OpenClaw memory plugin misclassifies cron distillation tasks as transcript-ingest and injects ingest-reply-assist

[Suidge]

Summary

OpenViking's OpenClaw memory plugin can misclassify cron maintenance prompts (specifically Daily Memory Distillation / Weekly Synthesis style tasks) as transcript-ingestion input and inject the ingest-reply-assist prompt block.

This is an upstream plugin bug because these cron maintenance tasks are not memory-ingestion transcripts, yet they receive transcript-specific prompt augmentation and can produce unwanted mid-run chatter.

Environment

• OpenClaw with plugins.slots.contextEngine = openviking
• OpenViking runtime env: ~/.openviking/.venv
• OpenViking version: 0.2.11
• OpenClaw plugin path under test:
  • examples/openclaw-memory-plugin/index.ts
  • examples/openclaw-memory-plugin/text-utils.ts
• Daily distillation cron job payload:

Daily memory maintenance. Read skills/opencortex/references/distillation.md for full instructions and follow them. Workspace: /Users/neoshi/.openclaw/workspace

What happens

The plugin applies:

<ingest-reply-assist>
The latest user input looks like a multi-speaker transcript used for memory ingestion.
...
</ingest-reply-assist>

...even though the task is a cron maintenance job, not a transcript-ingestion request.

Observed log pattern:

openviking: ingest-reply-assist applied (reason=transcript_like_ingest, speakerTurns=2, chars=454)

Impact

Two different impacts were observed:

1. Hard failure mode

   • A 03:00 distillation run produced only a one-line opening message and then stopped without any tool calls.
   • Session transcript contained only the cron prompt + one assistant line.

2. Residual degraded mode

   • After a local mitigation, a 07:54 manual backfill run still received ingest-reply-assist, but it completed successfully.
   • It executed the full workflow:
     • read distillation.md
     • created/cleaned lockfile
     • read/write/archive files
     • appended debrief to archived daily log
     • final summary matched actual tool execution
   • However, transcript-ingest prompt injection still remained, and mid-run chatter was still possible.

So the critical execution path can recover, but the plugin still misclassifies cron maintenance prompts and injects the wrong prompt block.

Why this looks like a plugin bug

The current transcript detection is too broad for OpenClaw cron-maintenance traffic.

These maintenance prompts may contain structured metadata / speaker-like formatting, causing them to look transcript-like under current heuristics, even though semantically they are control/maintenance tasks.

This should be filtered upstream in the plugin rather than relying on downstream prompt resilience.

Minimal expected behavior

For cron maintenance prompts such as:

• Daily memory maintenance
• Weekly synthesis
• prompts that explicitly read maintenance reference docs
• prompts that include cron markers / workspace execution instructions

The plugin should not inject ingest-reply-assist.

Suggested fix direction

Add an explicit exclusion path in the OpenClaw memory plugin before transcript-ingest assistance is applied, e.g. detect cron-maintenance/task-run prompts and skip ingestReplyAssist for them.

A stronger solution would likely rely on session/task context rather than only transcript-text heuristics.

Evidence summary

• Failed run: only one assistant opening line, zero tool calls
• Successful backfill run: 20+ tool calls, 111s duration, workflow completed
• Both were still subject to ingest-reply-assist injection
• Therefore the root bug is not gone; it is just less destructive after local mitigation

Requested outcome

Please treat cron-maintenance prompts as a first-class exclusion for transcript-ingest assistance in the OpenClaw memory plugin.