Summary

maintain() calls rewriteTranscriptEntries() which rewrites the JSONL session file via branch-and-reappend, changing the file size, mtime, and entry IDs. However, conversation_bootstrap_state is only updated during bootstrap(), not after maintenance rewrites.

On the next gateway restart, bootstrap sees the stale checkpoint:

• Fast path 1 (exact size+mtime) fails — file shrank by bytesFreed
• Fast path 2 (append-only) fails — requires size > storedSize, but file is smaller
• Falls through to reconcileSessionTail() full content-based reconcile
• On conversations with many identical (role, content) pairs (empty assistant messages, repeated tool outputs — common in agent workflows), the occurrence-counting anchor matches at the wrong position → thousands of duplicate messages imported

Reproduction

1. Active LCM conversation with repeated patterns (empty assistant messages, identical tool outputs)
2. maintain() runs transcript GC (called after bootstrap, turns, compaction)
3. rewriteTranscriptEntries() shrinks the JSONL — checkpoint now stale
4. Gateway restart → both fast paths fail → full reconcile → anchor mismatch → flood
5. Flood inflates context → doom loop (compactFullSweep unbounded while(true) loop causes 50-minute session lockout (doom loop) #268) → cascading failure

Impact

19+ flood events on a single conversation (17K messages) between March 29 – April 5, 2026. Worst case: 5,733 duplicate messages imported in one second, context_items exploded 41×, caused fleet-wide outage.

Fix

Two parts:

1. Primary: Update checkpoint after maintain() rewrite

After rewriteTranscriptEntries() returns changed: true, call upsertConversationBootstrapState() with new file size/mtime/hash. Ensures next bootstrap hits the fast path.

2. Defense-in-depth: Import cap in reconcileSessionTail()

If reconcile would import more than 20% of existing DB count (min 50), abort and log an error instead of importing. Prevents catastrophic floods even if the checkpoint fix somehow fails.

Related Issues

• compactFullSweep unbounded while(true) loop causes 50-minute session lockout (doom loop) #268 (doom loop) — floods trigger the doom loop
• reconcileSessionTail: exponential message accumulation when session contains many identical messages (empty assistant, NO_REPLY, generic tool outputs) #271 (reconcileSessionTail accumulation) — same root cause, was closed prematurely
• fix: stop conv replay pollution in lossless-claw #263 (replay pollution) — fixed afterTurn dedup but not bootstrap path

Environment

• OpenClaw 3.28, lossless-claw 0.6.1
• macOS 26.3.1, M2, 24GB