Skip to content

v0.3.0

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 24 Aug 11:44
· 6 commits to main since this release
v0.3.0
721c0f2

What's new in v0.3.0

Added

  • Vision policy for text-only backends (text_only = true on oauth/api_key backends + [vision] section) — text-only upstreams (vLLM DSV4-Flash etc.) hard-400 on image content, and clients re-send the full transcript, so one screenshot poisons every later request in the session. Spock now rewrites image content before the request leaves: mode = "strip" (default) replaces images with an omission note; mode = "describe" captions each screenshot via an OpenAI-compatible VL sidecar (llama-server + mmproj) and inlines the caption as text. Any sidecar failure degrades to strip — a request never dies here. Covers /v1/messages (incl. anthropic passthrough and KV sessions) and the verbatim /v1/chat/completions ingress. Captions cached in memory only (sha256 of image + prompt); no per-request image limit — one failed sidecar call strips the rest of that request (bounded stall), a healthy sidecar captions every image. Retroactively un-sticks poisoned sessions: the next request goes out clean. glm-5.3 name-matcher flatten unchanged as a safety net.
  • POST /v1/responses search shim — grok-build web_search posts OpenAI Responses ({base}/responses + hosted web_search tool). Search-only: run [web_search] (Brave/Serper/SearXNG/DDG) and return a completed Responses object with output_text + url_citation. No tool / search disabled / empty query → 400. Not a general Responses proxy (no chat/completions fallback).
  • macOS Settings Catalog effort column — auto / 1 / 0 next to context; Save & Apply writes supports_reasoning_effort so Grok Build /model Tab shows effort on non-heuristic rows (Qwen, local GGUF)
  • llama-server KV sessions (kv_sessions = true on an api_key backend) — Claude Code traffic parks a named master via native /completion, children POST /fork with parent_session_id, leave via POST /close_session. /v1/chat/completions is never the session path. Missing routes or unknown session error the request (no silent cold prefill). Inherit proof: HTTP cache_n / prompt_n. Headers: x-spock-session, x-spock-parent-session, x-spock-close-session. cache_control on the shared prefix names the master.
  • Generalized OAuth providers — registry-driven spock login|logout <provider> / spock providers (xai, kimi, qwen); status + menus follow the table
  • Backend types: oauth / api_key / anthropic (Settings labels: OAuth / API Key / Anthropic)
  • Kimi Code OAuth (provider = "kimi", https://api.kimi.com/coding/v1, KimiCLI UA + X-Msh headers)
  • Qwen OAuth (optional) — provider = "qwen" chat.qwen.ai device flow + PKCE S256; uses token resource_url when present (qwen-code path)
  • Qwen Cloud (qwencloud.com) documented as type = "api_key" on dashscope-intl…/compatible-mode/v1 + DASHSCOPE_API_KEY (Token Plan / Max Preview)
  • Token files: ~/.config/spock/oauth-<provider>.json (imports legacy grok-test / kimi-cli / qwen-code paths once; logout clears them)

Changed

  • Breaking: bare spock login removed — use spock login xai (or kimi / qwen)
  • Proxy never opens a browser mid-request; refresh is single-flight per provider
  • Long streaming generations: idle read/write timeouts (1h) instead of total 600s request caps; stream clients skip server-tools buffer path; backend map lock dropped before upstream I/O
  • macOS app: Local Network + Bonjour keys so LAN backends work from Spock.app
  • Product framing: multi-backend Anthropic Messages proxy (not Grok-only)

Fixed

  • Mid-conversation role:"system" reminders 400 on template-strict upstreams: Claude Code delivers some system-reminders (e.g. the Agent-tool types list) as role:"system" messages inside messages. Qwen3.5+ jinja chat templates (SGLang / vLLM / llama.cpp) hard-400 with System message must be at the beginning. — LAN Qwen looked dead (GPU idle) while every request died at template validation. Generic (OpenAI-compat) backends and the llama-server KV path now fold non-leading system messages into user turns wrapped in <system-reminder> tags (the client's own idiom); xAI / Kimi keep passthrough.
  • z.ai GLM 5.3 text-only content: Claude Code screenshots become OpenAI image_url parts; GLM-5.3 400s messages.content.type is invalid, allowed values: ['text']. Flatten images to a text note for glm-5.3* only — vision backends unchanged.
  • Accepted sockets inherit O_NONBLOCK on Darwin: listen socket is nonblocking so the accept loop can poll shutdown. macOS/accept() copies that flag onto the client fd; read_exact then returns WouldBlock (os error 35) as soon as the kernel buffer is empty. Claude Code reports ECONNRESET; Grok Build reports reqwest error stream: error sending request. Force blocking + idle timeouts on every accepted socket. /v1/chat/completions now logs route and emits a stream error event instead of dying silent.
  • Catalog /v1/models: non-empty catalog is served from local entries only. Live backend /models probes no longer block Grok's ~5s catalog fetch (empty /model picker, unknown footer).
  • Advisor protocol: run the nested review; never leak advisor as client tool_use (No such tool available: advisor). Flatten leftover server_tool_use / advisor_tool_result to text on the OpenAI-compat path (xAI 400). VSCodium 2.1.226 webview cannot render those block types (Unsupported content type: server_tool_use as a chat line) — emit labeled text instead. Strip <|eos|>. Log advisor start/end/duration.
  • WebSearch on stream: Claude Code nested web_search_20250305 calls now run Spock server-tool emulation (was skipped when stream:true); emit real server_tool_use + web_search_tool_result blocks; SSE keepalives during long rounds
  • Login… from the menu bar runs as a direct child process (no Terminal/AppleScript stall)
  • Menu + CLI surface already-logged-in OAuth state clearly
  • Slow LAN (llama-server) streams no longer die at ~10 minutes while the backend still generates
  • Mid-SSE llama.cpp tool-call diff aborts (Invalid diff: now finding less tool calls) labeled as upstream llama-server, not Spock
  • Streamed OpenAI tool_calls merge by index (Qwen/Token Plan): empty id arg chunks no longer open extra tool_use blocks — fixes Claude Code Bash getting empty/broken commands on Fable→Qwen
  • Route backend:model client ids to that backend when it exists (stop falling through to profile default / xAI)
  • Docs: Qwen Cloud Coding Plan (coding-intl…, sk-sp-…) vs Token Plan (token-plan.ap-southeast-1…, separate key) — qwen3.8-max-preview is Token Plan only

Install (macOS App — Apple Silicon)

  1. Download Spock-0.3.0-darwin-arm64.zip
  2. Unzip → drag Spock.app to Applications
  3. First open: right-click → Open if Gatekeeper blocks (ad-hoc signed)
  4. Menu bar → Login xAI… or paste an API key in Settings

Claude Code: set ANTHROPIC_BASE_URL=http://127.0.0.1:8048

CLI

Platform Asset
macOS arm64 spock-0.3.0-darwin-arm64.tar.gz
macOS x64 spock-0.3.0-darwin-x64.tar.gz
Linux x64 spock-0.3.0-linux-x64.tar.gz
Linux arm64 spock-0.3.0-linux-arm64.tar.gz
Windows x64 spock-0.3.0-windows-x64.zip 
tar -xzf spock-0.3.0-darwin-arm64.tar.gz
sudo mv spock-0.3.0-darwin-arm64/spock /usr/local/bin/
spock serve

Config: ~/.config/spock/config.toml
OAuth: ~/.config/grok-test/auth.json · or API key / XAI_TOKEN

SHA-256: see checksums.txt.

Full Changelog: v0.2.0...v0.3.0

Assets 15
Loading