@broberg/ai-sdk 0.25.1 → 0.27.0

package/README.md

@@ -87,6 +87,56 @@ const ai = createAI({
 Sinks: `upmetricsSink` (canonical), `discordSink`, `sqliteSink`, `multiSink`,
 `noopSink`. A failing sink never crashes a call.
 
+### Cost-tracking is on by default (v0.24+)
+
+You don't have to wire a sink. If `UPMETRICS_API_KEY` is in the env, a bare
+`createAI()` auto-attaches the upmetrics sink — this exists because most
+call-sites passed no sink, leaving ~91% of Mistral spend invisible. No key in
+the env → no sink, no crash (ship-dark).
+
+```ts
+createAI();                      // key in env → tracked; no key → nothing happens
+createAI({ costSink: mySink });  // explicit sink wins
+createAI({ costSink: null });    // explicit OPT-OUT — see below
+```
+
+| Env var | Effect |
+|---|---|
+| `UPMETRICS_API_KEY` | **The only switch.** Present → tracking on. |
+| `UPMETRICS_AGENT_NAME` | Row label. Set it — the fallback is `npm_package_name`, which is empty for a process not started via an npm/bun script (a `node dist/…` or Docker service logs as `unknown`). |
+| `UPMETRICS_BASE_URL` | Ingest host. Defaults to `https://upmetrics.org`. |
+| `UPMETRICS_COMPLIANCE` | `1` sets compliance mode. `Usage` carries no prompt/response text either way. |
+
+**Adopting it in a repo — two things to do first, or the adoption corrupts the
+numbers it was meant to reveal:**
+
+1. **Opt out wherever you already report your own costs.** Pass
+   `costSink: null`. Otherwise the SDK adds a *second* reporting path on top of
+   yours and every call is counted **twice**, in production, with no error
+   anywhere. (Found by `buddy`, who aggregates to `cli_usage` and pushes hourly.)
+
+2. **Keep the key out of your test run.** Test runners auto-load `.env` (Bun
+   does; Vitest with a dotenv setup does), so any suite that reaches a
+   `createAI()` will POST **fabricated usage into production telemetry**. Note
+   the trigger is *importing* a module that builds the client — a module-level
+   `createAI()` arms it without anything calling it. Strip the whole prefix:
+
+   ```ts
+   // test-setup.ts  — bunfig.toml: [test] preload = ["./test-setup.ts"]
+   for (const k of Object.keys(process.env)) {
+     if (k.startsWith("UPMETRICS_")) delete process.env[k];
+   }
+   ```
+
+   Strip the **prefix**, not a list of names — a guard you must remember to
+   update is one that silently rots. And assert the *effect* (no `UPMETRICS_*`
+   visible to a test) with a control proving the probe can go red; a test that
+   checks "the key is undefined" passes while protecting nothing on a machine
+   that never had the key.
+
+Adoption is **per repo, opt-in** — there is no fleet-wide push. Turn it on when
+a repo has spend worth watching, with both guards in place.
+
 ## License
 
 FSL-1.1-Apache-2.0

package/dist/{chunk-IZG5UZH5.js → chunk-LUPUAAHZ.js}

@@ -67,6 +67,12 @@ var PRICING = {
   "gemini:gemini-2.5-flash": { inputPer1M: 0.3, outputPer1M: 2.5, version: V },
   // flash-lite is the default `video` tier (F019) — cheap native video understanding.
   "gemini:gemini-2.5-flash-lite": { inputPer1M: 0.1, outputPer1M: 0.4, version: "2026-06-04-or-xref" },
+  // Vertex AI (F038) — the EU-resident route to the SAME Gemini models, so Google's
+  // published Gemini token prices apply. Listed separately because cost lookups key on
+  // `provider:model`: without these rows an EU vision/video call would silently log
+  // $0, which is worse than no tracking (a confident wrong number).
+  "vertex:gemini-2.5-flash": { inputPer1M: 0.3, outputPer1M: 2.5, version: V },
+  "vertex:gemini-2.5-flash-lite": { inputPer1M: 0.1, outputPer1M: 0.4, version: "2026-06-04-or-xref" },
   // Mistral (direct, La Plateforme). Official prices from mistral.ai/pricing
   // (2026-06-04, per Christian's CD report). EU/Paris-hosted — the designated
   // GDPR-safe provider for client/personal-data workloads (see F015). NB:
@@ -105,4 +111,4 @@ export {
   PRICING,
   getPrice
 };
-//# sourceMappingURL=chunk-IZG5UZH5.js.map
+//# sourceMappingURL=chunk-LUPUAAHZ.js.map

