@weldall/opencode

Weldall as an opencode plugin.

Weldall is the access layer in front of company systems. It holds the catalog of what each user is permitted to do, and it performs every API call on their behalf, so no access token reaches the agent.

This package ships one skill. When a request concerns a company system, a business record, or an action on one, the agent reads the live catalog, matches a skill, and follows the document Weldall returns. It does not answer those requests from memory and does not invent an API call.

Requirements

• opencode, tested against 1.18
• The Weldall CLI at 0.14.0 or newer, on the PATH of the shell opencode runs commands in, and signed in with weldall login

The version floor is hard. The skill reads the catalog with weldall skills list --agentic, which arrived in CLI 0.14.0, and it carries no fallback to an older flag. On an older CLI the lookup fails and the agent reports the failure rather than guessing.

Install

opencode plugin @weldall/opencode

Add -g to install into ~/.config/opencode/opencode.json instead of the project. Restart opencode afterwards: config is read once at startup and is not reloaded.

What it does

opencode has no manifest field for a package's skills, so plugin.js puts this package's skill directory on config.skills.paths at startup. That is the whole plugin. It registers no tools, no hooks of its own, and no MCP server.

The catalog differs per user and an administrator can change it at any moment, so the skill reads it on every run rather than remembering it between turns. A skill the user lacks scopes for is reported along with the scopes it needs, never attempted anyway. A catalog that came back incomplete is reported as incomplete, never as proof that a capability is absent.

The procedure is one file in the repository, and every host package is generated from it, so the text does not fork per host.

License

FSL-1.1-ALv2. Source at seibert-external/weldall.