L2M

Appearance

Common Issues ​

SECRET_MASTER_KEY_BASE64 environment variable is required ​

Set SECRET_MASTER_KEY_BASE64 in .env with a base64-encoded 32-byte value.

Webhook 404 on /webhook-test/:path ​

  • Ensure webhook_input node path matches exactly.
  • Ensure workflow containing that webhook path is saved.

MCP tool discovery returns unexpected tools ​

  • Check MCP endpoint/base URL in mcp_tool node.
  • Re-run discovery after endpoint change.
  • Verify you are not using stale node config.

Agent hits max iterations ​

  • Reduce tool set size.
  • Improve prompt constraints for tool usage and stop conditions.
  • Increase context window/model limits where possible.
  • Raise max iterations only when needed.

Context-length/token overflow ​

  • Do not expose very large tool catalogs to the model.
  • Use selective tool inclusion.
  • Keep tool descriptions and schemas concise.
  • Keep Persist Tool Messages disabled on memory unless you explicitly need full tool transcripts.
  • If you use attached MCP Tool nodes, remove legacy agent_orchestrator.config.mcpServers entries from older workflows.

Local LLM crash / segmentation fault after tool-heavy runs ​

  • This usually indicates context exhaustion in the local inference server.
  • Reduce Tool Message Max Chars and max iterations.
  • Limit attached tools to the minimum needed for the task.
  • Clear old session history or use a fresh session_id before retrying.

Follow-up prompt did not reuse previously fetched MCP data ​

  • Ensure follow-up uses the same session_id.
  • Ensure Agent has memory namespace stable across runs.
  • Keep persistToolMessages off unless you explicitly need raw tool transcripts in prompt context.
  • Runtime now provides session_cache_list and session_cache_get; follow-up prompts can explicitly ask agent to reuse cached results before making new MCP calls.

Output Parser fails with Invalid JSON even though response includes JSON ​

  • In Output Parser, keep mode=json_schema and change parsingMode:
    • start with lenient
    • use anything_goes only when upstream outputs are highly inconsistent
  • Set inputKey to a concrete path (or moustache path), for example:
    • debug.agent_answer
    • {{debug.agent_answer}}
    • messages[5].content
  • Inspect parserTrace in node output to see which parse strategy was attempted and why it failed.

Connector nodes run in demo fallback unexpectedly ​

  • Check secret selection is not None (demo fallback).
  • Validate required endpoint/index/database fields.
  • Run Test Connection in node config.