mcsc — Multi-CLI Subagent Configurator

mcsc (mcps/mcsc/) is AOS's own cross-harness delegation engine. It is the mandatory path for delegating work to another CLI harness inside an AOS-installed repo — not a Claude-Code-only convenience.

Why this exists

Three separate, harness-specific delegation paths already exist (the Claude Code plugins antigravity:delegate, opencode:opencode-rescue, codex:codex-rescue — see AGENTS.md's "Delegating to an external CLI"). Those are fine for interactive delegation from a Claude Code session, but they are Claude-Code-only, and none of them tell agenttrail's live board what is happening. mcsc's adapters (mcps/mcsc/packages/core/src/adapters/{agy,codex,opencode}.js) wrap each CLI once, work from any harness, and call emitTrail() around every delegated run — SessionStart/SessionEnd, plus live PreToolUse/PostToolUse where the underlying CLI supports structured/streaming output — so a delegated run shows up on the board the moment it starts, without any extra setup.

Skipping mcsc doesn't just lose convenience — it makes the delegated agent invisible to anyone watching the live map, and to anyone in this repo who asks "what are the other agents doing right now."

How to use it

Once installed (mcsc in mcp_config.json, entry point mcps/mcsc/server.js), it registers up to four MCP tools — one per installed CLI, only exposed when that CLI is detected as ready in the local inventory, and never offered back to a caller that's already running as that same CLI (MCSC_CALLER env var, avoids self-delegation loops):

• delegate_agy — { prompt, model? } → Antigravity / Gemini. READ-ONLY until tested: use it for research, review and analysis, never for edits. The adapter can pass --mode accept-edits (write: true), but that path is untested and not part of the contract.
• delegate_opencode — { prompt, model?, variant? } → OpenCode
• delegate_codex — { prompt, model? } → Codex
• delegate_smart — { task_type, prompt } → reads rulebook.yaml (or the bundled default) and picks the best CLI for that task type automatically

Call the tool that matches the harness you want, with a complete, self-contained prompt — the adapter underneath handles the CLI's actual flags, streams tool-call telemetry to agenttrail as it runs, and returns the final output. variant (OpenCode only) maps to --variant (reasoning effort).

Depth limit: one level, like a fork

Every adapter sets MCSC_DEPTH to the caller's depth plus one in the child's environment. An mcsc server that starts with MCSC_DEPTH >= 1 lists no tools and refuses every call. A delegated agent therefore cannot delegate again, so agy -> codex -> agy chains are impossible. MCSC_CALLER alone only excluded the caller's own CLI and was overwritten by each adapter.

What it is not

• Not a replacement for the Claude Code delegation plugins for a quick, interactive, single-turn ask inside a Claude Code session — those remain fine for that case.
• Not itself a review/QA step — verify a delegated result the same way AGENTS.md already says to: read the actual diff, never trust the returned status field alone.