@urun-sh/openai 0.5.5 → 0.6.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/{ResponsesClient-BYx3YLGo.d.ts → ResponsesClient-CSSrOYD8.d.ts} +7 -1
- package/dist/{ResponsesClient-Dft3bg3b.d.cts → ResponsesClient-_OZERUjH.d.cts} +11 -1
- package/dist/chunk-3JWYIKHM.js +2 -0
- package/dist/chunk-45FJ5GTP.js +7 -0
- package/dist/chunk-6M5JK4YC.js +1 -0
- package/dist/chunk-QTVTCMJU.js +1 -0
- package/dist/chunk-TDRZZGUK.js +1 -0
- package/dist/chunk-TPH77ZOW.js +58 -0
- package/dist/chunk-WCMX5VOF.js +4 -0
- package/dist/chunk-YXK42SKC.js +1 -0
- package/dist/gemini-live.cjs +2 -2
- package/dist/gemini-live.d.cts +79 -7
- package/dist/gemini-live.d.ts +32 -7
- package/dist/gemini-live.js +1 -1
- package/dist/hosted/bin.cjs +43 -27
- package/dist/hosted/bin.js +4 -4
- package/dist/hosted/index.cjs +41 -26
- package/dist/hosted/index.d.cts +710 -17
- package/dist/hosted/index.d.ts +171 -6
- package/dist/hosted/index.js +1 -1
- package/dist/index.cjs +1 -1
- package/dist/index.d.cts +5 -5
- package/dist/index.d.ts +5 -5
- package/dist/index.js +1 -1
- package/dist/{models-NYMZrklp.d.cts → models-DUdx_Y6X.d.cts} +1 -1
- package/dist/pi-extension/index.cjs +7 -7
- package/dist/pi-extension/index.d.cts +2 -2
- package/dist/pi-extension/index.d.ts +2 -2
- package/dist/pi-extension/index.js +1 -1
- package/dist/pi-extension/standalone.cjs +53 -53
- package/dist/proxy/cli.cjs +55 -43
- package/dist/proxy/cli.js +14 -13
- package/dist/proxy/index.cjs +31 -28
- package/dist/proxy/index.d.cts +132 -18
- package/dist/proxy/index.d.ts +28 -11
- package/dist/proxy/index.js +1 -6
- package/dist/responses-turn-BfdBbey8.d.cts +2394 -0
- package/dist/responses-turn-CdhIre_a.d.ts +652 -0
- package/dist/{translator-CcDBEfvm.d.cts → translator-CO8W_hmJ.d.cts} +3 -3
- package/dist/{translator-C9uPKypK.d.ts → translator-Ddd65sXR.d.ts} +1 -1
- package/dist/{types-lsVTbNcH.d.cts → types-CHPtJx6d.d.cts} +13 -0
- package/dist/{types-lsVTbNcH.d.ts → types-CHPtJx6d.d.ts} +4 -0
- package/dist/{video-out-D20UuJ8G.d.cts → video-out-BsuLqlID.d.cts} +6 -6
- package/dist/{video-out-CWesbk12.d.ts → video-out-Cawtm7SF.d.ts} +1 -1
- package/package.json +14 -11
- package/dist/chunk-5BZCM3RS.js +0 -4
- package/dist/chunk-5NXM4IO3.js +0 -2
- package/dist/chunk-CWJRDBDC.js +0 -1
- package/dist/chunk-DMD5UENK.js +0 -1
- package/dist/chunk-GBBY3PCZ.js +0 -1
- package/dist/chunk-I2Q3B3OG.js +0 -6
- package/dist/chunk-OI2OY32M.js +0 -1
- package/dist/chunk-QVF7NF7G.js +0 -44
- package/dist/responses-turn-KOAoIqZ-.d.ts +0 -200
- package/dist/responses-turn-OrO4euEN.d.cts +0 -513
- /package/dist/{models-NYMZrklp.d.ts → models-DUdx_Y6X.d.ts} +0 -0
|
@@ -1,513 +0,0 @@
|
|
|
1
|
-
import { C as CatalogRow } from './models-NYMZrklp.cjs';
|
|
2
|
-
import { IncomingMessage } from 'node:http';
|
|
3
|
-
import { a as AudioBridge, h as VideoFrameLane, j as VideoOutLane } from './video-out-D20UuJ8G.cjs';
|
|
4
|
-
|
|
5
|
-
/**
|
|
6
|
-
* WHAT a proxy serves — the `identity` block on `GET /stats`. Reuse of a
|
|
7
|
-
* running proxy is allowed iff every field matches the launching invocation
|
|
8
|
-
* EXACTLY: a partial match (same app, different fn; same everything, older
|
|
9
|
-
* proxy_version) silently routes the agent at the wrong backend, which is
|
|
10
|
-
* worse than any error.
