UNPKG

@blundergoat/goat-flow

Version:

AI coding agent harness and local dashboard for Claude Code, OpenAI Codex, Google Antigravity, and GitHub Copilot - setup audits, guardrails, structured skills, deny hooks, and persistent learning loops.

26 lines 1.52 kB
import type { AuditContext } from "./types.js"; /** * Decide whether drift should auto-run without --check-drift. * * Multi-agent projects leave satellite skill dirs (`.agents/skills/`, * `.claude/skills/`, etc.) stale after a single-agent migration completes. * The existing drift machinery detects `manifest.stale_names` orphans but * is off by default, so `audit --agent claude` on a project that also ships * AGENTS.md exits "pass" while the Codex / Antigravity skill dirs still hold * pre-v1.2 names. When more than one agent instruction file is present on * disk we run drift automatically. Evidence: n=4 migrations reviewed * 2026-04-20 all had stale satellite dirs surviving a "pass" audit. * * The signal is computed from the manifest-backed instruction paths rather * than `ctx.agents`, which has already been narrowed by `--agent` upstream. * Using the filtered list would hide the multi-agent signal exactly when it * matters - the single-agent-filter case is the one stale satellites exploit. * * Single-agent projects preserve the prior opt-in behaviour. * * @param ctx - audit context; only `ctx.fs` is consulted, to existence-test manifest instruction files * @returns true when more than one agent instruction file is present on disk, signalling a multi-agent * project where stale satellite skill dirs could otherwise survive a "pass" audit; false otherwise */ export declare function shouldAutoRunDrift(ctx: AuditContext): boolean; //# sourceMappingURL=audit-drift-policy.d.ts.map