Summary

Gateway announce queue timeout (hardcoded at 60s) causes message memory loss during sub-agent announcement delivery. Messages successfully deliver to Telegram but are never written to the agent's session transcript, creating a split-brain where users see messages the agent has no memory of sending.

Environment

• OpenClaw Version: 2026.2.13
• Platform: macOS (local mode)
• Agent Configuration: 3 sub-agents (p-research, p-media, p-social)

Root Cause

When sub-agent announcements take longer than 60 seconds to process:

1. ✅ Messages successfully deliver to Telegram (partial success)
2. ❌ Transcript write-back fails (data loss)
3. Result: External delivery succeeds, internal history recording fails

Evidence

Gateway Error Log (~/.openclaw/logs/gateway.err.log):

2026-02-15T01:30:08.587Z announce queue drain failed for agent:main:telegram:group:-1003810792546:topic:1: Error: gateway timeout after 60000ms

Gateway Delivery Log (~/.openclaw/logs/gateway.log):

2026-02-15T10:30:13.055+09:00 Good news — the sub-agent system is working now...
2026-02-15T10:30:19.871+09:00 Great — p-media just finished testing...
2026-02-15T10:30:27.550+09:00 Excellent — p-research just finished...

Session Transcript (~/.openclaw/agents/main/sessions/*.jsonl):

• All 5 messages from 10:30 AM: MISSING
• User's replies quoting those messages: PRESENT
• Agent's later acknowledgment of the issue: PRESENT

Steps to Reproduce

1. Configure 3 sub-agents with announcement delivery to main agent
2. Spawn all 3 simultaneously:

   sessions_spawn({ task: "...", agentId: "p-research" });
   sessions_spawn({ task: "...", agentId: "p-research" });
   sessions_spawn({ task: "...", agentId: "p-media" });

3. Wait for all to complete (~10-15 seconds)
4. Check gateway.err.log for "announce queue drain failed" error
5. Verify session transcript missing announcement messages
6. Verify messages were delivered to Telegram/channel

Impact

Severity: CRITICAL

• Agent loses conversation continuity
• Creates contradictions (agent says different things than what it actually sent)
• Cannot reference previous delegated work
• May repeat already-completed tasks
• Breaks user trust (agent genuinely doesn't know what it told user)

Example scenario:

• Agent sends: "Research complete: Microsoft Maia 200, DeepSeek-OCR 2, NVIDIA Jetson T4000"
• User asks: "What were those AI products you found?"
• Agent replies: "I don't have any record of finding AI products"
• User is confused/frustrated because they received the original message

Configuration Attempted

Tried adding configurable timeout via gateway.config.patch:

{"gateway": {"announceTimeoutMs": 120000}}

Result: Rejected - "invalid config"

Conclusion: Gateway timeout is hardcoded in source, not user-configurable.

Recommended Fixes

Option 1: Increase Hardcoded Timeout (Quick Fix)

Change gateway announce timeout from 60s → 120s or higher in source code.

Pros: Simple, immediate relief
Cons: Still may fail with complex scenarios, doesn't address root issue

Option 2: Separate Delivery from Transcript Write (Recommended)

Decouple Telegram/channel delivery from session history write:

• If delivery succeeds but transcript fails → retry transcript write independently
• Don't let transcript write failure prevent successful delivery
• Add transcript write retry logic with exponential backoff

Pros: Proper fix, handles partial failures gracefully
Cons: Requires architectural change

Option 3: Make Timeout Configurable (Best Long-term)

Add gateway.announceTimeoutMs to config schema:

{
  "gateway": {
    "announceTimeoutMs": 120000
  }
}

Pros: Users can tune based on workload, no source code changes needed for adjustments
Cons: Doesn't fix root cause, just delays it

Option 4: Async Transcript Write

• Don't block announcement delivery on transcript write completion
• Queue transcript writes separately with independent retry logic
• Process transcript queue asynchronously

Pros: Eliminates blocking, improves reliability
Cons: Most complex implementation

Temporary Workarounds

Until fix is available:

1. Serialize sub-agent spawns - don't batch multiple simultaneously
2. Use sub-agents during active hours - user sees announcements in real-time
3. Monitor gateway logs - watch for "announce queue drain failed" errors
4. Verify delivery - ask user "Did you see the results?" after spawning

Additional Context

Lane wait warnings preceded the timeout, suggesting queue congestion:

2026-02-15T01:30:01.214Z [diagnostic] lane wait exceeded: 
  lane=session:agent:main:telegram:group:-1003810792546:topic:1 
  waitedMs=52602 queueAhead=0

This indicates the queue was under heavy load before the timeout occurred.

Related Files

Full bug report with detailed logs available at:

• Investigation log: memory/2026-02-15.md
• Comprehensive report: BUG_REPORT_message_memory_loss.md

Can provide upon request.