feat(cc-bridge): native Anthropic Messages adapter for Claude Code

[chaoshou-coder]

PR description — feat/anthropic-messages-adapter → baseRT

What / Why

Adds cc-bridge: a native Anthropic Messages API adapter that lets Claude
Code talk to a local basert serve directly (/v1/messages ↔
/v1/chat/completions), with native SSE streaming (thinking / text / tool_use
blocks), keep-alive pings, and cancellation propagation.

The protocol is generated natively instead of via LiteLLM, because LiteLLM was
measured to fail in two compounding ways (see
docs/proposals/native-anthropic-adapter.md and TECHNICAL.md §1):

1. LiteLLM's openai→anthropic non-streaming conversion always returns empty
   content (use_chat_completions_api: true drops content when provider returns reasoning_content BerriAI/litellm#27492, open).
2. Claude Code drops a streaming connection after ~36s without a first byte
   and retries non-streaming — straight into bug 1. baseRT's ~35s prefill of a
   23.5k-token prompt triggers this on every heavy request.

Security posture

• --master-key is required: every inbound /v1/messages and
  count_tokens request must present it (x-api-key or Authorization: Bearer), otherwise rejected with an Anthropic-style 401 before the body
  is parsed.
• Binds to 127.0.0.1 by default; 0.0.0.0 requires an explicit --host.
• The harm-classifier short-circuit (--allow-classifier-bypass) is off by
  default and only matches a narrow heuristic (single user message, no
  tools, non-streaming, non-local model name). Disabling Claude Code's
  online harm detection is an operator decision, documented in TECHNICAL.md
  §3.7.
• Debug dumping (--debug-dir) is off by default.
• Token accounting is honest: /v1/messages/count_tokens and streaming
  message_start usage report an upper-bound estimate, never a fabricated 0.

What is cc-bridge?

A standalone deployment component, explicitly part of the baseRT ecosystem
(own directory, own venv, own tests). The engine-side gaps it works around are
tracked as upstream issues, not papered over in this PR:

• basert serve: support loading an external chat template (--chat-template) #42 --chat-template (external chat templates) — removes the
  system-merging workaround
• basert serve: allow disabling or limiting thinking per request #43 per-request thinking control — removes --max-tokens-override
• basert serve: concurrent-client handling (continuous batching) #44 concurrent-client behavior — removes the timeout-based survival logic

Why issues and PR are opened together

The gaps this adapter works around (external chat templates, per-request
thinking control, single-slot queueing) are product requirements that belong
in the engine, not workarounds frozen into a bridge. Opening issues #42–#44
side-by-side with this PR is deliberate: the adapter is an honest bridge
today, and each issue tracks the engine change that would let the
corresponding workaround be removed. I would rather file a requirement and
ship the piece I can build than file a requirement alone — and I am happy to
do the engine-side work too, if maintainers point me at the right files,
knowing it will take me longer than it would take you.

Tests

cc-bridge/tests/ — 38 pytest cases, all green locally:

• request conversion golden cases (system merging, thinking, tools,
  tool_result ordering, sampling passthrough)
• authentication (pure function + endpoint-level 401s, auth-before-parse)
• token estimation (non-negative, monotonic, CJK weighting)
• SSE lifecycle end-to-end through a fake upstream (normal stream, empty
  stream via [DONE] and abrupt close, truncation → max_tokens,
  classifier-bypass gating)

Run: cd cc-bridge && python3 -m venv .venv && .venv/bin/pip install -r requirements-dev.txt && .venv/bin/python -m pytest

Note: baseRT's CI is C++/Rust only; we deliberately did not add a Python job
in this PR. If the maintainers want it covered in CI, we'll add a job in a
follow-up.

Intentional deviations from the review

• cc-bridge/misc/chat-template-fix.md and the vendored froggeric v21.3
  template are kept as reference material for issue chore(docs): repoint repo references prabod → basecompute #1 (with a full MIT
  license header), rather than deleted.
• No Python CI job yet (see above); tests are local-only.

A note from the contributor

This is my first open-source contribution, and I want to be upfront about
what kind of contributor I am. I do not know C or Rust, and my Python is
basic. What I can reliably offer is careful work on the parts that do not
need deep engine knowledge: integration glue like this adapter, tests,
triage, and documentation — I plan to follow up with a Chinese translation of
the website and repository docs, and more issues and PRs after this one. I
have the time to keep contributing, and I would rather earn my place here as
someone who ships working, reviewed code than as someone who only files
requests. Any process or quality feedback is genuinely welcome — I expect to
make beginner mistakes, and I will learn from them.

Thank you for building baseRT for Apple Silicon users. That is why I am here.

Real-session evidence

A full multi-turn Claude Code tool-use session ran against this branch on
2026-08-12 (01:21–01:34, 29 requests, ~13 minutes): WebSearch-style tools,
tool_calls/tool_result loops with context growing from 22.4k to 26.0k
prompt tokens, streaming and non-streaming paths, and one client disconnect
at 01:34:19 that baseRT logged as [disconnect] client closed connection; aborted stream after 1 token(s) — the adapter's abort propagation working
end-to-end.

The complete request/response records (raw /v1/messages bodies and exact
SSE streams, plus the matching basert serve log) are packaged as
cc-bridge-tool-run-logs.tar.gz (560 KB) — I will attach it in a comment to
this PR. Representative basert log lines:

[01:22:56] <-- 200 Qwen-AgentWorld-35B-A3B prompt=22520 completion=1460 prefill=716 t/s decode=76 t/s 50751ms tool_calls
[01:25:33] <-- 200 Qwen-AgentWorld-35B-A3B prompt=24346 completion=209 prefill=662 t/s decode=75 t/s 39566ms tool_calls
[01:31:19] <-- 200 Qwen-AgentWorld-35B-A3B prompt=25950 completion=1098 prefill=602 t/s decode=70 t/s 58904ms stop
[01:34:19] --> POST /v1/chat/completions model=Qwen-AgentWorld-35B-A3B stream max_tokens=32000
[disconnect] client closed connection; aborted stream after 1 token(s) (model=Qwen-AgentWorld-35B-A3B)

Changelog

CHANGELOG.md gets an [Unreleased] entry; the root README gains a cc-bridge
section.