v0.21.1: Opt-in Claude Code identity injection for OAuth tokens (#74) (#75)

Anthropic gates premium models (Sonnet/Opus) behind subscription/OAuth
tokens (sk-ant-oat*) unless the request leads with the official Claude
Code identity system block. The real client always sends it; raw API/SDK
callers omit it and get a bare rate_limit_error on those models while
Haiku works.

Add NADIRCLAW_CLAUDE_CODE_IDENTITY (default off). When enabled,
/v1/messages and the OAuth completion path prepend
"You are Claude Code, Anthropic's official CLI for Claude." as the first
system block — only for Bearer/OAuth tokens, only when not already
present, preserving any caller-supplied system prompt after it. Recorded
as claude_code_identity on the request log.

Also fix the OAuth completion path forwarding role:"system" messages
inside the messages array (Anthropic requires system as a top-level
field); system/developer turns are now collected into the top-level
system field.

Co-authored-by: Nadir <info@getnadir.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: 3 people

CHANGELOG.md

@@ -4,6 +4,14 @@ All notable changes to NadirClaw will be documented in this file.
 
 ## [Unreleased]
 
+## [0.21.1] - 2026-06-25
+
+### Added
+- **Opt-in Claude Code identity injection for OAuth tokens** (`NADIRCLAW_CLAUDE_CODE_IDENTITY=1`) — Anthropic gates premium models (Sonnet/Opus) behind subscription/OAuth tokens (`sk-ant-oat*`) unless the request leads with the official Claude Code identity system block. The real Claude Code client always sends it; raw API/SDK callers omit it and get a bare `rate_limit_error` on those models while Haiku works (#74). When enabled, `/v1/messages` and the OAuth completion path prepend `"You are Claude Code, Anthropic's official CLI for Claude."` as the first `system` block — only for Bearer/OAuth tokens (no effect on `sk-ant-api*` keys), only when not already present, preserving any caller-supplied system prompt after it. The decision is recorded as `claude_code_identity` on the request log. Default off, since it changes the system prompt the model sees.
+
+### Fixed
+- **OAuth completion path sent `system` turns as chat messages** — the direct Anthropic OAuth call in `/v1/chat/completions` forwarded `role: "system"` messages inside the `messages` array, which Anthropic's `/v1/messages` API rejects (system must be a top-level field). System/developer turns are now collected into the top-level `system` field before forwarding (#74).
+
 ## [0.21.0] - 2026-06-24
 
