feat: add transparent quota-aware Claude subscription routing

[ejc3]

Summary

Adds claude-master, a launcher that keeps the installed native Claude Code client and its normal master login while routing only /v1/messages and /v1/messages/count_tokens through separately authorized Claude subscription profiles.

• Uses the current claude executable on PATH; there is no version pin, model pin, or local model allowlist.
• Creates every inference profile through the normal Anthropic OAuth browser login. It never copies or replaces the native master credential.
• Preserves the native model, payload, tools, active context, conversation/session identity, betas, and end-to-end headers. Only authorization and selected account/device identity change, followed by required CCH re-signing.
• Preserves Anthropic status, body representation, duplicate/unknown response headers, compression, and incremental streaming bytes. SSE is not parsed or reconstructed, and a request is never replayed after response bytes begin.
• Reads each profile weekly usage/reset state and selects the usable quota that resets soonest. The configured order is the deterministic fallback when usage is unavailable or tied.
• Reserves the final 10% of an account for bound opaque continuations when another account has capacity.
• Decides switchability from the current request payload, not a compaction/version header. Self-contained text and client tool history can move; encrypted/signed thinking, containers/files, unfinished server tools, signed compaction, and unclassifiable payloads stay bound.
• Gives self-contained subagents independent bindings while opaque children inherit the parent account.
• Does not rotate subscriptions for request-, model-, authentication-, validation-, or transport-scoped failures, and never falls back to the master account for inference.

Login and run

go build -o claude-master ./cmd/claude-master

./claude-master login claude-primary
./claude-master login claude-secondary

./claude-master run claude-primary \
  --next-profile claude-secondary \
  -- --remote-control

Native Claude options go after --, for example:

./claude-master run claude-primary \
  --next-profile claude-secondary \
  -- --model claude-sonnet-5-5 --fallback-model haiku --remote-control

The repository helper is:

/home/ubuntu/CLIProxyAPI-claude-master/try-claude-master.sh --remote-control

The checked-in scripts/claude-master-gym/ harness drives two disjoint tmux sessions on claude-sonnet-5-5, applies /effort ultracode interactively, and asks each lane to fan out to two bounded read-only subagents.

Verification

• go test ./...
• PR race matrix for Claude Master, auth, native header/error, and executor boundaries
• native Claude Code 2.1.285 CONNECT/TLS/header/SSE smoke under the race detector
• exact raw gzip, duplicate/unknown header, direct-error, split-stream, quota, reserve, affinity, and subagent-binding tests
• bash -n and shellcheck for the launcher/gym scripts
• go build -o test-output ./cmd/server
• go build -o claude-master-test-output ./cmd/claude-master

Full design and operating instructions are in docs/claude-master.md.

[github-actions]

This pull request targeted main.

The base branch has been automatically changed to dev.