OpenClaw Agent Harness

OpenClaw

OpenClaw is an AI agent harness built on top of pi-mono.md, extending its file read and compaction machinery with defense-in-depth context management.

File Read Management

• Inherits Pi’s 2,000 lines / 50KB hard cap
• Additional caps for bootstrap files (one-time context loaded at session start):
  • 12,000 chars per file
  • 60,000 chars total budget
  • Exceeding budget: 75% head / 25% tail split (you see the beginning and end, middle is cut)
• Tool results: 16,000 chars or 30% of context window, whichever is smaller
  • If the tail looks “important” (errors, JSON close braces, summary keywords): head+tail mode
  • Otherwise: head only
• Philosophy: defense in depth — Pi’s truncation first, then bootstrap caps, then tool result budgets

Session Compaction

• Trigger: history exceeds 50% of the context window (maxHistoryShare, default 0.5)
• History split into equal-mass token chunks; oldest chunk dropped
• Dropped content: staged multi-pass LLM summarization with a merge step
• Summary placement: synthetic message prepended to kept tail
• Tool-call safety: repairToolUseResultPairing fixes orphaned tool results after chunk dropping; splitMessagesByTokenShare avoids cutting inside a tool-call/result pair
• Pre-compaction flush: silent agentic turn lets the agent persist state to memory files before history disappears
• Second layer: non-destructive in-memory pruning of tool results (soft-trim, then hard-clear) on a 5-minute cache TTL — protects the persistent conversation while reclaiming context for the current request

Subagent Handling

• Fresh isolated sessions by default — no parent transcript
• Fork mode: copies parent’s transcript into child, but only for same-agent spawns
• Workspace context filtered to a minimal allowlist: AGENTS.md, TOOLS.md, SOUL.md

相关页面

harbor, letta-code