| name | sigma-prune |
| description | Find loaded-but-unused MCP servers and plugins that bloat every Claude context, and reversibly disable the dead weight to save tokens + cost. Use when the context feels heavy, tool lists are huge, or you want to trim what's loaded. Triggers: "too many tools", "context is bloated", "what MCP/plugins am I not using", "save context", "prune unused", "sigma prune", or a periodic context-hygiene pass.
|
| origin | sigma |
sigma-prune
Trim the recurring context tax. Every enabled plugin + connected MCP server injects
its tool schemas / skill descriptions into every turn; unused ones are pure
overhead. The pure inventory/usage/ranking logic lives in cli/prune.py; the file
reads + reversible write in cli/prune_run.py. This skill is the judgment layer.
What counts as a candidate
- Loaded — an enabled plugin (
enabledPlugins) or a connected MCP server
(user ~/.claude.json or project .mcp.json).
- Idle — zero
tool_use / Skill invocations in the usage window (default: all
scanned transcripts; narrow it with --recent-files N to prune servers idle
lately even if heavily used long ago).
- Ranked by REAL context weight. A server's weight scales with its distinct tool
count (schema width, observed across history), not a flat per-kind constant — a
100-tool server dwarfs a 2-tool one instead of ranking identical. Unknown tool count
→ conservative per-kind fallback.
- Rarely-used = low-confidence (opt-in).
--idle-threshold N also surfaces items
used ≤N times, flagged ⚠ rarely used — a judgment call for the human, never an
auto-disable. Default 0 = unused-only (unchanged).
Two non-negotiable laws
- Never prune on absent evidence. No transcripts to scan → surface nothing.
Missing usage data is treated as "used", not "unused" — the conservative default
(the inverse of guessing, mirroring gate-defaults-WAKE / verdict-defaults-FAIL).
- Reversible, never destructive. Disable = flip
enabledPlugins[name]=false via
an immutable settings merge (every other key preserved, like statusline). The
item stays installed on disk. Re-enable = flip it back. User-level MCP servers in
~/.claude.json are surfaced for a manual edit — prune never auto-edits that
file.
When to keep something unused
- It's a safety/just-in-case tool you reach for rarely but critically.
- It was only just installed (no usage yet ≠ unused).
- Disabling it would break an active workflow in another project (settings.json is
global — a plugin idle here may be hot elsewhere). Prefer project-scoped judgment.
Compose, don't duplicate
Bundle-hygiene trio: sigma-scout grows the bundle, sigma-prune trims it,
sigma-cost sizes the token cost of carrying it. Orthogonal to RTK (proxy token
cut) and caveman (output terseness) — prune cuts input/context weight, the layer
neither of those touches.
Flags
--check (read-only CI gate, exit 1 on bloat), --yes (disable all prunable
plugins), --files N (transcript scan = schema width), --recent-files N (usage
window — prune items idle in the last N), --idle-threshold N (also surface items
used ≤N times as low-confidence). Default is confirm-per-item, unused-only.
