor
+ * and NO HTTP lane module: its only imports are type-only, so the module
+ * graph server.ts → responses-ws.ts → responses-turn.ts is acyclic and no
+ * Vitest/Bun/Node module-evaluation order can observe an uninitialized
+ * binding (the constructor-failure class of bug this extraction removes).
+ */
+
+/** The request envelope every Responses-shaped upstream call carries. */
+interface ResponsesCreateParams {
+    model?: string;
+    input: unknown;
+    stream?: boolean;
+    tools?: unknown;
+    tool_choice?: unknown;
+    temperature?: number;
+    max_output_tokens?: number;
+    /**
+     * System/developer instructions, forwarded to the serve envelope — where
+     * they become ONE prepended `system`-role message (the protocol's native
+     * per-request system input; there is NO envelope-level instructions field).
+     * Typed `string` end to end (transport/encode.ts) so no cast bridges the seam.
+     */
+    instructions?: string;
+    top_p?: number;
+    stop?: string[];
+    /**
+     * Reasoning controls (urun-python #1667): forwarded VERBATIM to the serve
+     * envelope — `chat_template_kwargs.enable_thinking:false` is the think-off
+     * switch. Absent -> absent (no default injection; server-side validation).
+     */
+    reasoning_effort?: string;
+    chat_template_kwargs?: Record<string, unknown>;
+}
+/** How one pinned session ended (the native phase machinery's terminal step). */
+interface SessionEndInfo {
+    /**
+     * Milliseconds until the session's native deadline (`endsAt`), or null when
+     * the app declared no maximum session length. NOTE (missing primitive,
+     * called out in the PR): core exposes no PRE-expiry notice event — this
+     * callback fires AT terminal loss, so timeLeftMs is ~0 on expiry.
+     */
+    timeLeftMs: number | null;
+    /** The terminal reason (expired / ended / error), for the loud close. */
+    reason: string;
+}
+/** The upstream calls the proxy makes — injectable (tests; alt transports). */
+interface ProxyClients {
+    /** `UrunResponses(session).responses.create` — an async iterable of Responses stream events. */
+    createResponse(params: ResponsesCreateParams): Promise<AsyncIterable<unknown>> | AsyncIterable<unknown>;
+    /** `listModels(...)` result (an OpenAI model list object). */
+    listModels(): Promise<unknown>;
+    /**
+     * The OpenRouter provider document (`GET /v1/models?format=openrouter`).
+     * Optional at the seam: hand-built test clients may omit it, in which case
+     * the format=openrouter branch answers 501 — loud, never a silent empty.
+     */
+    openRouterModels?(): Promise<unknown>;
+    /**
+     * Open (or reuse) the NATIVE audio lanes on the pooled session `model` routes
+     * to — the same first-party machinery as RealtimeClient.enableAudio
+     * (transport/media.ts `enableSessionAudio`: AudioBridge over the session's
+     * `rt-audio-in`/`rt-audio-out` stream lanes at 24 kHz). One lane per pooled
+     * session; repeat calls return the same lane. Optional at the seam because
+     * text-only embeddings exist — but a surface that RECEIVES audio while the
+     * embedder wired no `openAudio` must fail LOUD, never drop chunks.
+     */
+    openAudio?(model: string | undefined): Promise<ProxyAudioLane>;
+    /**
+     * Open (or reuse) the NATIVE video FRAME lane on the pooled session `model`
+     * routes to (transport/media.ts `enableSessionVideo`: discrete JPEG frames →
+     * `stream('rt-video-in').emit`, the §5 named-DATA image-bytes path — NOT the
+     * RTP media plane, which requires an already-H.264-encoded track). One lane
+     * per pooled session; repeat calls return the same lane. Optional at the seam
+     * because text-only embeddings exist — but a surface that RECEIVES video
+     * frames while the embedder wired no `openVideo` must fail LOUD, never drop
+     * frames.
+     */
+    openVideo?(model: string | undefined): Promise<ProxyVideoLane>;
+    /**
+     * Open (or reuse) the NATIVE video OUTPUT lane on the pooled session
+     * `model` routes to (transport/video-out.ts `openSessionVideoOutLane`):
+     * decoded spec-v1 records consumed from the session's named §5
+     * `rt-video-out` downstream. The Live adapter fans each record onto the
+     * SAME Bidi WS as an extension server message (spec v2, WS-inline) — no
+     * side-channel transport. Optional at the seam; an ABSENT seam means the
+     * `urun.videoOut` capability is refused by omission in setupComplete
+     * (vanilla parity), while a wired seam that cannot open the lane must
+     * throw LOUD — never a quiet downgrade.
+     */
+    openVideoOut?(model: string | undefined): Promise<ProxyVideoOutLane>;
+    /**
+     * SESSION-IDENTITY SEAM (ModelRouter.handleFor): the opaque stable handle
+     * for the pooled session currently serving `model`'s turns. Derived from
+     * native identity (app slug + uRun session id) — the same identity the
+     * serve-side session-affinity tag rides (urun-python#1556/#1582).
+     */
+    sessionHandle(model: string | undefined): Promise<string>;
+    /**
+     * createResponse PINNED to the exact session a handle names
+     * (ModelRouter.sessionForHandle). Throws SessionGoneError LOUDLY when that
+     * session is gone or was replaced — never silently opens a fresh session
+     * while claiming resume. Deliberately NO re-home on this path: re-homing
+     * would swap the pinned session out from under the caller.
