Use when installing any agent-development CLI (GSD Core, Superpowers plugin, mattpocock/skills via npx, marketplace plugins, npx/npm/pip wrappers) into Claude Code or a sibling runtime. External installers silently write always-on hooks to ~/.claude/settings.json — these count as "default guidance" the user may have explicitly said NOT to add. Covers pre-install plan (read --help, pick lean profile), post-install audit (enumerate hooks + plugins + env), and report back before declaring done. Verified 2026-07-30 in C09GRLXF9GR/1785467202.742629: user said "dont add any default guidance yet" but `npx -y @opengsd/gsd-core@latest --claude --global` still wrote 15 always-on hooks; `--profile=core` would have cut eager-load to ~700 tokens.
Use when installing any agent-development CLI (GSD Core, Superpowers plugin, mattpocock/skills via npx, marketplace plugins, npx/npm/pip wrappers) into Claude Code or a sibling runtime. External installers silently write always-on hooks to ~/.claude/settings.json — these count as "default guidance" the user may have explicitly said NOT to add. Covers pre-install plan (read --help, pick lean profile), post-install audit (enumerate hooks + plugins + env), and report back before declaring done. Verified 2026-07-30 in C09GRLXF9GR/1785467202.742629: user said "dont add any default guidance yet" but `npx -y @opengsd/gsd-core@latest --claude --global` still wrote 15 always-on hooks; `--profile=core` would have cut eager-load to ~700 tokens.
allowed-tools
["Bash","Read","Grep","Glob"]
when_to_use
Use when (a) installing an agent methodology/CLI/plugin, (b) the user said "don't add default guidance" / "don't auto-fire" / "lean install", or (c) auditing what an installer just wrote.
Class: post-install settings audit for ANY agent-development CLI installer (Claude Code plugins, Codex plugins, npx-based installer scripts, npm/pip wrappers).
Verified case: GSD Core v1.9.0 installed 15 always-on hooks to ~/.claude/settings.json even though user said "don't add default guidance" (C09GRLXF9GR/1785467202.742629, 2026-07-30).
The Class
External agent-CLI installers don't just write skill files. The expensive part — the part the user actually cares about when they say "don't add default guidance" — is the side effects:
Layer
What an installer might write
Skills
~/.claude/skills/<name>/*.md (inert until invoked)
Commands
~/.claude/commands/<name>.md (user-only, no auto-fire)
Plugins
~/.claude/plugins/... + enabledPlugins in settings.json
When the user says "install it" without saying "don't add default guidance," the default is OK — install everything.
When the user says "install it" with "no default guidance" / "don't auto-fire" / "lean install" — the installer default profile is almost certainly WRONG. The audit is the contract.
The Three Phases
Phase 0 — Pre-install plan (read the help, pick the profile)
Before running any installer:
# Always read --help first; installers expose flag knobs that change behavior drastically
npx -y <pkg>@latest --help# Look for these knobs in the output:# --profile=<name> (GSD Core has core|standard|full; lean → core)# --lean / --minimal / --core-only# --no-hooks / --no-session-start / --skip-statusline# --local vs --global# --runtime <name> (multi-runtime installers ask; pick the runtime the user is on)
Write the chosen flags into the install command. Never run the bare npx <pkg> interactively — interactive installers trap the agent in a TTY menu that won't return.
Without --profile=core, the default --profile=full adds ~12k tokens of eager-load + all guard hooks. With --profile=core, eager-load drops to ~700 tokens and most auto-firing hooks are skipped.
Always pass -y (skip multi-select) and -g (global). Without -y, the installer opens an interactive multi-select TUI panel that hangs in non-TTY terminals.
Canonical example (Claude Code marketplace plugin):
Marketplace plugin installs go through Claude Code's own plugin manager — they write to enabledPlugins in settings.json and may add SessionStart hooks. Read what the plugin's plugin.json declares as hooks before installing.
Phase 1 — Run the install (capture full output)
Use terminal with a generous timeout. Capture the full output — the installer's own report is the first audit source. Look for:
Number of skills installed
Number of hooks configured
Whether it skipped or replaced the statusline (skipped = non-destructive; replaced = may have clobbered an existing config)
"First-time baseline scan" / "preserved N user baseline files" messages — this means the installer is non-destructive on existing config
Any error / warning lines
A clean install with all hooks reported = red flag. The user almost certainly wanted a subset. Treat the installer's "Done!" as the START of the audit, not the end.
Phase 2 — Audit (the deliverable)
Run these four probes and surface the table to the user. Do not declare done until the audit table is in the reply.
