Opt-in Claude Code identity injection for OAuth subscription tokens (#74)

[doramirdor]

Summary

Implements the opt-in Claude Code identity injection flag floated in #74. After v0.21.0 fixed the credential path, the reporter (@ThunderEagle) retested and confirmed: Haiku works, but Sonnet/Opus still return an immediate rate_limit_error through both nadirclaw test and a direct POST /v1/messages. Haiku-OK-but-premium-gated on the same subscription token is the classic signature of Anthropic 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.

What this does

New env flag NADIRCLAW_CLAUDE_CODE_IDENTITY=1 (default off). When enabled, NadirClaw prepends

You are Claude Code, Anthropic's official CLI for Claude.

as the first system block on outgoing Anthropic requests, on both surfaces:

• /v1/messages (the Anthropic-native path Claude Code + the reporter use)
• the direct Anthropic OAuth path inside /v1/chat/completions

Guard rails:

• OAuth only — injected only for Bearer / sk-ant-oat* tokens. No effect on sk-ant-api* API keys (those use x-api-key).
• Idempotent — no-op if the request already leads with the identity block (so real Claude Code traffic is untouched).
• Non-destructive — any caller-supplied system prompt is preserved, appended after the identity block; system is normalized to Anthropic's block-array form.
• Observable — the decision is recorded as claude_code_identity on the request log.
• Default off, since it changes the system prompt the model sees.

Drive-by fix

The direct Anthropic OAuth path in /v1/chat/completions was forwarding 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.

Testing

• New TestClaudeCodeIdentityInjection unit tests for the helper (no-system / string / block-array / already-present / noop cases).
• New endpoint tests asserting injection happens only for OAuth + flag-on, and not for flag-off or API-key tokens.
• Full test_server.py, test_oauth.py, test_tool_calling.py, test_streaming_fallback.py, test_thinking_passthrough.py pass (91 tests).

For the reporter to try

pip install -U nadirclaw        # once released, or install from this branch
export NADIRCLAW_CLAUDE_CODE_IDENTITY=1
nadirclaw serve
# then retest claude-sonnet-4-6 / opus through /v1/messages

If Sonnet/Opus start returning completions with the flag on, this was the cause. If they still rate_limit_error, that points to plan-tier model-access gating (Pro vs Max) on the subscription itself rather than anything actionable in NadirClaw.

Refs #74 — keeping the issue open until the reporter confirms the flag resolves the gating (or rules it out as plan-tier).

🤖 Generated with Claude Code