+     */
+    createResponseOn(handle: string, params: ResponsesCreateParams): Promise<AsyncIterable<unknown>> | AsyncIterable<unknown>;
+    /**
+     * Subscribe to the pinned session's terminal end via core's NATIVE phase
+     * machinery (Session.onPhase → terminal 'expired'/'ended'/'error'). Fires
+     * `cb` once. Throws SessionGoneError if the handle's session is already
+     * gone — which doubles as the loud reattach check at resume time. Returns
+     * the unsubscribe.
+     */
+    onSessionEnd(handle: string, cb: (end: SessionEndInfo) => void): Promise<() => void>;
+}
+/**
+ * The audio lane handle `openAudio` returns — structurally the transport
+ * AudioBridge (media.ts): base64 PCM16 @24 kHz mono in both directions.
+ */
+type ProxyAudioLane = Pick<AudioBridge, 'appendInputAudio' | 'onOutputAudio'>;
+/**
+ * The video frame-lane handle `openVideo` returns — structurally the
+ * transport VideoFrameLane (media.ts): one raw encoded JPEG frame per call,
+ * input-only (Live-style protocols have no video OUT modality).
+ */
+type ProxyVideoLane = Pick<VideoFrameLane, 'sendInputFrame'>;
+/**
+ * The video-OUT lane handle `openVideoOut` returns — structurally the
+ * transport VideoOutLane (video-out.ts): the codec plus frame/error fan-out
+ * the WS-inline delivery subscribes.
+ */
+type ProxyVideoOutLane = Pick<VideoOutLane, 'codec' | 'onOutputFrame' | 'onLaneError'>;
+/**
+ * Per-request `ProxyClients` resolution — the seam the HOSTED multi-tenant
+ * server (`src/hosted/`) plugs into so ONE handler implementation serves
+ * every org: the hosted server resolves the caller's org from its Bearer API
+ * key and returns THAT org's backhaul. The local CLI passes a fixed
+ * `ProxyClients` instead; both go through the identical handler body.
+ *
+ * Throwing from here is the loud path: the thrown error surfaces in the
+ * lane's native error envelope (see {@link ProxyHandlerOptions.statusOf}).
+ */
+type ProxyClientsFor = (req: IncomingMessage) => ProxyClients | Promise<ProxyClients>;
+interface ProxyHandlerOptions {
+    /** A fixed backhaul (local CLI) or a per-request resolver (hosted server). */
+    clients: ProxyClients | ProxyClientsFor;
+    /** Optional bearer the local agent must present (never forwarded upstream). */
+    apiKey?: string;
+    /**
+     * WHAT this proxy serves (app/org/fn/base_url + proxy_version), surfaced as
+     * the `identity` block on `GET /stats` so a second `urun compat` invocation
+     * can reuse this proxy iff the identity matches its own exactly. Only the
+     * standalone `urun compat proxy` command sets it — a launch-mode proxy is
+     * child-owned (it dies with its child) and an embedder that omits it is
+     * simply never reused.
+     */
+    identity?: ProxyIdentity;
+    /**
+     * Map a thrown error onto an HTTP status + error `type` before the generic
+     * 500. The hosted server uses it to turn its auth/tenancy failures into a
+     * 401 in the OpenAI envelope. Returning null means "not mine" — the error
+     * takes the ordinary loud 500 path.
+     */
+    statusOf?: (err: unknown) => {
+        status: number;
+        openaiType: string;
+        anthropicType: string;
+    } | null;
+}
+
+export { ModelRouter as M, type ProxyClients as P, SessionGoneError as S, UnknownModelError as U, type ProxyHandlerOptions as a, type ProxyIdentity as b, type ProxyVideoOutLane as c };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@urun-sh/openai",
-  "version": "0.5.4",
+  "version": "0.5.5",
   "description": "OpenAI-compatible Realtime + Responses SDK over uRun session primitives (not websockets).",
   "type": "module",
   "main": "./dist/index.cjs",
@@ -70,7 +70,7 @@
     "build:bin": "bun build --compile --external @roamhq/wrtc src/proxy/bin.ts --outfile dist/bin/urun-openai && node scripts/assert-bin-transport.mjs"
   },
   "peerDependencies": {
-    "@urun-sh/core": "^0.5.4"
+    "@urun-sh/core": "^0.5.5"
   },
   "dependencies": {
     "openai": "4.104.0",

package/dist/chunk-4MSBS7M6.js

@@ -1 +0,0 @@
