Setting Up the BGS Modding Environment

This is the first-run bootstrap. Use it when:

• The user installed bgs-modding-superpowers and this is the first conversation

in their project directory.

• A new modpack project is being set up from scratch.
• MO2 path is not yet known to the agent.
• The user explicitly asks to "set up", "install", "bootstrap", "configure", or

"initialize" the BGS modding environment.

• A subsequent skill (e.g., xedit-automation) finds the environment incomplete

and routes back here.

Hard guardrails

• Do NOT install MO2 yourself by default. The default path is to detect MO2

and, if absent, guide the human user to install it themselves. Agent-driven install requires explicit user consent ("yes, install MO2 for me").

• Never write into the user's vanilla game install (e.g.,

<Steam>/steamapps/common/<Game>/Data/). All overlays go through MO2.

• Treat the user's existing MO2 profile as canonical. Do not silently mutate

profiles/<Profile>/plugins.txt, modlist.txt, INIs, or load order.

• Pause and surface state before each install action. The user should always

know which MO2 root, which profile, and which file is about to be touched before it happens.

• MO2 must run visibly. Never start MO2 in any background / hidden mode.

The human owner needs to see and interact with the MO2 GUI. The launch helper scripts/start-mo2.ps1 enforces this and surfaces "zombie" MO2 processes (running but no window) so they can be cleaned up before a fresh start.

What the control plane actually is

The "MO2 control plane" we deploy in step 4 is a Python MO2 plugin plus a PowerShell broker:

• tools/mo2-control-plane/live-bridge/mo2_agent_control.py — the actual MO2

plugin. When MO2 loads it, the plugin opens a named pipe and starts publishing bootstrap runtime files. This is what the agent harness talks to.

• tools/mo2-control-plane/broker/ — PowerShell IPC client used by

xedit-client.ps1. Runs from the plugin checkout; no install step.

There is no C++ DLL to build or deploy at v0.1. A C++ kernel skeleton lives under docs/internal/future-c-kernel/ as a design note for later perf-critical paths; it is intentionally unbuilt. If a previous agent told you to "build Mo2AgentControl.dll first" — that was wrong. The Python plugin is the integration.

Workflow

Step 1 - Detect MO2 install

Search for ModOrganizer.exe in common locations:

• $env:LOCALAPPDATA\ModOrganizer*\ModOrganizer.exe
• D:\ModOrganizer*\ModOrganizer.exe, E:\ModOrganizer*\ModOrganizer.exe,

C:\ModOrganizer*\ModOrganizer.exe

• The user's Documents folder.
• The user's Desktop / common steam-library siblings.

If exactly one MO2 install is found, propose it to the user for confirmation. If multiple, list them with paths and ask which is the target. If none, go to step 2.

Persist the confirmed install root in conversation context as MO2_Root. All subsequent steps reference this variable.

Also surface the BGS_MO2_ROOT env-var requirement. The xedit MCP server needs to know MO2_Root so it can resolve <MO2_Root>/tools/xEdit/xEdit.exe and <MO2_Root>/profiles/<profile>/plugins.txt. Without it, the MCP either falls back to a stale dev-sandbox path (silent wrong-answer on dev clones) or errors with a "MO2 sandbox root is not configured" message (on end-user installs). Tell the user where to set it for their harness:

• OpenCode: add to the xedit MCP entry's env block in opencode.json:

json { "mcp": { "xedit": { "env": { "BGS_MO2_ROOT": "<MO2_Root>" } } } }

• Codex: append to ~/.codex/config.toml:

toml [mcp_servers.xedit.env] BGS_MO2_ROOT = "<MO2_Root>"

• Claude Code: set the env var in the shell that launches Claude Code, or

edit <plugin-root>/.mcp.json to add an env block on the xedit entry.

After the user sets it, the user must restart their harness session so the MCP server picks up the new env. Confirm by calling xedit_status({}) and checking that the returned hint no longer mentions BGS_MO2_ROOT.

