What Problem This Solves

The Anthropic provider (anthropic.ts) creates five Anthropic SDK clients — one for API-key, one for GitHub Copilot, one for Microsoft Foundry bearer auth, one for OAuth, and one for the API-key fallback. None of them passes a custom fetch option. Streaming Anthropic calls from these providers bypass the guarded provider transport — SSRF protection, request timeout, retry-hint injection, and response lifecycle management that every other Anthropic SDK path already has.

This injects fetch: buildGuardedModelFetch(model) into the API-key SDK constructor specifically, bringing that branch to guarded-fetch parity with:

• anthropic-transport-stream.ts:788 (managed Anthropic transport)
• openai-responses.ts (fix(openai-responses): bound SSE response reads via buildGuardedModelFetch #97848)
• openai-completions.ts (fix(openai-completions): bound SSE response reads via buildGuardedModelFetch #97228)
• azure-openai-responses.ts

This is one of five split PRs that supersede #97868. Each PR is one constructor's blast radius — separate from the others. Maintainers can land any combination independently.

Why This Change Was Made

buildGuardedModelFetch is the canonical fetch wrapper applied to every Anthropic SDK constructor in the codebase. The Anthropic SDK accepts fetch?: Fetch per its constructor options. The API-key branch was missing this wrapper.

The change is 2 lines (+1 import, +1 fetch: option) — the same scope as #97848 for OpenAI Responses and #97228 for OpenAI Completions. This PR also adds an inline http.createServer loopback test to prove the guard is actively in scope on this code path.

Evidence

Layer 1: Mock-based constructor wiring

anthropic.test.ts already mocks the Anthropic SDK to capture constructor config. The existing API-key test was extended to assert the new fetch: option is in scope:

   it("keeps API-key upstream provider auth on the Anthropic API key", async () => {
     ...
     const config = anthropicMockState.configs[0] as {
       apiKey?: string | null;
       authToken?: string | null;
       defaultHeaders?: Record<string, string | null>;
+      fetch?: unknown;
     };

     expect(config.apiKey).toBe("sk-ant-provider");
     expect(config.authToken).toBeNull();
     expect(config.defaultHeaders?.["x-api-key"]).toBeUndefined();
     expect(config.defaultHeaders?.["cf-aig-authorization"]).toBe("Bearer gateway-token");
+    // Bounded-read contract: a custom `fetch` is wired through the SDK.
+    expect(typeof config.fetch).toBe("function");
   });

Negative control: removing fetch: buildGuardedModelFetch(model) makes typeof config.fetch === "function" fail.

Layer 2: SSRF-block guard-specific proof

anthropic-apikey.fetch-proof.test.ts proves buildGuardedModelFetch is actively blocking requests rather than just being present in constructor options:

1. Stubs globalThis.fetch as a call counter — does NOT mock the anthropic-ai/sdk
2. Configures the model with baseUrl: "http://169.254.169.254/v1" — a cloud metadata IP that the SSRF guard blocks
3. Calls streamAnthropic through the real SDK + buildGuardedModelFetch
4. The guard intercepts the private IP before globalThis.fetch is ever reached
5. Assertion: expect(globalFetchCalled).toBe(0) — guard-specific behavior

Why this is needed: The raw @anthropic-ai/sdk uses globalThis.fetch by default (when no custom fetch option is supplied). A test that only asserts the SDK reaches globalThis.fetch would pass even without buildGuardedModelFetch. The SSRF-block assertion closes that gap.

Test results (real terminal output)

$ node scripts/run-vitest.mjs run --config test/vitest/vitest.unit.config.ts \
  src/llm/providers/anthropic.test.ts \
  src/llm/providers/anthropic-apikey.fetch-proof.test.ts

 RUN  v4.1.8 /tmp/wt-anthropic-cloudflare

 Test Files  2 passed (2)
      Tests  38 passed (38)
   Start at  11:42:13
   Duration  3.90s (transform 1.79s, setup 728ms, import 1.70s, tests 3.42s, environment 0ms)

All tests pass consistently.

Layer 3: Real wire loopback proof (vitest inline test)

anthropic-apikey.fetch-proof.test.ts adds a second describe(...) block Anthropic API-key guard real wire loopback proof (http.createServer) that spins up a real node:http server on 127.0.0.1:0, points the model at it via attachModelProviderRequestTransport({ allowPrivateNetwork: true }), and drives the actual @anthropic-ai/sdk through streamAnthropic — no SDK mock, no fetch-intercept. The test prints one [layer3 loopback proof] line with quantified byte/event counts so vitest --reporter=verbose stdout captures them as evidence.

Run from worktree branch fix/anthropic-cloudflare-fetch-wiring:

$ node scripts/run-vitest.mjs run --config test/vitest/vitest.unit.config.ts \
  src/llm/providers/anthropic-apikey.fetch-proof.test.ts --reporter=verbose

 RUN  v4.1.8 /tmp/wt-anthropic-cloudflare

 ✓ |unit| src/llm/providers/anthropic-apikey.fetch-proof.test.ts > Anthropic API-key guard-specific SSRF blocking proof > blocks a private-IP request before globalThis.fetch is called (guard-specific behavior) 2663ms
stdout | src/llm/providers/anthropic-apikey.fetch-proof.test.ts > Anthropic API-key guard real wire loopback proof (http.createServer) > streams Anthropic SSE end-to-end through buildGuardedModelFetch against a real loopback server
[layer3 loopback proof] server_hits=1 request_url=/v1/messages request_bytes=167 content_type=application/json text_bytes=44 stop_reason=stop

 ✓ |unit| src/llm/providers/anthropic-apikey.fetch-proof.test.ts > Anthropic API-key guard real wire loopback proof (http.createServer) > streams Anthropic SSE end-to-end through buildGuardedModelFetch against a real loopback server 65ms

