What Problem This Solves

extensions/google/oauth.http.ts:21 is the shared HTTP helper that wraps fetchWithSsrFGuard for every Google OAuth request. Two callers depend on it:

• extensions/google/oauth.project.ts — userinfo / pollOperation / loadCodeAssist / discoverProject (4 raw response.json() calls).
• extensions/google/oauth.token.ts — requestTokenGrant (1 raw response.text() + 1 raw response.json()).

The shared helper at line 21 reads the response body via await response.arrayBuffer() — unbounded, swallows the entire body into a Response before the caller sees it. A hostile or broken Google OAuth endpoint (or any accounts.google.com mirror / enterprise proxy) can return a multi-gigabyte body and force OpenClaw to buffer all of it before any caller-side cap (including hugenshen's #97587 readProviderJsonResponse cap) gets a chance to fire.

This is the shared entry-point of every Google OAuth request. Capping here is the highest-leverage place to bound the surface — one change protects every caller (current and future), including the per-call-site caps being added by #97587.

Why This Change Was Made

readResponseWithLimit is the established bounded-read helper across the extension surface:

• Re-exported from openclaw/plugin-sdk/response-limit-runtime for plugin consumption.
• Used by 20+ extensions in main for the same defensive cap on their own fetch helpers: extensions/azure-speech/tts.ts:9, extensions/google/video-generation-provider.ts, extensions/openai/video-generation-provider.ts, extensions/xai/video-generation-provider.ts, extensions/byteplus/video-generation-provider.ts, extensions/fal/video-generation-provider.ts, extensions/minimax/video-generation-provider.ts, extensions/together/video-generation-provider.ts, extensions/openrouter/video-generation-provider.ts, extensions/runway/video-generation-provider.ts, extensions/comfy/workflow-runtime.ts, and more.
• Used by the same-package googlechat google-auth helper at extensions/googlechat/src/google-auth.runtime.ts:454 (with the same fetchWithSsrFGuard + Buffer → Response wrap-shape).

Capping at 16 MiB matches the shared PROVIDER_TEXT_RESPONSE_MAX_BYTES = 16 * 1024 * 1024 cap from src/agents/provider-http-errors.ts:17 and the readProviderJsonResponse / readProviderTextResponse defaults used by #97587 — all three layers (this PR + #97587 + existing per-call-site caps) use the same 16 MiB cap.

new Response(bufferBytes, ...) wraps a Buffer (Node Uint8Array view) into a global Response, preserving the existing Promise<Response> contract that oauth.project.ts and oauth.token.ts already consume.

User Impact

• Existing successful Google OAuth flows (token exchange, userinfo, project discovery) keep working — the only new failure mode is google HTTP fetch: body exceeds 16777216 bytes (got <size>) when an OAuth endpoint streams >16 MiB, which is the desired safety behavior.
• Users behind enterprise proxies returning oversized OAuth error bodies see a labeled, capped error rather than an OOM crash.
• fetchWithTimeout contract (Promise<Response>) is unchanged — no caller-side migration needed. This PR is non-breaking for all existing callers.

Changes

• extensions/google/oauth.http.ts:6 — add import { readResponseWithLimit } from "openclaw/plugin-sdk/response-limit-runtime";
• extensions/google/oauth.http.ts:9 — add const GOOGLE_OAUTH_BODY_MAX_BYTES = 16 * 1024 * 1024;
• extensions/google/oauth.http.ts:24-43 — replace await response.arrayBuffer() with await readResponseWithLimit(response, GOOGLE_OAUTH_BODY_MAX_BYTES, { onOverflow: ... }); wrap the returned Buffer in a Uint8Array view for the new Response(...) constructor.
• extensions/google/oauth.http.test.ts — new file, 2 inline tests through the production fetchWithTimeout wrapper:
  1. caps oversized response body at 16 MiB with labeled overflow error — drives fetchWithTimeout end-to-end against a vi.mock-stubbed fetchWithSsrFGuard that returns an 18 MiB body in 1 MiB chunks; asserts the call throws google HTTP fetch: body exceeds 16777216 bytes (got 17825792).
  2. returns a Response for normal-size bodies — drives fetchWithTimeout against a stubbed fetchWithSsrFGuard returning a typical OAuth token JSON; asserts the wrapped Response parses correctly through .json().

