openclaw - 💡(How to fix) Fix Feature: tiered context compaction (tool-output-first compression) [2 comments, 2 participants]

Motivation

Large tool outputs from exec, web_fetch, and similar tools consume context disproportionately. A single web_fetch call can return 20–50KB of markdown, and exec outputs from build logs or database queries can be even larger. When context limits are reached, the current compaction strategy summarizes the entire conversation at once — losing conversational nuance, user intent, and decision history that is far more valuable per-token than raw tool output.

This creates a poor trade-off: the agent forgets why it was doing something in order to preserve what a tool returned verbatim.

Additionally, when compaction fails (e.g., due to API errors or rate limits), repeated automatic retries waste API calls without making progress.

Proposed Solution

1. Micro-compaction pre-pass

Add a lightweight compaction stage that runs before full context compaction triggers. This pre-pass would:

• Identify tool result blocks (exec output, web_fetch content, read file contents) above a configurable threshold (e.g., >2KB).
• Summarize each tool output individually, preserving key facts and discarding verbatim content.
• Replace the original tool result with the summary + a marker like [compacted: tool output summarized, original was ~18KB].
• Leave conversational turns (user messages, assistant reasoning) untouched.

This buys significant context headroom without sacrificing conversational coherence.

2. Circuit breaker for compaction failures

Add a circuit breaker that stops auto-compaction after N consecutive failures (default: 3). Behavior:

• After N failures, pause auto-compaction for a configurable cooldown period.
• Log a warning so the user/operator is aware.
• Resume attempts after the cooldown, or allow manual trigger.

This prevents runaway API waste when the summarization endpoint is down or rate-limited.

Technical Notes

• The micro-compaction pass would operate on the message history, targeting messages with role: "tool" or tool result content blocks.
• It could use a smaller/cheaper model for summarization (e.g., Haiku) since the task is straightforward extraction, not reasoning.
• The threshold for triggering micro-compaction vs. full compaction could be configured independently:
  • Micro-compaction at 70% context usage
  • Full compaction at 85% context usage
• Existing [compacted] markers in the codebase suggest partial infrastructure already exists.

Alternatives Considered