You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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.
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.
"+ 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).
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.
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)