uefn-mcp-lab-guard git:20261016.5b4e0c6A

Immutable. This exact content never changes and is served at /api/v1/blob/ce094c549365de9c.

---
name: uefn-mcp-lab-guard
description: UEFN MCP lab guard: path forms (GUID vs LAB_ROOT), shell, what the guard refuses. Use before any MCP tool that WRITES, or any script that sets LAB_ROOT/LAB_AGGRESSIVE.
last_reviewed: 2026-10-16
sources: [original]
---
# uefn-mcp-lab-guard

## Rules
1. Asset paths and tool arguments: ALWAYS GUID form `/<guid>/<project>/...` (here `/09e4e8d3-83f9-4ee0-9777-072ed3984380/root_ai_lab`). The editor refuses `/root_ai_lab` as an asset path (`outside this project's valid asset roots`).
2. `LAB_ROOT` (the guard's environment variable): the form depends on the guard. Read the header of `Claude/hooks/mcp_lab.py` in the project and use the documented form (after T-014: `/root_ai_lab`; wrong form = every write refused with `path outside /<guid>/root_ai_lab`). Do not confuse it with the argument form.
3. Shell: use PowerShell for anything that touches `LAB_ROOT` or paths starting with `/`. Git Bash rewrites `/09e4...` in `C:/Program Files/Git/...` and the lab does not start; if unavoidable `MSYS_NO_PATHCONV=1`.
4. Activation: `$env:LAB_AGGRESSIVE="1"; $env:LAB_ROOT=...` set by `Content`.
5. Ledger `Claude/logs/explore-ledger.jsonl`: every call has `r` = ok / rejected / refused / no_answer. Cite a recipe only from an `ok` line with the exact arguments.
6. The guard refuses: strings that start with `/` outside the lab and outside the accepted source roots (`/Script`, `/Engine`, `/Niagara`, `/Game`, `/FortniteGame`) per create/duplicate/place/add; drive paths in writes; editor quit/restart. It is blind inside `execute_tool_script` (it only sees the outer call).

## Don't
- Bypassing a refusal: report the exact message and stop.
- Using paths from another project (`root_pillars`), closing or killing a running editor (relaunch only through `uefn_session.py recover`, opt-in), writing outside the lab, exploiting B-002 (`values` as a JSON string).
- Improvising when a recipe fails: exact error and stop.

## Recipes (recipes.json, status tool-ok or proven)
- `AssetTools/find_assets` with `folder_path` and `asset_type` (filtered read in the lab).
- `AssetTools/list_folders` with a non-recursive `root_path`.
- `GetLogEntries#filter_noise`: pattern `^(?!.*LogModelContextProtocol)`.

## To try (not verified)
- Duplicates from other projects with `extra-root` (e.g. `/mech-drone`): unverified, requires `extra-root` in LAB-PROJECT.txt.

## Known editor traps (UEFN 42.20)

- Never call `get_lod_count`, `get_number_verts`, `get_num_triangles`, `get_num_sections` or `agg_geom.export_text()` on a static mesh: they crash the editor.
- Creating an asset with an existing name opens a modal "Overwrite Existing Object" that blocks the MCP: use a new name or check first.
- A tool tagged `[experimental]` is ask-the-owner-first. A call that hangs while UEFN is in the background: report it, do not loop.
- With Revision Control disconnected `StartSession` fails (Validation failed): reconnect it first.