--- name: diff-addons description: Verify profile isolation by diffing the declared inventory (script) against what's actually loaded in this Claude session (system-reminder) --- This skill is a **profile isolation validator**. It runs in three phases and produces a diff report. Follow the phases in order, do not skip. ## Phase 1 — Declared inventory (from the filesystem) Run the helper and parse its JSON output. This is what the profile *declares* should be loaded: !`node "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/marketplaces/claudeworks/plugins/profiles-manager}/scripts/list-addons.js" --json 2>&1` Parse the JSON into three sets: `declared.skills`, `declared.commands`, `declared.agents`. Each is a sorted array of `plugin:name` strings. Remember the `profile`, `pluginCount`, `counts.excluded`, and `excludedSkills`/`excludedCommands`/`excludedAgents` fields — you'll need them for the final report. ## Phase 2 — Actual inventory (from your session context) Now enumerate what's **actually available to you in this running session**. The authoritative source is the `` block near the start of the conversation that lists available skills. That block is ground truth — the Claude Code runtime injected it after profile assembly completed, so it reflects what was actually loaded. **Strict rules for Phase 2 — read, do not recall:** - Read directly from the system-reminder text in your context. Do not rely on memory of what a "typical" profile has. - Convert every listed skill to the same `plugin:name` format used in Phase 1. Items in the system reminder that already have a `plugin:name` prefix stay as-is. Items without a plugin prefix (e.g. `update-config`, `keybindings-help`, `simplify`, `loop`) are **builtin or globally-installed** — prefix them with the sentinel `(global):` so they diff cleanly. - Do the same for commands (user-invocable slash commands grouped under "user-invocable skills" or similar) and for agents (under the Agent tool's `subagent_type` list). - **Profile-scoped commands exception**: the declared inventory may include entries prefixed `(profile):` (e.g. `(profile):workflow`, `(profile):workflow-debug`, `(profile):intro`). These are slash commands the profile wrote into its own `config/commands/` directory — Claude Code loads them as bare command names, same as `(global):` items. When you see a bare command in the session whose short name matches a declared `(profile):` entry, prefix it with `(profile):` (not `(global):`) so the diff matches. Only fall back to `(global):` when no `(profile):` declaration uses that name. - If a category (skills, commands, agents) is not visible in your context, say so explicitly in the report instead of guessing. Build three sets: `session.skills`, `session.commands`, `session.agents`. ## Phase 2.5 — Extraction self-check (required before Phase 3) Phase 2 is an LLM extraction step and is prone to skim errors that silently drop declared items from `session.*` sets — producing false positives in "Missing from session". Run both checks below before diffing. If either catches a discrepancy, fix the extraction and re-run. 1. **Count cross-check.** For each of the three categories, count the raw bullet lines in the system-reminder block (ignoring sub-bullets and descriptions) and compare to the size of the extracted set. A mismatch means you dropped or doubled items — re-extract that category. 2. **Declared-item recheck.** For every declared entry (`plugin:name`) that would land in "Missing from session" after the diff, scan the raw system-reminder text for both the short name (`name`) and the full form (`plugin:name`). If either appears anywhere in the reminder, the item IS in the session — add it to the correct `session.*` set and re-run the diff. If the recheck rescued any items, add a one-line note to the report directly under the `Declared` / `Session` summary: `Extraction self-check: rescued N item(s) that Phase 2 missed.` Otherwise, omit the line entirely. ## Phase 3 — Diff and report Compute three deltas per category: - **Matched** = `declared ∩ session` — items in both. Report the count only, never the list. - **Missing from session** = `declared ∖ session` — declared in the profile but absent from your context. These are **load failures** — profile assembly wrote the files but Claude Code didn't pick them up. Every entry here is a bug worth investigating. - **Unexpected in session** = `session ∖ declared` — present in your context but not in the profile's declaration. These are **isolation leaks** — items reaching your session from some source the profile didn't ask for (usually `~/.claude/` bleeding through because the runtime reads it regardless of `CLAUDE_CONFIG_DIR`). Every entry here is a profile-separation bug. Items prefixed `(global):` are always unexpected from the profile's perspective — group them under their own subheading so a user can see "global builtins leaking in" separately from "non-builtin leaks," because the latter is a more urgent signal. ## Output format Produce a terse markdown report. Example shape — fill with your real diff: ``` # Profile isolation check — Declared: skills, commands, agents ( excluded) Session: skills, commands, agents ## ✓ Matched Skills: / Commands: / Agents: / ## ⚠ Missing from session (declared but not loaded) - plugin:name - plugin:name (none → "All declared items are loaded.") ## ⚠ Unexpected in session (loaded but not declared) ### Non-builtin leaks (investigate) - plugin:name ### Global builtins global skills, global agents (expected — not an isolation issue) (none → "No unexpected items.") ``` **Global builtins are expected noise** — Claude Code always injects its own skills and agents regardless of profile. Do NOT enumerate them individually. Just show a one-line count. Only enumerate non-builtin leaks (those are the actionable items). **Verdict line at the end**: if Missing is empty AND the "Non-builtin leaks" subsection is empty, print `✓ Profile isolation verified.`. If either has entries, print `✗ Profile isolation issue: missing, unexpected (non-builtin).` and nothing else on that line. Do not enumerate the profile's excluded items (the helper's JSON includes them for your reference only — they are noise at scale for mega-bundle plugins like antigravity-awesome-skills and must not appear in the report body).