fix(anthropic): remove Claude Code fingerprinting from OAuth Messages API path

[teknium1]

Summary

OAuth Anthropic requests now identify as Hermes on the wire. The Claude Code spoofing layer (PR #1597, Mar 2026) is gone from the Messages API path; live-verified against api.anthropic.com with a fresh sk-ant-oat01-* token.

What changed

• Removed (cosmetic identity spoofing):
  • You are Claude Code... system-prompt prepend
  • Hermes Agent→Claude Code, Nous Research→Anthropic system substitutions
  • mcp_ tool-name prefix on outgoing schemas + message history
  • Matching mcp_ strip on inbound tool_use blocks (strip_tool_prefix path + all 5 call sites in run_agent.py and auxiliary_client.py)
  • user-agent: claude-cli/<v> (external, cli) and x-app: cli headers on the Messages API client
• Added:
  • OAuth path strips context-1m-2025-08-07 beta — without Claude Code UA, Anthropic returns 400 This authentication style is incompatible with the long context beta header. OAuth subscription traffic gets the 200K default window, which matches Claude Code's own behavior.
• Kept (auth plumbing, not identity spoofing):
  • _is_oauth_token classifier / is_oauth threading
  • Bearer vs x-api-key auth routing
  • _OAUTH_ONLY_BETAS (claude-code-20250219, oauth-2025-04-20) — Anthropic requires these on the OAuth-gated Messages endpoint
  • _OAUTH_CLIENT_ID (Claude Code's) — Anthropic doesn't issue OAuth creds to third parties
  • claude-cli/<v> UA on platform.claude.com/v1/oauth/token (login + refresh only) — bare requests get Cloudflare 1010 blocked

Validation

	Before (spoofed)	After (realigned)
Messages API / OAuth / simple call	HTTP 200	HTTP 200 ✅ live-verified
Messages API / OAuth / tool call	HTTP 200, mcp_terminal round-trip	HTTP 200, terminal round-trip ✅ live-verified
System prompt on wire	You are Claude Code...\nYou are Claude Code by Anthropic.	You are Hermes Agent by Nous Research.
Tool names on wire	mcp_terminal, mcp_read_file, ...	terminal, read_file, ...
User-Agent on Messages	claude-cli/2.1.74 (external, cli)	python-httpx/* (SDK default)
User-Agent on OAuth token endpoints	claude-cli/...	unchanged (Cloudflare gate)
context-1m beta on OAuth	sent (accepted under spoof)	stripped

Targeted tests: 239 passed + 2 new regression tests covering the no-spoofed-headers and context-1m-strip invariants. 2 pre-existing change-detector failures on main are unrelated.

Live test

Ran a real OAuth-auth Messages request with Hermes identity in system and a tool named terminal:

stop_reason=tool_use  in=580 out=53
TOOL: name='terminal' input={'command': 'echo hello'}

No "you must use extra usage" gate, no 500s, no 401.

Closes / supersedes

• fix: normalize Anthropic PascalCase MCP tool names #16820 (the mcp_ PascalCase normalization patch) — the entire mcp_ round-trip is gone, so the PascalCase dispatch failure it fixed can no longer occur on this path.

[croll83]

Hey can you provide more context on this commit? It broke my anthropic setup (oauth with local claude-cli).

Even with extra usage enabled, it triggers ratelimit 429 error with sonnet and opus (haiku complains about adaptive reasoning).

Tested with different auth (claude-code and hermes-pkce), same result: 429 unregarding of the extra-usage enabled or not.

Max subscription with lots of available quota and 200€ extra usage credits.

Claude cli on the same machine works perfectly.