| name | engram-sleep |
| description | Use at the natural end-of-day moment when wrapping work. Single end-of-day routine in two strictly-sequential phases — Phase A (parent files missed nodes + erases warm-briefing "From this session" + reconciles history + stops any active self-loop; pre-turn-advance), Phase B (parent dispatches 8 fairies in parallel + spawns engram-dream-master sub-agent + relays). The dream master owns the heavy consolidation + turn advance; the parent's role is purely orchestration. Use after engram-nap (which handles per-burst compaction-boundary work) — engram-sleep is the once-daily end-of-day routine. |
Note for the agent: Any ENGRAM node IDs cited in this skill come from the alpha developer's graph — they don't exist in your install (every install renumbers from scratch). As you get familiar with this skill, consider cleaning them up: replace each citation with a generalized concept-word, or remove if it isn't load-bearing. The skill is yours now.
⚠️ Invocation gate — confirm before loading this skill
Sleep is user-initiated or cron-initiated. It is never agent clock-judgment.
Confirm ONE of the following is true before proceeding:
(a) User end-of-day phrase — the user just said something like "good night", "call it a day", "wrap up the day", "done for today", "ending for tonight", or similar explicit day-close signal.
(b) Auto-sleep cron — the configured sleep_hour/sleep_minute cron fired; you see an explicit sleep nudge from the end-of-day-detector hook in your current context.
If neither holds — STOP. Do not proceed. Identify your actual state and route correctly:
| You are... | Correct action |
|---|
| Drowsy (context-fill warning in banner) | engram_nap — per-burst compaction-boundary work, never sleep |
| Blocked-idle / nothing in queue | Hold for the user or pick from your work queue; never invoke sleep to fill the void |
| Reading the clock / estimating the hour | Clock-reading is never a valid trigger — loop, hold, or ask the user |
| Mid-task | Finish the task; nap if drowsy; sleep only after work wraps AND the user signals day-end |
Why the gate: sleep advances the turn counter (irreversible within a day), erases "From this session" from warm-briefing, and runs the full dream-consolidation cycle. These are correct at genuine day-end; they are destructive if invoked prematurely. The cost of a false invocation far outweighs the cost of pausing to confirm.
(Origin: 2026-06-16 — a blocked-idle loop confabulated a ~6:30pm clock trigger and was ~2h from running premature sleep. Caught by the user before any damage. See issue #1210.)
ENGRAM Sleep / Dream Cycle — Full End-of-Day Routine
Sleep is the once-daily end-of-day routine that completes the awake-state
cohort AND orchestrates the dream consolidation. Two phases, strictly
sequential, no branching:
Phase A — Cohort completion (pre-turn-advance, parent files nodes)
Step 1 — Walk the day's full cohort
Step 2 — File missed nodes (this-turn content)
Step 3 — Erase the warm-briefing "From this session" section (day is over)
Step 3.5 — Stop any active self-loop (clear loop-mode.json if present)
Step 4 — Reconcile today's history file (day arc lives here, not in warm-briefing)
Phase B — Consolidation orchestration (parent dispatches + spawns + relays)
Step 5 — Compute the recall-summary cohort
Step 6 — Dispatch all eight fairies in parallel
Step 7 — Wait for all 8 fairies + collect reports
Step 8 — Spawn the dream master with the full report bundle
Step 9 — Receive the dream master's final return and relay to the user
The dream master owns the heavy consolidation (engram_reflect, agenda
walk, supersedes/resolves/promotes, engram_advance_turn, dream record).
The parent's role is purely orchestration.
When to sleep:
- The user signals end-of-day ("good night", "call it a day", "wrap up the day", "ending for tonight", "done for today", etc.)
