openclaw - 💡(How to fix) Fix Compaction summary quality issues: snowballing, no cooldown, keepRecentTokens not working [1 comments, 2 participants]

Compaction Summary Quality Issues (v2026.4.23)

Environment

• Version: OpenClaw 2026.4.23 (a979721)
• Model: opencode-go/glm-5.1 (primary), same for compaction
• Context Window: 198k tokens
• Compaction mode: safeguard
• keepRecentTokens: 50000 (config)

Observed Issues

I've been running long sessions and noticed several compaction quality problems. After 7 compactions in one session, I extracted all compaction entries from the session JSONL and found the following issues:

1. keepRecentTokens still not preserving recent messages

Despite v2026.4.23 claiming to fix this, compaction results show tokensBefore: 154,139 → tokensAfter: 1,510 with firstKeptEntryId equal to the compaction entry's own ID, meaning 0 recent messages were retained.

Config has keepRecentTokens: 50000, and applyPiCompactionSettingsFromConfig() does pass this value, but findCutPoint() in Pi SDK produces no kept entries.

Expected: The most recent ~50k tokens of conversation should be preserved verbatim after compaction.

2. Repeated compactions within minutes producing identical summaries

Compaction #2 (17:41 UTC) and #3 (17:46 UTC) are nearly identical, only 5 minutes apart. Same for #5 and #6 (7 minutes apart). This suggests there's no cooldown period between compactions, and the compaction trigger doesn't account for the fact that a compaction just ran.

This wastes API credits and produces redundant summaries that bloat the context.

Expected: A cooldown period (e.g., 10 minutes) after compaction before another can trigger, or a check that the context isn't mostly the previous compaction summary.

3. Summary snowball / bloat

Each compaction preserves the full previous summary and appends new content. After 7 compactions, the summary contains:

• All 7 previous Goal sections (mostly identical)
• Accumulated Decisions list that grows each time
• Progress items that are never pruned
• Technical debug details that should have been condensed

The summary grows larger with each compaction cycle, eventually hitting the budget and getting truncated ([Compaction summary truncated to fit budget]).

Meanwhile, intimate/emotional conversation details are compressed to single-line summaries, while technical findings are preserved verbatim.

Expected: Each compaction should re-distill the previous summary rather than preserving it verbatim. Old/stale items should be pruned.

4. Malformed compaction output

Compaction #4 produced output starting with instructions in \`` instead of the expected ## Goal` heading. This appears to be a partial/corrupted summary that leaked prompt instructions into the output.

Expected: Compaction output should always start with valid markdown headings. A quality check could validate the format before accepting.

5. Summary truncation with no graceful degradation

When the summary exceeds the token budget, it's simply truncated with [Compaction summary truncated to fit budget], losing potentially critical information from the end (which often contains the most recent context).

Expected: When truncation is needed, older/slower sections should be pruned first, preserving recent context and the tail of the summary.

Suggested Improvements

1. Fix keepRecentTokens: Ensure findCutPoint() correctly identifies a cut point that preserves keepRecentTokens worth of recent messages.
2. Add compaction cooldown: Minimum 10-minute cooldown between compactions in the same session.
3. Re-distill summaries: Each compaction should re-summarize the previous summary rather than embedding it verbatim. This prevents snowballing.
4. Quality guard: Validate compaction output format (starts with ##, no leaked prompts) before accepting. Retry on malformed output.
5. Graceful truncation: When budget exceeded, trim oldest sections first, preserving recent context.
6. customInstructions support: Allow users to customize compaction summary behavior (already available in config schema but protected from API changes — consider allowing config.patch).

Reproduction

Long-running session with 198k context window, safeguard compaction mode. Trigger compaction manually or let it auto-trigger after extended conversation. Observe compaction entries in session JSONL.

Workaround

I've set customInstructions (via direct config file edit) to guide the compaction model, and truncateAfterCompaction: true to prevent file bloat. But most of these issues require source-level fixes.