Trackline — AI skill for Claude Code

trackline

**An alignment layer for AI agents.**

Trackline watches whether an AI agent's actions still match the task, the rules and the evidence it was given. It works across two surfaces:

• Locally, it runs beside coding agents like Claude Code, Codex and Cursor, catching drift while the work is still happening.
• In production, it ingests agent traces, scores alignment, and flags runs where the agent appears to have gone off task.

The same core engine powers both: **normalise what the agent did, compare it against intent, produce an evidence-backed verdict.**

An agent that goes off task does not crash. It edits files nobody mentioned, ignores the rules file it read an hour ago, and the build stays green. Tests check that code does what it was written to do; they have no opinion on whether it is the code you asked for. Trackline is the layer that notices.

---

Build status

Trackline is being built in phases. This section is the only part of this README that changes as they land.

Component	Status
Eval gate (CI regression gate)	✅ Working. Documented below.
Core alignment engine	Foundation phase
Local surface (coding agents)	Not started
Production surface (trace ingest)	Not started

The eval gate is the first working component and will fold into the broader alignment engine as its CI-side entry point. It is real, tested and gating merges today; everything else is honest roadmap.

Design notes and the reasoning behind the phases: [yerinsabraham.com/engineering/nothing-notices-when-an-agent-drifts](https://yerinsabraham.com/engineering/nothing-notices-when-an-agent-drifts).

**The measurements behind the design are in [`docs/experiments/`](docs/experiments/),** with the code and raw data that produced them:

	Question	Answer
1	When a hook blocks an agent and explains why, does the agent correct itself?	**Yes, 11