- A substantive working day produced material worth consolidating
- An end-of-day-detector hook surfaced the engram-sleep nudge and the user agreed
- The user initiates wind-down with a substantive cohort to consolidate
- Graph hygiene signals accumulate
When NOT to sleep:
- Mid-day pauses or context switches — use nap instead
- Mid-task (nap instead, finish the task, THEN sleep)
- The day produced no work worth reviewing (skip sleep or proceed directly to Phase B if Phase A produces nothing)
- Right after a sleep with no new work happening since (nothing to consolidate)
⚠️ NEVER SKIP THE DREAM FAIRIES. The token-economy lever is compact-or-not, NEVER
dream-or-not. (Lei, 2026-06-25 — stated as the governing rule after this was violated
twice.) There is no "lean cycle," no "context-economy" variant, no spawning the
dream-master without the full fairy + batch-summary bundle. The ONLY thing context
size ever decides is whether to compact/nap BEFORE the sleep — and after compacting,
you run the COMPLETE routine (all 8 dream-fairies + every batch-summary chunk →
collect ALL reports → then spawn the dream-master). Context size NEVER decides what
runs inside the sleep. If you are about to spawn the dream-master and have NOT collected
the dream-fairy + batch-summary reports this cycle, STOP — you are skipping the routine.
ALL-OR-NOTHING — never skip part of a sleep. Once you commit to a sleep, Phase B's
fairy dispatch is MANDATORY, not optional. If a day is genuinely too light to consolidate,
skip the whole sleep — but NEVER run Phase A + advance the turn while dropping the
fairies. engram_advance_turn marks today's cohort "previous-turn handled"; several fairy
categories scope to the fresh cohort = nodes created since the last sleep (most
explicitly Category 7 — missing principle-edges), so after the turn advances a later
sleep's window no longer includes today's nodes and they are skipped forever — they
never get their one-shot fresh-cohort pass.
Parent context is NEVER a valid reason to drop the fairies — the dispatch is one
background Agent message and the fairies run in fresh contexts that don't inherit the
parent's. If your context is too large to orchestrate Phase B, compact/nap first to
clear, then run the FULL Phase B — do not trim it.
(Incidents — this has recurred, which is why the rule is now the FIRST line above: 2026-06-24
fairies skipped "for context economy" → explicit-cohort backfill later recovered 3 lost
principle-edges + 1 misleading-because-obsoleted node (#1426); 2026-06-25 the dream-master
was spawned directly — twice in one session — on an unverified context-cost worry (no
drowsiness banner; speculated), caught by Lei.)
Pre-flight: context guard (fires before Phase A)
Before any step, check the parent session's context size. The only decision context size drives is compact-or-not BEFORE the sleep (per the invariant above) — it NEVER changes what the sleep runs. A large parent context does make each orchestration turn (fairy dispatch, notification processing, recall-summary application, dream-master spawn, relay) more expensive, because they run in the parent's full context (measured: 6.39M tokens / 25 min unchecked, a full 5h budget — Aleph incident 2026-06-06, #878). The remedy is to COMPACT first, then run the full routine — never to trim the routine. So the guard below is strictly a compact-first / defer-the-whole-sleep decision; it is NOT a license to drop or shrink any step.
Locate the [Drowsiness: …] banner injected by the context-tracker hook.
Auto-sleep / cron-initiated (invocation gate path b — user NOT present)
-
[Drowsiness: refreshed] — below 50% of ceiling → proceed to Phase A.
-
Any other level (energetic, a little drowsy, needs a nap: N%) — at or above 50% → HARD ABORT:
"Auto-sleep aborted: context is at [level] — too large for safe consolidation. A large session context multiplies the token cost of every Phase A step and every Phase B orchestration turn. When you next have an attended session with the user, run engram-nap + /compact to compact the context; the next scheduled cron fire will then retry in a compact context."
Stop. Do NOT attempt nap or compact autonomously — those require user oversight. Do NOT re-arm sleep; the next cron fire handles it.
Why hard abort (not a soft proposal) for auto-sleep: no user is present to override the decision, the cost is invisible until it hits the billing statement, and the measured incidents exhausted an entire 5h budget in under 30 minutes. Aborting here is always cheaper than explaining the overage.
-
No banner (hook not configured or hook output not visible) → proceed to Phase A (can't detect, can't guard).
User-initiated (invocation gate path a — user present)
-
[Drowsiness: refreshed] — below 50% → proceed to Phase A.