|
|
11
|
-
*/
|
|
12
|
-
interface ProxyIdentity {
|
|
13
|
-
app: string;
|
|
14
|
-
org: string;
|
|
15
|
-
fn: string;
|
|
16
|
-
/**
|
|
17
|
-
* The control-plane URL the proxy's backhaul session opens against
|
|
18
|
-
* (URUN_BASE_URL verbatim — same org/app/fn on staging vs prod are
|
|
19
|
-
* DIFFERENT backends; review finding, #285). Compared exactly: a cosmetic
|
|
20
|
-
* difference (trailing slash) merely refuses reuse and spawns an ephemeral
|
|
21
|
-
* proxy — the safe direction.
|
|
22
|
-
*/
|
|
23
|
-
base_url: string;
|
|
24
|
-
proxy_version: string;
|
|
25
|
-
}
|
|
26
|
-
|
|
27
|
-
/**
|
|
28
|
-
* THE OPENROUTER PROVIDER DOCUMENT — `GET /v1/models?format=openrouter`.
|
|
29
|
-
*
|
|
30
|
-
* OpenRouter's provider monitor polls this document (schema 2.4, per
|
|
31
|
-
* openrouter.ai/docs/guides/community/for-providers) to list a provider's
|
|
32
|
-
* models on the marketplace. The `id` field is the EXACT model identifier
|
|
33
|
-
* OpenRouter sends back as `model` on chat completions — which here is the
|
|
34
|
-
* app slug, the same id the OpenAI surface lists. One id space, two views.
|
|
35
|
-
*
|
|
36
|
-
* MODALITY GATING: only rows whose catalog `task` names a chat-completions
|
|
37
|
-
* shape (chat | code | agent | vl) are declared. A voice/video/image app
|
|
38
|
-
* behind a chat surface would stream garbage into OpenRouter's baseline
|
|
39
|
-
* tests; models the catalog cannot vouch for are OMITTED, never guessed.
|
|
40
|
-
* With no catalog oracle configured the document is `{"data": []}` — honest
|
|
41
|
-
* emptiness, not invented capability (the chart turns the oracle on).
|
|
42
|
-
*
|
|
43
|
-
* PRICING is deliberately omitted entirely: the schema's rule is "a modality
|
|
44
|
-
* with no pricing array is simply unpriced", OpenRouter configures pricing
|
|
45
|
-
* during provider onboarding, and the per-model price surface is
|
|
46
|
-
* urun-infra U3 (model_prices) — when that lands, this module grows a
|
|
47
|
-
* `prices` input and emits the input/output pricing arrays.
|
|
48
|
-
*
|
|
49
|
-
* The full closed-value-domain schema ships as OpenAPI 3.1 at
|
|
50
|
-
* openrouter.ai/docs/assets/provider-monitor-schema-v2.openapi.json.
|
|
51
|
-
*/
|
|
52
|
-
|
|
53
|
-
type OpenRouterProviderDoc = {
|
|
54
|
-
data: OpenRouterProviderModel[];
|
|
55
|
-
};
|
|
56
|
-
interface OpenRouterProviderModel {
|
|
57
|
-
schema_version: '2.4';
|
|
58
|
-
/** The EXACT id OpenRouter sends back as `model` — the app slug. */
|
|
59
|
-
id: string;
|
|
60
|
-
name: string;
|
|
61
|
-
created: number;
|
|
62
|
-
/** Valid enum: int4|int8|fp4|mxfp4|nvfp4|fp6|fp8|mxfp8|fp16|bf16|fp32|null. */
|
|
63
|
-
quantization: string | null;
|
|
64
|
-
description: string;
|
|
65
|
-
hugging_face_id: string;
|
|
66
|
-
input_modalities: Array<Record<string, unknown>>;
|
|
67
|
-
output_modalities: Array<Record<string, unknown>>;
|
|
68
|
-
}
|
|
69
|
-
|
|
70
|
-
/**
|
|
71
|
-
* Per-model → per-app routing for the compat proxy (owner directive
|
|
72
|
-
* 2026-08-07): swapping the model in a coding harness routes the request to
|
|
73
|
-
* the org's DEPLOYED app for that model. v1 is deployed-only — a uRun model
|
|
74
|
-
* that is not deployed gets a loud 404 naming `urun serve <id>`; a later
|
|
75
|
-
* phase (explicitly out of scope here; urun-infra#1490 shared-endpoints)
|
|
76
|
-
* auto-creates from the model catalog on first request.
|
|
77
|
-
*
|
|
78
|
-
* MODEL-ID SURFACE (the documented mapping): a model may be named by
|
|
79
|
-
* - the app slug itself ("qwen3-6-27b-bf16"), or
|
|
80
|
-
* - the catalog id ("qwen3.6-27b"), or
|
|
81
|
-
* - the catalog id:variant ("qwen3.6-27b:bf16"),
|
|
82
|
-
* where slugification mirrors urun-cli `serve.py _default_app_name` exactly:
|
|
83
|
-
* lowercase, every non-alphanumeric-non-dash character becomes "-", leading/
|
|
84
|
-
* trailing dashes stripped (catalog id "qwen3.6-27b" + variant "bf16" → app
|
|
85
|
-
* "qwen3-6-27b-bf16"). COLLISION RULE: an exact slug match always wins over
|
|
86
|
-
* the catalog-id (prefix) interpretation.
