Why a Harness

A scaffold gives you a correct starting point. The hard problem is keeping the architecture correct through hundreds of edits — especially when an AI agent is making many of them. stackr's answer is the context harness: the set of files, rules, and hooks that deliver the project's conventions to your editor and your agents, and mechanically enforce the parts that can be checked.

The problem it solves

Conventions written in a README are advisory. The moment work starts, drift creeps in: business logic slips into a route handler, a second service grows a user table, error handling diverges, a mobile component hardcodes a color. A human reviewer might catch some of it; an AI agent generating code at speed will reproduce whatever it isn't actively told not to.

The harness exists to make the conventions present at the moment of action — and to fail loudly when the mechanically-checkable ones are violated.

Push, not pull (and no MCP server)

stackr delivers context by writing it into the repository as files your tools already read — nested AGENTS.md, editor rule files, Claude skills. The agent encounters the right rule because it's editing a matching file, not because it remembered to ask a server.

That's a deliberate choice over an MCP server the agent must query:

• No runtime dependency. The rules are plain text on disk — always available, no process to run, no port to manage.
• Nothing to forget. Push delivery doesn't rely on the agent deciding to fetch context; the editor surfaces it by glob match.
• Always in sync. Every artifact is regenerated from one source whenever the project shape changes, so the rules can't go stale against the code.

One source of truth

Every agent-facing artifact is rendered from a single table — CONTEXT_RULES in src/config/context-map.ts — by one generator. The nested AGENTS.md, the CLAUDE.md bridge, the Cursor and Windsurf glob rules, the Claude skills, and the ast-grep rules all derive from the same rule definitions. They cannot disagree about which rules apply or what they say. And because both create-stackr and stackr add service drive that one generator, the artifacts never drift from the live set of services.

Salience vs. enforcement (an honest split)

The harness has two distinct jobs, and stackr is careful not to overstate the first:

Salience makes the agent more likely to get it right; enforcement makes a checkable mistake fail. Codegen is a third lever — stackr add entity sidesteps both by generating correct-by-construction code.

It's measured, not asserted

stackr ships an evaluation suite (eval/) that runs coding tasks under three conditions — context off, salience (rules present), and enforcement (rules

• reinject hook) — and scores the agent's diff with the project's own ast-grep rules. The two headline deltas, Δ off→salience and Δ salience→enforcement, separate "does push delivery help" from "does the check-and-reinject loop add reliability on top." The point isn't a marketing number — it's that the harness's value is treated as something to measure and improve, not a claim.

The pieces