Long Horizon

long-horizon

One goal, two human gates, one long run. The human and the agent iterate on the spec until it is signed; then the run pushes the whole ledger autonomously for hours inside an authorization envelope; then the human reviews a report and adjusts. The orchestrating context dispatches and verifies — it does not execute. Durable state lives in .long-horizon/<goal-slug>/, and three value gates hold throughout: admission (no verifiable Done-when, no entry), evidence (progress exists only when verified), exit (stopping is designed, not a failure).

Language: mirror the user in chat; workspace files follow the project's documentation language.

Load-on-use: every /name in this document is an invocation — before doing that work, load that skill's full document (the Skill tool, or read its SKILL.md); never act from its one-line description.

Workspace — .long-horizon/<goal-slug>/

• One directory per goal, slug-named; one goal active at a time. On any resume, identity-check first: read goal.md — if it describes a different goal than yours, stop and ask.
• goal.md — frozen spec: goal, non-goals, invariants, Done-when with a runnable check per clause, and the authorization envelope. Changing it is a human-gated re-spec, never a side effect.
• ledger.json — acceptance items {id, desc, verify, passes, evidence, deps, discovered-from}. Every item is born passes: false.
• log.md — append-only: decisions, dead ends, dispatch notes. Failed attempts stay in — they are evidence.
• next.md — resume pointer: what is in flight, what is ready next.
• report.html — live oversight view; contract below.
• The workspace stays out of git: gitignore it (and add formatter ignores) at admission — a tracked workspace pollutes every work-branch diff. Durability comes from code commits, graduation, and the report. On graduation, distill through /to-ctx, then archive or delete the directory.

Gate 1 — Spec (interactive; iterate as long as it takes)

1. Run the goal through /ctx-grill. Iterate freely with the user until the spec holds:
   • every Done-when clause names a runnable check — no check, no entry;
   • non-goals and invariants written;
   • the authorization envelope — what the run may do unattended. Default: committing to work branches, opening PRs, pushing its own branches — allowed; merging, deleting outside the workspace, sending anything external, spending — forbidden. Tighten or loosen only here, explicitly.
   • the verification setup — ask once: which checks the orchestrator reruns itself, which acceptance items get a black-box verifier agent, and the verifier model tier (default: a mid-tier model);
   • the initial ledger, with dependencies.
2. Refuse goals that fit one sitting — do them normally instead.
3. On sign-off: write the workspace and start the run.

The run — autonomous, hours

• Orchestrate, don't execute. Load /fan-out and dispatch ledger items as task contracts to subagents; keep only the spec, the ledger, evidence summaries, and decisions in your own context — workers absorb the exploration noise. This is structural, not optional: a whole ledger's execution does not fit one context window.
• Schedule off the DAG: dispatch the ready set (all deps passed); independent items in parallel — worktree isolation when they mutate the same repo; coupled clusters go to one worker.
• Verify before flipping: worker claims are never evidence. Mechanical checks (tests, lint, build) — rerun them yourself. Behavioral acceptance — a black-box verifier subagent (the spec'd tier) that sees only the artifact and the acceptance criteria.
• A failure buys one rewrite, never a repeat: when a dispatched item fails — the worker errors out, or verification rejects its claim — redispatch it once with the contract rewritten around the failure evidence: what was attempted, which check failed, how it failed. Change the angle or change the worker; never resend the same contract verbatim. One tactical retry per item per dispatch round; if the retry fails too, the item counts as unmoved for stagnation accounting and the failure goes into log.md — it does not earn a third attempt in that round.
• Every flip refreshes the report's data block — the report is the human's live dashboard while the run is in flight.
• Discovery is expected: new problems become new ledger items (discovered-from) and join the schedule. A question that would change the spec's direction becomes a blocked item with the question written down — and the run continues on every branch that does not depend on it.
• If the harness has a native goal mechanism, feed it: "all ledger items pass or blocked, or stop after N hours".
• Stop when: all items pass; everything remaining hangs on blocked; two dispatch rounds, tactical retries included, move nothing (stagnation — never a third identical attempt); or budget runs out. Always end with the final report and a /handoff.

Gate 2 — Review

The human reads the report and comments. Adjustments become spec amendments or ledger edits — the only path by which acceptance rows may be reworded or deleted. The next run resumes from the workspace, identity check first.

Interruption is survivable, not the rhythm

Assume any moment may be the last: context death, crash, budget cut. A fresh orchestrator reads the workspace, identity-checks goal.md, and resumes from ledger.json + next.md. That is what the files are for — not for handing the goal back to the human every few items.

Report — a contract, not a template

report.html is model-authored: design it however communicates best, and redesign it freely as models improve — it is disposable and regenerable from state. Non-negotiable:

• A view, never a store: generation is one-way; everything shown must already exist in the state files. Must show: the goal and its Done-when; every ledger item with status; the dependency structure drawn as a graph — critical path and blocking cut-points visible at a glance, not only a list; blocked items and their questions, prominently; key decisions; last-updated and session count.
• Self-contained: a single file that opens from file:// and makes no external requests.
• Data apart from rendering: one embedded data block; routine refreshes touch only that block, not the markup.

Rules

• The envelope is hard in every phase: nothing irreversible, nothing outside it.
• Ledger rows are write-protected for the agent: flip passes, append evidence, add new items; rewording or deleting rows happens only in review, by the human.
• passes flips on verification, not on code written.
• A chat correction that contradicts a workspace file is written back in the same turn — otherwise the stale file wins.
• Dependency direction: may use /ctx-grill, /fan-out, /handoff, and /to-ctx; long-horizon sits at the top of the graph and nothing invokes it.