mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-09-19 01:09:55 +08:00
Trevin Chow
1681de57e7
Installing the compound-engineering plugin to OpenCode produced zero slash commands: convertClaudeToOpenCode only consumed plugin.commands (a commands/ directory), but the plugin ships skills only. OpenCode users had no /ce-work, /ce-plan, etc. even though every skill installed correctly. The post-#967 skills-only, root-native plugin makes this total — the plugin now ships no commands at all. Add convertSkillsToCommands(): one slash-command stub per skill, carrying the skill's description (and argument-hint when present) as frontmatter and a body that loads/executes the named skill with $ARGUMENTS passed through. Skills with disable-model-invocation or user-invocable: false are excluded. Stubs are deduped against explicit commands/ entries and each other by normalized path key (colons -> slashes) so foo:bar and a skill named foo/bar don't both write commands/foo/bar.md; explicit commands win. applyPermissions now force-allows the skill tool (clearing any patterned skill(...) rule) whenever stubs are emitted, so a skills-only plugin doesn't get permission.skill = deny. Cleanup on skill rename/delete rides the existing managed-install manifest: stubs appear in bundle.commandFiles, so currentCommands tracks them and stale stubs are swept on reinstall. Parser/types gain user-invocable so a skill can opt out of being exposed as a slash command. Drift test derives the expected stub set independently from source frontmatter. Reapplies PR #776 onto post-#967 root-native paths and current types; that PR was anchored to the old plugins/compound-engineering/skills/ tree and its hand-built test literals predated the root/skillPath fields. Co-authored-by: Luca Battistini <lucabattistini@users.noreply.github.com>