groundtruth-rules-ai

Groundtruth — AI rule extraction (opt-in; not on by default)

Groundtruth's default rule compiler is deterministic — it only pulls the literal backtick forms (`X` not `Y`, never `X`) from your docs, with no model. This command adds an opt-in model pass: fan out an agent to read your rule docs in prose and propose the deterministic rules the literal extractor missed — then route them through the EXACT SAME safety pipeline (grounding + the /groundtruth-rules human approval gate). Nothing here arms a rule, and the model never runs unless you invoke this command.

Do this

1. Fan out the reading. Use the Task tool to spawn subagent(s) — the Explore agent is ideal — to read this repo's rule docs: CLAUDE.md, AGENTS.md, SCHEMA.md, ARCHITECTURE.md, docs/*.md, every **/.claude/skills/**/SKILL.md, every **/.claude/agents/*.md, .cursorrules, .windsurfrules (whichever exist). If there are many docs, fan out one subagent per doc group and run them in parallel. If $ARGUMENTS names a path, focus there. Ask each subagent to return candidate rules only in this exact shape:

   { "id": "no-<slug>", "kind": "forbid_in_added", "file_re": "<regex>", "line_re": "<regex>", "message": "<the rule, one line>", "positive_example": "<a code line the rule SHOULD flag>" }
   

   • file_re — which files the rule applies to (\\.sql$, \\.(ts|tsx)$, …); a broad source set if unsure.
   • line_re — the forbidden token/pattern; anchor identifiers with \\b. JS RegExp syntax only — no PCRE/Python inline flags like (?i) (matching is already case-insensitive); a pattern that doesn't compile in JS is inert.
   • positive_example — a minimal line the rule is meant to catch. It PROVES the rule can fire: the grounder rejects (routes to review, never armable) any candidate whose line_re doesn't match its own example. Grounding otherwise only proves a rule doesn't hit existing code — never that it fires at all.