Skip to content

Rayline 0.4.2

Choose a tag to compare

@chilang chilang released this 07 Nov 21:38
· 100 commits to main since this release
cd11a0b

Rayline 0.4.2 (build 0.4.2+cd11a0bdc1e2, source cd11a0b).

Install:

curl -fsSL https://get.rayline.ai/install.sh | bash
brew install rayline-ai/tap/raydock

Windows (PowerShell): irm https://get.rayline.ai/install.ps1 | iex

These are the same signed files served from https://get.rayline.ai/cli/v0.4.2+cd11a0bdc1e2/. SHA256SUMS is signed with minisign (SHA256SUMS.minisig); its Homebrew tarball entries keep the homebrew/ path prefix of that layout. Rayline.dmg and Rayline-macosx_11_0_arm64.zip are Developer ID signed and notarized.

Release notes

New

  • Hybrid mode runs on Windows. raydock hybrid start, status, stop and run no longer refuse Windows. A second hybrid start waits on a real start lock, the hybrid state directory and its token, ready, supervisor, config and log files are readable by the current user only, and rld runs detached, so closing the console that started it no longer stops it. hybrid stop shuts rld and the execution core down gracefully and removes their token and ready files; the timeout and hard-kill fallback are unchanged. hybrid run codex and hybrid run pi find npm's .cmd launchers. Windows CI now runs the hybrid lifecycle tests. macOS and Linux behaviour is unchanged.
  • raydock router start runs rld detached on Windows, as hybrid does. The router and proxy that router start, raydock claude --local and the other harness launches start no longer hold the caller's standard handles or live in its console, so raydock router start | cat and $(raydock router start) return as soon as start finishes, and closing the console that ran router start no longer stops rld. macOS and Linux behaviour is unchanged.

Fixed

  • raydock hybrid run claude works on cloud routes. The execution core (rayline-core) is pinned to a newer source commit whose model_pin Messages transport forwards the system rows Claude Code 2.x sends inside messages and no longer refuses a request body that merely contains the string "xhigh". Before, every Claude Code 2.x turn on a cloud route failed with 400 model_pin Messages cannot preserve in-array system or developer messages. The root effort fields are still stripped before dispatch, and in-array developer rows are still refused (see Known issues).
  • Hybrid tells Claude Code the gateway's context window. raydock hybrid run claude now sets CLAUDE_CODE_MAX_CONTEXT_TOKENS to the window the gateway advertises, so Claude Code no longer assumes 200k for rayline-auto, auto-compacts early or warns about an unrecognized model on every launch. Other harnesses already got their window through their config files.
  • The local deployment lowers an over-cap output allowance. A harness that asks for more output than the local deployment's cap (a quarter of the local window, 24,576 tokens at the default 96k), such as Hermes with its fixed 128,000, has its local turns lowered to the cap instead of refused with output allowance exceeds the admitted deployment limit. Cloud deployments still refuse an allowance over their ceiling.
  • Streamed Responses turns on an Anthropic route record their cache-read tokens. When a Codex /v1/responses turn is served through an anthropic_messages endpoint and both sides stream, the local router now closes the metrics row with the cached part of the prompt as well as the prompt and output counts, so raydock top's cache column covers the shape an interactive turn normally takes. The row also gains the prompt total the cache hit rate is measured against, so that column reads as a rate rather than a blank. Those counts were dropped for stream-to-stream turns only; the buffered shapes and the Chat Completions and Messages paths already recorded them. An upstream that reports no cache tokens still leaves the count unset rather than reporting a zero. Response bodies and the usage the client receives are unchanged.
  • Python 3.13+ clients trust the proxy's certificates. Proxy leaf certificates now carry an Authority Key Identifier, filled from the CA's own Subject Key Identifier, so it matches CAs already on disk. Python 3.13+ rejects a leaf without one (CERTIFICATE_VERIFY_FAILED: Missing Authority Key Identifier), so any Python client that trusted the proxy CA failed its TLS handshake.

Changed

  • Responses cost details use the upstream spelling. The execution core now names the provider charge's cost_details halves upstream_inference_input_cost and upstream_inference_output_cost on Responses (Chat Completions keeps the prompt/completions spelling). Raydock reads either spelling.

Docs

  • The routing-modes examples follow the rayline-router removal: configs pin current model ids, the retired Rcl modes are documented as behaving like Rc, and a Hybrid mode section links the setup guide.
  • The SDK examples run as shipped: local/python builds its client with the SDK's own HTTP client, the local examples ask a question a raw API call can answer, and the examples allow 2048 output tokens so reasoning models print an answer.
  • The hybrid setup guide states that Qwen3.8-27B needs a machine with at least 32 GB of memory, and that a runtime which reports no model name gets a warning and is assumed to be Qwen3.8-27B.

Known issues

  • Codex through hybrid still fails on cloud model_pin Messages routes. The execution core refuses the in-array developer messages Codex sends on that transport, so raydock hybrid run codex turns that land on such a route fail. Local turns and Claude Code are not affected.

What's Changed

  • fix(local-router): streamed Responses turns record their cache-read tokens by @dorota-memex in #161
  • fix(cli): hybrid tells Claude Code the gateway's context window by @dorota-memex in #162
  • fix(hybrid): the local deployment lowers an over-cap output allowance by @dorota-memex in #164
  • examples: update routing modes after the rayline-router removal by @madamowski in #166
  • docs(hybrid): state the memory Qwen3.8-27B needs and what the model check does by @madamowski in #167
  • hybrid: support Windows by @madamowski in #168
  • release: pin rayline-core to 20f2a965 (model_pin Messages accepts Claude Code 2.x) by @madamowski in #169
  • Make the SDK examples run as shipped; add an Authority Key Identifier to proxy leaf certs by @madamowski in #170
  • router: start rld detached on Windows, as hybrid does by @madamowski in #171
  • chore(release): bump Rayline to 0.4.2 by @madamowski in #172

Full Changelog: v0.4.1...v0.4.2