long-running-agent-workflow

Long-Running Agent (LRA) Workflow

Load this skill when you are about to work on a large project that will span multiple sessions / context windows and you need continuity, atomic handoffs, and recovery from broken states.

The skill gives you a tiny CLI (scripts/lra_cli.py) that maintains a .lra/ directory: a machine-readable feature-list.json (atomic features with acceptance criteria and status) and a human/agent-readable progress.txt (session log). The protocol turns "one big vague task" into a sequence of small, fully-tested, check-pointed features.

---

Overview — the problem

AI agents working across many context windows hit three failure modes:

• Context amnesia — each new session has no memory of prior work.
• One-shot tendency — trying to do too much at once, leaving half-done features.
• Incomplete features — work spans sessions with no clear acceptance gate, so "done" is never verified.

LRA fixes this with: structured init, one atomic feature per session, an explicit test gate before done, and a checkpoint after every feature so the next session can recover.

---

When to use

• Long, multi-session projects (hours/days, many context windows).
• Any task where you might lose context between runs.
• Triggers: lra, checkpoint, feature list, long running, продолжи работу над проектом, долгая сессия, план фич, статус проекта.

If the task is small and finishes in one session, you do not need this skill — just do the work.

---

Prerequisites

• A git repository (so checkpoints can be committed and recovered).
• Python 3 on PATH (the CLI is pure stdlib, no dependencies).
• Run the CLI from the project root (it creates/reads .lra/ there).

---

Instructions

Phase 1 — Init

python3 scripts/lra_cli.py init "Short project description"

Creates .lra/feature-list.json ({"project": ..., "features": [], "created": <date>}) and .lra/progress.txt with a header. Refuses (exit 1) if .lra/ already exists, so you never clobber an in-progress project.