opencode-lmstudio v1.0.0-rc.2

opencode-lmstudio v1.0.0-rc.2

1.0.0-rc.2 hardens the typed v1 integration around LM Studio's native REST
API, effective loaded context, OpenCode's real npm package resolver, and ACP
stdio. It is published under npm next; stable npm latest remains on
0.3.1.

Track rollout status and report results in
#34.

Install

Pin the exact candidate in opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-lmstudio@1.0.0-rc.2"]
}

The moving prerelease channel is opencode-lmstudio@next. Restart OpenCode
after changing plugin versions.

Compatibility

• OpenCode 1.17.7 or newer
• LM Studio 0.4.0 or newer
• Native GET /api/v1/models on the configured LM Studio server
• OpenAI-compatible /v1/chat/completions for inference
• Node.js 22.22.2, 24.15.0, or a supported version from 26 onward; or
  Bun 1.3.5 or newer

The plugin does not fall back to the previous /api/v0 discovery shape. Stable
0.3.1 remains available separately.

Native model mapping

• Preserves native model key values and uses display_name in OpenCode.
• Includes type: "llm"; excludes loaded and unloaded embedding records.
• Maps capabilities.vision to image input and attachment support.
• Keeps tool calls enabled for LM Studio's native and default tool-use modes,
  and reports each model's mode through structured discovery diagnostics.
• Uses max_context_length for unloaded, on-demand models.
• Uses loaded instance config.context_length when active.
• Uses the minimum context when multiple instances share a model key.
• Regenerates plugin-owned models and whitelists when the server list changes.
• Preserves explicit model overrides and whitelists.

LM Studio does not report a distinct generation limit, so the plugin's output
reserve remains explicit policy: one quarter of effective context, capped at
8,192 tokens. trained_for_tool_use is used as a native-vs-default diagnostic,
not as a reason to disable default tool support. Reasoning settings are not
translated into a non-equivalent OpenCode field. The full rationale is in
docs/v1-contract.md.

Release evidence

• Native v1 schema validation against LM Studio 0.4.16+2 live responses
• Nemotron and GLM vision records without model-name heuristics
• Unloaded, singly loaded, and multiply loaded context regressions
• Authenticated discovery/inference and 401 no-fallback coverage
• Changing-model-list and loaded/unloaded embedding regressions
• Source and packed-package real OpenCode CLI smoke tests
• OpenCode 1.17.7 minimum and 1.17.9 current compatibility targets
• Official ACP SDK 0.21.0 initialize/new-session and JSON-RPC stdout gate
• Linux and macOS ACP gates plus a Windows package/config/chat smoke; OpenCode
  1.17.9 ACP itself currently hangs on the GitHub Windows runner and is not
  claimed as passing
• Direct packed-candidate and public npm spec resolution through OpenCode's
  clean HOME/XDG cache
• Five Microsoft TUI Test views rendered by browser xterm.js
• Dependency audit, coverage thresholds, package inspection, OIDC trusted
  publishing, and npm provenance

Feedback requested

Please report:

• LM Studio and OpenCode versions;
• operating system;
• whether models appear and chat succeeds;
• loaded context values if a model is already active;
• authenticated-server behavior, if used; and
• whether ACP/editor startup remains clean.

Sanitize tokens, private URLs, model paths, configuration, and logs before
posting in #34.

Contributors

Community contributors and the pull requests or issue work incorporated into
the v1 line are recorded in
CONTRIBUTORS.md.