Nerf

Nerf: Context Budget Control

This skill routes to the nerf-server MCP. All operations are handled by deterministic MCP tool calls — do NOT implement any logic in this skill file.

Routing

Parse the user's input and call the corresponding MCP tool:

Every routing entry MUST pass session_id. See "Session ID" below.

| User Input | MCP Tool | Arguments | |------------|----------|-----------| | /nerf | nerf_status | { "session_id": "" } | | /nerf status | nerf_status | { "session_id": "" } | | /nerf mode | nerf_mode | { "session_id": "" } | | /nerf mode | nerf_mode | { "mode": "", "session_id": "" } | | /nerf darts | nerf_darts | { "session_id": "" } | | /nerf darts | nerf_darts | { "soft": , "hard": , "ouch": , "session_id": "" } | | /nerf | nerf_budget | { "ouch": , "session_id": "" } | | /nerf scope | nerf_scope | { "session_id": "" } |

Parsing Rules

• Accept k suffix: 200k → 200000, 1.5m → 1500000
• A bare number (e.g., /nerf 200k) routes to nerf_budget, not nerf_darts
• Mode names are exact: not-too-rough, hurt-me-plenty, ultraviolence

Session ID (required for all calls)

You MUST pass the real Claude Code session ID on every nerf_* call. The MCP server has a multi-strategy resolver (env var → per-project transcript scan → md5 fallback), but the explicit override is the only path that's free of filesystem-state assumptions. Passing session_id keeps the skill working even if the server's heuristic resolver regresses or runs in an environment where transcript paths differ.

Find the session ID from the transcript path in conversation context. If not available, omit session_id entirely — do NOT attempt to derive it from shell commands or filesystem reads (history.jsonl and similar files are racy across concurrent CC sessions and would silently feed the wrong UUID). Falling through to the server's resolver chain is safe; injecting a wrong value is not.

Important

• Do NOT perform arithmetic — the MCP server handles all calculations
• Do NOT read or write config files — the MCP server manages config state
• Do NOT estimate context usage — the MCP server shells out to the analyzer
• Present the MCP tool's response as-is — it returns pre-formatted output

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Wave-Engineering
• Source: Wave-Engineering/claudecode-workflow
• License: MIT

Install and usage instructions live in the source repository linked above.