package/dist/chunk-LUPUAAHZ.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/cost/pricing.ts"],"sourcesContent":["// Versioned per-(provider, model) pricing. F3.6 populates the table + adds tests\n// + MiniMax coverage. F3.1 ships the type + lookup with an empty table, so\n// computeCost returns 0 for every model until F3.6 lands (calls still complete).\nexport interface PricingEntry {\n  /** USD per 1M input tokens. */\n  inputPer1M: number;\n  /** USD per 1M output tokens. */\n  outputPer1M: number;\n  /** USD per 1M cache-read tokens (falls back to input rate if unset). */\n  cacheReadPer1M?: number;\n  /** USD per 1M cache-write/creation tokens (falls back to input rate if unset). */\n  cacheWritePer1M?: number;\n  /** Pricing snapshot version (date or tag) so stale entries are detectable. */\n  version: string;\n}\n\n// USD per 1M tokens. Anthropic cache multipliers follow the standard model:\n// cache-read ≈ 0.1× input, cache-write ≈ 1.25× input. Verified against the\n// pricing tables in cms (packages/cms-ai/src/providers) + trail (model-lab).\n// MiniMax M2.7 is an estimate pending confirmation against OpenRouter's live\n// price page — flagged in its version string.\nconst V = \"2026-06-02\";\n// Mistral prices come straight from mistral.ai/pricing (per Christian's CD report).\nconst MS = \"2026-06-04-mistral.ai\";\n\n/** Keyed `${provider}:${model}`. Exported so the catalogue-research job (F014)\n *  can enumerate every priced entry and diff it against the live provider lists. */\nexport const PRICING: Record<string, PricingEntry> = {\n  // Anthropic (direct API). DEFAULT_TIER_MAP: fast/cheap=haiku, smart/vision=sonnet, powerful=opus.\n  \"anthropic:claude-haiku-4-5\": {\n    inputPer1M: 0.8,\n    outputPer1M: 4.0,\n    cacheReadPer1M: 0.08,\n    cacheWritePer1M: 1.0,\n    version: V,\n  },\n  \"anthropic:claude-sonnet-4-6\": {\n    inputPer1M: 3.0,\n    outputPer1M: 15.0,\n    cacheReadPer1M: 0.3,\n    cacheWritePer1M: 3.75,\n    version: V,\n  },\n  \"anthropic:claude-opus-4-8\": {\n    inputPer1M: 15.0,\n    outputPer1M: 75.0,\n    cacheReadPer1M: 1.5,\n    cacheWritePer1M: 18.75,\n    version: V,\n  },\n\n  // OpenAI. embedding default tier = text-embedding-3-small (no output tokens).\n  \"openai:text-embedding-3-small\": { inputPer1M: 0.02, outputPer1M: 0, version: V },\n  \"openai:text-embedding-3-large\": { inputPer1M: 0.13, outputPer1M: 0, version: V },\n  \"openai:gpt-4o\": { inputPer1M: 2.5, outputPer1M: 10.0, version: V },\n  \"openai:gpt-4o-mini\": { inputPer1M: 0.15, outputPer1M: 0.6, version: V },\n  // Whisper is priced per minute, not per token — not representable here; transcribe\n  // (F5.6) computes its own cost. Listed as 0 so token-based compute never charges it.\n  \"openai:whisper-1\": { inputPer1M: 0, outputPer1M: 0, version: V },\n\n  // OpenRouter (meta-router — model slugs include the upstream vendor). Slugs use\n  // dots (claude-sonnet-4.6) to match OpenRouter's live ids; the dashed forms\n  // never matched a real call. Caught by the F014 catalogue research.