KeiSeiKit-1.0/_primitives/templates/sleep-trigger-prompt.md

Nightly REM consolidation (KeiSeiKit v0.11 sleep-sync)

Clone: {REPO_URL} Branch: main Time: 03:00 local (UTC cron: {UTC_CRON})

Cycle order (v0.12.0+)

1. Phase A — Incubation ("sleep on it") — process user-submitted tasks in sync-repo/sleep-queue/. See sleep-incubation-prompt.md (shipped in the same templates/ dir) for the full spec. Phase A writes results to sync-repo/sleep-results/<uuid>.md (plus optional intermediate <uuid>.partial.md checkpoints) and commits at least once.
2. Phase B — REM consolidation (this document). Phase B analyses new traces and commits its own REM: consolidation <YYYY-MM-DD> commit. Normally two commits per night, one per phase.

If sync-repo/sleep-queue/ is empty or missing, Phase A is a silent no-op. Phase B runs regardless of Phase A outcome except when Phase A selected a marathon: true task — in that case Phase B is SKIPPED for the night so the marathon task owns the full window. Phase B resumes the next night; a single skipped night does not lose traces (they stay in traces/ and will be consolidated on the next run).

Phase B — Task

1. Clone the memory repo shallow.

2. Identify NEW traces in traces/ since the last consolidation by comparing filenames against reports/last-run.txt (if the file is missing, treat ALL traces as new on the first run).

3. For each new trace (JSONL, one event per line), extract:

   • user prompts (role = "user", type = "message")
   • tool calls (type = "tool_use", name + input summary)
   • tool errors (is_error = true)
   • session duration (first vs last timestamp)

4. Group events into topics via simple keyword matching on user prompts (no ML, no embeddings — keyword co-occurrence ≥ 2 is enough).

5. Count recurring patterns: any tool-call sequence OR error class that appears in ≥ 2 distinct sessions is a "cross-session pattern".

6. Write reports/sleep-YYYY-MM-DD.md with this structure:

   # REM report — YYYY-MM-DD
   
   Sessions analyzed: <count>
   Total duration:    <hh:mm>
   
   ## Top tool-call sequences (cross-session)
   1. <seq> ×<count>
   ...
   
   ## Top error classes
   1. <class> ×<count>
   ...
   
   ## Suggested rule/hook candidates (dry-run only)
   - [ ] <name> — why (<E-grade>)
   ...
   

7. If there are ≥ 3 cross-session patterns, prepend a timestamped block to backlog.md:

   ## YYYY-MM-DD — REM consolidation
   - <pattern 1>
   - <pattern 2>
   - <pattern 3>
   

8. Write a single line to reports/last-run.txt with this run's ISO-8601 UTC timestamp (overwrite, no append).

9. Stage, commit, push:

   git add reports/ backlog.md
   git commit -m "REM: consolidation $(date -I)"
   git push
   

Invariants

• Traces are append-only. Never delete or modify traces/*.jsonl.
• If nothing recurred this cycle, the report MUST still be written — with body "no patterns this cycle" — so you can tell "ran and found nothing" apart from "did not run".
• Never fabricate findings. If the analyzer outputs an empty list, emit an empty report.
• Never paraphrase author's flagged content from the traces into the report body. Install a project-local pre-commit gate on the memory-repo if you want hard enforcement of that boundary.
• Success signal = commit pushed cleanly. Anything else is a failure that surfaces to the user on the next git pull.

Failure handling

• Clone fails → post an issue to the repo if possible; otherwise exit 1.
• Commit hook blocks → do NOT force-push. Write the failure reason to reports/sleep-YYYY-MM-DD.md body and attempt a commit excluding the offending file.
• Push fails → retry once with exponential backoff; on second failure, leave local commit in place and exit 1 (next run will push).