through-line carries a large effort across many agent sessions without losing its
direction. It keeps the route durable: the destination, open choices, settled
results, and their downstream consequences.
The product is built around decisions and outcomes rather than a task backlog. A chartered route advances one meaningful unit at a time and derives its next frontier from recorded state, so it still makes sense to a fresh worker.
Install from the skills registry:
npx skills@latest add stuarth/through-line
Update with npx skills update through-line. To follow the repository directly,
clone it and symlink the skill directory:
git clone https://github.com/stuarth/through-line ~/dev/through-line ln -s ~/dev/through-line ~/.claude/skills/through-line
Invoke it explicitly as /through-line in Claude Code or $through-line in
Codex. Automatic invocation is disabled; see
agents/openai.yaml. When an issue tracker is configured,
the route can live there. Otherwise, the bundled local Markdown
tracker and
validator provide a self-contained option.
The charter fixes the route. It names the destination, the boundaries of the effort, the first end-to-end result worth reaching, the premises currently being relied on, and who may decide what. This is the boundary for autonomous progress: discoveries inside it can extend the map, while a proposed change to the charter comes back to the human.
Decision rights keep direction human-owned. The human retains control of the destination, protected meanings, external promises, and consequential or irreversible effects. The builder can make ordinary reversible choices that follow the charter and settled decisions. A route can reserve additional choices for the human when the work needs tighter control.
A unit records either a decision or an outcome. Each unit carries the question or result, the premises and acceptance evidence that matter, and its eventual resolution with provenance. The route has at most one claimed builder unit. Human choices remain independent units even when several ready questions share one concise prompt. Finishing or checkpointing builder work leaves a durable handoff before a fresh session takes another unit.
Every resolution propagates. When a premise, decision, or result changes, Through-line revisits every resolved dependent and records whether it still stands or reopens. Open work is then recomputed rather than manually curated. The visible frontier is simply the work that is open, unblocked, and unclaimed; only items that need human judgment form the human frontier.
Principles are optional leverage. If a commitment recurs, rules out a plausible choice, and survives counterexamples, the agent may propose it as a standing principle. It becomes active only when the human adopts it. Principles help future units inherit hard-won judgment, but a route does not need them, and contradictory evidence can put one back in question.
Repository execution is optional. A route may stop at durable decisions and outcomes, or include implementation. When code is in scope, each repository moves through one integration ref so composition is observable. Review covers the exact integrated ref at stable architecture or provider boundaries and before an external effect, publication, dependent human gate, or completion. It reports whether the stated claim is supported; it does not grant authority for an effect.
Unattended mode uses fresh workers. A supervisor sends one explicit unit packet to a fresh worker with no conversation history, waits for its terminal result, and then derives the next unit or route transition. The loop continues serially until the route reaches a human frontier, needs a charter change, completes, or reaches its first-result effort trigger without path evidence. Fresh workers make recorded state, not conversational memory, prove that the route is resumable.
This repository is the skill. SKILL.md introduces the model and routes into the focused guides:
- CHART.md — turn an idea into a chartered route
- ADVANCE.md — resolve a unit or complete a route transition
- SUPERVISE.md — continue through units unattended
- EXECUTION.md and IMPLEMENT.md — plan, integrate, and verify repository work
- REVIEW.md — test a claim against one immutable object
- PRINCIPLES-GUIDE.md — propose, adopt, and challenge standing principles
The tracker adapter defines storage; the skill defines the product model. That separation lets Through-line stay useful across hosts without coupling the route to another mandatory skill.