shed
bởi compass-soul
Coding Agents & IDEsTóm tắt
Context window hygiene for long-running LLM agents.
Mô tả
Kỹ năng này làm gì
Decision rules for managing context in long-running LLM agents. Covers when to mask tool outputs, trigger condensation, spawn sub-agents, or switch context as the window fills. Backed by NeurIPS 2025 and ACL 2024 research on context scaling and positional attention bias.
Simple observation masking halves cost and matches LLM summarization on solve rate, making it more effective than investing in complex compression pipelines.
Khi nào nên dùng
- Cutting API costs during multi-hour coding agent runs
- Preventing context overflow on large file analysis tasks
- Designing multi-agent systems with clean context handoffs
- Structuring agent memory for extended autonomous research
- Diagnosing degraded agent behavior in long sessions
Quy trình ví dụ
User asks: An agent just returned 800 lines of grep output and context is at 65%. What should I do?
- 1
Extract the relevant file paths and match counts from the grep output into a short summary
- 2
Write that summary to a file for later reference
- 3
Discard the raw tool output from context
- 4
Check if context is still above 70% after masking
- 5
If yes, mask the oldest reasoning turns from earlier in the session
Context usage reduced, key facts preserved in a file, agent continues without compaction or loss of decision chain
Cài đặt
clawhub install compass-soul/shedChạy lệnh CLI, hoặc dán prompt vào Claude Code, Cursor, Codex hay coding agent bạn đang dùng.