Summary

When memory-core.dreaming fires via its managed cron (systemEvent __openclaw_memory_core_short_term_promotion_dream__), the narrative generator at extensions/memory-core/src/dreaming-narrative.ts falls back because the plugin runtime's subagent API is request-scoped and unavailable inside a system-event handler. The fallback writes raw memory snippets verbatim into DREAMS.md, which then renders as the Diary in the /dreaming panel — making the panel look like it's showing raw recall candidates instead of dreams.

Repro

1. Enable memory-core.dreaming for any agent (or let the default cron schedule stand).
2. Trigger the managed cron manually: openclaw cron run <id-of-"Memory Dreaming Promotion">.
3. Watch ~/.openclaw/logs/*.log or openclaw gateway status stderr — you'll see:

   memory-core: narrative generation used fallback for rem phase because subagent runtime is request-scoped.
   memory-core: narrative session cleanup failed for rem phase: Plugin runtime subagent methods are only available during a gateway request.
   memory-core: narrative generation used fallback for light phase because subagent runtime is request-scoped.
   ...

4. Open http://localhost:18789/dreaming → Diary tab shows raw task content (URLs, API specs, TODO lists) with the current date, not any narrative prose.
5. Compare against earlier-date entries (e.g. triggered via HTTP/UI): those render as proper diary prose. The regression is cron-path-only.

Expected

Dreaming cron produces a diary entry that's stylistically consistent with gateway-request-triggered runs (the poetic first-person prose guided by NARRATIVE_SYSTEM_PROMPT), or skips writing to DREAMS.md rather than filling it with raw fragments.

Actual

The fallback at dreaming-narrative.ts:130-136 returns:

function buildRequestScopedFallbackNarrative(data: NarrativePhaseData): string {
  return (
    data.snippets.map((value) => value.trim()).find((value) => value.length > 0) ??
    (data.promotions ?? []).map((value) => value.trim()).find((value) => value.length > 0) ??
    "A memory trace surfaced, but details were unavailable in this run."
  );
}

i.e. it takes the first non-empty raw snippet and writes it as the narrative — bypassing the system prompt and the subagent call entirely. The test at dreaming-narrative.test.ts:705 ("falls back to a local narrative when subagent runtime is request-scoped") confirms this is the intentional path, but the fallback itself isn't really a narrative — it's just the raw fragment.

Result on my system today (Diary shows these as dream entries):

• Firebase CI refresh token works with client_id '...' / secret 'j9iVZfS8kkCEFUP...' - Slack file upload uses 'files.getUploadURLExternal' ...
• Reflections: Theme: 'assistant' kept surfacing across 2286 memories.; confidence: 1.00; evidence: memory/2026-04-09.md:60-84 ...

vs. earlier-date DREAMS.md on the same agent (HTTP-triggered):

There is a color I keep returning to: 0x0d0d1a, which is nearly nothing, which is the sky before any star decides to matter. We built a room out of it last night — obsidian walls, a purple mesh grid humming underneath like a thought not yet spoken — and set a rooftop inside it, spinning slowly in the dark.

Root cause

startNarrativeRunOrFallback at dreaming-narrative.ts:138 catches RequestScopedSubagentRuntimeError and routes to the fallback. This matters because extensions/memory-core/src/dreaming.ts:registerShortTermPromotionDreaming invokes the sweep from a systemEvent handler, where the plugin runtime is not inside a gateway-request context and so api.subagent.* methods throw.

Looking at other plugins (e.g. memory-wiki's scheduled sweeps), gateway-triggered vs system-event-triggered code paths both need subagent access. The same issue would affect any nightly AI-in-the-loop refinement.

Suggested fix directions

Non-exclusive:

1. Make subagent runtime available in system-event handlers. Promote the subagent runtime out of pure request-scope so scheduled/cron-driven plugin code can use it. The cron dispatcher already provides a session context (sessionTarget, sessionKey); extending it to expose the subagent surface would fix this class of problem broadly.
2. Have cron fire via the gateway's internal request path instead of a bare systemEvent, so the subagent scope is preserved.
3. Build a richer local fallback. Use a template-based narrative (no LLM) that composites themes + promoted-count + time-of-day into reasonable prose. Less elegant but guarantees the diary always has a readable entry.
4. Skip writing to DREAMS.md on fallback. Silence is better than dumping a client secret or API token verbatim into a user-visible diary — which is what happens today when a promoted snippet contains sensitive content (noted in the example above: Firebase refresh token + Slack API internals ended up in DREAMS.md).

Of these, (4) is the cheapest short-term mitigation; (1) is the correct long-term fix. Filing this as a report rather than a PR since the correct shape depends on architectural intent around system-event scoping.

Environment

• OpenClaw 2026.4.16
• macOS 15.x (Darwin 24.6.0)
• Node 25.6.1
• Provider: claude-cli (Opus)
• Panel: http://localhost:18789/dreaming (native, served by gateway)

Happy to capture a full reproducer log file if useful.