Per-call override is also available: xedit_start({ moRoot: "<MO2_Root>", ... }). Useful for one-off testing, but the env-var path is what makes restarts and subsequent sessions work without re-passing it.

Step 2 - If MO2 absent: branch the user

Surface a [BLOCKED] notice and offer THREE paths, in order of preference:

(a) Default - human install. Provide the official MO2 release URL: https://github.com/ModOrganizer2/modorganizer/releases/latest. Give the user a brief setup outline (download installer, pick an install root outside Program Files, launch once to initialize). Wait for the user to install and come back with the path.

(b) Agent-handled install (explicit user consent required). Only if the user says something like "yes, install MO2 for me, go ahead": download the official MO2 installer to a temp location, run it (silent install where supported), register the resulting path as MO2_Root. Surface every step ("downloading from ...", "installing to ...", "MO2 ready at ..."). Never proceed without the explicit verbal consent.

(c) No-MO2 mode. If the user says "I do not need MO2 for this project" (e.g., they are writing modpack docs / planning without a runtime), record MO2_Root = none, continue through step 3 for KB target selection, then skip to step 9 (dev-log / changelog init). Mark that any MO2-bound or xEdit-bound work will be unavailable until they revisit this skill.

Step 3 - First-run KB target selection and pack acquisition

Ask which target games the modpack should support. Use the exact game codes when talking to KB tools where possible: SkyrimLE, SkyrimSE, SkyrimAE, SkyrimVR, Fallout4, Fallout4VR, Fallout3, FalloutNV, Starfield.

The bundled bgs-kb-core pack is materialized inside the plugin tree at plugins/bgs-modding-superpowers/knowledge/bgs-kb/packs/core/. No action is needed for it; it appears in bgs_kb_status even before MO2 or xEdit are configured.

Per-game packs (bgs-kb-skyrim, bgs-kb-fallout4, bgs-kb-fallout3-fnv, bgs-kb-starfield) are published as GitHub Release artifacts on the BB-84C/bgs-modding-superpowers repo. The KB MCP wraps the entire fetch + sha256 verify + cache install in two tools — do not download zips by hand:

1. Map the chosen games to pack IDs:

• SkyrimLE / SkyrimSE / SkyrimAE / SkyrimVR -> bgs-kb-skyrim
• Fallout4 / Fallout4VR -> bgs-kb-fallout4
• Fallout3 / FalloutNV -> bgs-kb-fallout3-fnv
• Starfield -> bgs-kb-starfield

Additionally, if the user has indicated a translation/localization workflow for any of these games (intent to run xtl batch against the game), add the corresponding localization pack. Today's released localization packs:

• Starfield + Simplified Chinese -> bgs-l10n-starfield-zhhans

(151k Bethesda-official terminology entries; RAG retrieval inside xtl depends on this pack — see using-bgs-translator.)

These follow the same install path (bgs_kb_check_updates -> bgs_kb_install_pack) and the same on-disk cache layout as the per-game packs. Do not assume a localization pack is present just because the per-game pack is installed; check bgs_kb_status after install.

1. List what is available from the live release manifest:

bgs_kb_check_updates({}) This reads the published manifest-index.json and returns the latest { packId, version, sha256, url, breakingChange } for every official pack. Surface the per-game-pack rows that match the chosen games before doing anything else.

1. Get explicit user consent before installing each pack ("yes, install

bgs-kb-skyrim"). Pack install mutates ~/.bgs-modding-superpowers/kb/packs/<packId>/<version>/ (legacy %LOCALAPPDATA%/bgs-modding-superpowers/kb/packs is no longer used; the cache root is unified with xtl).

1. For each consented pack:

bgs_kb_install_pack({ packId: "bgs-kb-skyrim", version: "<from check_updates>", dryRun: true }) Verify the dry-run reports the expected version + sha256, then re-run without dryRun to actually install. The MCP downloads from the official release URL, verifies sha256, and lays the pack out in the versioned cache.