\n  \"openrouter:anthropic/claude-sonnet-4.6\": { inputPer1M: 3.0, outputPer1M: 15.0, version: V },\n  // OpenRouter ground-truth $1/$5 — a markup over Anthropic-direct's $0.8/$4\n  // (the `anthropic:` entry above). Was masked while the slug used dashes.\n  \"openrouter:anthropic/claude-haiku-4.5\": { inputPer1M: 1.0, outputPer1M: 5.0, version: \"2026-06-04\" },\n  \"openrouter:google/gemini-2.5-flash\": { inputPer1M: 0.3, outputPer1M: 2.5, version: V },\n  // Ground-truth from OpenRouter /api/v1/models (was a 0.3 estimate; now 0.279).\n  \"openrouter:minimax/minimax-m2.7\": {\n    inputPer1M: 0.279,\n    outputPer1M: 1.2,\n    version: \"2026-06-04\",\n  },\n  // DeepSeek V4 (CN-hosted — NOT GDPR-safe; non-personal-data workloads only).\n  // On 2026-05-22 DeepSeek made the \"75% off\" promo the permanent official price.\n  // V4-Pro $0.435/$0.87 is ~34x cheaper than GPT-5.5 on output; flash is cheaper\n  // still. Numbers match OpenRouter /api/v1/models 1:1 (no router markup). A strong\n  // cheap route for fleet background work once `claude -p` is API-billed (15 Jun).\n  \"openrouter:deepseek/deepseek-v4-pro\": { inputPer1M: 0.435, outputPer1M: 0.87, version: \"2026-05-22-deepseek-official\" },\n  \"openrouter:deepseek/deepseek-v4-flash\": { inputPer1M: 0.0983, outputPer1M: 0.1966, version: \"2026-05-22-deepseek-official\" },\n  // DeepSeek DIRECT API (provider \"deepseek\", F030 non-PII secondary). Rates from\n  // api-docs.deepseek.com 2026-06-30 ($0.14/$0.28 per 1M; both map to deepseek-v4-flash).\n  // `deepseek-chat` (non-thinking) + `deepseek-reasoner` (thinking) DEPRECATE 2026-07-24.\n  // (The bare `deepseek-v4-flash` basename is already priced via the openrouter entry\n  // above — kept distinct here to avoid a basename collision in the F027 pricing-API.)\n  // Verify against a real key when it lands.\n  \"deepseek:deepseek-chat\": { inputPer1M: 0.14, outputPer1M: 0.28, version: \"2026-06-30-deepseek-direct\" },\n  \"deepseek:deepseek-reasoner\": { inputPer1M: 0.14, outputPer1M: 0.28, version: \"2026-06-30-deepseek-direct\" },\n\n  // Google Gemini (direct). Provider key is \"gemini\" — matches the adapter's\n  // usage.provider + the override.provider callers pass. (Image-gen models are\n  // priced per-image in the adapter, not here.)\n  \"gemini:gemini-2.5-flash\": { inputPer1M: 0.3, outputPer1M: 2.5, version: V },\n  // flash-lite is the default `video` tier (F019) — cheap native video understanding.\n  \"gemini:gemini-2.5-flash-lite\": { inputPer1M: 0.1, outputPer1M: 0.4, version: \"2026-06-04-or-xref\" },\n\n  // Vertex AI (F038) — the EU-resident route to the SAME Gemini models, so Google's\n  // published Gemini token prices apply. Listed separately because cost lookups key on\n  // `provider:model`: without these rows an EU vision/video call would silently log\n  // $0, which is worse than no tracking (a confident wrong number).