openclaw - 💡(How to fix) Fix compaction: reserveTokens default (15000) leaves insufficient headroom for structured summarization at typical session sizes [1 comments, 2 participants]

Summary

The default reserveTokens: 15000 threshold causes compaction to fire when only ~15k tokens remain in the context window. At this point, the summarization model itself is severely context-constrained and cannot produce a structured, task-preserving summary. Instead it falls back to a verbatim transcript tail — silently dropping all goal state, in-progress task queues, and approved work.

Observed behavior

Manual /compact triggered at 182,615 tokens with reserveTokens: 15000 (200k context window):

• Summarization model had ~15k tokens to produce a summary of 182k tokens of prior context
• Result: verbatim transcript replay stored as summary field — not an LLM-generated summary
• All structured state (goals, in-progress tasks, approved task queues) silently dropped
• Contrast: safeguard compaction that fired earlier at 123,845 tokens produced a fully structured summary (## Goal, ## Progress, ## Done, ## Pending)

The difference is not a code path difference between manual and safeguard compaction — it is purely a function of how much headroom the summarization model has when it runs.

Root cause

Compaction summarization is itself a model call. At 182k/200k fill, only 15k tokens remain for the summarization model to:

1. Read the prior conversation context (injected as input)
2. Generate a structured summary (output)

15k tokens is insufficient for both at any realistic session length. The summarizer is token-starved and cannot produce structured output. It effectively echoes back the tail of the transcript instead.

This means: the sessions most in need of a high-quality structured summary are exactly the sessions that receive the worst summary quality, because they only trigger compaction when they're already critically full.

User impact

• Every session that hits the reserveTokens: 15000 threshold (fires at ~92% fill on a 200k window) has been silently degrading compaction quality
• Approved task queues, in-flight investigations, and structured goal state are lost at exactly the moment context pressure is highest
• Users experience this as "the agent forgot" — with no visible signal that compaction quality was degraded
• This is not a new edge case: it is the default behavior for any user running a long or investigation-heavy session

Proposed fix

Raise the default reserveTokens threshold substantially. Based on testing:

• reserveTokens: 30000 (fires at ~85% fill / ~170k tokens on 200k window) — sufficient headroom for structured summarization at typical session lengths
• reserveTokens: 40000 (fires at ~80% fill / ~160k tokens) — more headroom, slightly more frequent compactions on heavy days

The structured safeguard-triggered compaction at 123k tokens confirms that ~77k tokens of remaining headroom produces a correct structured summary. The right default is somewhere between 15k (clearly insufficient) and 77k (clearly sufficient).

A minimum viable fix: reserveTokens: 25000–30000. This alone would have prevented the quality degradation observed.

Additional recommendation

Consider adding a warning or metric when the summarization result appears to be a verbatim transcript tail rather than a structured summary — e.g. if the summary contains raw tool results, JSON blobs, or turn-format markers (- User:, - Assistant:, - Tool result:). This would make the failure mode visible rather than silent.

Environment

• OpenClaw 2026.5.3-1 (2eae30e)
• Context window: 200k (venice/claude-sonnet-4-6)
• Default reserveTokens: 15000
• Session fill at manual compaction: 182,615 tokens (~91%)
• Session fill at safeguard compaction (good quality): 123,845 tokens (~62%)