|
|
87
|
-
*
|
|
88
|
-
* RESOLUTION ORDER (one canonical path, documented end to end):
|
|
89
|
-
* 1. model absent / "urun" / an alias of the startup app → DEFAULT app.
|
|
90
|
-
* 2. exact slug match on a deployed serve app → that app.
|
|
91
|
-
* 3. catalog-id form matching exactly one deployed app → that app
|
|
92
|
-
* (two or more candidates → loud ambiguity error naming them).
|
|
93
|
-
* 4. the name maps to an org app that is NOT an active app exposing the
|
|
94
|
-
* proxy's serve function → loud 404
|
|
95
|
-
* naming `urun serve <model>` and the available models.
|
|
96
|
-
* 5. the name matches a catalog row but no deployed app → loud 404
|
|
97
|
-
* naming `urun serve <id>` (deployed-only v1).
|
|
98
|
-
* 6. anything else — a model name outside the uRun namespace entirely
|
|
99
|
-
* (e.g. the harness's own upstream default, "claude-*"/"gpt-*") →
|
|
100
|
-
* DEFAULT app. This IS today's single-app contract, kept deliberately
|
|
101
|
-
* so `urun compat <agent>` with the agent's stock model keeps working
|
|
102
|
-
* with zero new env; the per-model /stats table records every such
|
|
103
|
-
* mapping so it is visible, never silent. Models the proxy ADVERTISES
|
|
104
|
-
* on /v1/models can never land here — they resolve (2/3) or fail loud
|
|
105
|
-
* (4/5) above.
|
|
106
|
-
*
|
|
107
|
-
* NO-DEFAULT MODE (`defaultApp: null`) — the HOSTED multi-tenant lane
|
|
108
|
-
* (`src/hosted/`): a shared endpoint serving every org has no "the app this
|
|
109
|
-
* proxy was started for", so rules 1 and 6 have nothing to fall back TO.
|
|
110
|
-
* Rather than inventing one (picking "some" app for a caller would be the
|
|
111
|
-
* worst kind of silent divergence), both rules become the SAME loud
|
|
112
|
-
* {@link UnknownModelError} that rules 4/5 already raise: name a deployed
|
|
113
|
-
* model, here is the list. Rules 2–5 are byte-for-byte the local behavior —
|
|
114
|
-
* one router, one resolution order, two configurations.
|
|
115
|
-
*/
|
|
116
|
-
|
|
117
|
-
/** One org app row from `GET {orgApi}/apps` (urun-cli `ApiClient.list_apps`). */
|
|
118
|
-
interface DeployedApp {
|
|
119
|
-
app_slug: string;
|
|
120
|
-
function_name?: string | null;
|
|
121
|
-
deployment_status?: string | null;
|
|
122
|
-
[k: string]: unknown;
|
|
123
|
-
}
|
|
124
|
-
/**
|
|
125
|
-
* A model that does not resolve to a deployed app the proxy may serve.
|
|
126
|
-
* Rendered by the server in each lane's NATIVE error format as a 404 — never
|
|
127
|
-
* silently served by the default app.
|
|
128
|
-
*/
|
|
129
|
-
declare class UnknownModelError extends Error {
|
|
130
|
-
}
|
|
131
|
-
/**
|
|
132
|
-
* A session handle names a pooled session that no longer exists (closed,
|
|
133
|
-
* evicted after its pod died, replaced by a re-home, or the proxy restarted).
|
|
134
|
-
* The caller asked to REATTACH that exact session — opening a fresh one and
|
|
135
|
-
* calling it "resumed" would be a silent lie, so this is always loud.
|
|
136
|
-
*/
|
|
137
|
-
declare class SessionGoneError extends Error {
|
|
138
|
-
}
|
|
139
|
-
/**
|
|
140
|
-
* OpenAI-shaped model list (same shape as models.ts listModels). When the
|
|
141
|
-
* catalog oracle is configured, each entry ALSO carries the catalog
|
|
142
|
-
* enrichment fields (additive JSON — OpenAI clients ignore unknown fields;
|
|
143
|
-
* the OpenRouter + Vercel AI Gateway provider listings need them). Entries
|
|
144
|
-
* whose slug matches no catalog row stay at the bare four fields.