\n  \"vertex:gemini-2.5-flash\": { inputPer1M: 0.3, outputPer1M: 2.5, version: V },\n  \"vertex:gemini-2.5-flash-lite\": { inputPer1M: 0.1, outputPer1M: 0.4, version: \"2026-06-04-or-xref\" },\n\n  // Mistral (direct, La Plateforme). Official prices from mistral.ai/pricing\n  // (2026-06-04, per Christian's CD report). EU/Paris-hosted — the designated\n  // GDPR-safe provider for client/personal-data workloads (see F015). NB:\n  // medium-3.5 is the premium \"Vibe\" coding tier ($1.5/$7.5); Large 3 ($0.5/$1.5)\n  // is the cheaper frontier general-purpose model despite the higher number.\n  \"mistral:mistral-large-latest\": { inputPer1M: 0.5, outputPer1M: 1.5, version: MS },\n  \"mistral:mistral-large-2512\": { inputPer1M: 0.5, outputPer1M: 1.5, version: MS },\n  \"mistral:mistral-medium-latest\": { inputPer1M: 1.5, outputPer1M: 7.5, version: MS },\n  \"mistral:mistral-medium-3.5\": { inputPer1M: 1.5, outputPer1M: 7.5, version: MS },\n  \"mistral:mistral-medium-3\": { inputPer1M: 0.4, outputPer1M: 2.0, version: \"2026-06-04-or-xref\" },\n  \"mistral:mistral-small-latest\": { inputPer1M: 0.1, outputPer1M: 0.3, version: MS },\n  \"mistral:mistral-small-2603\": { inputPer1M: 0.1, outputPer1M: 0.3, version: MS },\n  \"mistral:ministral-3b-latest\": { inputPer1M: 0.1, outputPer1M: 0.1, version: MS },\n  \"mistral:ministral-8b-latest\": { inputPer1M: 0.15, outputPer1M: 0.15, version: MS },\n  \"mistral:ministral-14b-latest\": { inputPer1M: 0.2, outputPer1M: 0.2, version: MS },\n  \"mistral:magistral-medium-latest\": { inputPer1M: 2.0, outputPer1M: 5.0, version: MS },\n  \"mistral:magistral-small-latest\": { inputPer1M: 0.5, outputPer1M: 1.5, version: MS },\n  \"mistral:devstral-latest\": { inputPer1M: 0.4, outputPer1M: 2.0, version: MS },\n  \"mistral:codestral-latest\": { inputPer1M: 0.3, outputPer1M: 0.9, version: MS },\n  \"mistral:open-mistral-nemo\": { inputPer1M: 0.15, outputPer1M: 0.15, version: MS },\n  // Moderation (F016.4) — per input token; output 0. (OCR is per-page in the adapter.)\n  \"mistral:mistral-moderation-latest\": { inputPer1M: 0.1, outputPer1M: 0, version: MS },\n  // Embeddings (F016.5) — per input token.\n  \"mistral:mistral-embed\": { inputPer1M: 0.1, outputPer1M: 0, version: MS },\n  \"mistral:codestral-embed\": { inputPer1M: 0.15, outputPer1M: 0, version: MS },\n};\n\nexport function getPrice(provider: string, model: string): PricingEntry | undefined {\n  const exact = PRICING[`${provider}:${model}`];\n  if (exact) return exact;\n  // Providers ship dated model snapshots, e.g. \"claude-haiku-4-5-20251001\".\n  // Strip a trailing -YYYYMMDD and retry the base lookup so a dated variant\n  // prices the same as its base model instead of falling through to 0 — a real\n  // paid call must never be logged as $0 (F012). Covers openrouter slugs too.\n  const base = model.replace(/-\\d{8}$/, \"\");\n  if (base !