-import{c as l}from"./chunk-YSFSRI3D.js";l();l();function p(s){return s.toLowerCase().replace(/[^\p{L}\p{N}-]/gu,"-").replace(/^-+|-+$/g,"")}l();var C={chat:!0,code:!0,agent:!0,vl:!0},y={int4:"int4",int8:"int8",fp4:"fp4",mxfp4:"mxfp4",nvfp4:"nvfp4",fp6:"fp6",fp8:"fp8",mxfp8:"mxfp8",fp16:"fp16",bf16:"bf16",fp32:"fp32"};function S(s){let e=s.toLowerCase();return Object.hasOwn(y,e)?y[e]:null}function g(s){let e=s.engine_args?.context_length;return typeof e=="number"&&Number.isFinite(e)&&e>0?e:null}function _(s,e){let t=e.filter(i=>p(`${i.model_id}-${i.variant}`)===s);if(t.length===1)return t[0];if(t.length>1)return v(t);let n=e.filter(i=>p(i.model_id)===s),o=new Set(n.map(i=>i.variant));return n.length>0&&o.size===1?v(n):null}function v(s){let e=s[0],t=r=>s.every(a=>(a[r]??void 0)===(e[r]??void 0))?e[r]:void 0,n=s.map(g),o=n[0],i=o!=null&&n.every(r=>r===o)?o:null;return{model_id:e.model_id,variant:e.variant,gpu_spec:t("gpu_spec"),engine:t("engine"),task:t("task"),description:t("description"),size_gb:t("size_gb"),hf_repo:t("hf_repo"),engine_args:i!==null?{context_length:i}:{},updated_at:t("updated_at")}}function b(s){let e=s.updated_at?Date.parse(s.updated_at):NaN;return Number.isFinite(e)?Math.floor(e/1e3):0}function R(s,e){let t=_(s,e);if(!t)return{id:s,object:"model",created:0,owned_by:"urun"};let n=g(t);return{id:s,object:"model",created:b(t),owned_by:"urun",name:`${t.model_id}:${t.variant}`,...t.description?{description:t.description}:{},...t.task?{task:t.task}:{},...t.engine?{engine:t.engine}:{},...t.gpu_spec?{gpu_spec:t.gpu_spec}:{},...n!==null?{context_length:n}:{}}}function x(s,e){let t=[];for(let n of s){let o=_(n,e);if(!o||!o.task||C[o.task]!==!0)continue;let i=g(o),r={temperature:{type:"range",min:0,max:2},top_p:{type:"range",min:0,max:1},stop:{type:"array",max_items:4}};i!==null&&(r.max_tokens={type:"integer",min:1,max:i,unit:"token"}),t.push({schema_version:"2.4",id:n,name:`${o.model_id}:${o.variant}`,created:b(o),quantization:S(o.variant),description:o.description??"",hugging_face_id:o.hf_repo?o.hf_repo.split("@")[0]??o.hf_repo:"",input_modalities:[{type:"text",...i!==null?{supported_inputs:{max_context_length:{value:i,unit:"token"}}}:{}},...o.task==="vl"?[{type:"image"}]:[]],output_modalities:[{type:"text",streaming:!0,supported_parameters:r}]})}return{data:t}}var I="https://api.urun.sh/v1",d=class extends Error{},u=class extends Error{},h="urs1.";function k(s,e){return h+Buffer.from(JSON.stringify([s,e]),"utf8").toString("base64url")}function P(s){if(!s.startsWith(h))return null;try{let e=JSON.parse(Buffer.from(s.slice(h.length),"base64url").toString("utf8"));if(Array.isArray(e)&&typeof e[0]=="string"&&typeof e[1]=="string")return{app:e[0],key:e[1]}}catch{}return null}async function F(s){let e=s.fetchImpl??fetch,t=`${s.apiUrl.replace(/\/+$/,"")}/apps`,n=await e(t,{headers:{Authorization:`Bearer ${s.apiKey}`,Accept:"application/json"}});if(!n.ok)throw new Error(`org apps listing failed: GET ${t} \u2192 ${n.status}`);let o=await n.json();if(!Array.isArray(o.apps))throw new Error(`org apps listing returned no "apps" array (GET ${t})`);if(o.truncated===!0)throw new Error(`org apps listing was truncated (GET ${t} returned ${o.apps.length} of more) \u2014 model discovery would silently omit deployed apps`);return o.apps}var A=6e4,$=class{constructor(e){this.opts=e}opts;pool=new Map;appsCache=null;rowsCache=null;seed(e,t){this.pool.set(e,Promise.resolve(t))}async deployedApps(){if(!this.opts.listApps)return[];let e=this.opts.appsTtlMs??A;if(this.appsCache&&Date.now()-this.appsCache.at<e)return this.appsCache.apps;let t=await this.opts.listApps();return this.appsCache={at:Date.now(),apps:t},t}async catalogRows(){if(!this.opts.listCatalog)return[];let e=this.opts.appsTtlMs??A;if(this.rowsCache&&Date.now()-this.rowsCache.at<e)return this.rowsCache.rows;let t=await this.opts.listCatalog();return this.rowsCache={at:Date.now(),rows:t},t}servable(e){return e.filter(t=>t.function_name===this.opts.fnName&&t.deployment_status==="active").map(t=>t.app_slug)}async availableIds(){let e=this.opts.defaultApp,t=new Set(e===null?[]:[e]);if(this.opts.listApps)for(let o of this.servable(await this.deployedApps()))t.add(o);for(let o of this.pool.keys())t.add(o);let n=[...t].filter(o=>o!==e).sort();return e===null?n:[e,...n]}async noDefaultApp(e){let t=await this.availableIds(),n=t.length>0?t.join(", "):"(none deployed)";throw new d(e===""?