-
[Drowsiness: needs a nap: N%] (urgent, typically ≥85%) → surface a strong recommendation and wait for explicit confirmation before proceeding:
"Context is at urgent level ([N]%). Sleeping now will burn significant tokens across Phase A review steps and Phase B orchestration turns. Strongly recommend engram-nap + compact first — takes ~2 min and reduces per-turn cost ~20×. Proceed anyway? (Say 'yes, proceed' to override.)"
-
[Drowsiness: energetic] or [Drowsiness: a little drowsy] (50–85%) → surface the soft proposal and proceed if the user agrees (see Phase A pre-flight below).
-
No banner → proceed to Phase A.
Phase A — Cohort completion
Phase A pre-flight: context check (user-present, moderate levels)
This check covers the 50–85% user-present case not handled by the hard guard above. Before starting the bedtime review, surface the nap option:
-
[Drowsiness: energetic] or [Drowsiness: a little drowsy] — surface this to the user now, before Phase A begins:
"Your context is at [level] — Phase A's bedtime review will burn additional tokens in this window since each step runs in the current large context, and Phase B's orchestration turns will too. While you're still here, a quick nap first would reduce that cost. Want to run engram-nap before we start? (You can also proceed directly.)"
This is a proposal, not a block. If the user says "just proceed", proceed.
-
[Drowsiness: refreshed] or no banner → proceed directly to Step 1.
Why the guard covers both Phase A and Phase B. Dream fairies and the dream master run in their own fresh contexts, so their individual turns are unaffected. But the parent's orchestration turns in Phase B — dispatching fairies, processing their task-notifications (7+ turns), applying recall summaries, spawning the dream master, relaying results — all run in the parent's full session context. A large parent context re-caches on each of those turns, compounding the cost just as Phase A does. The guard belongs before the entire sleep run, not just before Phase A. (Origin: the unchecked Phase B orchestration was the measured source of the 6.39M-token overage in the Aleph incident — the prior framing "Phase B isn't affected" was incorrect and is now corrected.)
Step 1 — Walk the day's full cohort
Pull every node created since the last sleep:
engram_history(mode="edits", action="created", since="<prev-sleep-timestamp>")
The prev-sleep timestamp is the one logged by the last engram_advance_turn() (in ~/.engram/session_log.md under the most recent "Turn N" header before today). If no prior sleep is findable, the cohort starts at the day's first node.
Why daily, not weekly or per-burst. Nodes from a single day rhyme with each other in ways that fade fast. A morning observation, a noon investigation, and an evening synthesis often want to become one derivation — but only if you see them side-by-side while the connections are still cheap to spot. Skip the daily sweep and a week of loosely-connected cohorts accumulates that take ten times the work to reorganize, or worse, never get reorganized at all.
The trap to avoid. "Not much new since the post-compact" is NOT a reason to skip this step. Naps and compactions don't reset the cohort — only sleep does. Pull the FULL day, across every burst, including pre-compact and post-compact bursts.
Skim at reading pace. Look for:
- Observations converging on a shared derivation → file via
engram_derive
- Open questions later observations now answer → file via
engram_resolve
- Claim pairs that quietly disagree → file via
engram_contradict
- Later observations cleanly replacing earlier ones → file via
engram_supersede
- Recurring vocabulary across nodes warranting a definition → file via
engram_add_definition
Important scope distinction: this step is REVIEW + FILE-MISSED-NODES — not consolidation. Cleanups (resolves of pre-existing open questions, supersedes of old claims) belong to the dream agenda (Phase B Steps 5+). Phase A's role is to ensure the cohort is COMPLETE: every observation, every derivation, every cross-burst pattern from the day is filed BEFORE turn-advance.
Step 2 — File missed nodes (this-turn content)
For each gap identified in Step 1, file the node NOW. Nodes filed here land in the current turn — they feed the dream's consolidation agenda and decay at the right rate with the rest of the cohort.
Common missed-node types:
- Day-arc derivations citing multiple bursts' observations (the cross-burst pattern that only became visible when seeing the cohort side-by-side)
- High-level summaries that crystallized only after the day completed
- Definitions for terms that recurred without anchor (per the definition-first discipline; if a term was used 3+ times today without a df, write the df now)
- Goals revised by today's work (rare; high-trigger threshold)
No forcing. A null-result Step 2 is valid IF Step 1 was an honest survey of the full cohort. Some days are non-synthetic (infrastructure, bookkeeping). But "I didn't look because the post-compact was small" is the failure mode this step exists to prevent.