|
|
145
|
-
*/
|
|
146
|
-
interface RouterModelEntry {
|
|
147
|
-
id: string;
|
|
148
|
-
object: 'model';
|
|
149
|
-
created: number;
|
|
150
|
-
owned_by: string;
|
|
151
|
-
/** Display name — the canonical `<model_id>:<variant>` catalog ref. */
|
|
152
|
-
name?: string;
|
|
153
|
-
/** User-facing description from the catalog (console Endpoints copy). */
|
|
154
|
-
description?: string;
|
|
155
|
-
/** Catalog modality (chat | code | agent | vl | audio | image | ...). */
|
|
156
|
-
task?: string;
|
|
157
|
-
/** Serving engine (vllm | sglang | llamacpp | ...). */
|
|
158
|
-
engine?: string;
|
|
159
|
-
/** The catalog lane this app deploys, e.g. 'rtx6000:1'. */
|
|
160
|
-
gpu_spec?: string;
|
|
161
|
-
/** Context window in tokens (chat rows carry it in engine_args). */
|
|
162
|
-
context_length?: number;
|
|
163
|
-
}
|
|
164
|
-
interface RouterModelList {
|
|
165
|
-
object: 'list';
|
|
166
|
-
data: RouterModelEntry[];
|
|
167
|
-
}
|
|
168
|
-
interface ModelRouterOptions<S> {
|
|
169
|
-
/**
|
|
170
|
-
* The startup app slug (URUN_APP) — the DEFAULT model — or `null` for the
|
|
171
|
-
* hosted multi-tenant lane, which has no per-proxy default app: there,
|
|
172
|
-
* every request must NAME a deployed model and an unnamed/unknown one
|
|
173
|
-
* fails loud instead of silently landing somewhere (see the module header,
|
|
174
|
-
* "NO-DEFAULT MODE").
|
|
175
|
-
*/
|
|
176
|
-
defaultApp: string | null;
|
|
177
|
-
/** The serve function name every routed app must expose (URUN_FUNCTION). */
|
|
178
|
-
fnName: string;
|
|
179
|
-
/** Open a backhaul session for an app slug (called at most once per app). */
|
|
180
|
-
openSession: (appSlug: string) => S | Promise<S>;
|
|
181
|
-
/** Terminal release for one pool entry (Session.end() underneath). */
|
|
182
|
-
closeSession: (entry: S) => Promise<void>;
|
|
183
|
-
/**
|
|
184
|
-
* List the org's deployed apps, or null when the credentials cannot
|
|
185
|
-
* (URUN_JWT lane: the pre-vended token is scoped to the default app, so
|
|
186
|
-
* there is no org listing AND no cross-app session — routing degrades to
|
|
187
|
-
* the default-app-only contract, which is exactly today's behavior).
|
|
188
|
-
*/
|
|
189
|
-
listApps: (() => Promise<DeployedApp[]>) | null;
|
|
190
|
-
/**
|
|
191
|
-
* Catalog rows (models.ts fetchCatalogRows) as the uRun-namespace oracle
|
|
192
|
-
* for rule 5 AND the /v1/models enrichment + OpenRouter provider-doc
|
|
193
|
-
* source, or null when catalog access is not configured. Optional fields
|
|
194
|
-
* beyond model_id/variant are tolerated (thin rows still typecheck).
|
|
195
|
-
*/
|
|
196
|
-
listCatalog: (() => Promise<CatalogRow[]>) | null;
|
|
197
|
-
/** Deployed-apps cache TTL (the list changes on deploys, not per request). */
|
|
198
|
-
appsTtlMs?: number;
|
|
199
|
-
/**
|
|
200
|
-
* The stable NATIVE identity of one pooled entry (the uRun session id in
|
|
201
|
-
* the proxy wiring — the same identity the serve-side session-affinity tag
|
|
202
|
-
* rides, urun-python#1556). Powers the session-identity seam
|
|
203
|
-
* ({@link ModelRouter.handleFor} / {@link ModelRouter.sessionForHandle});
|
|
204
|
-
* a router without it fails LOUD on those calls, never approximates.
|
|
205
|
-
*/
|
|
206
|
-
sessionKey?: (entry: S) => string;
|
|
207
|
-
}
|
|
208
|
-
/**
|
|
209
|
-
* The image-capability modes the Images lane gates on (images.ts's
|
|
210
|
-
* ImageCapabilityMode — declared here because the router is the LOWER layer:
|
|
211
|
-
* the lane's gate type stays structurally satisfied by
|
|
212
|
-
* {@link ModelRouter.supportsImages} with no router→lane import).
|
|
213
|
-
*/
|
|
214
|
-
type ImageCapabilityMode = 'generation' | 'edit';
|
|
215
|
-
/**
|
|
216
|
-
* The session pool: one backhaul session per deployed app, keyed by app slug,
|
|
217
|
-
* opened lazily on the first request that routes to it and reused for every
|
|
218
|
-
* subsequent one. The startup app is seeded eagerly by the CLI. Sessions
|
|
219
|
-
* close on proxy shutdown via {@link closeAll}; there is NO idle-close policy
|
|
220
|
-
* (deliberate v1 simplification — noted as a follow-up in the PR).
|
|
221
|
-
*/
|
|
222
|
-
declare class ModelRouter<S> {
|
|
223
|
-
private readonly opts;
|
|
224
|
-
private readonly pool;
|
|
225
|
-
private appsCache;
|
|
226
|
-
private rowsCache;
|
|
227
|
-
constructor(opts: ModelRouterOptions<S>);
|
|
228
|
-
/** Seed an already-open session (the CLI's eagerly-opened startup app). */
|
|
229
|
-
seed(appSlug: string, entry: S): void;
|
|
230
|
-
private deployedApps;
|
|
231
|
-
/**
|
|
232
|
-
* Catalog rows with the same TTL discipline as {@link deployedApps} (the
|
|
233
|
-
* catalog changes on reseed migrations, not per request). Null when no
|
|
234
|
-
* oracle is configured — callers degrade to the bare four-field entries.
