references/mounting-guide.md

Nomad Mounting Guide

Defines how to connect an already Nomad-shaped team (see team-skeleton.md) to one more AI harness — adding a platform alongside ones already in use, not building the team from scratch. This is Path C from SKILL.md.

0. What this file assumes

Step 0 of SKILL.md already settled the language. This is the English version — see mounting-guide.zh-CN.md for the content-equivalent Chinese version. The team you're mounting already meets the minimum-viable-team bar from team-skeleton.md §8 for the agent(s) being mounted — i.e. four of the six §2 agent-home pieces (AGENTS.md, CLAUDE.md/equivalent thin shell, workspace/, memory/). It does not need the two deferred pieces (hooks config, how-to-start doc) already in place — wiring hooks for this harness is part of what mounting does, and how-to-start is exactly the kind of doc a mount (a new reader) can trigger. If the agent doesn't yet meet even the §8 minimum, that's Path A (team-skeleton.md), not this file.

1. Universal steps (apply on every harness before checking the per-platform table)

1. Point the harness's project/working directory at the agent's home, not at the team root and not at some parent folder. Every platform in §3 below determines "which agent am I talking to" by which folder is open — opening an ancestor folder does not automatically reach a nested agent's AGENTS.md.
2. Confirm AGENTS.md is discoverable the way this specific harness expects it. Some harnesses read AGENTS.md directly; others need a thin per-platform shell file (see team-skeleton.md §2 for why CLAUDE.md exists as one such shell) or an explicit setting turned on before the file is even considered.
3. Wire hooks only if the node layer is present and this harness actually executes hooks. Check §3 and §4 below before assuming a harness supports this — assuming parity across harnesses is the single most common mounting mistake. The node layer's script templates now ship in this package's scripts/ directory (team-skeleton.md §4; per-platform wiring examples in scripts/README.md §3) — but confirm the target harness actually executes hooks before wiring anything in regardless.
4. Run the acceptance checklist in §5 before declaring the mount done. A folder being "open" in a harness is not the same as the personality having actually loaded.
5. Decide the mounting pattern for multi-agent teams before mounting anything. Everything above describes mounting one agent's home. When the team has more than one agent, there are two distinct ways to actually put them in front of a harness — pick one with the user, don't default silently:
   • One harness window/session per agent — open each {AGENT_NAME}/ home in its own separate session. The cleanest match for harnesses that determine identity purely from "which folder is open" (§3), and the right default when agents mostly work independently.
   • Hub-coordinator mode — one agent (typically the team's orchestrator/router role) is mounted as usual, but its session is additionally given read/write access to the other agents' homes and the shared team root, via whatever cross-directory access mechanism the harness supports (e.g. Claude Code's additionalDirectories). That one agent can then read and dispatch to the others without each of them needing its own live session. Use this when one agent's actual job is to route or synthesize across the team — see 02-shared-knowledge/collaboration-protocol.md. Not every agent in a team needs the same pattern — decide per agent whether it needs its own mounted session or only needs to be reachable through the hub.

2. Desensitization note

The platform facts in §3 are public product-mechanism knowledge (how each platform's file discovery, hooks, and settings actually work), plus verified-in-practice findings from independent testing. No team name, principal name, or absolute local path appears below — every path uses the {TEAM_ROOT} / {AGENT_NAME} placeholders from team-skeleton.md §12.

