gaslighter

name	gaslighter
description	Requirement completeness checker — nudges you to re-read the original request before finishing. Routes to eval and judge sub-skills.
when_to_use	Use when user says 'gaslighter', 'eval', 'evals', 'run evals', 'benchmark', 'judge', 'score', 'rate', or invokes /gaslighter directly.
model	haiku
allowed-tools	["Bash(CLAUDE_PLUGIN_DATA=\"${CLAUDE_PLUGIN_DATA}\" node \"${CLAUDE_PLUGIN_ROOT}/hooks/gaslighter-config-cli.js\" --get)","Skill(gaslighter:eval)","Skill(gaslighter:judge)","Skill(gaslighter:config)","AskUserQuestion()"]

Gaslighter — Requirement Completeness Guard

A Stop hook that asks you to verify you haven't missed anything before completing code changes. When nudged, genuinely re-read the original request — don't assume you're done.

Config Check

Current persisted config: !CLAUDE_PLUGIN_DATA="${CLAUDE_PLUGIN_DATA}" node "${CLAUDE_PLUGIN_ROOT}/hooks/gaslighter-config-cli.js" --get

Routing

Before routing to eval/judge (including the AskUserQuestion fallback for "Run evals" / "Judge run"): check the config JSON injected above. If it is {} or missing the mode key, ignore ARGUMENTS and route straight to Skill({ skill: "gaslighter:config" }) instead — the plugin isn't configured yet.

This gate does not apply when ARGUMENTS is config/configure/setup or help — config is already the destination, or isn't needed.

Route based on ARGUMENTS:

• eval, evals, benchmark → Skill({ skill: "gaslighter:eval", args: "{remaining args}" })
• judge, score, rate → Skill({ skill: "gaslighter:judge", args: "{remaining args}" })
• config, configure, setup → Skill({ skill: "gaslighter:config" })
• help → show the Modes and How It Works sections below
• anything else or empty → ask the user:

AskUserQuestion({
  questions: [{
    question: "What would you like to do?",
    header: "Action",
    multiSelect: false,
    options: [
      { label: "Run evals", description: "Run the gaslighter benchmark suite" },
      { label: "Judge run", description: "Score a completed eval run for completeness and overcorrection" },
      { label: "Configure", description: "Set mode and nudge cap, persisted to config.json" },
      { label: "Help", description: "Show gaslighter usage and configuration" }
    ]
  }]
})

Then route based on selection (Configure and Help bypass the config gate; Run evals / Judge run are subject to it):

• Run evals → Skill({ skill: "gaslighter:eval" })
• Judge run → Skill({ skill: "gaslighter:judge" })
• Configure → Skill({ skill: "gaslighter:config" })
• Help → show sections below

Modes

Mode and nudge cap are persisted in ${CLAUDE_PLUGIN_DATA}/config.json — set them via Skill({ skill: "gaslighter:config" }) (or /gaslighter:config) instead of re-exporting an env var every session.

Precedence (mode and nudge cap independently): env var > persisted config.json > mode default.

• lite (default) — hook on, soft nudge via additionalContext, up to 3 nudges/session by default
• full — hook on, hard block via decision: "block", unlimited nudges by default
• off — hook disabled, 0 nudges, always

Override per-session without touching the persisted config:

• GASLIGHTER_MODE — off / lite / full
• GASLIGHTER_MAX_NUDGES — a positive integer, or infinite/unlimited/-1 for no cap

How It Works

Each time you try to finish a response, the hook fires and asks you to re-check the original request. It fires at most once per stop, up to the mode's nudge cap (see Modes above). First nudge forces re-examination; subsequent nudges give you an escape hatch if you're genuinely done.