== model) return PRICING[`${provider}:${base}`];\n  return undefined;\n}\n"],"mappings":";AAqBA,IAAM,IAAI;AAEV,IAAM,KAAK;AAIJ,IAAM,UAAwC;AAAA;AAAA,EAEnD,8BAA8B;AAAA,IAC5B,YAAY;AAAA,IACZ,aAAa;AAAA,IACb,gBAAgB;AAAA,IAChB,iBAAiB;AAAA,IACjB,SAAS;AAAA,EACX;AAAA,EACA,+BAA+B;AAAA,IAC7B,YAAY;AAAA,IACZ,aAAa;AAAA,IACb,gBAAgB;AAAA,IAChB,iBAAiB;AAAA,IACjB,SAAS;AAAA,EACX;AAAA,EACA,6BAA6B;AAAA,IAC3B,YAAY;AAAA,IACZ,aAAa;AAAA,IACb,gBAAgB;AAAA,IAChB,iBAAiB;AAAA,IACjB,SAAS;AAAA,EACX;AAAA;AAAA,EAGA,iCAAiC,EAAE,YAAY,MAAM,aAAa,GAAG,SAAS,EAAE;AAAA,EAChF,iCAAiC,EAAE,YAAY,MAAM,aAAa,GAAG,SAAS,EAAE;AAAA,EAChF,iBAAiB,EAAE,YAAY,KAAK,aAAa,IAAM,SAAS,EAAE;AAAA,EAClE,sBAAsB,EAAE,YAAY,MAAM,aAAa,KAAK,SAAS,EAAE;AAAA;AAAA;AAAA,EAGvE,oBAAoB,EAAE,YAAY,GAAG,aAAa,GAAG,SAAS,EAAE;AAAA;AAAA;AAAA;AAAA,EAKhE,0CAA0C,EAAE,YAAY,GAAK,aAAa,IAAM,SAAS,EAAE;AAAA;AAAA;AAAA,EAG3F,yCAAyC,EAAE,YAAY,GAAK,aAAa,GAAK,SAAS,aAAa;AAAA,EACpG,sCAAsC,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,EAAE;AAAA;AAAA,EAEtF,mCAAmC;AAAA,IACjC,YAAY;AAAA,IACZ,aAAa;AAAA,IACb,SAAS;AAAA,EACX;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,uCAAuC,EAAE,YAAY,OAAO,aAAa,MAAM,SAAS,+BAA+B;AAAA,EACvH,yCAAyC,EAAE,YAAY,QAAQ,aAAa,QAAQ,SAAS,+BAA+B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAO5H,0BAA0B,EAAE,YAAY,MAAM,aAAa,MAAM,SAAS,6BAA6B;AAAA,EACvG,8BAA8B,EAAE,YAAY,MAAM,aAAa,MAAM,SAAS,6BAA6B;AAAA;AAAA;AAAA;AAAA,EAK3G,2BAA2B,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,EAAE;AAAA;AAAA,EAE3E,gCAAgC,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,qBAAqB;AAAA;AAAA;AAAA;AAAA;AAAA,EAMnG,2BAA2B,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,EAAE;AAAA,EAC3E,gCAAgC,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,qBAAqB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOnG,gCAAgC,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EACjF,8BAA8B,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EAC/E,iCAAiC,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EAClF,8BAA8B,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EAC/E,4BAA4B,EAAE,YAAY,KAAK,aAAa,GAAK,SAAS,qBAAqB;AAAA,EAC/F,gCAAgC,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EACjF,8BAA8B,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EAC/E,+BAA+B,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EAChF,+BAA+B,EAAE,YAAY,MAAM,aAAa,MAAM,SAAS,GAAG;AAAA,EAClF,gCAAgC,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EACjF,mCAAmC,EAAE,YAAY,GAAK,aAAa,GAAK,SAAS,GAAG;AAAA,EACpF,kCAAkC,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EACnF,2BAA2B,EAAE,YAAY,KAAK,aAAa,GAAK,SAAS,GAAG;AAAA,EAC5E,4BAA4B,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,GAAG;AAAA,EAC7E,6BAA6B,EAAE,YAAY,MAAM,aAAa,MAAM,SAAS,GAAG;AAAA;AAAA,EAEhF,qCAAqC,EAAE,YAAY,KAAK,aAAa,GAAG,SAAS,GAAG;AAAA;AAAA,EAEpF,yBAAyB,EAAE,YAAY,KAAK,aAAa,GAAG,SAAS,GAAG;AAAA,EACxE,2BAA2B,EAAE,YAAY,MAAM,aAAa,GAAG,SAAS,GAAG;AAC7E;AAEO,SAAS,SAAS,UAAkB,OAAyC;AAClF,QAAM,QAAQ,QAAQ,GAAG,QAAQ,IAAI,KAAK,EAAE;AAC5C,MAAI,MAAO,QAAO;AAKlB,QAAM,OAAO,MAAM,QAAQ,WAAW,EAAE;AACxC,MAAI,SAAS,MAAO,QAAO,QAAQ,GAAG,QAAQ,IAAI,IAAI,EAAE;AACxD,SAAO;AACT;","names":[]}

