Two layers: judgment and enforcement

deadreckon is split into two halves that are kept strictly apart. One half is judgment: the soft, changeable decisions about how to get good work out of a model. The other half is enforcement: the hard guarantees that must never bend, like "a run cannot be marked done without a valid signature."

Judgment lives in a skill. A skill is just a Markdown file (skills/<name>/SKILL.md) written in plain English. Think of it as the agent's playbook: it says what to ask the model for, which tools to prefer, and when to call the work finished. Because it is only text, you can open it in any editor, change how the agent behaves, and rerun, with nothing to recompile.

How a text file steers a run

The trick is simple. On every turn, deadreckon reads your skill straight off disk and pastes it into the prompt it sends the model, right beside your goal and a summary of the run so far. The skill is not configuration the binary interprets; it is text the model reads.

That makes the skill influence, not control. It can tell the model to prefer one tool, keep turns small, or document its decisions, and a capable model will follow it. What it cannot do is widen the sandbox, skip a lock, or sign its own gate, because none of that code reads the skill. The boundaries live in the binary and hold no matter what the skill says.

Enforcement lives in the Rust binary, the compiled program. It owns the locks, the atomic writes, the signed gate, and the sandbox: the things that have to be correct every single time. Inside, it is layered: a thin command-line layer on top, a runtime that drives the loop in the middle, and a core of durable building blocks at the base.