 ### Added

README.md

@@ -740,6 +740,17 @@ nadirclaw auth anthropic login
 nadirclaw auth setup-token
 ```
 
+#### Subscription tokens and premium-model access
+
+If you authenticate with a Claude **subscription** token (`sk-ant-oat*`, from `nadirclaw auth anthropic login` or `claude setup-token`) and find that **Haiku works but Sonnet/Opus return an immediate `rate_limit_error`**, Anthropic is likely gating premium models behind the official Claude Code identity. The real client always leads its requests with a fixed identity system block; raw API/SDK callers omit it. Opt in to have NadirClaw prepend it:
+
+```bash
+export NADIRCLAW_CLAUDE_CODE_IDENTITY=1
+nadirclaw serve
+```
+
+When enabled, NadirClaw prepends `"You are Claude Code, Anthropic's official CLI for Claude."` as the first `system` block on OAuth (`sk-ant-oat*`) requests — only when not already present, preserving any system prompt you sent after it. It has **no effect on API-key (`sk-ant-api*`) credentials** and is **off by default**, since it changes the system prompt the model sees.
+
 ### What happens
 
 Claude Code sends every request to Anthropic's API. With NadirClaw in front, each prompt is classified in ~10ms:

nadirclaw/__init__.py

@@ -1,3 +1,3 @@
 """NadirClaw — Open-source LLM router."""
 
-__version__ = "0.21.0"
+__version__ = "0.21.1"

nadirclaw/server.py

@@ -991,16 +991,31 @@ async def _call_litellm(
             if cred_provider == "anthropic" and "sk-ant-oat" in api_key:
                 import httpx
                 model_id = litellm_model.removeprefix("anthropic/")
-                anthropic_messages = [
-                    {"role": m["role"], "content": m["content"]}
-                    for m in call_kwargs.get("messages", [])
-                    if m.get("content") is not None
-                ]
+                # Anthropic /v1/messages requires system prompts as a top-level
+                # `system` field and only accepts user/assistant roles in the
+                # messages array — split system/developer turns out here.
+                system_blocks: list[str] = []
+                anthropic_messages = []
+                for m in call_kwargs.get("messages", []):
+                    if m.get("content") is None:
+                        continue
+                    if m["role"] in ("system", "developer"):
+                        content = m["content"]
+                        if isinstance(content, str):
+                            system_blocks.append(content)
+                        continue
+                    anthropic_messages.append({"role": m["role"], "content": m["content"]})
                 anthropic_body = {
                     "model": model_id,
                     "messages": anthropic_messages,
                     "max_tokens": call_kwargs.get("max_tokens", 1024),
                 }
+                if system_blocks:
+                    anthropic_body["system"] = "\n\n".join(system_blocks)
+                # OAuth tokens gate Sonnet/Opus behind the Claude Code identity
+                # block (#74); prepend it when opted in.
+                if settings.CLAUDE_CODE_IDENTITY:
+                    _inject_claude_code_identity(anthropic_body)
                 if call_kwargs.get("temperature") is not None:
                     anthropic_body["temperature"] = call_kwargs["temperature"]
                 req_extra = request.model_extra or {}
@@ -2296,6 +2311,49 @@ async def view_logs(
 _ANTHROPIC_UPSTREAM = "https://api.anthropic.com/v1/messages"
 _CLAUDE_OAUTH_BETA = "oauth-2025-04-20,claude-code-20250219"
 
+# The exact first system block the official Claude Code client sends. Anthropic
+# gates premium models (Sonnet/Opus) behind subscription OAuth tokens unless the
+# request leads with this identity string; raw API callers omit it and get a
+# bare rate_limit_error on those models (see issue #74). Opt-in via
+# settings.CLAUDE_CODE_IDENTITY — injected only for OAuth (sk-ant-oat*) tokens.
+_CLAUDE_CODE_IDENTITY = "You are Claude Code, Anthropic's official CLI for Claude."
+
+
+def _has_claude_code_identity(system: Any) -> bool:
+    """True if ``system`` already leads with the Claude Code identity block."""
+    if isinstance(system, str):
+        return system.lstrip().startswith(_CLAUDE_CODE_IDENTITY)
+    if isinstance(system, list) and system:
+        first = system[0]
+        if isinstance(first, dict):
+            return str(first.get("text", "")).lstrip().startswith(_CLAUDE_CODE_IDENTITY)
