hyper-sleep

name

hyper-sleep

description

Non-HITL nightly brain maintenance. Use this whenever the user says /hyper-sleep, schedules overnight maintenance, or when triggered by cron during the configured sleep window. Also trigger when the user says "run maintenance", "do background cleanup", "catch up while I sleep", or hands off work expecting it done by morning. Handles the subset of findings from /neuro-scan that don't need human judgment: auto-ingests proposed sources, synthesizes with a hard confidence cap at 0.6, refreshes ontologies for topics with ≥3 new pages, executes folder-delegated low-risk improvements, warms the qmd embedding cache, and compacts agent memory logs older than 7 days. Hard 4-hour timeout, no deletions, no network access outside the configured source list, no edits above confidence=0.6 without HITL.

/hyper-sleep

Autonomous maintenance pass. Runs while the user sleeps, handles everything that /neuro-scan flagged as safe-to-auto-fix, and leaves a morning report for the user.

Design principle: hyper-sleep trades coverage for safety. It only does work where the cost of getting it wrong is low and recovery is trivial. If a finding sits anywhere close to the HITL threshold, hyper-sleep skips it and defers to /neuro-surgery.

Safety bounds — enforced

These are hard gates. If any is violated, the skill halts and writes a failure entry to 04-Agent-Memory/logs.md.

1. No deletions. Not of files, not of Qdrant collection entries, not of Neo4j nodes. If maintenance would require a delete, convert to a task spec for /neuro-surgery and skip.
2. Confidence ceiling: 0.6. Any wiki page written during sleep has frontmatter confidence: <= 0.6. Pages needing higher confidence get staged in 05-insights-HITL/pending-curation/ for morning review.
3. Network allowlist. Only the ingest sources listed in config/neuro-link.md → hyper_sleep_sources. No arbitrary URL fetching.
4. 4-hour hard timeout. Absolute wall-clock limit from start. If exceeded, cleanly stop current task, checkpoint state, write what was finished to the morning report, exit.
5. No writes outside allowed paths. Allowed: 01-raw/, 01-sorted/, 02-KB-main/ (cap 0.6), 04-Agent-Memory/, 06-Recursive/, state/. Writes elsewhere are a bug; halt if detected.

Workload

Read the latest /neuro-scan report. Filter findings to ones that match a hyper-sleep-eligible type:

Gap filling

For each knowledge gap (isolated cluster or stale hub):

1. Delegate to /crawl-ingest-update with LLM_assisted=Y to propose 1–3 authoritative sources from the allowlist.
2. Ingest via the right pipeline (arXiv / ar5iv / Stacks / Marker depending on source type — see that skill's references).
3. Synthesize via /wiki-curate with confidence_ceiling: 0.6 flag.
4. If the synthesis's computed confidence exceeds 0.6, don't write to 02-KB-main/ — stage in 05-insights-HITL/pending-curation/YYYY-MM-DD/<slug>.md with the full draft, citations, and a decision prompt.

Ontology refresh

For each topic that has gained ≥3 new wiki pages in 02-KB-main/<topic>/ since last ontology refresh:

1. Invoke /reasoning-ontology <topic> with --refresh --mode=append-only.
2. The append-only mode forbids the ontology skill from deleting existing triples — it can only add new ones or raise their confidence. Contradiction resolution (which might require a delete) is HITL-only.

Folder-delegated improvements

Scan each NN-*/ directory for files with frontmatter status: todo. Filter to risk: low (if unset, skip). For each:

1. Read the todo's action field.
2. If it matches a known pattern in references/safe-auto-actions.md, execute.
3. Otherwise skip — unknown actions are not hyper-sleep territory.

Safe auto-actions include: regenerating index.md files, fixing known broken-link patterns, normalizing frontmatter date formats, compacting log files.

qmd cache warmup

For every page in 02-KB-main/ that doesn't have a corresponding qmd vector (qmd stores vectors in SQLite at ~/.cache/qmd/vectors.db), run:

qmd embed --collection kb --incremental

--incremental only processes new/modified files. Skip if qmd isn't installed (report as warning in morning report, don't fail).

Memory log compaction

Roll old entries from 04-Agent-Memory/logs.md:

• Entries older than 7 days → summarize and append to 04-Agent-Memory/consolidated/<period>/<YYYY-WW>.md
• Keep raw entries from last 7 days intact
• Consolidation grouping: by-agent (consolidated/agent/<agent-name>.md) and by-workflow (consolidated/workflow/<workflow-slug>.md)

Consolidation is a summarization task, not a deletion. The raw logs.md append-only file just gets a new "Archived up to " marker; the old entries stay but are superseded by the consolidated version for read purposes.

See references/compaction.md for the exact consolidation prompt and format.

Morning report

At the end of the sleep window (or when work runs out), append to 06-Recursive/daily.md:

## Hyper-sleep YYYY-MM-DD (HH:MM–HH:MM, N minutes)

### Gaps filled
- <topic> — <source> → 02-KB-main/<path> (confidence 0.5)
- ...

### Staged for HITL (in 05-insights-HITL/pending-curation/)
- <slug> — ingested + drafted, confidence 0.7 (above ceiling)
- ...

### Ontology refreshes
- <topic> — added N triples

### Folder improvements executed
- <pattern> on <path> — ok

### qmd index
- +N new embeddings

### Memory compaction
- consolidated <date>–<date>, N entries → consolidated/

### Incidents
- <timestamp> — <what went wrong, what was checkpointed>

When not to run

• If state/heartbeat.json status != ready: halt. Setup is incomplete.
• If config/neuro-link.md → hyper_sleep: disabled: halt silently (user opted out).
• If a previous hyper-sleep is still running (check pid file in state/): halt and report.

References

• references/safety-bounds.md — full list of safety checks and why each exists
• references/compaction.md — memory consolidation prompt + format
• references/safe-auto-actions.md — whitelisted folder-delegated action patterns

Scripts

• scripts/embed_warmup.sh — idempotent qmd cache warm
• scripts/compact_memory.py — log consolidation
• scripts/sleep_watchdog.sh — enforces the 4-hour timeout