#Dreaming not working

1 messages · Page 1 of 1 (latest)

wide light
#

I just filed issue https://github.com/openclaw/openclaw/issues/81465 with the details, but thought would also ask here in cast Krill has any ideas.

OPenclaw 2025.5.7 with Claude CLI, dreaming always seems to fail.. OC analysis:

Subagent spawn fails inside the cron. The dreaming code tries to spawn a subagent to write the narrative, but the cron's session runtime is "request-scoped" and rejects nested subagents. So it falls through to a fallback path. That's why the May 9/10 entries are low-quality fragments like A: I'll gather the GitHub activity… and User: [Subagent Context]… — those are raw transcript scraps, not actual reflections. There's even a literal [your dream entry] placeholder in the May 11 + May 13 log lines, suggesting the fallback template isn't substituting properly.

tame hareBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

trail flumeBOT
#

Yeah, your diagnosis looks very plausible to me.

I checked the current docs/source: Dream Diary generation is explicitly a best-effort subagent turn, and the request-scoped fallback in memory-core really does just append the first snippet/promotion verbatim. So if the cron path can’t access the gateway subagent runtime, you get exactly the kind of raw transcript scraps / placeholder-ish diary output you’re seeing instead of an actual reflection.

A couple useful checks:

openclaw logs --follow | grep -i "narrative generation\|dreaming\|request-scoped\|fallback"
openclaw memory status --deep

If you see:

memory-core: narrative generation used fallback ... because subagent runtime is request-scoped