+        if isinstance(first, str):
+            return first.lstrip().startswith(_CLAUDE_CODE_IDENTITY)
+    return False
+
+
+def _inject_claude_code_identity(body: Dict[str, Any]) -> bool:
+    """Prepend the Claude Code identity block to an Anthropic ``/v1/messages`` body.
+
+    Normalizes ``system`` to the block-array form Anthropic expects and inserts
+    the identity as the first block, preserving any caller-supplied system
+    prompt after it. No-op (returns False) if the identity is already first.
+    Returns True if the body was modified.
+    """
+    system = body.get("system")
+    if _has_claude_code_identity(system):
+        return False
+    identity = {"type": "text", "text": _CLAUDE_CODE_IDENTITY}
+    if system is None:
+        body["system"] = [identity]
+    elif isinstance(system, str):
+        body["system"] = [identity, {"type": "text", "text": system}] if system else [identity]
+    elif isinstance(system, list):
+        body["system"] = [identity, *system]
+    else:
+        body["system"] = [identity]
+    return True
+
 
 def _anthropic_messages_to_chat(messages: List[Dict[str, Any]]) -> List[ChatMessage]:
     """Convert Anthropic message blocks to our internal ChatMessage shape.
@@ -2531,6 +2589,12 @@ async def anthropic_messages(
 
     headers = _anthropic_auth_headers(raw, body)
 
+    # OAuth subscription tokens (Bearer) gate Sonnet/Opus behind the Claude Code
+    # identity system block (#74). Inject it for OAuth requests when opted in.
+    identity_injected = False
+    if settings.CLAUDE_CODE_IDENTITY and "Authorization" in headers:
+        identity_injected = _inject_claude_code_identity(body)
+
     import httpx
     from fastapi.responses import StreamingResponse
 
@@ -2539,6 +2603,7 @@ async def anthropic_messages(
         "requested_model": requested_model,
         "selected_model": upstream_model,
         "streaming": stream,
+        "claude_code_identity": identity_injected,
         **analysis_info,
     }
 

nadirclaw/settings.py

@@ -353,6 +353,19 @@ def LOG_SYSTEM_PROMPTS(self) -> bool:
         """When True, log (redacted+truncated) system prompts to the request log."""
         return os.getenv("NADIRCLAW_LOG_SYSTEM_PROMPTS", "").lower() in ("1", "true", "yes")
 
+    @property
+    def CLAUDE_CODE_IDENTITY(self) -> bool:
+        """When True, prepend the Claude Code identity system block to Anthropic
+        OAuth (subscription / ``sk-ant-oat*``) requests.
+
+        Anthropic gates premium models (Sonnet/Opus) behind subscription tokens
+        unless the request's first system block is the official Claude Code
+        identity string — the real client always sends it, raw API callers don't.
+        Opt-in because it changes the system prompt seen by the model. No effect
+        on API-key (``sk-ant-api*``) credentials. See issue #74.
+        """
+        return os.getenv("NADIRCLAW_CLAUDE_CODE_IDENTITY", "").lower() in ("1", "true", "yes", "on")
+
     @property
     def HSTS(self) -> bool:
         """When True, emit Strict-Transport-Security header. Opt-in for HTTPS deployments."""

tests/test_server.py

@@ -127,6 +127,46 @@ def test_extract_text_from_anthropic_response(self):
         assert _extract_text_from_anthropic_response(payload) == "hello world"
 
 
+class TestClaudeCodeIdentityInjection:
+    """The opt-in Claude Code identity system block injection (#74)."""
+
+    IDENTITY = "You are Claude Code, Anthropic's official CLI for Claude."
+
+    def test_inject_into_body_without_system(self):
+        from nadirclaw.server import _inject_claude_code_identity
+        body = {"model": "claude-opus-4-7", "messages": []}
+        assert _inject_claude_code_identity(body) is True
+        assert body["system"] == [{"type": "text", "text": self.IDENTITY}]
+
+    def test_inject_prepends_before_string_system(self):
+        from nadirclaw.server import _inject_claude_code_identity
+        body = {"system": "Be terse."}
+        assert _inject_claude_code_identity(body) is True
+        assert body["system"] == [
+            {"type": "text", "text": self.IDENTITY},
+            {"type": "text", "text": "Be terse."},
+        ]
+
+    def test_inject_prepends_before_block_array_system(self):
+        from nadirclaw.server import _inject_claude_code_identity
+        body = {"system": [{"type": "text", "text": "Be terse."}]}
+        assert _inject_claude_code_identity(body) is True
