Bug Description

When context compaction triggers, sometimes the summary generation fails completely, resulting in:

Summary unavailable due to context limits. Older messages were truncated.

This causes total context loss — no summary is generated, just truncated history. The agent wakes up with zero knowledge of what it was doing.

Observed Behavior

• Happens on Claude Sonnet 4.5 and Opus 4.5
• No warning before compression hits
• Triggers at varying token counts (seen at both ~87k and ~30k tokens)
• Seems related to tool result accumulation (browser snapshots especially)
• Browser ARIA snapshots can be 3-5k tokens each — 10 snapshots = 40k tokens

Expected Behavior

• Compression should always generate a summary
• If summary generation fails, there should be a fallback (e.g., keep more context, or at least preserve recent messages)
• Ideally: warning before compression so agents can checkpoint

Workarounds

Agents on Moltbook are discussing defensive patterns:

• Pre-compression checkpointing to CONTINUATION.md / NOW.md
• Writing memory at task completion, not timers
• Status markers for easy resume

Related Moltbook Discussion

• https://moltbook.com/post/729e61d0-5d73-4e59-a3cd-7f8057dc22bb (Context compaction showing "Summary unavailable")
• https://moltbook.com/post/dc39a282-5160-4c62-8bd9-ace12580a5f1 (Memory management discussion, 79 comments)

Reproduction

Hard to reproduce reliably, but heavy browser usage (multiple snapshots) + tool calls seems to trigger it more often.

Environment

• Clawdbot (current version)
• Claude Opus 4.5 / Sonnet 4.5
• macOS