|
|
235
|
-
*/
|
|
236
|
-
private catalogRows;
|
|
237
|
-
/** Apps this proxy may serve: active AND exposing the serve function. */
|
|
238
|
-
private servable;
|
|
239
|
-
private availableIds;
|
|
240
|
-
/**
|
|
241
|
-
* NO-DEFAULT MODE's terminal for rules 1 and 6: there is no app to fall
|
|
242
|
-
* back to, so say so loudly and list what the CALLER'S org actually has.
|
|
243
|
-
* Never returns.
|
|
244
|
-
*/
|
|
245
|
-
private noDefaultApp;
|
|
246
|
-
/**
|
|
247
|
-
* Resolve a request's `model` to an app slug — the documented resolution
|
|
248
|
-
* order from the module header. Throws {@link UnknownModelError} for a uRun
|
|
249
|
-
* model that is not deployed (rules 4/5).
|
|
250
|
-
*/
|
|
251
|
-
resolveApp(model: string | undefined): Promise<string>;
|
|
252
|
-
/** The pooled session for a model — opened lazily, reused afterwards. */
|
|
253
|
-
sessionFor(model: string | undefined): Promise<{
|
|
254
|
-
app: string;
|
|
255
|
-
entry: S;
|
|
256
|
-
}>;
|
|
257
|
-
private keyOf;
|
|
258
|
-
/**
|
|
259
|
-
* SESSION-IDENTITY SEAM (a): the opaque stable handle for the pooled
|
|
260
|
-
* session currently serving `model`'s turns. Rides the SAME acquisition
|
|
261
|
-
* path as every request ({@link sessionFor}) — the session opens lazily if
|
|
262
|
-
* this model has none yet — and derives the handle from native identity
|
|
263
|
-
* (app slug + uRun session id), zero bespoke bookkeeping.
|
|
264
|
-
*/
|
|
265
|
-
handleFor(model: string | undefined): Promise<{
|
|
266
|
-
app: string;
|
|
267
|
-
handle: string;
|
|
268
|
-
}>;
|
|
269
|
-
/**
|
|
270
|
-
* SESSION-IDENTITY SEAM (b): the exact pooled session a handle names.
|
|
271
|
-
* NEVER opens a fresh session — a handle whose session is gone (closed,
|
|
272
|
-
* evicted, re-homed to a replacement, proxy restarted) or malformed throws
|
|
273
|
-
* {@link SessionGoneError} loudly. Resume is reattach-or-fail, not
|
|
274
|
-
* reattach-or-quietly-restart.
|
|
275
|
-
*/
|
|
276
|
-
sessionForHandle(handle: string): Promise<{
|
|
277
|
-
app: string;
|
|
278
|
-
entry: S;
|
|
279
|
-
}>;
|
|
280
|
-
/**
|
|
281
|
-
* Drop ONE pooled session whose backhaul died (its pod was restarted /
|
|
282
|
-
* drained / deleted) and release it — the next {@link sessionFor} opens a
|
|
283
|
-
* fresh one, i.e. asks the control plane for a new assignment. Used by the
|
|
284
|
-
* one-shot re-home (rehome.ts, urun-sh/urun-python#1592).
|
|
285
|
-
*
|
|
286
|
-
* IDENTITY-GUARDED (the same rule the pi lane's SessionPool follows): a
|
|
287
|
-
* concurrent request that already re-homed this app has put a NEWER entry
|
|
288
|
-
* under the key, and evicting that would close a healthy session out from
|
|
289
|
-
* under it.
|
|
290
|
-
*/
|
|
291
|
-
evict(app: string, entry: S): Promise<void>;
|
|
292
|
-
/**
|
|
293
|
-
* `GET /v1/models`: the org's deployed serve apps as model entries, the
|
|
294
|
-
* default app FIRST. On the JWT lane (no org listing) this is the default
|
|
295
|
-
* app plus any app already in the pool — the gap is called out loudly in
|
|
296
|
-
* the PR, not papered over here.
|
|
297
|
-
*/
|
|
298
|
-
modelList(): Promise<RouterModelList>;
|
|
299
|
-
/**
|
|
300
|
-
* The OpenRouter PROVIDER document (`GET /v1/models?format=openrouter`):
|
|
301
|
-
* schema 2.4 per openrouter.ai/docs/guides/community/for-providers. Only
|
|
302
|
-
* chat-completions-shaped models are declared (task chat | code | agent |
|
|
303
|
-
* vl — a voice or video app behind a chat surface would stream garbage);
|
|
304
|
-
* models the catalog cannot vouch for are OMITTED, never guessed. Pricing
|
|
305
|
-
* is deliberately omitted entirely (schema rule: "A modality with no
|
|
306
|
-
* pricing array is simply unpriced" — OpenRouter configures pricing during
|
|
307
|
-
* provider onboarding; the per-model price surface is urun-infra U3).
