executing-plans
executing-plans
Use when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available
context-shuntOffload large / multi-file reads to a cheap worker model so raw files never enter Claude's context (token savings)
Independent GitHub snapshot · Automated classification · Original documentation governs
Reading big files into context is the most expensive thing an agent does for the
least reasoning value. The shunt hands those reads to a cheap worker model
(bulk-read) that answers a question about the files and returns a compact
summary. The raw bytes never enter this session's context.
This is orthogonal to whole-turn routing (srooter / route-task): those pick the
model for the turn; the shunt trims what a tool call pulls into context when
the turn is legitimately here.
bulk-read "<question>" file....get_code_snippet(qualified_name): free and exact.A shunt answer is for understanding, not for producing a diff.
bulk-read "how does token refresh work?" src/auth/session.ts src/auth/refresh.ts
bulk-read "which config keys are read at startup?" $(git ls-files 'config/*.yaml')
bulk-read prints structured bullets citing path:line, or
NOT FOUND IN PROVIDED FILES. A token-savings report goes to stderr.
Env vars or ~/.claude/shunt.conf (see templates/shunt.conf):
SHUNT — on/off master switch for the PreToolUse hook.SHUNT_MIN_LINES — large-read threshold (default 350).SHUNT_MODE — suggest (default) / block / off for the hook's large-read action.SHUNT_GRAPH_NUDGE — on/off once-per-session graph nudge.SHUNT_MODEL — worker command (default deepseek --flash; also gemini-api --flash-lite, qwen3, glm).The context-shunt-gate PreToolUse hook enforces the thresholds; this skill tells
you when to reach for bulk-read yourself.