Google services for the Aphrody Bun fork

@aphrody/bun-google uses the built-in bun:google transport. It needs the fork containing that module. The package has no npm runtime dependencies.

@aphrody/bun-google/edge-proxy owns the disk caching proxy used by Google Fonts. It uses Bun HTTP, file and compression APIs directly, with no Aphrody checkout, Yolo runtime or runtime SDK dependency. Its loopback tests cover cache fills, concurrent misses, Brotli variants, immutable assets and rejected routes.

import { createGoogleServices } from "@aphrody/bun-google";

// Supply a provider owned by your existing authenticated application.
const google = createGoogleServices({ credentialProvider });
const document = await google.docs.get(documentId, signal);
await google.docs.batchUpdate(
  documentId,
  reviewedRequests,
  document.revisionId!,
  signal,
);

The credential provider returns a bearer token for each request. The transport uses Bun fetch in process, enforces Google HTTPS or loopback endpoints, rejects redirects, and preserves cancellation. Credentials are never discovered implicitly. Existing OS credential storage and ADC adapters retain ownership.

For dedicated email, Google user identity and keyless Cloud automation, follow the dedicated-account runbook. OVH forwarding, Workspace/Gmail authorization and Cloud IAM qualify separately. The runbook records the verified OVH forwarder and Windows keyless Cloud setup; the official verification email was delivered. Signup awaits an authentication code deliberately withheld by the connector; no dedicated Google user, Gmail mailbox or password has been created.

| Surface | Executable contract | | ------------------------------- | ---------------------------------------------------------------- | | createClient, googleRuntime | Native REST transport and public configuration | | drive | List/get/export files; bounded page iterator | | docs | Get document; update with required reviewed revision | | gemini | JSON generateContent; streaming SSE Response | | chromium, obscura, dns | Exact aliases of existing runtime owners | | crates/boq | Standalone aphrody-boq Rust codec, source provenance preserved |

Windows Credential Manager target reads and Antigravity Cloud Code/Vertex models, generation and streaming use the native in-process SDK. The explicit credential provider returns token strings to JavaScript; the desktop owner retains refresh and store writes. Interactive PKCE login keeps the existing CLI owner.

Drive reconciliation, resumable model uploads, Fonts proxy and Gemini Web adapters were transferred under aphrody/. Their complete engine and release qualification remains separate from this REST SDK. Protected credentials and cookie stores retain their original locations. The transferred @aphrody/google facade is a root Bun workspace package. Its Google transport resolves through @aphrody/bun-google; it does not depend on the location of a neighbouring Aphrody checkout. This package does not claim complete Google API or JSC/V8 compatibility. It performs no automatic retries of writes.

Contracts follow the official Drive list, Drive export, Docs batchUpdate and Gemini generateContent references.

Commands and source catalog

Native bun search

The optional runtime feature --google-search=on registers bun search directly with the existing Rust request, cookie, HTTP and SERP owners. Enable it before the execution arguments when building the checkout:

bun bd --google-search=on search --help
bun bd --google-search=on search --json --num 3 --timeout-ms 10000 -- "site:example.org evidence"
bun bd --google-search=on search --anonymous --json --num 3 -- "public evidence"

Without this feature, help remains available and a search exits with an explicit provider-unavailable error. Source implementation and fixture coverage do not qualify an installed binary; run the changed-binary tests and authenticate only with the qualified release.

The default command requires account cookies from ~/.aphrody/google-cookies.json. An explicit absolute APHRODY_GOOGLE_COOKIES_JSON environment path selects a different protected jar. Keep the jar private to its owner (mode 0600 on Unix); cookie values are never command arguments or output. The command does not import browser profiles, perform login or fall back to anonymous search when credentials are unavailable. --anonymous disables account-cookie reads explicitly.

Flags are --json, --rich, --anonymous, --num 1..20, --start 0..1000, --hl <language>, --gl <country>, --domain <google-host>, --safe active|off and --timeout-ms 1..120000. The default deadline is 30 seconds and the default result limit is 10. --domain selects the Google search host; use a site: query to constrain discovered sources. Put -- before the query to stop option parsing. The transport is native fetch, with an 8 MiB response cap and a bounded redirect chain restricted to Google HTTPS destinations. There is no browser or external CLI fallback.

JSON output contains query, authenticated, transport, servedFromCache and organic rows (position, title, url, snippet), with optional totalResults and correctedQuery. --rich adds structured SERP evidence. Success exits 0; invalid requests, unavailable credentials, blocked requests, transport failures and elapsed deadlines exit 2 with a redacted error on stderr. They produce no successful JSON response. SIGINT cancels the request and exits 130. Callers should also bound subprocess output and cancellation.