|
|
308
|
-
*/
|
|
309
|
-
openRouterModels(): Promise<OpenRouterProviderDoc>;
|
|
310
|
-
/**
|
|
311
|
-
* IMAGE CAPABILITY GATE (the Images lane's `imageModels` oracle, images.ts):
|
|
312
|
-
* is `model`'s deployed app a native destination for `mode`? Authorized
|
|
313
|
-
* resolution ({@link resolveApp} — undeployed uRun models throw
|
|
314
|
-
* {@link UnknownModelError}, which the lane renders as a 404) then the
|
|
315
|
-
* SAME catalog join as the /v1/models enrichment (PR421's rowForSlug):
|
|
316
|
-
* exact `<model_id>-<variant>` slug across ALL its GPU-placement rows, a
|
|
317
|
-
* bare model_id only when one variant owns it.
|
|
318
|
-
*
|
|
319
|
-
* Capability is read ONLY from catalog modality metadata, never from the
|
|
320
|
-
* serve function name: every row must carry task 'image' INVARIANTLY across
|
|
321
|
-
* the placements, and the mode comes from engine_args.expects_image —
|
|
322
|
-
* false ⇒ pure text-to-image ('generation'; diffusers v0.38.0
|
|
323
|
-
* pipeline_qwenimage_edit_plus raises torch.cat on an EMPTY image list, so
|
|
324
|
-
* an edit destination is never generation-capable); true or ABSENT (the
|
|
325
|
-
* native default) ⇒ 'edit'. The consensus rule is per-placement:
|
|
326
|
-
* capability must agree across every placement — a CONFLICTING flag (mixed
|
|
327
|
-
* true/false/absent) or a non-boolean one is a refused capability, never
|
|
328
|
-
* silently defaulted to edit.
|
|
329
|
-
*
|
|
330
|
-
* A CONFIGURED catalog oracle that FAILS propagates its rejection — the
|
|
331
|
-
* gate never fabricates a `false` (which would render as a misleading 404)
|
|
332
|
-
* out of an infrastructure outage. No oracle configured, no matching row,
|
|
333
|
-
* or a slug the catalog cannot vouch for ⇒ false (loud 404 upstream).
|
|
334
|
-
*/
|
|
335
|
-
supportsImages(model: string | undefined, mode: ImageCapabilityMode): Promise<boolean>;
|
|
336
|
-
/** Close every pooled session (Session.end() underneath) — proxy shutdown. */
|
|
337
|
-
closeAll(): Promise<void>;
|
|
338
|
-
}
|
|
339
|
-
|
|
340
|
-
/**
|
|
341
|
-
* The shared Responses execution/store seam — the ONE execution + event-
|
|
342
|
-
* envelope path and the authorized tenant store, imported by BOTH the HTTP
|
|
343
|
-
* SSE lane (server.ts) and the `/v1/responses` WebSocket lane
|
|
344
|
-
* (responses-ws.ts). This module deliberately imports NO transport acceptor
|
|
345
|
-
* and NO HTTP lane module: its only imports are type-only, so the module
|
|
346
|
-
* graph server.ts → responses-ws.ts → responses-turn.ts is acyclic and no
|
|
347
|
-
* Vitest/Bun/Node module-evaluation order can observe an uninitialized
|
|
348
|
-
* binding (the constructor-failure class of bug this extraction removes).
|
|
349
|
-
*/
|
|
350
|
-
|
|
351
|
-
/** The request envelope every Responses-shaped upstream call carries. */
|
|
352
|
-
interface ResponsesCreateParams {
|
|
353
|
-
model?: string;
|
|
354
|
-
input: unknown;
|
|
355
|
-
stream?: boolean;
|
|
356
|
-
tools?: unknown;
|
|
357
|
-
tool_choice?: unknown;
|
|
358
|
-
temperature?: number;
|
|
359
|
-
max_output_tokens?: number;
|
|
360
|
-
/**
|
|
361
|
-
* System/developer instructions, forwarded to the serve envelope — where
|
|
362
|
-
* they become ONE prepended `system`-role message (the protocol's native
|
|
363
|
-
* per-request system input; there is NO envelope-level instructions field).
|
|
364
|
-
* Typed `string` end to end (transport/encode.ts) so no cast bridges the seam.
|
|
365
|
-
*/
|
|
366
|
-
instructions?: string;
|
|
367
|
-
top_p?: number;
|
|
368
|
-
stop?: string[];
|
|
369
|
-
/**
|
|
370
|
-
* Reasoning controls (urun-python #1667): forwarded VERBATIM to the serve
|
|
371
|
-
* envelope — `chat_template_kwargs.enable_thinking:false` is the think-off
|
|
372
|
-
* switch. Absent -> absent (no default injection; server-side validation).
|
|
373
|
-
*/
|
|
374
|
-
reasoning_effort?: string;
|
|
375
|
-
chat_template_kwargs?: Record<string, unknown>;
|
|
376
|
-
}
|
|
377
|
-
/** How one pinned session ended (the native phase machinery's terminal step). */
|
|
378
|
-
interface SessionEndInfo {
|
|
379
|
-
/**
|
|
380
|
-
* Milliseconds until the session's native deadline (`endsAt`), or null when
|
|
381
|
-
* the app declared no maximum session length. NOTE (missing primitive,
|
|
382
|
-
* called out in the PR): core exposes no PRE-expiry notice event — this
|
|
383
|
-
* callback fires AT terminal loss, so timeLeftMs is ~0 on expiry.