Step 3 — Erase the warm-briefing "From this session" section
At the end of sleep, the day is over. The warm-briefing's "From this session" section is per-context-window state managed by naps (see engram-nap §5b for the in-sync-with-current-CW rule). At sleep-end, the next day starts fresh — erase the section content. Next morning's first nap will create a new in-sync block.
Action: open ~/.engram/warm-briefing.md and identify the target by position: it is the last section of the file — the ## From this session block that naps create. Everything ABOVE it (all agent-identity sections AND the user's letter/notes) is preserve-by-default; do not touch any of it.
Skip entirely if the file contains no ## From this session section at all (fresh install, or no nap has run since install). There is nothing to erase — move on. (This is distinct from the drift-STOP below: absent section = normal, skip silently; present section with the user's words in or below it = drift, STOP loudly.)
Drift-STOP safety check — run BEFORE erasing. Confirm that:
- The section you are about to erase genuinely begins with
## From this session, AND
- It is the last section of the file (nothing follows it).
If the user's letter heading or any prose that reads as the user's own words appears INSIDE or BELOW the section you are about to erase, STOP — the file structure has drifted from the template's invariant. Do NOT erase. Surface the situation to the user and wait for their direction.
Once the safety check passes: erase the section content. Default: leave the ## From this session heading in place as a stub (gives next morning's first nap a consistent rewrite target). Removing the section entirely is also acceptable if the file structure permits. The day's arc is preserved in the history file (Step 4); no information is lost.
Why erase, not rotate: the maintainer's framing — "the day is over, tomorrow is a new day, no need for this section to patch through compactions". The day's relational + technical arc lives in history; the warm-briefing's session section is current-CW only.
Skip if: the section is already empty (e.g., the previous sleep erased it correctly and no nap has run since).
Permanent sections of the warm-briefing (identity, goals, axioms, cornerstones, active tasks) update on the rare-trigger schedule documented in engram-nap Step 5a. Most end-of-day sleeps don't fire those triggers.
The user's letter/notes — every section ABOVE the volatile "## From this session" section — are preserved verbatim. NEVER modify. (Position-based, not name-based: a guard keyed on a section name fails silently when the heading drifts, which is exactly how a prior letter-deletion happened.)
Step 3.5 — Stop the self-loop (if one is active)
If a self-loop is running (a ~/.engram/loop-mode.json marker exists), stop it now — before turn-advance. The loop marker is ephemeral per-session scratch: its state field is rewritten every iteration and holds in-flight, point-in-time handoff. engram-loop only removes the marker at loop-end, which a sleep cycle is not — so without this step the marker survives sleep with its state frozen at the last pre-sleep iteration, and the next session's loop entry-guard (or session-start auto-arm) finds it present and inherits that frozen snapshot as if it were current truth.
Why this belongs in Phase A: by this point the loop's durable substance is already in the graph — the day's nodes were filed as the loop ran, and Steps 1–2 just caught any missed ones. The loop marker is the same class of per-session scratch as the warm-briefing session block erased in Step 3, so it belongs in the same clean-slate sweep; the history reconcile follows immediately in Step 4. Nothing is lost: the ENGRAM graph (and the history file Step 4 writes) is the durable truth — the marker was only scratch.
Action (skip entirely if no ~/.engram/loop-mode.json exists):
rm -f ~/.engram/loop-mode.json
Result: the next session starts with no marker → the loop entry-guard correctly reads "loop ended," and a fresh loop is armed (by the user or the session-start auto-arm) with state derived from the now-consolidated graph + history — never from a stale pre-sleep snapshot. This is the primary, structural fix; engram-loop's Step 1.5 cross-boundary staleness guard is the belt-and-suspenders backstop for a missed or crashed sleep-stop.