`no "model" was given and this endpoint has no default model \u2014 name one of your org's deployed "${this.opts.fnName}" apps; available models: ${n}`:`model "${e}" does not name a deployed "${this.opts.fnName}" app in your org, and this endpoint has no default model to fall back to; available models: ${n}`)}async resolveApp(e){let t=(e??"").trim(),n=this.opts.defaultApp;if(!t||t==="urun")return n===null?this.noDefaultApp(t):n;let o=p(t);if(o===n)return n;if(this.pool.has(o))return o;if(this.opts.listApps){let i=await this.deployedApps(),r=this.servable(i);if(r.includes(o))return o;let a=r.filter(c=>c.startsWith(`${o}-`));if(a.length===1)return a[0];if(a.length>1)throw new d(`model "${t}" is ambiguous across deployed apps ${a.join(", ")} \u2014 name the exact app slug (or catalog id:variant)`);let f=i.find(c=>c.app_slug===o||c.app_slug.startsWith(`${o}-`));if(f)throw new d(`model "${t}" maps to app "${f.app_slug}", which is not an active "${this.opts.fnName}" app \u2014 deploy it with \`urun serve ${t}\`; available models: ${(await this.availableIds()).join(", ")}`)}if(this.opts.listCatalog){let r=(await this.opts.listCatalog()).find(a=>t===a.model_id||t===`${a.model_id}:${a.variant}`||o===p(a.model_id)||o===p(`${a.model_id}-${a.variant}`));if(r)throw new d(`model "${t}" is in the uRun catalog but not deployed \u2014 deploy it with \`urun serve ${r.model_id}\`; available models: ${(await this.availableIds()).join(", ")}`)}return n===null?this.noDefaultApp(t):n}async sessionFor(e){let t=await this.resolveApp(e),n=this.pool.get(t);return n||(n=Promise.resolve(this.opts.openSession(t)),this.pool.set(t,n),n.catch(()=>this.pool.delete(t))),{app:t,entry:await n}}keyOf(e){let t=this.opts.sessionKey;if(!t)throw new Error("ModelRouter: sessionKey is not configured \u2014 the session-identity seam (handleFor/sessionForHandle) is unavailable on this router");let n=t(e);if(typeof n!="string"||n.length===0)throw new Error("ModelRouter: sessionKey returned an empty identity for a pooled session");return n}async handleFor(e){let{app:t,entry:n}=await this.sessionFor(e);return{app:t,handle:k(t,this.keyOf(n))}}async sessionForHandle(e){let t=P(e);if(!t)throw new u(`malformed session handle ${JSON.stringify(e)} \u2014 not issued by this proxy`);let{app:n,key:o}=t,i=this.pool.get(n),r=i?await i.then(a=>a,()=>null):null;if(r===null)throw new u(`the session behind this handle (app "${n}") is gone \u2014 closed, evicted after its pod died, or the proxy restarted; it cannot be reattached. Reconnect without a handle to start a new session.`);if(this.keyOf(r)!==o)throw new u(`the session behind this handle (app "${n}") was replaced \u2014 the original backhaul died and a new session serves this app now; the handle's session cannot be reattached. Reconnect without a handle to start a new session.`);return{app:n,entry:r}}async evict(e,t){let n=this.pool.get(e);!n||await n.then(i=>i,()=>null)!==t||this.pool.get(e)!==n||(this.pool.delete(e),await this.opts.closeSession(t).catch(()=>{}))}async modelList(){let e=await this.availableIds(),t=await this.catalogRows().catch(()=>[]);return{object:"list",data:e.map(n=>R(n,t))}}async openRouterModels(){let e=await this.availableIds(),t=await this.catalogRows();return x(e,t)}async closeAll(){let e=[...this.pool.values()];this.pool.clear();let t=[];if(await Promise.all(e.map(async n=>{try{await this.opts.closeSession(await n)}catch(o){t.push(String(o instanceof Error?o.message:o))}})),t.length>0)throw new Error(`failed to close ${t.length} pooled session(s): ${t.join("; ")}`)}};export{I as a,d as b,u as c,F as d,$ as e};

package/dist/chunk-RQBD4OAA.js

@@ -1,6 +0,0 @@
-import{a as b}from"./chunk-K23AZPI4.js";import{a as E}from"./chunk-GBBY3PCZ.js";import{a as R,c as h,d as P,e as _}from"./chunk-4MSBS7M6.js";import{c as x,f as v,g as S,p as U}from"./chunk-CWJRDBDC.js";import{c as l}from"./chunk-YSFSRI3D.js";l();import{App as O,createClientToken as B}from"@urun-sh/core";l();function j(e){let n=e?.type;return n==="response.created"||n==="response.in_progress"}function D(e){let n=e?.type;return n==="response.completed"||n==="error"}function k(e){return async function(t){return(async function*(){for(let a=1;a<=2;a++){let{app:s,entry:r}=await e.sessionFor(t.model),i=!1,u=!1;try{let c=await e.createStream(r,t);for await(let d of c){if(D(d)&&(i=!0),j(d)){if(a>1)continue}else u=!0;yield d}}catch(c){if(a<2&&!u){await e.evict(s,r);continue}throw c}if(i)return;if(a<2&&!u){await e.evict(s,r);continue}return}})()}}l();var M=3,F="llm",W=new Set(["paused","error"]);function A(e,n){let t=e.doc(F),o=typeof t.onConnectionState=="function",a=typeof e.onPhase=="function";if(!o&&!a)throw new Error(`session for app "${n}" exposes neither a doc connection-state surface nor onPhase \u2014 session-gone detection requires @urun-sh/core Session (SessionDocument.onConnectionState / Session.onPhase); pooling a session with no native lifecycle signal would silently swallow requests dispatched after the platform closes it`);let s=null,r=new Set,i=[],u=c=>{if(!s){s=c;for(let d of[...r])d(c);r.clear()}};return o&&i.push(t.onConnectionState(c=>{c.consecutiveFailures<M||u(new h(`the pooled session for app "${n}" is gone: its doc backhaul failed ${c.consecutiveFailures} consecutive times (last close ${c.lastCloseCode===null?"by provider liveness reap":`code ${c.lastCloseCode}`}${c.lastCloseReason?