Agent contracts

The vendor documentation each adapter follows. Formats change; recheck these pages when an agent releases a new configuration format (see maintenance).

Skills

• Agent Skills specification: a Skill is a folder, not only its Markdown file. References, assets and scripts are preserved. Metadata is discovered before the body is loaded, so a Library outside discovery folders keeps unrelated Skills out of every session.
• Claude Code skills: project Skill folders and plugin Skills have separate provenance. Claude-specific features such as manual-only invocation are kept where documented; conversion refuses a manual-only Skill when the target cannot express the same policy.
• OpenCode skills: its own folders and the shared .agents/skills folder.
• Kiro skills and configuration scopes: Skills and steering are separate capabilities.
• Antigravity discovery: CLI and IDE versions differ in workspace support.
• Grok skills: .grok/skills, AGENTS.md and project .grok/config.toml.
• Skills CLI: Rigshelf records bunx skills add recipes for compatible repositories. Listing and installing are separate; installing needs network access and is always explicit.

MCP

• Codex MCP distinguishes env_vars, env_http_headers and bearer_token_env_var; a literal ${VAR} string is not equivalent. Remote executor selection and helper commands stay native and are reported rather than dropped.
• OpenCode MCP uses local/remote records, command arrays and an environment field, with {env:VAR} variables. Claude MCP JSON cannot be copied as is.
• Kiro MCP uses mcpServers; conversion never broadens permission grants.
• Antigravity MCP documents cwd and tool allow/deny lists in workspace .agents/mcp_config.json, but not ${VAR} interpolation. Definitions that need it use the Rigshelf bridge instead of a guessed schema.
• Grok MCP.

The optional bridge forwards stdio, HTTP and SSE servers, resolves declared environment references and enforces tool allow/deny lists at discovery and call time. Native formats come first. Sampling, elicitation and experimental MCP features are outside the bridge; keep a native integration when you need them.

Plugins and listings

• Marketplace install commands are provider-specific Plugin recipes. A Plugin is never reduced to a plain Skill, because that would lose its hooks, MCP servers and settings.
• A skills.sh listing means a directory entry exists. It is not publisher verification or a security review.
• Source evidence keeps reachability, publisher ownership, listing availability, static findings, content digest and review date as separate fields. Rigshelf never shows a generic “verified safe” badge.

Packaging

• Bun standalone executables: embedded assets, explicit runtime paths, cross-target builds and disabled automatic configuration loading.
• Bun install backends: release installs use the copyfile backend because notice collection rejects linked files.
• Workflow actions are pinned to exact commits; checkout does not persist credentials.