sleep

name	sleep
description	Run a sleep cycle to consolidate and improve skills. Use when the user says 'sleep', 'run a sleep cycle', or 'consolidate skills'. Checks pressure levels and dispatches the N1→N2→N3→REM pipeline.
context	fork
hooks	{"SessionStart":[{"matcher":"startup","hooks":"[Truncated]"}],"Stop":[{"matcher":"","hooks":"[Truncated]"}],"Setup":[{"matcher":"maintenance","hooks":"[Truncated]"}]}
allowed-tools	["Bash","Read","Write","Edit","Glob","Grep","Agent"]

Sleep — The Orchestrator

This is the program.md of the sleep skills suite. The human iterates on this document to tune the autonomous skill consolidation system. The sleep skills themselves iterate on OTHER skills. This file is the one the human improves.

The idea: give an AI agent your skills and let it experiment autonomously overnight. It generates edge cases, tests patches, checks if the skill improved, keeps or discards, and repeats. You wake up in the morning to a dream journal and (hopefully) better skills. — adapted from Karpathy's autoresearch

---

What This System Optimizes

Primary metric: patch_acceptance_rate — the fraction of proposed patches the user actually applies. Higher is better. A low rate means dreams aren't producing useful insights.

Secondary metrics:

• coverage_score — fraction of real usage patterns covered by skill instructions
• cruft_score — unnecessary complexity in skills (lower is better)
• insight_rate — patches proposed per dream (target: 0.05-0.15)

Constraints:

• Never auto-apply patches. User always decides.
• Patches must be 1-3 lines. Larger findings are feature requests.
• Don't bloat skills. Simplicity is a feature.
• Don't break working skills. A patch that helps edge cases but degrades common cases is net negative.

---

The Two Processes

Process S — Homeostatic Pressure

Every conversation is metabolic activity. Pressure accumulates per-skill:

pressure(t) = 1 - e^(-t / τ)

• t = weighted conversations since last sleep
• τ = skill-specific time constant (default 20, auto-tuned by insight_rate)
• Saturates toward 1.0 — you can't accumulate infinite sleep debt

Thresholds: <0.2 rested, 0.2-0.4 drowsy, 0.4-0.7 moderate, 0.7+ heavy.

Process C — Circadian Rhythm

A Desktop scheduled task fires at a fixed cadence (default: nightly). Opens its own session. Runs autonomously. The SessionStart hook provides a softer check on every session start.

Interaction

Pressure	Gate	Depth
Low	Closed	Nothing
Low	Open	Micro-nap (N1 only)
Moderate	Open	Standard (2 cycles)
High	Any	Deep (3-4 cycles, NEVER STOP until rested)

---

Cycle Architecture

N1 (drift) → N2 (spindle) → N3 (consolidate) → REM (dream)
                                                      │
                                                [eval loop]
                                                mutate → test → measure
                                                keep if improved
                                                discard if not
                                                      │
                                                ◄─────┘ next cycle

Each successive cycle: N1 scans wider, N2 filters less, N3 gets shorter, REM gets longer. Mirrors biological sleep architecture.

Cycle Depth Profiles

Depth	Cycles	Stages	REM Eval
Micro-nap	1 partial	N1 only	No
Light	1	N1→N2	No
Standard	2	Full pipeline	Yes, 2-3 evals per dream
Deep	3-4	Full pipeline, extended REM	Yes, full eval loop

---

Autonomous Overnight Mode

NEVER STOP. Once the sleep loop begins (via Desktop scheduled task or explicit deep sleep command), do NOT pause to ask the human. The human may be asleep or away. Continue cycling until:

1. All tracked skills are below 0.2 pressure, OR
2. Maximum cycle count is reached (default: 4), OR
3. Session context is approaching compaction threshold

If you run out of fragment material, scan older conversations. If no conversations exist, generate synthetic fragments and dream about those. If all dreams score 0, that's fine — the skills are robust. Log it and move to the next skill.

Log everything. Every dream scenario, every eval result, every patch proposed or discarded. Append to results.tsv. Commit dream journals to the dreams branch if git is available.