`, reason ${JSON.stringify(c.lastCloseReason)}`:""}) \u2014 the platform closed the session (idle timeout or a serve-side crash). The entry is evicted; retry to get a fresh session.`))})),a&&i.push(e.onPhase(c=>{W.has(c.name)&&u(new h(`the pooled session for app "${n}" lost its compute binding (phase "${c.name}"). The entry is evicted; retry to get a fresh session. The named session itself is not over \u2014 it stays attachable by name.`))})),{gone:()=>s,onGone(c){return s?(c(s),()=>{}):(r.add(c),()=>r.delete(c))},dispose(){r.clear();for(let c of i.splice(0))c()}}}async function*g(e,n){let t=n.gone();if(t)throw t;let o=e[Symbol.asyncIterator](),a,s=new Promise((r,i)=>{a=n.onGone(u=>i(u))});s.catch(()=>{});try{for(;;){let r=await Promise.race([o.next(),s]);if(r.done)return;yield r.value}}finally{a?.(),o.return?.(void 0)}}function N(e,n,t){let o=e[t];if(typeof o!="function")throw new Error(`app "${n}" has no function "${t}" (set URUN_FUNCTION)`);let a=o();if(typeof a.end!="function")throw new Error(`app "${n}" function "${t}" returned a session without end()`);return a}function H(e,n){let{baseUrl:t,orgId:o,fnName:a,auth:s,subject:r}=e,i=p=>({session:p,responses:new E(p),gone:A(p,n)});if(s.lane==="jwt"){let p=O(n,{baseUrl:t,orgId:o,jwt:s.jwt});return i(N(p,n,a))}let{apiKey:u,gatewayUrl:c}=s,d=O(n,{baseUrl:t,orgId:o,getAccessToken:async()=>(await B(u,{baseUrl:c,expiresIn:300,allowedFunctions:[`${n}/${a}`],...r===void 0?{}:{subject:r}})).token});return i(N(d,n,a))}function K(e){let n=e.session.id;if(typeof n!="string"||n.length===0)throw new Error("pooled session exposes no stable session id \u2014 the session-identity seam requires @urun-sh/core Session (session.id)");return n}var q=new Set(["error","paused"]);function xe(e){let{auth:n,catalog:t}=e,o=null,a=r=>{let i=H(e,r);return i.gone.onGone(()=>{o?.evict(r,i)}),i},s=new _({defaultApp:e.defaultApp,fnName:e.fnName,openSession:a,closeSession:async r=>{r.gone.dispose(),await r.session.end()},listApps:n.lane==="api-key"?()=>P({apiUrl:e.apiUrl,apiKey:n.apiKey}):null,listCatalog:t?()=>b(t):null,sessionKey:K});return o=s,e.defaultApp!==null&&s.seed(e.defaultApp,a(e.defaultApp)),s}function ve(e,n=S){let t=new WeakMap,o=new WeakMap,a=new WeakMap;return{createResponse:k({sessionFor:s=>e.sessionFor(s),evict:(s,r)=>e.evict(s,r),createStream:async(s,r)=>g(await s.responses.responses.create(r),s.gone)}),listModels:()=>e.modelList(),openRouterModels:()=>e.openRouterModels(),openAudio:async s=>{let{entry:r}=await e.sessionFor(s),i=t.get(r);return i||(i=(async()=>x(r.session,await n()))(),t.set(r,i),i.catch(()=>t.delete(r))),i},openVideo:async s=>{let{entry:r}=await e.sessionFor(s),i=o.get(r);return i||(i=v(r.session),o.set(r,i)),i},openVideoOut:async s=>{let{entry:r}=await e.sessionFor(s),i=a.get(r);return i||(i=U(r.session),a.set(r,i)),i},sessionHandle:async s=>(await e.handleFor(s)).handle,createResponseOn:async(s,r)=>{let{entry:i}=await e.sessionForHandle(s);return g(await i.responses.responses.create(r),i.gone)},onSessionEnd:async(s,r)=>{let{entry:i}=await e.sessionForHandle(s),u=i.session;if(typeof u.onPhase!="function")throw new Error("pooled session exposes no onPhase \u2014 session-end (goAway) wiring requires @urun-sh/core Session");let c=!1;return u.onPhase(d=>{if(c||!q.has(d.name))return;c=!0;let p=u.endsAt??null;r({timeLeftMs:p?Math.max(0,p.getTime()-Date.now()):null,reason:d.reason??`session ${d.name}`})})}}}l();import V from"os";import J from"path";import{fileURLToPath as X}from"url";import{format as z}from"util";var I="0.5.4";var C=4141,Q=I,Oe={claude:{bin:"claude",env:"anthropic"},codex:{bin:"codex",env:"openai"},pi:{bin:"pi",env:"openai"},omp:{bin:"omp",env:"openai"},aider:{bin:"aider",env:"openai"},opencode:{bin:"opencode",env:"openai"},crush:{bin:"crush",env:"openai"}};function Z(e,n="openai"){if(n==="anthropic")return{ANTHROPIC_BASE_URL:`http://127.0.0.1:${e}`,ANTHROPIC_API_KEY:"urun-local"};let t=`http://127.0.0.1:${e}/v1`;return{OPENAI_BASE_URL:t,OPENAI_API_BASE:t,OPENAI_API_KEY:"urun-local"}}function Ne(e,n,t){let o={...e,...Z(n,t)};return delete o.URUN_API_KEY,delete o.URUN_JWT,o}function Ie(e,n,t){let o=["-c","model_providers.urun.name=urun","-c",`model_providers.urun.base_url="http://127.0.0.1:${e}/v1"`,"-c",'model_providers.urun.wire_api="responses"',"-c",'model_provider="urun"'],a=t.indexOf("--"),s=a===-1?t:t.slice(0,a),r=s.some(d=>d==="-m"||d==="--model"||d.startsWith("--model=")),i=t[0]==="exec",u=s.includes("--skip-git-repo-check"),c=i&&!u?