anthropic adapter: baseUrl ending in /v1/messages doubles the path, input_json_delta is not scoped to tool_use, and Anthropic-compatible relays need tool-call quirk tolerance #765

Client or integration

Codex CLI

Provider or upstream service

AgentRouter (agentrouter.org), an Anthropic Messages-compatible relay. Two of the four problems below are provider-agnostic and affect the anthropic adapter for any host.

OpenCodex version

2.7.43

Endpoint or capability

POST /v1/messages — streaming tool calls, and baseUrl normalisation.

Current behaviour

1. baseUrl ending in /v1/messages produces a doubled path (provider-agnostic).

const base = provider.baseUrl.replace(/\/v1\/?$/, "");
const url = `${base}/v1/messages`;

The regex only strips a trailing /v1. A baseUrl of https://host/v1/messages — a natural thing to paste from provider documentation — becomes https://host/v1/messages/v1/messages and every request 404s. Affects any Anthropic-compatible provider, not just this one.

2. input_json_delta is applied outside tool_use blocks (provider-agnostic).

} else if (delta.type === "input_json_delta" && typeof delta.partial_json === "string") {

There is no check that the current content block is a tool_use. Per the Messages streaming spec input_json_delta only occurs within tool_use; a provider that emits it elsewhere (or emits deltas after a block closes) corrupts the accumulated tool arguments.

3. tool_use blocks arrive without an id. The adapter forwards block.id ?? "", so an empty tool-call id reaches the Codex client, which then cannot correlate the tool result.

4. Streams close without terminal frames. This relay sometimes ends an SSE stream after content blocks without message_stop / message_delta, so the adapter never emits a completion event and the client hangs waiting for one.

Additionally, tool_use.input sometimes arrives as a JSON string rather than an object, occasionally with a malformed prefix. JSON.stringify(block.input ?? {}) then produces a doubly-encoded string that the client rejects.

Expected behaviour

1. baseUrl normalisation should tolerate /, /v1 and /v1/messages suffixes and produce exactly one /v1/messages.
2. input_json_delta should only accumulate while the current block is a tool_use, per the Anthropic streaming spec.
3. A tool_use block with no usable id should get a synthesised one (toolu_…) so tool results can be correlated, rather than an empty string being forwarded.
4. A stream that ends after content without terminal frames should still produce a completion event instead of leaving the client waiting.
5. tool_use.input that arrives as a JSON string should be parsed before being re-serialised, so the client receives a single-encoded object.

Items 1 and 2 look like plain correctness fixes worth applying unconditionally. Items 3-5 are quirk tolerance and could reasonably be gated per provider — locally I gate them on the endpoint hostname, but a capability flag on the provider config would be cleaner and would not hard-code a vendor name in the adapter.

Minimal redacted request or reproduction

Configure an Anthropic-compatible provider and route a tool-using model through it:

{
  "id": "<provider-id>",
  "type": "anthropic",
  "baseUrl": "https://<host>/v1/messages",
  "authMode": "key",
  "apiKey": "<redacted>"
}

1. With that baseUrl, every request 404s (problem 1). Changing it to https://<host> works around it.
2. With a working baseUrl, ask the model to call a tool. The tool call arrives with an empty id and/or doubly-encoded arguments (problems 3 and 5).
3. Some responses end without message_stop, and the client waits indefinitely (problem 4).

Actual response or error

Problem 1 — the composed URL:

POST https://<host>/v1/messages/v1/messages  ->  404

Problem 3/5 — what reaches the client:

{ "type": "tool_call_start", "id": "", "name": "shell" }
{ "type": "tool_call_delta", "arguments": "\"{\\\"command\\\":[\\\"ls\\\"]}\"" }

Expected:

{ "type": "tool_call_start", "id": "toolu_<generated>", "name": "shell" }
{ "type": "tool_call_delta", "arguments": "{\"command\":[\"ls\"]}" }

Problem 4 — the stream simply stops after the last content_block_stop, with no message_delta or message_stop.

Upstream documentation

• Anthropic Messages streaming, including input_json_delta occurring only within tool_use blocks and the message_delta / message_stop terminators: https://docs.anthropic.com/en/docs/build-with-claude/streaming
• Tool use, including the toolu_ id format used to correlate tool_use with tool_result: https://docs.anthropic.com/en/docs/build-with-claude/tool-use

Suggested mapping or implementation notes

What I run locally, in case it is useful:

// 1. Tolerate /, /v1 and /v1/messages suffixes.
export function anthropicMessagesUrl(baseUrl: string): string {
  const base = baseUrl.replace(/\/(?:v1(?:\/messages)?)?\/?$/, "");
  return `${base}/v1/messages`;
}

// 3. Synthesise a correlatable id when the provider omits one.
function ensureAnthropicToolUseId(id: unknown): string {
  return typeof id === "string" && id.trim() !== ""
    ? id
    : `toolu_${crypto.randomUUID().replace(/-/g, "")}`;
}

// 5. Accept a JSON string, recovering from a malformed prefix, and emit one object.
function normalizeAnthropicToolInput(input: unknown): string {
  let value = input;
  if (typeof input === "string") {
    const raw = input;
    try {
      value = JSON.parse(raw);
    } catch {
      for (let index = raw.lastIndexOf("{"); index >= 0; index = raw.lastIndexOf("{", index - 1)) {
        try { value = JSON.parse(raw.slice(index)); break; } catch { /* try an inner object */ }
      }
    }
  }
  return value && typeof value === "object" && !Array.isArray(value)
    ? JSON.stringify(value)
    : "{}";
}

For problem 2 the fix is just adding the block-type guard:

} else if (currentBlockType === "tool_use" && delta.type === "input_json_delta" && ...) {

For problem 4 I track whether any content was seen and emit the completion event when the stream ends without a terminator.

One convention note worth documenting. Any per-provider predicate that can throw on a malformed baseUrl must be resolved lazily inside buildRequest, not eagerly in createAnthropicAdapter. I initially resolved mine in the factory and it broke tests/anthropic-hardening.test.ts -> rethrows a malformed baseUrl when prompt caching is enabled, which expects a rejected promise rather than a constructor throw. usesNativeAnthropicEndpoint already gets this right by being called inside the request path; making that explicit would help anyone adding a provider check.

Additional context and attachments

Happy to send a PR. Items 1 and 2 are two-line unconditional fixes; for 3-5 I would rather add a small provider capability flag than hard-code a hostname, and would welcome a hint on the shape you would accept.

Checks

• I searched existing provider and compatibility issues.
• The request and response were redacted.
• The expected behaviour is based on an upstream specification or a concrete client requirement.