package/dist/index.d.ts

@@ -1697,7 +1697,10 @@ declare const aiConfigSchema: z.ZodObject<{
         transport: "http" | "subprocess";
     }>>>;
     providers: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<ProviderAdapter, z.ZodTypeDef, ProviderAdapter>>>;
-    costSink: z.ZodOptional<z.ZodType<CostSink, z.ZodTypeDef, CostSink>>;
+    /** Omit → cost-tracking auto-wires from the upmetrics env (F034). Pass `null`
+     *  to opt OUT explicitly — for a consumer that already reports its own costs
+     *  and would otherwise be counted twice (F034.3). An object is used as-is. */
+    costSink: z.ZodOptional<z.ZodNullable<z.ZodType<CostSink, z.ZodTypeDef, CostSink>>>;
     budget: z.ZodOptional<z.ZodObject<{
         perCallUsd: z.ZodOptional<z.ZodNumber>;
         rollingUsd: z.ZodOptional<z.ZodNumber>;
@@ -1725,7 +1728,7 @@ declare const aiConfigSchema: z.ZodObject<{
         transport: "http" | "subprocess";
     }>> | undefined;
     providers?: Record<string, ProviderAdapter> | undefined;
-    costSink?: CostSink | undefined;
+    costSink?: CostSink | null | undefined;
     budget?: {
         perCallUsd?: number | undefined;
         rollingUsd?: number | undefined;
@@ -1741,7 +1744,7 @@ declare const aiConfigSchema: z.ZodObject<{
         transport: "http" | "subprocess";
     }>> | undefined;
     providers?: Record<string, ProviderAdapter> | undefined;
-    costSink?: CostSink | undefined;
+    costSink?: CostSink | null | undefined;
     budget?: {
         perCallUsd?: number | undefined;
         rollingUsd?: number | undefined;
@@ -2045,8 +2048,8 @@ declare const falStubAdapter: ProviderAdapter;
  *  wires the live adapters. */
 declare const stubProviders: Record<string, ProviderAdapter>;
 
-declare const VERSION: "0.25.1";
-declare const SDK_TAG: "@broberg/ai-sdk@0.25.1";
+declare const VERSION: "0.27.0";
+declare const SDK_TAG: "@broberg/ai-sdk@0.27.0";
 
 /** Built-in defaults. Every entry is overridable via AiConfig.defaults or a
  *  per-call override.

package/dist/index.js

@@ -8,7 +8,7 @@ import {
 } from "./chunk-IT7HNKLY.js";
 import {
   getPrice
-} from "./chunk-IZG5UZH5.js";
+} from "./chunk-LUPUAAHZ.js";
 