+        assert body["system"][0] == {"type": "text", "text": self.IDENTITY}
+        assert body["system"][1] == {"type": "text", "text": "Be terse."}
+
+    def test_inject_is_noop_when_identity_already_first(self):
+        from nadirclaw.server import _inject_claude_code_identity
+        body = {"system": [{"type": "text", "text": self.IDENTITY + " extra"}]}
+        assert _inject_claude_code_identity(body) is False
+        assert len(body["system"]) == 1
+
+    def test_inject_is_noop_when_string_system_already_identity(self):
+        from nadirclaw.server import _inject_claude_code_identity
+        body = {"system": self.IDENTITY}
+        assert _inject_claude_code_identity(body) is False
+        assert body["system"] == self.IDENTITY
+
+
 class TestMessagesEndpoint:
     """The /v1/messages Anthropic-compatible proxy endpoint."""
 
@@ -191,6 +231,71 @@ async def post(self, url, headers=None, json=None):
         # OAuth token → Bearer header
         assert captured["auth"] == "Bearer sk-ant-oat01-test"
 
+    @staticmethod
+    def _capturing_client():
+        """Return (FakeClient, captured) recording the forwarded JSON body."""
+        import httpx
+        captured = {}
+
+        class _FakeResponse:
+            status_code = 200
+            headers = {"content-type": "application/json"}
+            def json(self):
+                return {"id": "msg_1", "model": captured.get("model"),
+                        "content": [{"type": "text", "text": "ok"}],
+                        "usage": {"input_tokens": 3, "output_tokens": 1}}
+
+        class _FakeClient:
+            def __init__(self, *a, **kw): pass
+            async def __aenter__(self): return self
+            async def __aexit__(self, *a): return False
+            async def post(self, url, headers=None, json=None):
+                captured["model"] = json.get("model")
+                captured["system"] = json.get("system")
+                captured["auth"] = headers.get("Authorization") or headers.get("x-api-key")
+                return _FakeResponse()
+
+        return httpx, _FakeClient, captured
+
+    def test_identity_injected_for_oauth_when_enabled(self, client, monkeypatch):
+        monkeypatch.setenv("NADIRCLAW_CLAUDE_CODE_IDENTITY", "1")
+        httpx, _FakeClient, captured = self._capturing_client()
+        with patch("nadirclaw.credentials.get_credential", return_value="sk-ant-oat01-test"), \
+             patch.object(httpx, "AsyncClient", _FakeClient):
+            resp = client.post("/v1/messages", json={
+                "model": "claude-opus-4-7", "max_tokens": 10,
+                "messages": [{"role": "user", "content": "hi"}],
+            })
+        assert resp.status_code == 200
+        assert captured["auth"] == "Bearer sk-ant-oat01-test"
+        assert captured["system"][0]["text"].startswith("You are Claude Code")
+
+    def test_identity_not_injected_when_disabled(self, client, monkeypatch):
+        monkeypatch.delenv("NADIRCLAW_CLAUDE_CODE_IDENTITY", raising=False)
+        httpx, _FakeClient, captured = self._capturing_client()
+        with patch("nadirclaw.credentials.get_credential", return_value="sk-ant-oat01-test"), \
+             patch.object(httpx, "AsyncClient", _FakeClient):
+            resp = client.post("/v1/messages", json={
+                "model": "claude-opus-4-7", "max_tokens": 10,
+                "messages": [{"role": "user", "content": "hi"}],
+            })
+        assert resp.status_code == 200
+        assert captured["system"] is None
+
+    def test_identity_not_injected_for_api_key_token(self, client, monkeypatch):
+        """Even with the flag on, an sk-ant-api key uses x-api-key — no injection."""
+        monkeypatch.setenv("NADIRCLAW_CLAUDE_CODE_IDENTITY", "1")
+        httpx, _FakeClient, captured = self._capturing_client()
+        with patch("nadirclaw.credentials.get_credential", return_value="sk-ant-api-test"), \
+             patch.object(httpx, "AsyncClient", _FakeClient):
+            resp = client.post("/v1/messages", json={
+                "model": "claude-opus-4-7", "max_tokens": 10,
+                "messages": [{"role": "user", "content": "hi"}],
+            })
+        assert resp.status_code == 200
+        assert captured["auth"] == "sk-ant-api-test"  # x-api-key path
+        assert captured["system"] is None
+
     def test_upstream_error_is_passed_through(self, client):
         import httpx