Skip to content

v1.4.2

Choose a tag to compare

@github-actions github-actions released this 16 Aug 16:06
· 103 commits to main since this release
c44bdba

v1.4.2

修复 / Fixed

  • OpenCode Go 的 Qwen 识图模型不再被网关换成 MiniMax M3:用户反馈 opencode-go/qwen3.6-plus 作识图模型时实际调用的是 MiniMax M3——根因是已安装的 pi-ai 目录把 Qwen3.6 Plus 分类为 OpenAI 协议(/v1/chat/completions),而 OpenCode Go 官方只在 Anthropic /v1/messages 上提供该模型(同病相怜的还有 minimax-m2.7),错误端点上的请求被网关落到默认模型。插件新增内置目录纠错(catalogCorrections,默认开启):当解析出的目录条目仍指向错误协议时,该后端改由插件直连正确协议应答(Anthropic Messages + 渠道自身凭据);上游目录修复、或用户把路由指向自己的网关后,纠错自动失效,回到正常 harness 路径。新增 12 个测试覆盖纠错判定、Anthropic 请求构造、整轮链路的端到端接管与自动失效。
  • OpenCode Go Qwen vision models are no longer served as MiniMax M3: users reported that picking opencode-go/qwen3.6-plus as the vision backend actually invoked MiniMax M3 — the installed pi-ai catalog classifies Qwen3.6 Plus as OpenAI protocol (/v1/chat/completions), but OpenCode Go only serves it over the Anthropic /v1/messages endpoint (minimax-m2.7 is misclassified the same way), so the gateway falls back to a default model on the wrong endpoint. The plugin now ships built-in catalog-routing corrections (catalogCorrections, on by default): while the resolved catalog entry still points at the wrong protocol, the backend is answered directly over the corrected one (Anthropic Messages with the channel's own credential); the moment upstream fixes the catalog — or the user points the route at their own gateway — the correction disarms itself and the normal harness path resumes. 12 new tests cover the decision table, the Anthropic request shape, and end-to-end chain take-over and auto-disarm.
  • Oh-DSH Desktop 内置 DSH 0.1.0-rc.5 上的启动崩溃已修复(#124:rc.5 的条目激活顺序是服务驱动的,插件可能在官方 llm-deepseek 行之前 apply;同步执行的 keep-alive 检查把尚未就绪的官方路由误判为「已停用」,抢先注册 deepseek-official 目录项,随后官方行注册时抛出 DUPLICATE_DIRECTORY,把整个运行时带崩(Oh-DSH Desktop 里表现为 DSH runtime exited before readiness)。接管决策现在推迟到启动稳定窗口(2s)之后再做:官方路由已注册则完全放手;确实停用/缺失才接管,stealth 与 keep-alive 路径行为不变。新增回归测试覆盖「迟到的官方行不得被误判」场景,并在真实 rc.5 宿主构建上完成端到端验证。
  • Startup crash on Oh-DSH Desktop's bundled DSH 0.1.0-rc.5 is fixed (#124): on rc.5, entry activation is service-driven, so the plugin could apply before the stock llm-deepseek row; the synchronous keep-alive check then misread the not-yet-applied official route as dead, registered the deepseek-official directory entry first, and the stock row's own registration threw DUPLICATE_DIRECTORY, killing the whole runtime (surfacing as DSH runtime exited before readiness in Oh-DSH Desktop). The takeover decision now runs after a short boot settle window (2s): a registered stock route means hands off; a still-dead route means the row is genuinely absent/disabled and the takeover is safe. Stealth and keep-alive behavior are otherwise unchanged. A new regression test covers the late-stock-row ordering, verified end-to-end against a real rc.5 host build.
  • 「+ 自动识图」twin 组保留选择器里的推理等级(#103:切到 twin 组后宿主会丢掉会话选择器里选好的 reasoningEffort,多步 turn 只有第一步会思考。现在推理等级完全归选择器管:twin 元数据一比一镜像源路由,wrapper 流按「委托 provider + 模型」记忆最近显式的 reasoningEffort 并在后续步骤缺失时补上(选 max 一直 max,选 off 忠实跟随);视觉链保持不消耗推理预算。顺带修复配置层不一致:wrapper 路由、视觉链路由与 agent/request 钩子改由解析后的 settings 文档响应式挂载/卸载,卡片开关即时生效(含 #120 引导完成时机修复)。
  • "+ Auto Vision" twins preserve the picker's reasoning effort (#103): switching to a twin group made the host drop the reasoningEffort chosen in the session picker, so only the first step of a multi-step turn reasoned. The reasoning level now belongs entirely to the picker: twin metadata mirrors the source route one-to-one, and the wrapper stream remembers the latest explicit reasoningEffort per delegated provider+model and re-applies it when later steps omit it (max stays max, off stays off); the vision chain keeps reasoning effort undefined so it never spends the reasoning budget. Also fixes a config-layer inconsistency: wrapper routes, the chain route and the agent/request hook now mount/unmount reactively from the resolved settings document, so card switches take effect immediately (including the #120 walkthrough-finish fix).