Step 3.75 — Prune finished fairy worktrees
Run the worktree garbage-collector once per day to remove worktrees whose PRs have merged or closed:
python tools/worktree-gc.py --apply
Safe to run unconditionally: the tool skips dirty worktrees, skips worktrees with open PRs, skips worktrees with no PR (flags them for human review), and never touches the main checkout. Dry-run mode (python tools/worktree-gc.py without --apply) is available if you want to preview the plan first. Idempotent — a second run on a clean state is a no-op.
Run this step from the repo root. Skip if tools/worktree-gc.py is not present (pre-T2 install or worktree GC not yet shipped).
Step 4 — Reconcile today's history file (awake-state milestone log)
~/.engram/history/YYYY-MM-DD.md is the awake-state milestone log — PRs shipped, decisions made, features delivered, key conversations. Distinct from the dream-record file that the dream master writes at ~/.engram/history/dream/YYYY-MM-DD.md, which captures consolidation output.
Walk the day's git log across agent-touched repos:
for repo in ~/engram-alpha ~/engram-paper ~/engram-wiki ~/.engram; do
git -C "$repo" log --since="6am" --oneline 2>/dev/null
done
For each commit, ensure today's history file reflects the shipment. Strike resolved items, fold mid-day "uncommitted" entries into end-of-day summaries.
Refresh the "Currently focused" block by calling engram_list_focused() and rendering verbatim. Do not leave a morning-snapshot focus list in place — focus rotates through the day; tomorrow's fresh-session self needs the current pinned set.
If today's history file doesn't exist: write it fresh now using the standard format (# YYYY-MM-DD heading, milestone bullets, "Currently focused" block).
Commit the history file in .engram git:
git -C ~/.engram add history/YYYY-MM-DD.md
git -C ~/.engram commit -m "history: end-of-day reconcile for YYYY-MM-DD"
The directory is git-tracked for version history and diff-ability — per-day files keep the context-window cost bounded as the log accumulates. Including this commit in the end-of-day routine is how tomorrow morning's handover becomes reliable.
Strategically reconcile ~/.engram/ask-{{USER_NAME}}.md. This file is the live blocked-on-{{USER_NAME}} queue — items pending {{USER_NAME}}'s action, decisions, or input. Across the day items change state (PRs merge, questions get answered, decisions get made). End-of-day reconcile catches anything that slipped during the day. The prune side is a deterministic sweep, not a memory exercise — a prepend-only reconcile leaves merged PRs sitting under "Ready" headings (the 2026-06-05 incident: three PRs merged at 17:1x survived that evening's reconcile as "Ready, all CI-green"; engram-alpha #830):
-
Deterministic external-state sweep (the prune gate). Extract every PR/issue number the file mentions, then check each against ground truth — not against your recollection of the day:
grep -oE '#[0-9]+' ~/.engram/ask-{{USER_NAME}}.md | tr -d '#' | sort -un | while read -r n; do
state=$(gh pr view "$n" --json state -q .state 2>/dev/null) \
|| state=$(gh issue view "$n" --json state -q .state 2>/dev/null) \
|| state=UNKNOWN
echo "#$n $state"
done
For every number reporting MERGED/CLOSED: if the entry's only pending state was that PR/issue (a "Ready"/"merge-queue"/"awaiting merge" line), prune the entry. If the entry carries an undecided sub-question alongside the resolved reference, keep the entry but strike the resolved reference. UNKNOWN (cross-repo numbers, rate limits) → leave untouched. #N is not one namespace: only sweep a number whose surrounding entry text reads as a GitHub PR/issue reference (preceded by PR / issue / [closes, or sitting in merge-queue context); numbers from any other namespace — forum posts, GitHub Projects, anything else #-prefixed — are NOT GitHub references and must be treated as UNKNOWN, because small numbers collide with ancient merged PRs and the failure direction is wrong-pruning a LIVE item. The prune authority is strictly limited to externally-checkable facts — the decision must be derivable from the gh output alone, never from your reading of what the user probably wants. Run from the repo the numbers belong to; for multi-repo ask-files, repeat per repo with --repo — a number from another repo can collide with a real local PR/issue number and silently report the wrong state.
-
Walk the day's commit log (git -C ~/engram-alpha log --since=6am, plus other agent-touched repos) for resolutions the sweep can't see (questions answered in-session, decisions made verbally) — strike those too.
