intent-drift

Pre-spec steering layer for AI coding agents — capture product intent, invariants, and decision criteria before implementation, hand off to spec-driven workflows, observe intent drift (warn-only), and write learnings back to keep intent and code aligned.

Evidence-based engine for detecting intent drift in AI-assisted software development. Pluggable providers, weighted scoring, and explainable alignment reports.

Intent drift detection & taxonomy for LLMs — how multi-turn jailbreaks (gradual drift, persona anchoring, trust escalation) evade single-turn scanners, and how to detect them.

Claude Code plugin that reviews a PR/branch against what it was asked to build — plan/spec + the Claude session transcript + git history. Deterministic secrets/CVE/dep-hygiene pre-flight (PyPI+npm+OSV), 5 parallel review subagents, risk triage, custom rules & learnings. Local, free, stdlib-only — code never leaves your machine.

Local deterministic CLI for flagging drift between human intent, agent plans, and proposed actions.

Executable intent contracts for coding agents. Detect and block silent intent drift in Claude Code, Codex and other coding agents.