@synthryn/sypi-discovery

Optional external harness discovery for SyPi. It scans other agent harnesses' config directories: claude, codex, gemini, opencode, cursor, windsurf, cline, copilot (github), vscode, and a root mcp.json. It imports their skills, MCP servers, markdown slash commands, and exec-format tools. SyPi-native configuration always wins on name collisions, and project trust gates the whole import: an untrusted or unknown project stays inactive unless you opt in with allow untrusted.

What it does and does not load:

• Skills: Each discovered skill root directory is fed to the core skills loader after the SyPi roots. SyPi names always win.
• MCP servers: Deduplicated external servers are fed to sypi-mcp. A server the active profile already defines stays profile-owned.
• Commands: Discovered Markdown slash commands, including cline rules and copilot instructions) become /name commands that expand $1..$N, $ARGUMENTS, and ${N:-default} and run the markdown as a prompt. A name a builtin or another extension already claims is skipped and reported in /discovery.
• Tools: Exec-format scripts (.sh, .bash, .py, .js with shebang) become tools that shell out to the script. TS/JS module tools are counted and reported but not loaded in v1.
• Hooks and extensions: Detected and counted in /discovery only. Foreign hooks never execute. Foreign extension packages never load.

One /discovery command reports status: active or inactive with its reason, per-source and per-kind counts, issues, skipped collisions, and detected but unexecuted hooks and extensions. It provides quick toggles (enable, disable, allow, deny, rescan). A Settings row under integrations (category integrations → External harness discovery) toggles the master switch, the allow-untrusted bypass, the four kind rows, and all ten source rows (both / project / user / off).

Configuration lives in the profile discovery block. Absent blocks default to enabled with every source and kind on, while untrusted projects remain blocked. When discovery is disabled, no scanning occurs. The command and Settings rows remain live so discovery can be re-enabled.

It needs no account, API key, token, password, network, or external binary; installing it requires npm, disclosed in catalog consent.