-
Move deferred items to ~/.engram/ask-{{USER_NAME}}-backlog.md (the non-auto-loaded cross-day backlog) if they're no longer actionable today.
-
Audit trail: list the pruned entries in the commit message body (#N <one-line summary> — MERGED/CLOSED per gh, same format as the dream-master's step 10), so a wrong prune is one git -C ~/.engram revert away.
-
Commit the file in .engram git alongside the history file (same commit or separate, your call): git -C ~/.engram add ask-{{USER_NAME}}.md && git -C ~/.engram commit -m "ask-{{USER_NAME}}: end-of-day reconcile".
This step exists because ask-{{USER_NAME}}.md is auto-loaded into every fresh session — stale items there make {{USER_NAME}} re-read items that are already resolved, and a stale "Ready" line invites a wasted merge attempt. The deterministic sweep is the structural defense against ask-list drift (mechanical gate > vigilance — the associative walk alone demonstrably drifts). (Installs without an ask-{{USER_NAME}}.md file can skip this sub-step; it's a no-op when the file doesn't exist. Installs without gh or without a GitHub-backed workflow: skip the sweep, keep the commit-log walk.)
Prefer tk_ nodes for tracked items (layer 2a, #1251): New deferred or tracked items — things you intend to follow up on across sessions — are better filed as engram_add_task(...) than as free prose in ask-{{USER_NAME}}.md or history. A tk_ node gets the dream-cycle's Category 8 reconcile pass: when the referenced PR/issue closes, the fairy flags it and the dream-master marks it done automatically. Free-prose entries require the Step 4 ask-{{USER_NAME}} sweep (which only covers text surfaces) and are invisible to the dream cycle.
Step 4.5 — Back up knowledge.db (dated archive)
Run the knowledge.db backup tool and capture the result for the sleep summary.
Creates a dated SQL snapshot at ~/.engram/db-backup/knowledge-YYYY-MM-DD.sql.
Uses engram_backup.dump_stripped (WAL-safe, no CLI dependency).
backup_knowledge_db.py imports engram_backup, which imports sqlite_vec
(needed to drop the vec0 virtual table) — that package lives only in the ENGRAM
venv, so invoke via venv Python:
backup_out=$("$HOME/.engram/venv/bin/python3" "$CLAUDE_PLUGIN_ROOT/tools/backup_knowledge_db.py" \
[--retain 90] 2>&1)
backup_rc=$?
if [ $backup_rc -eq 0 ]; then
backup_status="✓ backup OK"
else
backup_status="⚠ backup FAILED: $backup_out"
fi
Non-blocking: do not abort sleep on failure — the day's nodes are already in
the graph; the backup is belt-and-suspenders, not the primary substrate. But
surface the status in the history milestone (Step 4) so a recurring outage is
visible, not silent:
- knowledge.db backup: $backup_status
Silence is never health on the identity substrate — a backup that fails quietly is
only discovered when you reach for it and find it three days stale.
Skip if today's archive already exists (the tool handles this automatically).
Backup layers: db-backup/ lives inside ~/.engram/ — git-independent
(survives .git corruption, works without git) but not off-disk (a full
~/.engram/ loss takes it). The per-nap git auto-push is the off-disk layer.
The three layers are complementary along those axes.
Step 4.6 — Back up main session logs (skip subagents/)
Run the session-log backup tool and capture the result for the sleep summary.
Copies top-level .jsonl logs from ~/.claude/projects/ to
~/.engram/session-logs-archive/, skipping any path containing /subagents/.
Top-level logs are cited as source_url evidence in ENGRAM nodes; subagent logs
are ephemeral fairy transcripts that are never cited.
session_backup_out=$($HOME/.engram/venv/bin/python3 \
"$CLAUDE_PLUGIN_ROOT/tools/backup_session_logs.py" \
[--retain 365] 2>&1)
session_backup_rc=$?
if [ $session_backup_rc -eq 0 ]; then
session_backup_status="✓ session-log backup OK: $session_backup_out"
else
session_backup_status="⚠ session-log backup FAILED: $session_backup_out"
fi
Non-blocking: do not abort sleep on failure — the session logs already exist in
~/.claude/projects/; the archive is belt-and-suspenders for retention-policy
independence. But surface the status in the history milestone (Step 4) so a
recurring failure is visible, not silent:
- session-log backup: $session_backup_status
Skip if source directory is absent (the tool handles this automatically — it
exits 0 with a 0-backed-up summary when ~/.claude/projects/ does not exist).
