Summary

Fixes #26458 — After a successful safeguard compaction, a second compaction fires immediately due to stale usage.totalTokens from kept assistant messages, destroying all preserved conversation history.

Root Cause

Pi-coding-agent's prompt() method checks _findLastAssistantMessage() before each prompt and calls _checkCompaction() with that message's usage.totalTokens. After compaction, kept assistant messages still carry their pre-compaction usage values (e.g., 184K tokens), which makes shouldCompact() return true even though the session is already compacted to ~1-5K actual tokens. This triggers a second compaction that finds 0 real messages to keep.

Evidence

• 22 double compaction events across 8 sessions on a real installation
• 100% reproduction rate: every compaction is followed by a spurious re-compaction
• Every double compaction preserves 0 conversation messages (complete amnesia)
• All 22 cases confirmed: the stale totalTokens from kept assistants exceeds the shouldCompact threshold

Changes

1. compaction-safeguard.ts — Cancel compaction with no real messages

Added a guard at the top of session_before_compact that cancels compaction when messagesToSummarize has no real conversation messages (user, assistant, or toolResult). In the double compaction scenario, only custom/compaction entries remain after the first compaction, so this guard prevents the destructive second compaction.

2. attempt.ts — Skip cache-ttl insertion after compaction

When compaction occurs during an attempt, the openclaw.cache-ttl custom entry is no longer appended. Previously, this entry was the only non-compaction entry in the session, which bypassed prepareCompaction()'s existing guard (checking if the last entry is a compaction type).

Tests

Added 3 test cases to compaction-safeguard.test.ts:

• Cancels compaction when messagesToSummarize is empty
• Cancels compaction when messagesToSummarize has only custom entries
• Proceeds normally when real conversation messages exist

Note on upstream fix

The root cause is in pi-coding-agent's prompt() → _checkCompaction(), which should either skip the check when the last assistant predates the latest compaction, or use estimateContextTokens() on current messages instead of stale usage.totalTokens. The fixes in this PR are defense-in-depth on the OpenClaw side to prevent the destructive outcome regardless of the upstream behavior.

Greptile Summary

Fixed double compaction bug that destroyed preserved conversation history after successful compaction. The issue occurred when stale usage.totalTokens from kept assistant messages triggered immediate re-compaction, with 22 confirmed incidents across 8 sessions showing 100% reproduction rate.

Two defensive fixes were implemented:

• Added early-return guard in compaction-safeguard.ts that cancels compaction when no real conversation messages exist to summarize
• Modified attempt.ts to skip cache-ttl custom entry insertion after compaction, preventing bypass of existing double-compaction guards

Test coverage includes three new test cases validating the guard behavior for empty messages, custom-only entries, and normal conversation flow.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The fix addresses a critical bug with a well-documented root cause (22 real incidents with 100% reproduction). Both changes are defensive guards that prevent destructive behavior without altering normal compaction flow. Test coverage validates all edge cases, and the code includes clear comments explaining the issue and fix.
• No files require special attention

_{Last reviewed commit: 25c7355}

_{(3/5) Reply to the agent's comments like "Can you suggest a fix for this @greptileai?" or ask follow-up questions!}