文档 / Docs

  • Oh-DSH Desktop 安装说明(#124:README 中英双语新增 Oh-DSH Desktop 章节——DSH_HOME=~/.ohdsh 的 desktop profile 安装命令、内置 rc.5 运行时的版本注意事项、错误安装导致无法启动时的恢复步骤、插件市场流程与内置 @oh-dsh/vision 的共存说明。
  • Oh-DSH Desktop install guide (#124): both READMEs now document the Oh-DSH Desktop path — the DSH_HOME=~/.ohdsh desktop-profile install command, the bundled rc.5 runtime caveat, recovery steps for a broken profile, the marketplace flow, and coexistence with the bundled @oh-dsh/vision.
  • 多插件对比与致谢恢复并逐条核实(#122 / #123:恢复早前 README 重写中删除的六项目社区插件对比表与致谢一节,并按各家 2026-08 的 README 逐条核实措辞(sidecar 默认端点、provider 路由描述、modlens 登录态、toolkit 差异、tool-vision 桥接瀑布等)。
  • Multi-plugin comparison and acknowledgements restored and fact-checked (#122 / #123): the six-project comparison table and the Acknowledgements section lost in an earlier README rewrite are back, with every claim re-verified against each project's README as of 2026-08 (sidecar default endpoint, provider routing description, modlens login reuse, toolkit differences, tool-vision bridge hook, and more).

Commits

  • Merge pull request #126 from ysr666/release/v1.4.2
  • release: prepare v1.4.2
  • Merge pull request #124 from ysr666/fix/keep-alive-takeover-race
  • Merge pull request #125 from ysr666/fix/opencode-go-qwen-protocol
  • docs: correct the new-test count in the changelog
  • fix: answer opencode-go qwen3.6-plus vision calls over the Anthropic endpoint
  • docs: add Oh-DSH Desktop install section (EN/ZH)
  • Merge pull request #123 from ysr666/fix/readme-comparison-wording
  • docs: fact-check comparison table against the six plugins' current READMEs
  • Merge pull request #122 from ysr666/fix/restore-readme-comparison-acks
  • docs: restore multi-plugin comparison and acknowledgements to both READMEs
  • fix: defer the deepseek-official takeover past the boot settle window
  • Merge pull request #121 from ysr666/fix/issue-103-twin-reasoning-effort
  • fix: scope the effort memory per provider+model (#103)
  • fix: reasoning level stays a picker concern — drop the config knob (#103)
  • fix: reconcile legacy routing mounts with the settings document
  • test: strict-upstream audit for fabricated capability hints (#103)
  • fix: ignore unsupported reasoning levels and tighten wrapper UI copy (#103)
  • fix: reasoning default for + Auto Vision twins (issue #103)
  • fix: finish the vision walkthrough on save, model pick, or settings close (#120)
  • docs: complete the v1.4.1 changelog with all changes since v1.4.0 (#119)