Preflight
A tool that scans a project and produces one canonical manifest for AI coding assistants, inferred commands with confidence and risk, extracted agent rules, and consistency warnings.
Assistants burn tokens reconciling scattered repo signals.
README, package.json, Makefile, pyproject.toml, CI YAML, workspace layouts, and ad-hoc docs all describe a project differently. AI coding assistants waste tokens reconciling them, and still miss contradictions.
Merge every signal into one structured document.
Preflight scans a project and emits a single JSON or Markdown manifest: inferred commands with confidence, risk, context, and supporting evidence; structured warning objects for contradictions and coverage gaps; project graphs for monorepos; extracted agent rules (AGENTS.md, CLAUDE.md, Copilot instructions); entrypoints; and a short agent-bootstrap brief plus a JSON-schema contract for downstream tools.
Massive context compression, benchmarked publicly.
On a public warning-corpus benchmark, Preflight compressed orientation context dramatically, e.g. next.js from 40,912 tokens of raw signal down to a 770-token bootstrap brief (98.1% reduction), and ruff by 98.8%, turning sprawling repos into a single orientation document an assistant can read in one shot.