🐛 LLM timeout (60s) happens silently — no visible log in CMD

[jlin53882]

Problem Summary

When smartExtraction LLM calls fail (timeout, rate limit, network error, etc.), the failure is completely invisible to the user. The agent simply hangs for ~60 seconds with no CMD output, making it impossible to diagnose.

What User Experiences

User sends a message → agent "thinks" for ~60 seconds → then responds normally (failure was swallowed)

No warning, no error, nothing in the CMD window.

---

Root Cause Analysis

Three-layer silent failure chain:

1. OpenAI SDK retry (src/llm-client.ts:63)

• OpenAI SDK v6.21.0 has maxRetries: 2 (default)
• timeout: 30000 ms per request
• On any LLM failure (timeout, 429, 500, network error): retries 2 more times
• Total wait: 30s × 3 = ~60–90 seconds

2. Error caught and downgraded (src/llm-client.ts:113-118)

} catch (err) {
  // Graceful degradation — return null so caller can fall back
  log(`memory-lancedb-pro: llm-client [${label}] request failed...`);
  return null;
}

The log function here is passed from index.ts and maps to api.logger.debug — which is completely silent in default CMD output.

3. No visible indication to user

• The agent is blocked waiting for the LLM call to finish (with retries)
• CMD shows nothing because all relevant logs are debug level
• User thinks the agent is "thinking" rather than "retrying a failed LLM call"

---

Affected Code Locations

File	Line	Issue
src/llm-client.ts	116	log() for request failures → debug level (silent)
src/llm-client.ts	62	timeout: 30000 per request, maxRetries: 2 (OpenAI SDK default)
index.ts	1663	log: (msg) => api.logger.debug(msg) — log function wired to debug
index.ts	1679–1681	SmartExtractor initialized with log: api.logger.info, debugLog: api.logger.debug

---

Proposed Fix

1. High priority: Upgrade LLM failure logs to warn or error

• In src/llm-client.ts:116, change log(...) → console.error(...) or wire to a warn-level logger
• LLM request failures are user-facing runtime failures, not debug diagnostics

2. Consider exposing maxRetries and timeout in config

• Allow users to configure retry behavior (e.g., disable retries or reduce timeout)
• This would help in air-gapped or rate-limited environments

3. Consider adding a smartExtraction.logLevel config option

• Users who don't want smart extraction verbose logs can silence them
• Users debugging issues can raise the level

---

Environment

• Plugin: memory-lancedb-pro@1.1.0-beta.9
• OpenAI SDK: ^6.21.0
• LLM config: smartExtraction.llm.model = models/gemini-3.1-flash-lite-preview
• Symptom observed: 60s hang per message when Gemini rate limited or slow

---

Reported by: james53882
Date: 2026-03-23

[AliceLJY]

Hi @jlin53882, thanks for the detailed root cause analysis — you nailed the three-layer silent failure chain.

Good news: this has been addressed in v1.1.0-beta.10:

1. Auto-recall timeout is now configurable via autoRecallTimeoutMs (default still reasonably short, but you can tune it):

   { "plugins": { "entries": { "memory-lancedb-pro": { "config": { "autoRecallTimeoutMs": 8000 } } } } }

2. Timeout warnings are now logged — when auto-recall or smart extraction times out, you'll see a warning in the logs instead of silent swallowing.

3. Hook migration to before_prompt_build — the lifecycle hooks have been migrated from the deprecated before_agent_start to before_prompt_build, which resolves several timing/race condition issues.

To fix: Upgrade to beta.10:

openclaw plugins install memory-lancedb-pro@beta
openclaw gateway restart

Let us know if the timeout visibility improves for you after upgrading.

[AliceLJY]

已在 v1.1.0-beta.10 中解决：

1. auto-recall timeout 现在可配置（autoRecallTimeoutMs，默认从 3s 提升到 5s）
2. timeout 发生时会输出 warn 级别日志，不再静默失败
3. PR fix: upgrade LLM request failure logs from debug to warn level #336 正在进一步将 LLM request failure 的日志从 debug 升级到 warn

升级到 beta.10 后 CMD 中应该能看到 timeout 相关日志了。