 // src/routing/tier-map.ts
 var DEFAULT_TIER_MAP = {
@@ -1633,8 +1633,8 @@ function vertexAdapter(config = {}) {
     return config.region ?? process.env.GOOGLE_VERTEX_REGION ?? DEFAULT_REGION2;
   }
   function project() {
-    const p = config.project ?? process.env.GOOGLE_VERTEX_PROJECT;
-    if (!p) throw new Error("vertex adapter: project not set (config.project or env GOOGLE_VERTEX_PROJECT)");
+    const p = config.project ?? process.env.GOOGLE_VERTEX_PROJECT ?? resolveCredentials(config).project_id;
+    if (!p) throw new Error("vertex adapter: project not set (config.project, env GOOGLE_VERTEX_PROJECT, or project_id in the credentials)");
     return p;
   }
   async function accessToken() {
@@ -1710,7 +1710,47 @@ function vertexAdapter(config = {}) {
     usage.costUsd = perSec * (req.durationSec ?? 8);
     return { url: `vertex://${op.name}`, bytes: Buffer.from(videoB64, "base64"), mimeType: videoMime, usage };
   }
-  return { name: "vertex", animate };
+  async function vision(req) {
+    const token = await accessToken();
+    const proj = project();
+    const reg = region();
+    const systemParts = [];
+    const contents = [];
+    for (const m of req.messages) {
+      if (m.role === "system") systemParts.push(...partsFrom(m.content));
+      else contents.push({ role: m.role === "assistant" ? "model" : "user", parts: partsFrom(m.content) });
+    }
+    const body = { contents };
+    if (systemParts.length > 0) body.systemInstruction = { parts: systemParts };
+    const genConfig = {};
+    if (req.maxTokens !== void 0) genConfig.maxOutputTokens = req.maxTokens;
+    if (req.temperature !== void 0) genConfig.temperature = req.temperature;
+    if (Object.keys(genConfig).length > 0) body.generationConfig = genConfig;
+    const url = `https://${reg}-aiplatform.googleapis.com/v1/projects/${proj}/locations/${reg}/publishers/google/models/${req.spec.model}:generateContent`;
+    const res = await fetchImpl(url, {
+      method: "POST",
+      headers: { authorization: `Bearer ${token}`, "content-type": "application/json" },
+      body: JSON.stringify(body)
+    });
+    if (!res.ok) {
+      throw new Error(`vertex vision ${res.status}: ${(await res.text().catch(() => "")).slice(0, 300)}`);
+    }
+    const data = await res.json();
+    const text = (data.candidates?.[0]?.content?.parts ?? []).map((p) => p.text ?? "").join("").trim();
+    const inputTokens = data.usageMetadata?.promptTokenCount ?? 0;
+    const outputTokens = data.usageMetadata?.candidatesTokenCount ?? 0;
+    const usage = freshUsage({
+      provider: "vertex",
+      model: req.spec.model,
+      transport: "http",
+      capability: "vision",
+      inputTokens,
+      outputTokens
+    });
+    usage.costUsd = computeCost("vertex", req.spec.model, inputTokens, outputTokens);
+    return { text, usage };
+  }
+  return { name: "vertex", animate, vision };
 }
 
 // src/providers/deepl.ts
@@ -2601,14 +2641,17 @@ var aiConfigSchema = z.object({
   // Functions can't be deeply validated — z.custom asserts the TS type and
   // passes the value through untouched.
   providers: z.record(z.string(), z.custom()).optional(),
-  costSink: z.custom().optional(),
+  /** Omit → cost-tracking auto-wires from the upmetrics env (F034). Pass `null`
+   *  to opt OUT explicitly — for a consumer that already reports its own costs
+   *  and would otherwise be counted twice (F034.3). An object is used as-is. */
+  costSink: z.custom().nullable().optional(),
   budget: budgetSchema.optional(),
   availability: availabilitySchema.optional()
 });
 
 // src/version.ts
-var VERSION = "0.25.1";
-var SDK_TAG = "@broberg/ai-sdk@0.25.1";
+var VERSION = "0.27.0";
+var SDK_TAG = "@broberg/ai-sdk@0.27.0";
 
 // src/cost/sinks/upmetrics.ts
 function upmetricsSink(config) {
@@ -2728,7 +2771,7 @@ function defaultCostSink() {
 function createAI(config = {}) {
   const cfg = aiConfigSchema.parse(config);
   const providers = cfg.providers ?? defaultProviders;
-  const costSink = cfg.costSink ?? defaultCostSink();
+  const costSink = cfg.costSink === void 0 ? defaultCostSink() : cfg.costSink ?? void 0;
   const budget = cfg.budget ? new BudgetGuard(cfg.budget) : void 0;
   const estTokens = (s) => Math.ceil(s.length / 4);
   async function preflight(spec, estInTokens, estOutTokens) {