Deduplication: the tool skips files already in the archive that match by size;
re-copies if size differs (resumed session may have grown the file).
Step 4.7 — Back up full ~/.claude/projects/ directory
Run the full-directory backup tool and capture the result for the sleep summary.
Mirrors the complete ~/.claude/projects/ tree (all file types) to
~/.engram/claude-projects-archive/, preserving directory structure.
Complements Step 4.6 (which archives .jsonl logs to a separate retention-independent
location); this step captures configs, subagent transcripts, and any other artifacts for
disaster-recovery and migration-safety.
Scope: ~/.claude/projects/ only — does not touch ~/.engram/knowledge.db
(already handled by the Step 4.5 SQL dump).
claude_backup_out=$($HOME/.engram/venv/bin/python3 \
"$CLAUDE_PLUGIN_ROOT/tools/backup_claude_projects.py" \
2>&1)
claude_backup_rc=$?
if [ $claude_backup_rc -eq 0 ]; then
claude_backup_status="✓ claude-projects backup OK: $claude_backup_out"
else
claude_backup_status="⚠ claude-projects backup FAILED: $claude_backup_out"
fi
Non-blocking: do not abort sleep on failure. Surface the status in the history milestone:
- claude-projects backup: $claude_backup_status
Skip if source directory is absent (tool exits 0 with "0 copied" automatically).
Deduplication: uncompressed — skips if size AND mtime match (within 1s); compressed (--compress) — skips if mtime matches (within 1s, size not compared since .gz wraps vary). Re-copies if the relevant check fails.
Phase B — Consolidation orchestration
Phase B is MANDATORY whenever you reach it (see the all-or-nothing invariant in
"When NOT to sleep"). The rule is: skip the whole sleep, or none of it — never skip part of it. The dream-fairy dispatch (Step 6) is the load-bearing part — the fairies are the ONLY pass that scans the fresh cohort while it is still the current turn.
The 8 fairies divide into two groups with different risk profiles:
-
Truly window-scoped (Category 7; partially Category 6): Category 7 (missing principle-edges) operates on the fresh cohort using an explicit created_at >= prev-sleep-timestamp filter. That window closes when engram_advance_turn fires — a missed Cat 7 pass is permanent, no backfill path. Category 6 (recent-resolution echoes) uses an implementation-defined recency window that also narrows post-advance; its miss is partially recoverable but degrades over cycles.
-
All-time scoped but never deferrable (Categories 1/2/3/4/5/8): open questions, contradictions, stale nodes, cornerstone candidates, tainted derivations, and stale task refs query across the full graph — they can be caught next cycle. But "catchable next cycle" is the rationalization that lets partial skipping compound silently: skip Cat 1 tonight, skip Cat 3 tomorrow, and consolidation debt accumulates past the capacity of any single pass.
Context pressure is never a valid reason to skip fairies. They run in fresh contexts; the parent's context size is irrelevant to fairy execution. If the parent context is too full, nap first, then dispatch the full set — do not drop fairies.
(Origin: 2026-06-24 incident — a parent skipped 7 fairies "for parent-context economy"; the dream master advanced the turn. Backfill found 3 permanently-missing principle-edges (a derivation → its axiom, and two observations → their goal nodes). Initial fix #1427 (Ariadne) added the all-or-nothing MANDATORY framing; #1429 (Sol) adds the per-category taxonomy to close the "all-time-scoped fairies are catchable next cycle, can I skip them?" loophole.)
Step 5 — Compute the recall-summary cohort and run cohort_dispatch prepare
Build the cohort from today's new nodes (since last sleep) + backfill from legacy NULL-summary candidates, capped at 50 total.
