@luxass/opencode-personal

Collects OpenCode V2 plugins that are tuned to a single personal workflow rather than general enough to ship on their own.

Install

After publication, add the package to the existing plugins array in ~/.config/opencode/opencode.jsonc:

{
  "plugins": ["@luxass/opencode-personal"]
}

To review the checkout before publication, use its absolute directory path instead of the npm package name. Restart OpenCode after changes.

What it does

Crow connects OpenCode to CLIProxyAPI (CPA). It discovers models and their metadata from CPA and uses OpenCode's model catalog for price estimates.

The Git interceptor sets GIT_EDITOR=true, GIT_SEQUENCE_EDITOR=true, and GIT_MERGE_AUTOEDIT=no for shell invocations. It denies shell permission requests that contain --no-verify.

The Git interceptor preserves your commit signing settings for commits and rebases. It blocks common signing bypasses, including --no-gpg-sign, false values for commit.gpgsign, removal of that setting, and indexed Git configuration environment overrides. It also tells the model to stop and ask for help if signing fails. It does not enable signing where you disabled it. These shell checks are safeguards for common commands, not a complete security boundary.

The dependency policy tells the model to reuse existing code, installed dependencies, and platform APIs first. It requires explicit user approval before adding, replacing, or upgrading dependencies. Routine installs of existing dependencies remain allowed through Socket Firewall. This is a model instruction, not an enforced approval gate.

The package-manager interceptor requires Socket Firewall for common install and execution commands from pnpm, npm, Bun, Yarn, and npx. Retry an unwrapped command with sfw before the package manager, such as sfw pnpm install. The interceptor denies the command without rewriting it.

The worker configuration guard denies direct edits to worker-configuration.d.ts and common shell mutations of that file. Run wrangler types to regenerate it. These checks are safeguards for common operations, not a complete security boundary.

These defaults suit one workflow and are meant to be edited in place rather than configured through settings. Each feature is a native OpenCode Plugin.define object in its own source file. It registers its own hooks in setup and returns its cleanup function. There is no custom feature-definition API.

The package entry point keeps its plugin list and lifecycle handling inline. It loads Crow, the dependency policy, the Git interceptor, the package-manager interceptor, and the worker configuration guard.

To add a feature, define an OpenCode plugin with Plugin.define, import it, and add it to the list. Remove an entry to stop loading it. Plugins start in list order with the package's OpenCode context.

The entry point runs plugin cleanup functions in reverse order on unload or failed setup. It attempts every plugin cleanup even if one fails. The package exports ./crow, ./dependency-policy, ./git-interceptor, ./package-manager-interceptor, and ./worker-configuration-guard for independent use.

The Git and worker guards come from dots. The package-manager guard is a port of the Pi personal extension. Before enabling this package, move git-interceptor.ts and worker-configuration-guard.ts out of OpenCode's automatically loaded plugins/ directory. Keep backup copies until the review is complete. Remove any explicit registrations of those files too, to prevent duplicate loading.

Keep the Herdr integration separate. This package does not replace it.

Crow

Run opencode auth login crow in a project that loads this package. Select the API-key method. Enter the Crow base URL and API key when OpenCode asks. The default URL is http://127.0.0.1:8317. OpenCode saves the credential in its own storage. Crow does not read or copy Pi credentials.

You can also set CLIPROXYAPI_API_KEY before you start OpenCode. Set CLIPROXYAPI_BASE_URL to use another address. OpenCode chooses the active connection. The base URL environment variable overrides the saved URL.

Crow requests model metadata at /v1/models?client_version=opencode and sends inference requests to /v1/responses. Base URLs can include a path prefix or end with /v1.

Limits, input types, and reasoning levels come from CPA. Models without an ID or marked hidden are omitted. Models remain available without an OpenCode catalog match. Matching catalog entries supply price estimates. Models without a match have an empty cost list. These estimates do not describe CPA's actual charges.

Missing limits default to 128,000 context tokens and 16,384 output tokens. Missing input types default to text. Reasoning variants use CPA's effort names, with none displayed as off.

Crow refreshes models at startup and after credential changes. Run /cliproxyapi-refresh to request another refresh. Each refresh clears the previous model list. If discovery fails, the list stays empty until a successful refresh. Crow does not persist a model cache.

Crow uses OpenCode's native HTTP Responses streaming, retry handling, and local compaction. It adds no notifications, feature settings, or TUI component.

License

MIT.