DeepSeek Search 401 with a chat proxy: separate tool credentials and endpoints

Symptoms

Harness discussion #408 and OpenClaw issue #183830 describe a gateway-backed chat route working while search authentication fails. These separate reports support an integration failure pattern, not a measured failure rate or proof that every gateway is affected.

Cause

The pinned Harness module configures its auxiliary Messages-based search endpoint separately from chat. Its readiness check cannot prove a key is valid or the endpoint supports the required search tool. Do not confuse this model-turn search backend with the dedicated Search/Fetch module: the latter uses retrieval endpoints. A proxy-issued chat key sent to the official search service can produce 401.

Safe resolution

Configure chat and retrieval independently, and verify the intended HTTPS origin before attaching a credential. Never probe a credential against arbitrary endpoints to discover where it works. A local preflight can fail closed without sending a request:

function checkToolRoute(endpoint, credentialIssuerOrigin) {
  const route = new URL(endpoint);
  if (route.protocol !== "https:" || route.username || route.password) {
    throw new Error("Invalid tool route");
  }
  if (route.origin !== credentialIssuerOrigin) {
    throw new Error("Tool origin does not match credential issuer");
  }
  return route.href;
}

The issuer origin must come from trusted deployment configuration, not model output. This illustrative check is not a drop-in Harness/OpenClaw patch; it intentionally rejects aliases unless separately reviewed. Do not log keys. Treat authentication 401, balance 402 and throttling 429 as separate states rather than retrying all errors identically. See API key safety and LLM API selection.

OpenClaw PR #183831 merged an official-origin guard for model-key fallback on October 26. That narrow change does not validate explicit search keys or settle account-balance checks. Release status was rechecked during publication: v2026.10.8 was published at 11:56 UTC on October 27, after the earlier research report. Its tag did not contain the reviewed model-credential.ts path, and the fix commit was not an ancestor of that release in the comparison. Do not present the merged patch as a verified shipped remedy.

Verification

Reviewed official DeepSeek documentation and Harness source, plus OpenClaw issue, patch and release metadata on 2026-10-27. No model, Search or Fetch request, credential creation or paid call was executed.

Proposed controls: an official-origin route with the appropriate key, a proxy route that must not inherit an official tool credential, a missing key, and a deliberately mismatched issuer/origin. Inspect the resolved host and error classification without exposing headers. None of these network controls was executed here. Upstream tests and repository build checks do not establish live authentication success.

Limitations

The official tool guide prices Search and Fetch separately at US$0.75 per 1,000 successful requests. Charging requires HTTP 200 plus nonempty results/content; unsuccessful or empty retrievals are not billed. The described DeepSeek-backed server tools used through Harness, Claude Code or Codex begin per-request billing on November 1, 2026 at 00:00 Beijing time (01:00 KST). This is not a price change for every call made by those products. Track tool requests separately from model tokens, especially for the Messages-based search backend.

No actual Search API usage volume, all-proxy compatibility or fixed-release runtime was established. The new DeepSeek profile deliberately covers Search/Fetch rather than an unverified full model catalog. Search terms are candidates, not measured demand; no GetBeta resolution of this exact problem was verified.