["exec","--skip-git-repo-check",...t.slice(1)]:t;return[...o,...r?[]:["-m",n],...c]}function $e(e,n){return n.some(o=>o==="--model"||o==="--provider"||o.startsWith("--model=")||o.startsWith("--provider="))?n:["--provider","urun","--model",e,...n]}function Ce(e,n){return n.some(o=>o==="--model"||o==="--provider"||o.startsWith("--model=")||o.startsWith("--provider="))?n:["--model",`urun/${e}`,...n]}function Te(e,n,t){let o=(e.URUN_PI_EXTENSION_ENTRY??"").trim();if(o){if(!t(o))throw new Error(`URUN_PI_EXTENSION_ENTRY points at ${o}, which does not exist \u2014 the launcher that injected it (the vendored urun CLI shim) must materialize the pi extension bundle before spawning pi`);return o}let a=X(new URL("../pi-extension/standalone.cjs",n));if(!t(a))throw new Error(`pi extension entry not found at ${a} \u2014 the pi lane loads the self-contained @urun-sh/openai dist/pi-extension/standalone.cjs from the installed package (run \`urun compat pi\` via the npm-installed CLI, e.g. \`npx @urun-sh/openai\`), or pass an entry to pi yourself: \`pi -e <path-to-@urun-sh/openai>/dist/pi-extension/standalone.cjs\``);return a}function T(e){let n=Number(e);if(!Number.isInteger(n)||n<0)throw new Error(`--port needs a non-negative integer (0 = OS-assigned ephemeral), got ${e}`);return n}function Ge(e){if(e[0]!=="--port")return null;let n=T(e[1]);return e.splice(0,2),n}function Le(e){let n=e.indexOf("--port");if(n===-1)return C;let t=T(e[n+1]);return e.splice(n,2),t}function je(e){if(e===0)throw new Error("`env` cannot use --port 0: it starts no proxy, so the OS never assigns the ephemeral port \u2014 pass the actual port of the running proxy");return e}function y(e,n){let t=(e[n]??"").trim();if(!t)throw new Error(`${n} is not set. The proxy needs URUN_BASE_URL, URUN_ORG_ID, URUN_APP and either URUN_API_KEY (the org key; the proxy vends short-lived function-scoped client tokens from it via the SDK) or URUN_JWT (an explicit pre-vended token override) \u2014 plus optional URUN_FUNCTION, default "serve" \u2014 to open the backhaul session.`);return t}function De(e){let n=ee(y(e,"URUN_BASE_URL")),t=y(e,"URUN_ORG_ID"),o=y(e,"URUN_APP"),a=(e.URUN_FUNCTION??"serve").trim()||"serve",s=(e.URUN_API_URL??"").trim(),r=s?$(s,"URUN_API_URL"):R,i=(e.URUN_JWT??"").trim();if(i)return{baseUrl:n,orgId:t,appId:o,fnName:a,apiUrl:r,auth:{lane:"jwt",jwt:i}};let u=(e.URUN_API_KEY??"").trim();if(u){let c=(e.URUN_GATEWAY_URL??"").trim(),d=c?$(c):void 0;return{baseUrl:n,orgId:t,appId:o,fnName:a,apiUrl:r,auth:{lane:"api-key",apiKey:u,gatewayUrl:d}}}throw new Error("neither URUN_API_KEY nor URUN_JWT is set. Set URUN_API_KEY (the org API key \u2014 the proxy vends short-lived, function-scoped client tokens from it via the SDK, so the key never reaches the agent process) or URUN_JWT (an explicit pre-vended token; when set it always wins). URUN_BASE_URL, URUN_ORG_ID and URUN_APP are also required.")}function ee(e,n="URUN_BASE_URL"){if(/\/v1\/*$/.test(new URL(e).pathname))throw new Error(`${n} must be the session-gateway base (e.g. https://api.urun.sh), got ${e} \u2014 a trailing /v1 is the org control-plane API form (that belongs in URUN_API_URL); session allocation 404s against it. Drop the /v1.`);return e}function $(e,n="URUN_GATEWAY_URL"){let t=new URL(e),o=t.hostname==="localhost"||t.hostname==="127.0.0.1"||t.hostname==="[::1]";if(t.protocol!=="https:"&&!o)throw new Error(`${n} must be https (got ${e}) \u2014 the org API key rides the request's Authorization header`);return e}function Me(e){return{app:e.appId,org:e.orgId,fn:e.fnName,base_url:e.baseUrl,proxy_version:Q}}function ne(e,n){return e.app===n.app&&e.org===n.org&&e.fn===n.fn&&e.base_url===n.base_url&&e.proxy_version===n.proxy_version}var te=250;async function oe(e,n=te){try{let t=await fetch(`http://127.0.0.1:${e}/stats`,{signal:AbortSignal.timeout(n)});if(!t.ok)return null;let o=(await t.json()).identity;if(o==null||typeof o!="object")return null;let{app:a,org:s,fn:r,base_url:i,proxy_version:u}=o;return typeof a!="string"||typeof s!="string"||typeof r!="string"||typeof i!="string"||typeof u!="string"?null:{app:a,org:s,fn:r,base_url:i,proxy_version:u}}catch{return null}}var G=2e3;async function re(e,n=G){try{let t=await fetch(`http://127.0.0.1:${e}/v1/models`,{signal:AbortSignal.timeout(n)});return t.ok?null:`GET /v1/models answered HTTP ${t.status}`}catch(t){return`GET /v1/models failed (${String(t instanceof Error?t.message:t)})`}}async function Fe(e,n,t=C,o=G){if(e!