I tried SKILL.state in my coding agents and stopped after four days

Описание и применение

Автор протестировал SKILL.state в своих кодирующих агентах, но после четырёх задач прекратил эксперименты, поскольку обновление модели и проблемы с лентой кода сделали хранение состояния избыточным.

Задача

Does an explicit state file improve long‑task coding agents or add unnecessary overhead?

Результат

The SKILL.state approach improved brief clarity but introduced schema failures, invented timestamps, and linter issues; after moving to Opus 5.5 the context problem diminished, making the state file overhead unjustified.

Рабочий процесс

• 1. Create a JSON state file per task containing objective, phase, decisions, findings, and next action.
• 2. Update the state only through a CLI that validates patches against a schema.
• 3. Launch a fresh subagent for each step, providing it the skill, current state, and latest observation, with no chat history.

Доказательства

SKILL.state swaps an agent's growing chat history for a state file and a fresh step each time.

Источник

The briefs were good. Each subagent got a small, self-contained frame and needed no history.

Источник

Patches that failed the schema.

Источник

Invented timestamps.

Источник

After that, the median peak context was 162K tokens, and only 2 of 48 units of work were compacted at all.

Источник