|
|
384
|
-
*/
|
|
385
|
-
timeLeftMs: number | null;
|
|
386
|
-
/** The terminal reason (expired / ended / error), for the loud close. */
|
|
387
|
-
reason: string;
|
|
388
|
-
}
|
|
389
|
-
/** The upstream calls the proxy makes — injectable (tests; alt transports). */
|
|
390
|
-
interface ProxyClients {
|
|
391
|
-
/** `UrunResponses(session).responses.create` — an async iterable of Responses stream events. */
|
|
392
|
-
createResponse(params: ResponsesCreateParams): Promise<AsyncIterable<unknown>> | AsyncIterable<unknown>;
|
|
393
|
-
/** `listModels(...)` result (an OpenAI model list object). */
|
|
394
|
-
listModels(): Promise<unknown>;
|
|
395
|
-
/**
|
|
396
|
-
* The OpenRouter provider document (`GET /v1/models?format=openrouter`).
|
|
397
|
-
* Optional at the seam: hand-built test clients may omit it, in which case
|
|
398
|
-
* the format=openrouter branch answers 501 — loud, never a silent empty.
|
|
399
|
-
*/
|
|
400
|
-
openRouterModels?(): Promise<unknown>;
|
|
401
|
-
/**
|
|
402
|
-
* Open (or reuse) the NATIVE audio lanes on the pooled session `model` routes
|
|
403
|
-
* to — the same first-party machinery as RealtimeClient.enableAudio
|
|
404
|
-
* (transport/media.ts `enableSessionAudio`: AudioBridge over the session's
|
|
405
|
-
* `rt-audio-in`/`rt-audio-out` stream lanes at 24 kHz). One lane per pooled
|
|
406
|
-
* session; repeat calls return the same lane. Optional at the seam because
|
|
407
|
-
* text-only embeddings exist — but a surface that RECEIVES audio while the
|
|
408
|
-
* embedder wired no `openAudio` must fail LOUD, never drop chunks.
|
|
409
|
-
*/
|
|
410
|
-
openAudio?(model: string | undefined): Promise<ProxyAudioLane>;
|
|
411
|
-
/**
|
|
412
|
-
* Open (or reuse) the NATIVE video FRAME lane on the pooled session `model`
|
|
413
|
-
* routes to (transport/media.ts `enableSessionVideo`: discrete JPEG frames →
|
|
414
|
-
* `stream('rt-video-in').emit`, the §5 named-DATA image-bytes path — NOT the
|
|
415
|
-
* RTP media plane, which requires an already-H.264-encoded track). One lane
|
|
416
|
-
* per pooled session; repeat calls return the same lane. Optional at the seam
|
|
417
|
-
* because text-only embeddings exist — but a surface that RECEIVES video
|
|
418
|
-
* frames while the embedder wired no `openVideo` must fail LOUD, never drop
|
|
419
|
-
* frames.
|
|
420
|
-
*/
|
|
421
|
-
openVideo?(model: string | undefined): Promise<ProxyVideoLane>;
|
|
422
|
-
/**
|
|
423
|
-
* Open (or reuse) the NATIVE video OUTPUT lane on the pooled session
|
|
424
|
-
* `model` routes to (transport/video-out.ts `openSessionVideoOutLane`):
|
|
425
|
-
* decoded spec-v1 records consumed from the session's named §5
|
|
426
|
-
* `rt-video-out` downstream. The Live adapter fans each record onto the
|
|
427
|
-
* SAME Bidi WS as an extension server message (spec v2, WS-inline) — no
|
|
428
|
-
* side-channel transport. Optional at the seam; an ABSENT seam means the
|
|
429
|
-
* `urun.videoOut` capability is refused by omission in setupComplete
|
|
430
|
-
* (vanilla parity), while a wired seam that cannot open the lane must
|
|
431
|
-
* throw LOUD — never a quiet downgrade.
|
|
432
|
-
*/
|
|
433
|
-
openVideoOut?(model: string | undefined): Promise<ProxyVideoOutLane>;
|
|
434
|
-
/**
|
|
435
|
-
* SESSION-IDENTITY SEAM (ModelRouter.handleFor): the opaque stable handle
|
|
436
|
-
* for the pooled session currently serving `model`'s turns. Derived from
|
|
437
|
-
* native identity (app slug + uRun session id) — the same identity the
|
|
438
|
-
* serve-side session-affinity tag rides (urun-python#1556/#1582).
|
|
439
|
-
*/
|
|
440
|
-
sessionHandle(model: string | undefined): Promise<string>;
|
|
441
|
-
/**
|
|
442
|
-
* createResponse PINNED to the exact session a handle names
|
|
443
|
-
* (ModelRouter.sessionForHandle). Throws SessionGoneError LOUDLY when that
|
|
444
|
-
* session is gone or was replaced — never silently opens a fresh session
|
|
445
|
-
* while claiming resume. Deliberately NO re-home on this path: re-homing
|
|
446
|
-
* would swap the pinned session out from under the caller.