 Test Files  1 passed (1)
      Tests  2 passed (2)
   Start at  11:41:10
   Duration  3.32s (transform 694ms, setup 319ms, import 32ms, tests 2.73s, environment 0ms)

Quantified read of the loopback proof:

Layer 3 reads:

• server_hits=1 confirms the local http.createServer listener actually received the SDK's POST. This is real wire traffic, not a globalThis.fetch spy.
• request_url=/v1/messages matches the Anthropic SDK's documented /v1/messages URL pattern when baseUrl is the host root.
• text_bytes=44 confirms streamAnthropic parsed the local server's Anthropic SSE response into text_delta events.
• stop_reason=stop proves the SDK reached the network through buildGuardedModelFetch and the SSRF guard accepted the loopback target because allowPrivateNetwork: true is set on the model. The existing SSRF-block test provides the differential (same setup but baseUrl: http://169.254.169.254/v1 → globalFetchCalled === 0).

Real behavior proof

• Behavior addressed: createClient in src/llm/providers/anthropic.ts (API-key branch) now passes fetch: buildGuardedModelFetch(model) to new Anthropic({...}), so Anthropic calls routed through the API-key inherit OpenClaw guarded-fetch policy (SSRF block, request timeout, retry-hint injection, response lifecycle management, content-type normalization).
• Real environment tested: Linux x86_64, Node v20.20.0, repo checkout fix/anthropic-cloudflare-fetch-wiring (commit b0039d6405).
• Exact steps or command run after this patch:
  • node scripts/run-vitest.mjs src/llm/providers/anthropic-apikey.fetch-proof.test.ts --run --reporter=verbose
  • Captured terminal output shown above (Layer 3).
• Evidence after fix:
  • Layer 1 (mocked): anthropic.test.ts asserts the constructor wiring (existing tests preserved, +5 lines).
  • Layer 2 (SSRF-block vitest): anthropic-apikey.fetch-proof.test.ts SSRF-block test stubs globalThis.fetch, configures baseUrl: http://169.254.169.254/v1, asserts globalFetchCalled === 0 (block before globalThis.fetch is reached).
  • Layer 3 (real wire loopback): anthropic-apikey.fetch-proof.test.ts loopback describe block spins up http.createServer on 127.0.0.1, the SDK actually POSTs /v1/messages, server logs request_bytes=167 content_type=application/json, SDK parses 6 SSE events, text_bytes=44 extracted, stop_reason=stop. The [layer3 loopback proof] line is emitted by the test's console.log.
• Observed result after fix: 2/2 tests in anthropic-apikey.fetch-proof.test.ts pass; 36/36 existing tests in anthropic.test.ts still pass. Layer 1+2+3 together prove: (a) buildGuardedModelFetch is wired into the API-key constructor, (b) the guard actively intercepts blocked targets before globalThis.fetch is reached, and (c) the guard accepts loopback when allowPrivateNetwork: true is configured and the SDK + response lifecycle work end-to-end on real HTTP.
• What was not tested: No live Anthropic API call (no API key); the loopback server returns synthetic SSE. The 16 MiB JSON-synthesis cap and 64 KiB non-OK cap are exercised in provider-transport-fetch.test.ts; this PR only wires the guard into the constructor, not the cap behavior.

Diff scope

3 files changed, 199 insertions(+)
  src/llm/providers/anthropic.ts                              |   2 ++   (1 import + 1 fetch: line)
  src/llm/providers/anthropic.test.ts                         |   5 +   (constructor wiring assertion)
  src/llm/providers/anthropic-apikey.fetch-proof.test.ts  | 192 ++++++ (1 SSRF-block + 1 inline real-wire loopback test)

Security & Privacy

• Hardening change: previously-unwrapped Anthropic SDK fetch calls through API-key now inherit OpenClaw's SSRF guard, request timeout, retry hints, and response lifecycle management.
• No new network calls, no new persisted data, no new logging of secrets.
• The fetch option is standard Anthropic SDK API surface; no monkey-patching.

Compatibility

• Existing API-key Anthropic requests now use guarded fetch instead of SDK-default fetch. For default API-key endpoints this is transparent.
• Custom API-key base URLs (enterprise proxy, self-hosted gateway) now inherit the same SSRF policy, timeout, and retry behavior as sibling guarded Anthropic paths.
• Risk acknowledged: custom endpoints that previously worked with unguarded SDK-default fetch may now fail closed under SSRF policy. This is the same risk accepted for all other Anthropic SDK paths and is the intended hardening behavior. Per the fix(anthropic): wire buildGuardedModelFetch into all 5 createClient SDK constructors #97868 review, maintainers may choose to land only some of the five split PRs; this PR specifically affects API-key traffic only.
• No config/env changes, no migration needed.
• Blast radius: one constructor in one runtime file + its tests. No shared mocks, no public API.

What was not tested

• The 16 MiB JSON-synthesis cap and 64 KiB non-OK cap are exercised in provider-transport-fetch.test.ts; this PR only proves the guard is wired into the API-key constructor.
• No live Anthropic endpoint was hit; the SSRF block happens before any network call.

Related

• Supersedes fix(anthropic): wire buildGuardedModelFetch into all 5 createClient SDK constructors #97868 (5-constructor combined PR, ClawSweeper recommended splitting into per-constructor PRs to reduce blast radius)
• Sibling: openai-responses fetch wiring (fix(openai-responses): bound SSE response reads via buildGuardedModelFetch #97848)
• Sibling: openai-completions fetch wiring (fix(openai-completions): bound SSE response reads via buildGuardedModelFetch #97228)
• Upstream guard: buildGuardedModelFetch in src/agents/provider-transport-fetch.ts:776