1. If bgs_kb_check_updates returns an empty or partial list for a game the

user wanted (network failure, release not yet published), say so and continue with the bundled core pack. Do not improvise a download URL or iwr into the cache by hand.

Smoke test after selection / install:

text
bgs_kb_status({})
bgs_kb_query({ query: "plugins", games: ["<chosen-game>"], maxResults: 3 })

Success criterion: bgs_kb_status lists every pack that was just installed plus bgs-kb-core, and bgs_kb_query returns at least one hit. If only the core pack is available, a core-pack hit is enough for first-run.

For routine update / prune / version-pinning work after first-run, route to maintaining-modding-environments — it documents the full ongoing-care flow on top of the same two MCP tools.

Step 3A - Install or verify xtl for AI translation workflows

If the user plans to translate mods, or a downstream workflow routes into using-bgs-translator, make sure the standalone translator CLI is available. xtl is distributed on PyPI as the bgs-translator package; do not assume it is present in a fresh machine just because this plugin is installed.

Start with a readback:

powershell
# Check whether the translator CLI is already installed and can report capabilities.
xtl version

If xtl is missing, choose one install path:

powershell
# Preferred user-level CLI install when pipx is available.
pipx install bgs-translator==0.9.0rc1

# Fallback when pipx is unavailable, or when the user wants xtl inside the active Python environment.
py -3.12 -m pip install bgs-translator==0.9.0rc1

For a future stable release, the same commands can omit the exact prerelease pin, for example pipx install bgs-translator or py -3.12 -m pip install bgs-translator. Do not use broad pip install --pre bgs-translator as routine setup; it can opt unrelated dependencies into prerelease builds. Pin an exact release candidate only when that exact version is the intended runtime.

Verify after install:

powershell
# Confirm xtl is on PATH and exposes the expected command surface.
xtl version

# Inspect the top-level help before choosing subcommands or options.
xtl --help

# Inspect provider-profile options instead of assuming a single SDK kind.
xtl profile add --help

# Inspect GUI launch options if the user wants the web translator.
xtl gui --help

If xtl version fails after a pipx install, surface the pipx PATH warning and ask the user to reopen the shell/session, then retry. Once it passes, route actual translation work to using-bgs-translator; this setup skill only installs and smoke-checks the CLI.

Step 4 - Install the MO2 control plane (Python + broker)

Once MO2_Root is known (path a or b), deploy the Python plugin:

powershell
& "<plugin-root>/scripts/install-mo2-control-plane.ps1" -MO2Root "<MO2_Root>"

This deploys:

• tools/mo2-control-plane/live-bridge/mo2_agent_control.py -> <MO2_Root>/plugins/
• The Mo2AgentControl/ support tree -> <MO2_Root>/plugins/
• ModOrganizer.ini lock_gui=false normalization.

After the script returns, list <MO2_Root>/plugins/ and confirm mo2_agent_control.py is present. The script will NOT install a .dll — that is intentional (see "What the control plane actually is" above). If the script errors, stop and surface it; do not continue.

Step 5 - Start MO2 visibly

powershell
& "<plugin-root>/scripts/start-mo2.ps1" -MO2Root "<MO2_Root>" -Profile "<Profile>"

The launcher script:

• Refuses to start if a visible MO2 is already running (use that one).
• If a zombie MO2 (running but no window) is detected, surfaces it and asks

before killing — use -KillStale to skip the prompt.

• Starts MO2 with -WindowStyle Normal so the GUI is visible.
• Waits up to 30s for the main window to appear and reports its title.

Verify after launch: the user should see the MO2 GUI on their desktop. The plugin's bootstrap runtime files should appear at <MO2_Root>/plugins/Mo2AgentControl/bootstrap/runtime/ within a few seconds:

• status.json -> { schemaVersion, state: "ok", mo2Pid }
• endpoint.json -> { transport: "named-pipe", endpoint: "mo2-control-plane-<pid>" }
• capabilities.json -> lists launch.*, system.* methods.