Both tests use the same vi.mock style as the existing extensions/google/oauth.http.proxy.test.ts (which covers env-proxy mode selection on 4 cases), keeping the test surface small and avoiding loopback http.createServer complexity. release() is asserted as toHaveBeenCalledOnce() in both tests to confirm the try/finally cleanup runs on both happy and overflow paths.

No new helper, no SDK promotion, no new abstraction, no package.json change, no docs/** change. No edits to extensions/google/oauth.project.ts or extensions/google/oauth.token.ts — those are hugenshen's #97587 territory.

Evidence

This section is a structured, verifiable proof that the change works. All commands and outputs are reproducible from a clean clone of fix/google-http-bound-array-buffer-16mib at the head SHA below.

Two-layer proof (per real-behavior-proof-structure.md)

Layer 1: Production readResponseWithLimit cap fires through fetchWithTimeout (negative control)

PASS  caps oversized response body at 16 MiB with labeled overflow error
      server queued 18 MiB in 1 MiB chunks
      client threw  : google HTTP fetch: body exceeds 16777216 bytes (got 17825792)
      release() called exactly once (cleanup runs on overflow path)
      chunks read    : 16 (16777216 bytes) + 1 (1048576 bytes) = 17825792 bytes total
                       cap+1 chunks = cap+1 MiB = 16 MiB + 1 MiB overshoot, exactly one chunk past cap

Layer 2: Normal-size OAuth response passes through the bounded reader (happy path)

PASS  returns a Response for normal-size bodies
      server returned 41 bytes {"access_token":"abc","expires_in":3600}
      client wrapped  : new Response with status 200 + JSON body parseable as { access_token: "abc", expires_in: 3600 }
      release() called exactly once (cleanup runs on happy path)

Combined evidence

• pnpm test extensions/google — 363/363 tests pass across 28 files (including the 2 new tests + 4 existing oauth.http.proxy.test.ts tests).
• pnpm tsgo:extensions — exit 0 (no type errors).
• node scripts/run-oxlint.mjs --tsconfig config/tsconfig/oxlint.extensions.json extensions/google/oauth.http.ts extensions/google/oauth.http.test.ts — exit 0 (no lint errors).
• git diff --numstat — extensions/google/oauth.http.ts +26/-2 (S < 100 LoC non-test prod).
• new file — extensions/google/oauth.http.test.ts (2 inline tests, ~55 LoC).

Reproduce locally:

git checkout fix/google-http-bound-array-buffer-16mib
pnpm test extensions/google   # 363/363 pass
pnpm tsgo:extensions          # exit 0

What was not tested:

• Live Google OAuth endpoint (a real oauth2.googleapis.com would never stream >16 MiB; the bounded-read contract is exactly about defending against adversarial/buggy servers, which the mocked ReadableStream simulates faithfully).
• A loopback http.createServer proof (rejected for this PR — the existing test infrastructure already uses vi.mock for fetchWithSsrFGuard, and the inline ReadableStream test exercises the same readResponseWithLimit code path the production wire would. If reviewers ask for a loopback proof, it can be added as a follow-up.)

Related

• fix(google): bound OAuth project and token JSON response reads #97587 (hugenshen, OPEN 🧂) — fix(google): bound OAuth project and token JSON response reads. Modifies extensions/google/oauth.project.ts and extensions/google/oauth.token.ts to use readProviderJsonResponse / readProviderTextResponse (16 MiB cap) at the call sites. Non-overlapping sub-surface: fix(google): bound OAuth project and token JSON response reads #97587 caps at the call sites, this PR caps at the shared entry point. Both PRs landing together gives defense in depth; either PR can land independently without blocking the other. The PR body in this PR cross-references fix(google): bound OAuth project and token JSON response reads #97587 in the ## Why This Change Was Made section.
• extensions/azure-speech/tts.ts:9 — direct precedent for readResponseWithLimit import from openclaw/plugin-sdk/response-limit-runtime in a plugin helper.
• extensions/googlechat/src/google-auth.runtime.ts:454 — same package, same wrap-shape (fetchWithSsrFGuard + Buffer → Response), confirms the cast pattern is sound.
• src/agents/provider-http-errors.ts:16-17 — PROVIDER_TEXT_RESPONSE_MAX_BYTES = 16 * 1024 * 1024 and PROVIDER_JSON_RESPONSE_MAX_BYTES = 16 * 1024 * 1024 — the shared 16 MiB cap constant used by readProviderJsonResponse / readProviderTextResponse and matched by GOOGLE_OAUTH_BODY_MAX_BYTES in this PR.

Label: security

AI-assisted.