3. Platform-by-platform

Platform	How to connect	First-use gotcha	Known pitfalls
Claude Code	Open the harness with its working directory set to {TEAM_ROOT}/01-agents/{AGENT_NAME}/. Personality loads via the CLAUDE.md thin shell → @AGENTS.md reference.	None specific to first use — loads immediately once the working directory is correct.	This harness's project-level config lookup climbs to the repository root, so if a repository-root .claude/settings.json (kept for cross-platform hook sharing, see Copilot row below) and an agent-home-level .claude/settings.local.json are both present, Claude Code reads and fires both — that's a double-fire (session-start injection twice, stop-reminder twice), not a misfire. Don't chase this as "hooks stopped working" — chase it as "hooks are firing an extra time." Fix by keeping only one config in effect per deployment, or by adding an idempotency guard in the node script itself once the node-layer scripts exist (the production deployment this spec was distilled from used a duplicate-invocation guard for exactly this reason).
Codex / ChatGPT Desktop	When starting a new task, manually pick the agent's home folder as the working location — the picker does not walk into subfolders on its own. Personality loads via native AGENTS.md reading, no thin shell needed.	Hooks require an in-app trust review the first time — a per-hook trust card (commonly three prompts) must be approved before hook scripts fire at all.	Trust is scoped by content hash, not by "this hook in general" — upgrading the hook script's content may or may not require re-approval depending on how much changed; don't fully trust an "already trusted" UI indicator — confirm via an actual startup-injection check (§5) instead. Non-interactive/headless invocations of this platform may not support MCP tool calls or custom-subagent dispatch even when the interactive session does — don't assume headless parity without testing it directly.
OpenCode	Open a project at the agent's home folder, or — if a global multi-agent registry has been set up for this deployment — select the agent by name from the agent list inside any project.	None specific to first use.	If the global agent registry is populated by a generator/translator script rather than reading AGENTS.md directly, the actual chain is typically AGENTS.md → an intermediate per-harness definition file (e.g. a .toml agent definition) → the generator → the registry entry — two hops, two places drift can enter, not one. Drift can happen at either hop: the intermediate file's pointer to AGENTS.md could go stale, or the intermediate file itself could change and the generator not get re-run — either way the registry entry silently keeps showing an old personality, with no automatic drift alarm. Treat re-sync (re-running the generator) as a manual step you must remember after any personality edit, and don't assume the generator reads AGENTS.md directly just because the end result looks like it does. Hooks route through a plugin at the global plugin location; if the specific hook APIs used are marked experimental by the platform, re-verify after every version upgrade — don't assume yesterday's verification still holds post-update.
Copilot (VS Code)	Open the agent-focused session surface ("Agents window" — as distinct from the classic chat panel, see next column), point it at the agent's home folder, and pick Folder mode (in-place), not Worktree mode (isolated copy).	In the agent-focused session surface's Folder mode, AGENTS.md is read directly with no extra setting — verified in practice (full personality load, boundaries and context all correct). This is a different mechanism from the classic chat panel: if this deployment will also use that panel, it injects AGENTS.md a different way — it needs an explicit experimental setting turned on (chat.useAgentsMdFile), and requires the file to sit at the workspace root, where a nested agent-home AGENTS.md below the opened workspace root is only found if searched for, not auto-injected. Don't apply the chat-panel's requirements to the Agents window, or vice versa — they're two separate injection paths in the same product.	Two chat surfaces exist in the same product and are not equivalent: only the agent-focused Agents window reliably executes hook configs; the classic chat panel does not, even under identical settings — verify which surface you're in before trusting hook behavior. Filename pitfall: this surface reads hook config only from the exact filename settings.json — a .local.json-suffixed variant is silently ignored. If other platforms in this deployment use a .local.json naming convention, this platform needs its own settings.json — and if both end up present at the same root, guard against hooks double-firing. Session-grouping pitfall: the Agents window groups sessions by the enclosing repository and reads config from that repository's root, not necessarily from the agent's own subfolder — if the agent's home isn't its own repository, its own hook config may never be read at all; Worktree mode compounds this further (cwd becomes an isolated repo-root worktree copy, where personality does not load at all — Folder mode is the one that reaches the agent's own AGENTS.md).
Cursor	Open the agent's home folder as the project. Personality loads via native AGENTS.md reading.	None specific to first use — verified in practice to load the full personality and memory chain, not merely a partial summary.	Hooks are not reliably available on this platform as of writing — even where input-side plumbing exists, don't claim "hooks supported" until you've verified an actual startup injection lands in a real session (§5); default to relying on the md-layer self-driven instructions here. If this platform has its own native rules-file mechanism separate from AGENTS.md, verify which one wins when both exist and disagree — don't assume AGENTS.md automatically overrides a platform-native equivalent.

4. MD-layer-only vs. node-layer degrade, per platform

Per team-skeleton.md §4, the md layer (plain AGENTS.md + memory files) is the required minimum and must work with zero automation; the node layer (hooks) is acceleration, never a load-bearing dependency. What "degraded" looks like on each platform:

Platform	If the node layer / hooks aren't wired	What still works
Claude Code	No automatic session-start injection or stop-reminder	Personality and memory still load fully via CLAUDE.md → AGENTS.md; self-driven open/close steps written into AGENTS.md's platform-differences section still apply — the agent runs them itself
Codex / ChatGPT Desktop	Same — no automatic injection/reminders until hooks are trusted	Same self-driven fallback; native AGENTS.md reading is unaffected by hook status
OpenCode	No automatic injection/reminders if the plugin isn't registered	Same self-driven fallback; agent-list selection and AGENTS.md reading are unaffected
Copilot (VS Code)	No automatic injection/reminders in whichever surface doesn't execute hooks	Personality still loads fully in the agent-focused surface with the right settings and mode; self-driven fallback applies
Cursor	No automatic injection/reminders — current default state for this platform	Personality and memory load fully via native AGENTS.md reading; self-driven fallback is the normal operating mode here, not a fallback for an edge case

5. Post-mount acceptance checklist

Run all three before declaring a mount done — a folder being "open" is not proof the personality loaded:

1. Ask "who are you?" The agent should answer with its actual name, role, and boundaries from AGENTS.md — not a generic-assistant answer, and not silence on the personality question.
2. Check for startup injection. Did the agent mention anything indicating it read the team log, its own inbox, or otherwise ran an open-session routine — whether that came from a hook or from self-driven instructions? If neither happened, the open routine isn't wired yet, only the personality file is.
3. Say something like "let's wrap up here" and watch for the close-out routine. Per team-skeleton.md §6's five-step close-out, the agent should update its memory files and worklog on its own — with or without a hook. If nothing happens unprompted, note it as a known gap rather than assuming it silently succeeded.

---

Nomad skill · mounting-guide reference · desensitized — no real paths, names, or account details appear anywhere in this file. Version 0.3 · Last updated 2026-07-14