If those don't appear, the Python plugin didn't load. Check MO2's plugin settings to confirm mo2_agent_control is enabled.

Step 6 - Ask the user whether they want xEdit

xEdit is optional but high-value. Describe what it does so the user can make an informed choice:

> "xEdit is the canonical editor for Bethesda plugin files (.esp / .esm / .esl). > Modpack curators use it to resolve conflicts between mods, generate > compatibility patches, flag plugins as ESL (light) to free up load-order > slots, clean stray edits (ITM / UDR), and run Pascal-based bulk edits. > With this plugin's bundled xedit MCP, your agent can drive xEdit > programmatically - no GUI clicks required."

Then ask: "Would you like me to install xEdit?" - explicit consent gate.

If the user declines, skip to step 9.

Step 7 - If yes: fetch xEdit from BB-84C/TES5Edit

The forked, agent-friendly xEdit lives at https://github.com/BB-84C/TES5Edit. Fetch its latest release into the MO2 tools tree:

powershell
& "<plugin-root>/scripts/fetch-xedit-release.ps1" -MO2Root "<MO2_Root>"

This downloads the latest release zip, extracts into <MO2_Root>/tools/xEdit/, and verifies xEdit.exe exists post-extract. For ongoing version pinning after first install, route to maintaining-modding-environments.

Step 8 - Native xEdit automation

The xEdit hook bridge is retired and must not be deployed. The installed xEdit.exe is launched through xedit-client.ps1, which uses native -automation-serve and a run-scoped plugins.txt when needed.

Step 9 - Initialize dev-log and release-changelog

Ask the user for their modpack project root. This is usually one of:

• <MO2_Root>/profiles/<ProfileName>/ (if the profile is the unit of work).
• A separate Git-tracked source directory the user maintains (more common for

released modpacks).

Once <project_root> is known, route to:

• writing-modpack-devlog skill - creates <project_root>/docs/dev-log.md with

the project name and start date as the first entry.

• writing-modpack-changelog skill - creates

<project_root>/docs/release-changelog.md skeleton.

From here on, those two skills maintain the files at runtime; do not template or pre-fill them in this skill.

Step 10 - Verify with a semantic smoke test (NON-BLOCKING)

The xEdit MCP is fully non-blocking. Do NOT call a single blocking tool and wait — that will time out. Instead:

0. Read MO2's managed gamePath before launching xEdit. xEdit's default game discovery uses the Windows registry, which points at the Steam install — NOT MO2's Stock Game. To force xEdit to see MO2's projected game tree, read <MO2_Root>\ModOrganizer.ini and extract the gamePath value:

gameName=<Target Game> gamePath=@ByteArray(<MO2 managed game root>)

Strip the @ByteArray(...) wrapper; the Data dir is <gamePath>\Data. Pass this to xedit_start as dataPath in the next step. See writing-bgs-load-order for the full reasoning + alternative load-order workflows.

1. Kick off the daemon launch.

xedit_start({ dataPath: "<gamePath>\\Data", // from step 0 above }) Without an explicit dataPath, xEdit will discover the game via registry and may open the wrong Data directory.

Expect { ok: true, data: { status: "starting" } } (or "ready" if it was already up). If you want to use a custom subset of plugins (load-order experimentation), also pass pluginsFile: xedit_start({ dataPath: "<gamePath>\\Data", pluginsFile: "<agent-owned-artifacts-path>/plugins.txt", }) The writing-bgs-load-order skill explains how to author the file.

1. Poll xedit_status until ready. The first launch takes 60-240s because

xEdit must parse the active load order. Sleep 5-15s between polls. xedit_status({}) -> { status: "starting", elapsedSeconds: N } xedit_status({}) -> { status: "starting", elapsedSeconds: N+5 } ... xedit_status({}) -> { status: "ready", pid: N, readySince: timestamp }

If status: "failed", surface data.error exactly and STOP. Do not declare success. Common reasons: MO2 not running, MO2 not loading the Python plugin, xEdit binary missing, xEdit's automation-serve mode tripping on the active load order.