---

Dispatch Protocol

1. Read State

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/sleep_tracker.py status

2. Determine Targets

Rank skills by effective pressure (pressure + debt × 0.5). Select based on cycle depth: micro-nap = top 1, light = top 3, standard = top 5, deep = all above threshold.

3. Execute Pipeline

For each target skill, dispatch stage skills as subagents:

N1 (drift) — Spawn as Agent with drift instructions. Pass: target_skill, scan_depth, recency_bias, randomness, cycle_number. Collect: fragment bag.

N2 (spindle) — Spawn as Agent with spindle instructions. Pass: fragment bag, target_skill, selectivity, cycle_number. Collect: sorted manifest.

N3 (consolidate) — Spawn as Agent with consolidate instructions. Pass: sorted manifest, target_skill, prune_mode, cycle_number. Collect: structural audit. Record baseline metrics in results.tsv.

REM (dream) — Spawn as Agent with dream instructions. Pass: structural audit, surviving fragments, target_skill, intensity. Collect: dream journal with tested patches.

4. Record Results

For each completed cycle, append to ${CLAUDE_PLUGIN_ROOT}/results.tsv:

timestamp	skill	cycle	depth	fragments_scanned	fragments_survived	dreams_generated	patches_proposed	patches_tested	patches_passed_eval	pressure_before	pressure_after	notes

5. Update State

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/sleep_tracker.py reset [depth]

6. Loop or Report

If autonomous mode and pressure still elevated: begin next cycle. If pressure resolved or max cycles reached: write summary, present dream journals with tested patches to user.

---

Triggering

Automatic (recommended)

Desktop scheduled task — The circadian rhythm. Create via Claude Code Desktop:

Schedule: 0 3 * * * (3am daily, or whatever suits your rhythm)
Prompt: "Read ${CLAUDE_PLUGIN_ROOT}/sleep/SKILL.md. Run an autonomous
         sleep cycle on all skills above pressure threshold. Use deep
         sleep mode. Do not stop to ask for input. Log all results."
Permission mode: Auto accept edits
Worktree: enabled (isolate from working state)

Session-scoped /loop — Intra-session micro-naps for long sessions:

/loop 30m "Run sleep_tracker.py check. If any skill pressure > 0.4,
           run a light N1→N2 triage pass on the highest-pressure skill.
           Report findings briefly."

Manual

Command	Action
sleep	Auto-detect pressure, run appropriate cycle
sleep status	Show pressure gauges
nap [skill]	Light cycle on specific skill
deep sleep	Autonomous deep cycle, NEVER STOP
sleep history	Past results from results.tsv
snooze	Defer, accumulate debt
insomnia	Debug — full state dump

---

State

All state in ~/.claude/sleep_state.json. The tracker script handles all I/O. See schemas/sleep_state.schema.json.

The tracker is deliberately thin — pure math and JSON read/write. All intelligence lives in prompt hooks (relevance assessment) and agent hooks (fragment extraction, eval). The tracker never reasons about content.

---

Meta-Improvement

This orchestrator is itself subject to dreaming. If the dream skill is applied to sleep/SKILL.md, it can propose patches to the sleep process itself — adjusting thresholds, cycle profiles, dispatch order, or eval criteria. The human reviews these meta-patches with the same keep/discard discipline as any other patch.

Over time, the results.tsv becomes the training data for improving this document. Analyze it: which mutation strategies produce accepted patches? Which skills benefit most? Which cycle depths are most productive? Use those findings to refine these instructions.

---

Anti-Patterns

• Don't sleep too often. If you're sleeping every 5 conversations, τ is too low.
• Don't skip micro-naps. They're cheap and catch things early.
• Don't ignore debt. Pressure 0.3 + debt 0.5 needs deeper sleep than pressure 0.3 alone.
• Don't force insights. 60% of dreams should score 0. That's correct.
• Don't trust untested patches. REM eval exists for a reason. Mental simulation alone is insufficient — test patches before proposing them.