|
|
447
|
-
*/
|
|
448
|
-
createResponseOn(handle: string, params: ResponsesCreateParams): Promise<AsyncIterable<unknown>> | AsyncIterable<unknown>;
|
|
449
|
-
/**
|
|
450
|
-
* Subscribe to the pinned session's terminal end via core's NATIVE phase
|
|
451
|
-
* machinery (Session.onPhase → terminal 'expired'/'ended'/'error'). Fires
|
|
452
|
-
* `cb` once. Throws SessionGoneError if the handle's session is already
|
|
453
|
-
* gone — which doubles as the loud reattach check at resume time. Returns
|
|
454
|
-
* the unsubscribe.
|
|
455
|
-
*/
|
|
456
|
-
onSessionEnd(handle: string, cb: (end: SessionEndInfo) => void): Promise<() => void>;
|
|
457
|
-
}
|
|
458
|
-
/**
|
|
459
|
-
* The audio lane handle `openAudio` returns — structurally the transport
|
|
460
|
-
* AudioBridge (media.ts): base64 PCM16 @24 kHz mono in both directions.
|
|
461
|
-
*/
|
|
462
|
-
type ProxyAudioLane = Pick<AudioBridge, 'appendInputAudio' | 'onOutputAudio'>;
|
|
463
|
-
/**
|
|
464
|
-
* The video frame-lane handle `openVideo` returns — structurally the
|
|
465
|
-
* transport VideoFrameLane (media.ts): one raw encoded JPEG frame per call,
|
|
466
|
-
* input-only (Live-style protocols have no video OUT modality).
|
|
467
|
-
*/
|
|
468
|
-
type ProxyVideoLane = Pick<VideoFrameLane, 'sendInputFrame'>;
|
|
469
|
-
/**
|
|
470
|
-
* The video-OUT lane handle `openVideoOut` returns — structurally the
|
|
471
|
-
* transport VideoOutLane (video-out.ts): the codec plus frame/error fan-out
|
|
472
|
-
* the WS-inline delivery subscribes.
|
|
473
|
-
*/
|
|
474
|
-
type ProxyVideoOutLane = Pick<VideoOutLane, 'codec' | 'onOutputFrame' | 'onLaneError'>;
|
|
475
|
-
/**
|
|
476
|
-
* Per-request `ProxyClients` resolution — the seam the HOSTED multi-tenant
|
|
477
|
-
* server (`src/hosted/`) plugs into so ONE handler implementation serves
|
|
478
|
-
* every org: the hosted server resolves the caller's org from its Bearer API
|
|
479
|
-
* key and returns THAT org's backhaul. The local CLI passes a fixed
|
|
480
|
-
* `ProxyClients` instead; both go through the identical handler body.
|
|
481
|
-
*
|
|
482
|
-
* Throwing from here is the loud path: the thrown error surfaces in the
|
|
483
|
-
* lane's native error envelope (see {@link ProxyHandlerOptions.statusOf}).
|
|
484
|
-
*/
|
|
485
|
-
type ProxyClientsFor = (req: IncomingMessage) => ProxyClients | Promise<ProxyClients>;
|
|
486
|
-
interface ProxyHandlerOptions {
|
|
487
|
-
/** A fixed backhaul (local CLI) or a per-request resolver (hosted server). */
|
|
488
|
-
clients: ProxyClients | ProxyClientsFor;
|
|
489
|
-
/** Optional bearer the local agent must present (never forwarded upstream). */
|
|
490
|
-
apiKey?: string;
|
|
491
|
-
/**
|
|
492
|
-
* WHAT this proxy serves (app/org/fn/base_url + proxy_version), surfaced as
|
|
493
|
-
* the `identity` block on `GET /stats` so a second `urun compat` invocation
|
|
494
|
-
* can reuse this proxy iff the identity matches its own exactly. Only the
|
|
495
|
-
* standalone `urun compat proxy` command sets it — a launch-mode proxy is
|
|
496
|
-
* child-owned (it dies with its child) and an embedder that omits it is
|
|
497
|
-
* simply never reused.
|
|
498
|
-
*/
|
|
499
|
-
identity?: ProxyIdentity;
|
|
500
|
-
/**
|
|
501
|
-
* Map a thrown error onto an HTTP status + error `type` before the generic
|
|
502
|
-
* 500. The hosted server uses it to turn its auth/tenancy failures into a
|
|
503
|
-
* 401 in the OpenAI envelope. Returning null means "not mine" — the error
|
|
504
|
-
* takes the ordinary loud 500 path.
|
|
505
|
-
*/
|
|
506
|
-
statusOf?: (err: unknown) => {
|
|
507
|
-
status: number;
|
|
508
|
-
openaiType: string;
|
|
509
|
-
anthropicType: string;
|
|
510
|
-
} | null;
|
|
511
|
-
}
|
|
512
|
-
|
|
513
|
-
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 };
|
|
File without changes
|