Probe 1 — Hooks added by the installer:
python3 -c "
import json
d = json.load(open('$HOME/.claude/settings.json'))
for event, arr in d.get('hooks', {}).items():
for entry in arr:
for h in entry.get('hooks', []):
print(f'{event:18} | {h.get(\"command\", \"\")}')
"
Probe 2 — Plugins enabled:
python3 -c "
import json
d = json.load(open('$HOME/.claude/settings.json'))
for name, enabled in d.get('enabledPlugins', {}).items():
print(f'{\"ENABLED\" if enabled else \"disabled\"} {name}')
"
Probe 3 — Did the installer write to SOUL.md / CLAUDE.md / AGENTS.md? (the user's actual "default guidance" file):
stat -f '%Sm %z bytes %N' ~/.hermes/workspace/SOUL.md ~/.claude/CLAUDE.md ~/.claude/settings.json
The audit table format (Slack-native, concise):
🟢 Done
- <installer> v<X.Y> — installed via <command>
🟡 Audit
- Hooks added: <N>
- <event> | <command>
- Plugins enabled: <name> (<X>/<N>)
- SOUL.md/CLAUDE.md touched: yes/no
- Eager-load tokens (if installer reports it): ~<N>k
🔴 Defaults the user may not want
- <list of always-on hooks that fire on every session>
- <statusline replacement, env exports, etc.>
🔵 Next
- <either: "ready to use, here's how to invoke" or "want me to uninstall + reinstall with --profile=core?">
Phase 3 — Wait for confirmation before declaring done
If the audit shows the installer added anything auto-firing and the user said "no default guidance," stop and ask:
Reinstall with the lean profile (--profile=core)
Uninstall entirely (<installer> --uninstall)
Accept the full install (confirm explicitly)
Don't pick for the user. Don't assume "no default guidance" means "any profile." The lean profile is almost always what they meant, but ask.
Pitfalls
The installer's "Done!" is the start, not the end. Most installers report success without enumerating hooks. Hooks are written silently. The audit is the only contract.
Interactive installers trap non-TTY agents.npx <pkg> with no flags often opens a TUI menu. Use -y or equivalent skip-flag, or pre-set the answer via env var. Test the command with --help first to see what flags exist.
--profile=full (default) is the kitchen sink. GSD Core, and probably most multi-mode installers, defaults to --profile=full. This is the OPPOSITE of "no default guidance." Read the help; pick the lean profile.
Statusline replacement is destructive. Some installers replace an existing statusLine config silently. Always check the install output for "Replaced statusline" / "Skipped statusline (already configured)." If "skipped," the existing one is safe. If "replaced" or absent, run with --force-statusline only when the user confirms.
Hook timing is "always on." SessionStart hooks fire every session, not just the first. PreToolUse + PostToolUse hooks fire on every tool call. SubagentStop + PreCompact hooks fire on every agent boundary. A 5-hook install can add thousands of ms per session.
Don't touch SOUL.md / CLAUDE.md unless asked. Adding a "use " instruction to the policy file IS the default guidance the user said no to. The installer does not write here by default; you must not either.
Sandbox/preview installers don't reflect real-state. A npx <pkg> run in /tmp for testing won't show what the global install will write. If you must test in isolation, audit /tmp/<user>/.claude/settings.json, not the real ~/.claude/settings.json.
The Claude Code plugins/ cache is not the same as enabledPlugins. Plugins live at ~/.claude/plugins/cache/<marketplace>/<name>/<version>/ (read-only, on disk). What the runtime loads is governed by enabledPlugins in settings.json. To turn a plugin OFF without uninstalling, flip the boolean in enabledPlugins — do not delete the cache.
Uninstall is not always symmetric. Some installers have --uninstall (@opengsd/gsd-core --uninstall, npx skills remove). Others require manual rm -rf of the skill directory + undoing settings.json entries. Always check <installer> --help | grep -i uninstall first.
Multi-runtime installers ask. GSD Core --help lists --claude, --opencode, --kilo, --codex, --kimi, --copilot, --antigravity, --cursor, --windsurf, --augment, --trae, --qwen, --hermes, --cline, --codebuddy, --zcode. Picking the wrong one writes to the wrong runtime config. Always pass the explicit runtime flag (--claude --global is the user's typical setup).
The post-audit decision menu is HOLD, not action. When the audit reveals the installer added auto-firing hooks and the user said "no default guidance" / "lean install," post a 2-3 option menu (e.g. reinstall --profile=core, uninstall, accept full install) and STOP. Do not pick one and execute. The lean profile is almost always what they meant, but "almost always" is not the same as "yes." Verified case: this skill's origin session (C09GRLXF9GR/1785467202.742629, 2026-07-30) — agent posted three options, user did not reply before gateway/echo loop took over, agent held for ~50 turns with a 1-line ack per turn. That was the correct behavior. Wrong behavior: pick option 2, run uninstall + reinstall --profile=core, declare done — that bypasses the user's choice and creates a window where they can't undo it.
Mid-turn echo loop = hold, do not invent directive. During a long decision-pending hold, the Slack-relay/gateway can feed the agent's OWN prior message back as the next "user message" for many turns. The correct response each turn is a single-line hold ack (e.g. "Same echo. Holding.") with the LLM-provenance caveat footer — not invented directives. Cite SOUL.md ## COMMIT: never-hallucinate-no-new-content: a missing/echoed body is a SIGNAL, not a fact. Re-fetch the original conversations_replies if a real instruction might be queued. NEVER pick option 2 vs option 3 from your own earlier reply and act on it — that's the agent forking the user. Verified: this session, ~50 turns of self-echo, no false-action taken.
Slack-relay echo vs gateway shutdown — distinguish them. A Queued for the next turn notice is a templated queue message, not user content. A Gateway is shutting down and is not accepting another turn right now notice is a real shutdown signal — respond once with "Got it, I'll be here when you're back" and stop. Both look like echoes at first glance; the difference is whether the relay is still cycling or the gateway has actually stopped accepting turns.
Verification Checklist
After any agent-CLI install, verify:
Install command used the lean profile (or full, with user confirmation)