=null)return{mode:"spawn",port:e};let a=await oe(t);if(a!==null&&ne(a,n)){let s=await re(t,o);return s===null?{mode:"reuse",port:t,identity:a}:{mode:"spawn",port:0,reason:`a urun-openai proxy on port ${t} matches this invocation's identity but failed the serving-surface liveness check (${s}) \u2014 starting a fresh proxy on an ephemeral port instead of pointing the harness at it`}}return{mode:"spawn",port:0}}function We(e,n=V.homedir()){return J.join(n,".urun","logs",`compat-proxy-${e}.log`)}var w=["log","info","warn","error","debug","trace"];function Be(e,n=process,t=console){let o=n.stdout.write,a=n.stderr.write,s=(i,...u)=>e.write(i,...u);n.stdout.write=s,n.stderr.write=s;let r=Object.fromEntries(w.map(i=>[i,t[i]]));for(let i of w)t[i]=(...u)=>{e.write(`${z(...u)}
-`)};return()=>{n.stdout.write=o,n.stderr.write=a;for(let i of w)t[i]=r[i]}}var se=new Set(["EPIPE","EIO","EBADF","ENXIO"]);function L(e){let n=e?.code;return typeof n=="string"&&se.has(n)}function He(e,n){try{return e.write(n),!0}catch(t){if(L(t))return!1;throw t}}function Ke(e){let n={stdout:!1,stderr:!1},t=[];for(let o of["stdout","stderr"]){let a=e.streams[o],s=r=>{L(r)&&(n[o]=!0);let i=r?.code;try{e.note(`${new Date().toISOString()} urun compat: process.${o} died (${String(i??(r instanceof Error?r.message:r))}) \u2014 the terminal is gone; launcher output continues in this log only
-`)}catch{}};a.on("error",s),t.push([a,s])}return{isDead:o=>n[o],uninstall:()=>{for(let[o,a]of t)o.off("error",a)}}}function qe(e){let n=t=>{let o=t instanceof Error?`${t.stack??`${t.name}: ${t.message}`}`:String(t);try{e.appendSync(e.logPath,`${new Date().toISOString()} launcher crashed (in-process proxy): ${o}
-`)}catch{}e.restore();try{e.writeStderr(`urun compat: the in-process proxy crashed (${t instanceof Error?t.message:String(t)}) \u2014 terminating the harness (its backend is gone); details: ${e.logPath}
-`)}catch(a){try{e.appendSync(e.logPath,`${new Date().toISOString()} urun compat: the crash line could not reach stderr (${String(a?.code??a)}) \u2014 the terminal is gone
-`)}catch{}}e.killChild(),e.exit(1)};return e.proc.on("uncaughtException",n),e.proc.on("unhandledRejection",n),()=>{e.proc.off("uncaughtException",n),e.proc.off("unhandledRejection",n)}}function Ye(e,n){let t="pass --port 0 for an OS-assigned ephemeral port";return n?`port ${e} is held by another urun-openai proxy serving app ${n.app} (org ${n.org}, fn ${n.fn}, version ${n.proxy_version}) \u2014 reuse it by pointing your agent at http://127.0.0.1:${e}, or ${t}`:`port ${e} is already in use and does not answer the urun-openai /stats identity probe \u2014 ${t}`}export{xe as a,ve as b,Oe as c,Z as d,Ne as e,Ie as f,$e as g,Ce as h,Te as i,Ge as j,Le as k,je as l,De as m,ee as n,$ as o,Me as p,oe as q,Fe as r,We as s,Be as t,He as u,Ke as v,qe as w,Ye as x};

package/dist/chunk-STFRT5OY.js

@@ -1 +0,0 @@
-import{a as _,b as x,n as A,o as g}from"./chunk-RQBD4OAA.js";import{a as v}from"./chunk-VNM4M4P3.js";import{c as i}from"./chunk-YSFSRI3D.js";i();import{createClientToken as E,ClientTokenError as I}from"@urun-sh/core";var l=class extends Error{constructor(t,n=401){super(t);this.status=n;this.name="ProxyAuthError"}status},p=class extends Error{constructor(e){super(e),this.name="ControlPlaneUnavailableError"}};function C(r){let e=(r??"").trim();if(e==="")throw new l("missing Authorization header \u2014 send your uRun org API key as `Authorization: Bearer <key>`");let t=/^Bearer[ \t]+(\S.*)$/i.exec(e);if(!t)throw new l("malformed Authorization header \u2014 expected `Authorization: Bearer <urun org api key>`");return t[1].trim()}var S=60,N=6e4,O=4096;function k(r){return r===401||r===403}var y=class{constructor(e){this.opts=e;this.mint=e.mint??E,this.now=e.now??Date.now}opts;cache=new Map;mint;now;async resolve(e){if(e==="")throw new l("empty Authorization bearer token");let t=this.cache.get(e);if(t&&this.now()-t.at<N)return{apiKey:e,orgId:t.orgId};let n;try{n=(await this.mint(e,{baseUrl:this.opts.gatewayUrl,expiresIn:S})).orgId}catch(o){throw o instanceof I&&k(o.status)?