This authenticated CLI is separate from the public search SDK below. Public site requests must retain their anonymous policy and must never inherit its private jar or retry a public search with account cookies.

The ./search subpath exports createGooglePublicSearch. This bounded anonymous adapter calls the existing aphrody web search Rust SERP owner with --no-auth, --safe, fixed google.com and --transport fetch. It never falls back to a browser or loads account cookies. The existing owner may send its constant SOCS consent preference cookie; there are no user/session cookies. Bounds: four jobs, 20-second cancellation deadline, 1 MiB output, 1–20 results and offset at most 1000. Query/organic output preserves the M3 search response contract. It does not extend the authenticated bun:google endpoint allowlist or copy a SERP parser.

The public Bun download/site server routes /api/search only for declared site hosts. It forces web SafeSearch, rejects privileged parameters, drops incoming auth/cookie headers and returns a redacted 503 when Google blocks the request. Deployments install @aphrody/bun-google/search or explicitly configure its source module via GOOGLE_SEARCH_SDK_MODULE, and qualify GOOGLE_SEARCH_OWNER/aphrody. This is a capability-owner CLI adapter, not an in-process native search binding. The Windows anonymous live probe on 2026-10-11 was blocked by Google; fixture success does not establish production availability. No auth fallback was attempted.

Public crawl index and standalone CLI

@aphrody/bun-google/public-index exports createPublicIndexSearch({ root, collections }). It verifies approved public snapshot pointers, SHA-256 content, provenance, exact hosts and limits before building one in-memory SQLite FTS5 projection with isolated collection statistics. Searches return { body, etag, cache }; body includes query, organic, source: "public-crawl-cache", collection and generatedAt. The 128-entry query cache invalidates on changed snapshots. close() releases the projection. No private home index, browser profile, credentials or implicit database supplies missing configuration.

The separate bun-google-index-public binary and ./public-index-cli subpath delegate to that same SDK. They require an absolute export root and an absolute JSON policy file containing the collection configuration map:

{
  "google-public": {
    "pointer": "google-public.pointer.json",
    "hosts": ["developers.google.com"],
    "policySha256": "<64-character checksum of the approved crawl policy>",
    "owner": "aphrody-web-crawl"
  }
}

bun-google-index-public --help
bun-google-index-public --root /absolute/public-export --collections-file /absolute/public-policy.json --collection google-public --query "Bun Rust" --num 10 --start 0

Replace the example paths and policy checksum with the qualified producer's actual public export/configuration; this example is not a production receipt. Policy files are bounded to 64 KiB. Queries, collections and pagination use the SDK bounds; SIGINT/SIGTERM cancel the standalone command. Output is JSON with results, provenance, ETag and cache status. Errors are redacted. These public commands do not acquire Google credentials or contact a live search provider. unicode61 provides Unicode lexical tokenization; comprehensive language-specific stemming and Google-scale coverage are not claimed. A verified public snapshot must exist before these commands return results.

createAntigravityServices implements Cloud Code models/bootstrap, regional Vertex generation, Daily Cloud Code SSE decoding and userinfo. createAntigravityCredentialProvider reads the desktop OAuth envelope by explicit Windows target; it never refreshes or writes. The ./cli subpath exports runGoogleCommand with provider/transport/runtime/signal injection.

bun src/cli.ts config
bun src/cli.ts authstatus --target gemini:antigravity
bun src/cli.ts models --target gemini:antigravity
bun src/cli.ts generate --target gemini:antigravity --model gemini-2.5-flash --prompt "Hello"
bun src/cli.ts stream --target gemini:antigravity --prompt "Hello"

The npm binary is bun-google; package scripts expose the same commands. Config and imports never authenticate. Other CLI commands require an explicit target. Tokens are not accepted as arguments; no protected store is discovered, migrated or written. The provider returns a JS string, not an opaque token. Generic Gemini and Antigravity endpoints have separate caller-owned token contracts.

Transferred Rust owners under aphrody/crates are antigravity, google, gemini-web, gemini-runtime and model-antigravity; crates/boq is independent. The npm package packs SDK sources and Boq rather than the entire transferred application tree. Interactive PKCE login remains with the original process owner; native replacement requires faithful callback/refresh qualification.

Run bun check, bun lint, bun fmt and bun bd -j1 test ./packages/bun-google/test from the root. Tests use mocked stores and loopback HTTP, never live credentials. Source version 1.4.4, registry availability and native release receipts qualify separately; see publication boundary. Release graph catalogs immutable hashes, symbols and relative references with explicit unsupported inputs, truncation and unresolved calls. It is not full dynamic call coverage or JSC/V8 parity. AGENTS.md, CLAUDE.md, GEMINI.md, llms.txt are dedicated entrypoints.