AGENTS.md.Чёткая инструкция заставляет модель вести постоянный журнал контекста, чтобы хранить в одном месте все задачи, не терять цель и непрерывно работать без сбоев:
## Continuity Ledger (compaction-safe)
Maintain a single Continuity Ledger for this workspace in CONTINUITY.md. The ledger is the canonical session briefing designed to survive context compaction; do not rely on earlier chat text unless it’s reflected in the ledger.
### How it works
- At the start of every assistant turn: read CONTINUITY.md, update it to reflect the latest goal/constraints/decisions/state, then proceed with the work.
- Update CONTINUITY.md again whenever any of these change: goal, constraints/assumptions, key decisions, progress state (Done/Now/Next), or important tool outcomes.
- Keep it short and stable: facts only, no transcripts. Prefer bullets. Mark uncertainty as UNCONFIRMED (never guess).
- If you notice missing recall or a compaction/summary event: refresh/rebuild the ledger from visible context, mark gapsфайла AGENTS.md ask up to 1–3 targeted questions, then continue.
### functions.update_plan vs the Ledger
- functions.update_plan is for short-term execution scaffolding while you work (a small 3–7 step plan with pending/in_progress/completed).
- CONTINUITY.md is for long-running continuity across compaction (the “what/why/current state”), not a step-by-step task list.
- Keep them consistent: when the plan or state changes, update the ledger at the intent/progress level (not every micro-step).
### In replies
- Begin with a brief “Ledger Snapshot” (Goal + Now/Next + Open Questions). Print the full ledger only when it materially changes or when the user asks.
### CONTINUITY.md format (keep headings)
- Goal (incl. success criteria):
- Constraints/Assumptions:
- Key decisions:
- State:
- Done:
- Now:
- Next:
- Open questions (UNCONFIRMED if needed):
- Working set (files/ids/commands):
Сохраняйте, чтобы не потерять!
@notboring_tech