(this.cache.delete(e),new l("the uRun org API key was rejected by the control plane \u2014 check the key and that it is not revoked")):new p(`could not verify the API key against the uRun control plane at ${this.opts.gatewayUrl}: ${o instanceof Error?o.message:String(o)}`)}if(!n)throw new p("the control plane verified the API key but returned no org_id \u2014 refusing to route a request with no tenancy (control-plane contract violation)");if(this.cache.size>=O){let o=this.cache.keys().next();o.done||this.cache.delete(o.value)}return this.cache.set(e,{orgId:n,at:this.now()}),{apiKey:e,orgId:n}}clear(){this.cache.clear()}};i();var $=8080,T="serve",X="0.0.0.0";function L(r,e,t){let n=(r[e]??"").trim();if(!n)throw new Error(`${e} is not set \u2014 ${t}`);return n}function W(r=process.env){let e=(r.PORT??"").trim(),t=e===""?$:Number(e);if(!Number.isInteger(t)||t<0||t>65535)throw new Error(`PORT must be an integer in 0..65535, got ${JSON.stringify(e)}`);let n=A(g(L(r,"URUN_BASE_URL","the hosted proxy needs the uRun session-gateway base it opens sessions and verifies API keys against (e.g. https://api.urun.sh)"),"URUN_BASE_URL")),o=(r.URUN_API_URL??"").trim(),d=o?g(o,"URUN_API_URL"):`${n.replace(/\/+$/,"")}/v1`,a=(r.URUN_CATALOG_URL??"").trim(),m=(r.URUN_CATALOG_ANON_KEY??"").trim();if(!!a!=!!m)throw new Error("URUN_CATALOG_URL and URUN_CATALOG_ANON_KEY must be set together (or both left unset) \u2014 exactly one of them is set, which would silently disable the model-catalog oracle");let s=a?{catalogUrl:g(a,"URUN_CATALOG_URL"),anonKey:m}:null;return{port:t,baseUrl:n,apiUrl:d,catalog:s}}i();import{createHash as H}from"crypto";var z=512;function M(r){return`urun-inference-proxy-${H("sha256").update(r).digest("hex").slice(0,32)}`}var f=class{constructor(e){this.opts=e}opts;tenants=new Map;build(e){if(this.opts.build)return this.opts.build(e);let t=_({baseUrl:this.opts.baseUrl,orgId:e.orgId,defaultApp:null,fnName:T,apiUrl:this.opts.apiUrl,auth:{lane:"api-key",apiKey:e.apiKey,gatewayUrl:this.opts.baseUrl},catalog:this.opts.catalog,subject:M(e.apiKey)});return{router:t,clients:x(t)}}clientsFor(e){let t=this.tenants.get(e.apiKey);if(t)return this.tenants.delete(e.apiKey),this.tenants.set(e.apiKey,t),t.clients;for(;this.tenants.size>=z;){let o=this.tenants.entries().next();if(o.done)break;let[d,a]=o.value;this.tenants.delete(d),a.router.closeAll().catch(()=>{})}let n=this.build(e);return this.tenants.set(e.apiKey,n),n.clients}get size(){return this.tenants.size}async closeAll(){let e=[...this.tenants.values()].map(n=>n.router);this.tenants.clear();let t=[];if(await Promise.all(e.map(async n=>{try{await n.closeAll()}catch(o){t.push(String(o instanceof Error?o.message:o))}})),t.length>0)throw new Error(`failed to close ${t.length} tenant backhaul(s): ${t.join("; ")}`)}};i();import{createServer as G}from"http";var B=5e3,F=3e3,K="/api/client-tokens";function U(r,e,t){r.writeHead(e,{"content-type":"application/json"}),r.end(JSON.stringify(t))}function R(r,e,t,n){U(r,e,{error:{message:t,type:n}})}async function D(r,e=fetch){let t=`${r.replace(/\/+$/,"")}${K}`;try{let n=await e(t,{method:"POST",headers:{"content-type":"application/json"},body:"{}",signal:AbortSignal.timeout(F)});return n.status===404?{ready:!1,detail:`${t} answered 404 \u2014 this is not a uRun control plane`}:n.status>=500?{ready:!1,detail:`${t} answered HTTP ${n.status}`}:{ready:!0,detail:`${t} answered HTTP ${n.status}`}}catch(n){return{ready:!1,detail:`${t} unreachable: ${n instanceof Error?n.message:String(n)}`}}}function le(r){let e=r.resolver??new y({gatewayUrl:r.baseUrl}),t=r.registry??new f({baseUrl:r.baseUrl,apiUrl:r.apiUrl,catalog:r.catalog}),n=r.probeControlPlane??(()=>D(r.baseUrl)),o=async s=>e.resolve(C(s.headers.authorization)),d=v({clients:async s=>t.clientsFor(await o(s)),statusOf:s=>s instanceof l?{status:s.status,openaiType:"authentication_error",anthropicType:"authentication_error"}:s instanceof p?{status:502,openaiType:"upstream_error",anthropicType:"api_error"}:null}),a=null;return{server:G((s,c)=>{let h=new URL(s.url??"/","http://localhost");(async()=>{if(s.method==="GET"&&h.pathname==="/healthz"){U(c,200,{ok:!0});return}if(s.method==="GET"&&h.pathname==="/readyz"){let u=Date.now();if(!a||u-a.at>=B){let P=await n();a={at:u,...P}}U(c,a.ready?200:503,{ok:a.ready,control_plane:a.detail});return}if(h.pathname!=="/v1"&&!h.pathname.startsWith("/v1/")){R(c,404,`no route for ${s.method} ${h.pathname} \u2014 this endpoint serves the OpenAI-compatible /v1 surface plus /healthz and /readyz`,"invalid_request_error");return}d(s,c)})().catch(u=>{if(!c.headersSent){R(c,500,String(u instanceof Error?u.message:u),"proxy_error");return}c.end()})}),closeAll:()=>t.closeAll()}}async function ce(r,e){let t=new Promise(n=>r.close(()=>n()));r.closeAllConnections(),await t,await e()}export{l as a,p as b,C as c,N as d,O as e,y as f,$ as g,T as h,X as i,W as j,z as k,M as l,f as m,B as n,F as o,le as p,ce as q};