Summary

After upgrading to OpenClaw 2026.4.9 (0512059) on Linux, the assistant shows multi-channel degradation that matches user-visible symptoms across Discord, Telegram, and webchat:

• slash commands become noticeably lagged
• direct conversations degrade after a few turns and can effectively stop responding
• Telegram sessions repeatedly throw Cannot read properties of undefined (reading 'length')

This looks like at least two overlapping regressions:

1. Direct-session context / compaction instability
2. Discord interaction handling lag

Possibly a third recurring bug on Telegram:

3. lane task crash: Cannot read properties of undefined (reading 'length')

Environment

• OpenClaw: 2026.4.9 (0512059)
• Host: Linux 6.8.0-107-generic x64
• Node: v25.8.1
• Model in affected Discord DM: openai-codex/gpt-5.4
• Channels active: Discord, Telegram, webchat

Important exclusions

These were present in logs but are already known local false positives / intentional config and are not part of this report:

• memory-lancedb-pro invalid config: embedding: must have required property 'embedding'
• plugin disabled (disabled in config) but config is present
• Discord health monitor disabled intentionally as workaround for stale-socket restart loop

User-visible symptom

Observed live after update:

• slash commands are extremely lagged
• conversations end up dropping to no response after a few turns

Evidence

1) Discord direct session hits context overflow, then rotates/purges session state

2026-04-09T19:10:56.392-07:00 [agent] [context-overflow-diag] sessionKey=agent:<redacted>:discord:direct:<redacted> provider=openai-codex/gpt-5.4 source=assistantError messages=68 sessionFile=/home/<redacted>/.openclaw/agents/<redacted>/sessions/<redacted>.jsonl diagId=<redacted> compactionAttempts=0 observedTokens=unknown error=Context overflow: estimated context size exceeds safe threshold during tool loop.
2026-04-09T19:10:56.394-07:00 [agent] context overflow detected (attempt 1/3); attempting auto-compaction for openai-codex/gpt-5.4
2026-04-09T19:11:04.317-07:00 [agent] auto-compaction succeeded for openai-codex/gpt-5.4; retrying prompt
2026-04-09T19:11:06.575-07:00 [lcm] bootstrap: session file rotated for session <redacted>: "/home/<redacted>/.openclaw/agents/<redacted>/sessions/<redacted>.jsonl" → "/home/<redacted>/.openclaw/agents/<redacted>/sessions/<redacted>.jsonl"; resetting conversation <redacted>
2026-04-09T19:11:10.319-07:00 [lcm] bootstrap: purged 228 messages and all summaries for conversation <redacted>

This strongly matches the user-visible “few turns then no response” degradation in a direct chat.

Related upstream issue family:

• [Bug]: direct session can exceed configured contextWindow without timely compaction #50065 direct session can exceed configured contextWindow without timely compaction
• Auto-compaction not triggered during tool-use loops, leading to unrecoverable context overflow #24800 auto-compaction not triggered during tool-use loops

2) Discord interaction lane shows listener slowdown consistent with slash-command lag

2026-04-09T19:02:07.397-07:00 [EventQueue] Slow listener detected: InteractionEventListener took 1918ms for event INTERACTION_CREATE

This is weaker evidence than the direct-session overflow, but it does line up with the user-visible slash-command lag.

Related upstream issue family:

• [BUG] InteractionEventListener stalls up to 78s on Discord #57248 InteractionEventListener stalls up to 78s on Discord
• [Bug]: Discord main replies can appear late after result is already ready, causing timing confusion #55236 Discord main replies can appear late after result is already ready

3) Telegram direct lane is repeatedly crashing with reading 'length'

Representative lines from an extended journalctl slice:

2026-04-09T19:33:23.194-07:00 [diagnostic] lane task error: lane=session:agent:<redacted>:telegram:direct:<redacted> durationMs=10124 error="TypeError: Cannot read properties of undefined (reading 'length')"
2026-04-09T19:34:58.821-07:00 [diagnostic] lane task error: lane=session:agent:<redacted>:telegram:direct:<redacted> durationMs=51260 error="TypeError: Cannot read properties of undefined (reading 'length')"
2026-04-09T19:39:38.400-07:00 [diagnostic] lane task error: lane=session:agent:<redacted>:telegram:direct:<redacted> durationMs=6345 error="TypeError: Cannot read properties of undefined (reading 'length')"
2026-04-09T19:39:38.799-07:00 [telegram] sendMessage ok chat=<redacted> message=<redacted>
2026-04-09T19:39:59.634-07:00 [telegram] sendMessage ok chat=<redacted> message=<redacted>

This suggests Telegram is not fully dead, but the same direct lane is repeatedly throwing a length exception and partially recovering.

Potentially related:

• Main session prompt crash: Cannot read properties of undefined (reading 'length') in compaction token estimation #63612 Main session prompt crash: Cannot read properties of undefined (reading 'length') in compaction token estimation (closed as fixed, but this looks like a related or regressed shape-validation / compaction-estimation failure)
• [Bug]: Malformed tool result content block crashes context truncation in a session-permanent loop #34979 malformed content block crashes context truncation / estimateMessageChars

Why I think this matters

This is not just one noisy warning. On this host, 2026.4.9 shows a pattern of:

• direct-session instability in Discord
• real interaction lag for Discord slash-driven events
• repeated Telegram lane crashes with the same length error

The combined effect is exactly what the user reported: channel UX becomes laggy and then conversations stop feeling reliable.

Request

Please advise whether this should be split into:

1. a direct-session compaction / overflow regression for 2026.4.9
2. a Discord interaction lag regression
3. a Telegram lane-task reading 'length' regression

or whether maintainers think these all share one history / compaction normalization root cause.

If helpful, I can provide a sanitized wider journalctl slice, but the lines above are the shortest reproducer-quality evidence from this host.