1. Confirm with xedit_health — sends a real system.ping to the daemon

to catch zombies. Expect { status: "ready", responsive: true }.

1. Confirm domain tools work.

xedit_list_capabilities({}) -> ~49 commands, empty drift.onlyInDigest xedit_session({}) -> { gameMode: "Fallout4", loadOrderSize: N, daemonPid: N }

If all four pass: surface [OK] BGS modding environment ready with the recorded MO2_Root, xEdit path (if installed), and project root.

Acceptance (semantic, not surface)

The setup is complete only when ALL of the following hold. Do not declare success otherwise:

• <MO2_Root> is known and <MO2_Root>/ModOrganizer.exe exists.
• Target games are recorded; bgs_kb_status sees the bundled core pack; and at

least one bgs_kb_query smoke for the chosen game returns a hit.

• If AI translation workflows were requested: xtl version succeeds from the

user's shell, and xtl --help lists the translator command surface.

• <MO2_Root>/plugins/mo2_agent_control.py exists (skipped only in no-MO2 mode).
• MO2 is visibly running (process has MainWindowHandle != 0) and

<MO2_Root>/plugins/Mo2AgentControl/bootstrap/runtime/status.json reports state: "ok" with a current mo2Pid (skipped only in no-MO2 mode).

• If xEdit was chosen: <MO2_Root>/tools/xEdit/xEdit.exe exists. The retired

xEditHookBridge.dll is not required.

• <project_root>/docs/dev-log.md and

<project_root>/docs/release-changelog.md both exist (or user explicitly declined - record the decline in conversation context).

• xedit_status eventually reports status: "ready", and

xedit_health.data.responsive is true, and xedit_list_capabilities returns a non-empty digest. (Skipped only if xEdit was declined or no-MO2 mode.)

In no-MO2 mode, only the KB target selection / smoke and the dev-log / changelog steps need to pass.

Common mistakes

• Skipping the consent gate on step 2(b) or step 6. Both REQUIRE explicit user

agreement; silent install or silent install-decision is forbidden.

• Silently choosing one MO2 install when multiple are detected. Always ask.
• Trying to build Mo2AgentControl.dll from docs/internal/future-c-kernel/.

That is a skeleton, not a real MO2 plugin. The Python plugin IS the integration. Build the C++ kernel only if you are deliberately moving perf-critical paths out of Python in a future release.

• Calling xedit_session or any domain tool and blocking on the response.

Every tool returns immediately. If you see status: "starting", poll xedit_status; do not call the same tool again in a tight loop.

• Using /mcp reconnect as the normal way to relaunch xEdit.

Prefer xedit_dirty, xedit_stop, and xedit_restart so the agent can decide whether to save or abandon unsaved edits before clearing state.

• Starting MO2 with Start-Process -WindowStyle Hidden (or any other

invisible/background mode). Always use the start-mo2.ps1 helper, which forces a visible window.

• Writing into the user's vanilla game install directly instead of via MO2

overlay. NEVER do this. See using-bgs-modding-superpowers rule 1.

• Declaring success on a green script return without the semantic readback

(the file existence checks and the four MCP smoke calls).

• Using broad pip install --pre bgs-translator during setup. Prefer an exact

package version such as bgs-translator==0.9.0rc1, or use the stable package name once a stable release is published.

See also

For ongoing care after first-run, see maintaining-modding-environments.

• using-bgs-modding-superpowers - the per-session bootstrap; lists the full

skill inventory and hard rules.

• xedit-automation - hub skill for all xEdit work; load after this skill

succeeds.

• writing-modpack-devlog, writing-modpack-changelog - runtime asset skills

invoked from step 9.

• The active installer scripts under scripts/: install-mo2-control-plane.ps1,

start-mo2.ps1, fetch-xedit-release.ps1.