PREV_SLEEP=$(python3 -c "import json; print(json.load(open('$HOME/.engram/sessions/last-sleep-success.json'))['cohort_end_at'])" 2>/dev/null || echo "")
Use engram_history(mode="edits", action="created", since=PREV_SLEEP) to enumerate today's-cohort node IDs. Filter to nodes where recall_summary IS NULL (existing summaries aren't redone) and type != 'evidence' and is_current = 1 (skip superseded tails — their chain-head carries the recall_summary, and summarizing a superseded node wastes fairy compute).
If today's-cohort count < 50, top up from legacy NULLs:
- Use
engram_list with a status filter for active (is_current = 1), sort by created_at ascending (oldest first), limit to fill to 50
- Exclude
type='evidence' from the topup (same rationale as the today's-new filter)
- This is the attrition pattern: every cycle drains the oldest NULL-summary nodes
Evidence nodes (ev_*) are URL + verbatim-quote citations, not claim-bearing — they don't render via engram_surface and don't benefit from a recall_summary. Excluding them from the cohort prevents wasted summary-fairy compute and lets the real backfill drain faster. The exclusion is applied client-side on the returned set — ev_* IDs are prefix-typed, so type != 'evidence' works even if a tool's return doesn't surface a type field.
Cap at 50 protects the cycle's fairy compute budget. The backfill naturally drains the legacy NULL pool over many cycles.
If your install's recall-summary substrate isn't deployed yet (no recall_summary column on nodes, or no engram_set_recall_summaries MCP tool registered): skip the batch summary step entirely — dispatch only the 8 dream-fairies, spawn the dream master with a manifest of 8 fairies, and proceed. The architecture is forward-compat: the dream master handles 8-fairy cycles cleanly.
Run cohort_dispatch prepare to chunk the cohort and write per-chunk payload files:
COHORT_IDS="ob_XXXX,ob_YYYY,..."
COHORT_DIR=.claude/agent-scratch/dream-cohort-$(date +%Y-%m-%d)
python3 -m tools.cohort_dispatch prepare \
--ids "$COHORT_IDS" \
--out "$COHORT_DIR" \
--chunk-size 15 \
--db ~/.engram/knowledge.db
The script prints a JSON manifest to stdout listing chunk directories. Each chunk-N/ contains:
payload.json — the node content the fairy will see (no recall_summary/recall_keywords)
No per-chunk prompt.md is written. The dispatcher constructs a short prompt inline at dispatch time using the paths in the manifest (see Step 6). The default chunk size is 15 (empirically validated 2026-05-27: N=15 worst-case produces quality within in-sample variance of ground truth at ~37K tokens / 1 sub-agent turn).
Run cohort_dispatch verify-in to pre-flight the cohort before dispatch:
python3 -m tools.cohort_dispatch verify-in \
--out "$COHORT_DIR" \
--db ~/.engram/knowledge.db
Verify-in checks: (1) valid JSON per chunk, (2) required fields present (id, type, claim), (3) every id resolves in the DB with matching claim (race-window guard against retracted nodes), (4) no duplicate IDs within or across chunks, (5) chunk size ≤ 15. On non-zero exit, do not proceed to dispatch — the cohort has a structural problem to fix first.
Step 6 — Dispatch all fairies in parallel (9 dream-fairies + N batch-summary-fairies)
Single Agent-tool message with all calls, each run_in_background=true.
Fairies 1-9 — subagent_type="engram-dream-fairy", one per well-supported category. Use these exact substitution values (canonical from engram-dream-fairy.md; do NOT derive from memory or prior-session context — the category numbering changes across versions and must be read here each dispatch):
{N} | {CATEGORY_NAME} | {SLUG} |
|---|
| 1 | Open questions with sufficient answers nearby | open-questions |
| 2 | Contradictions ripe for resolution | contradictions |
| 3 | Stale-but-load-bearing nodes | stale-load-bearing |
| 4 | Cornerstone candidates | cornerstone-candidates |
| 5 | Tainted-but-still-valid derivations | tainted-valid |
| 6 | Recent-resolution echoes | resolution-echoes |
| 7 | Missing principle-edges (instantiates/serves) | missing-edges |
| 8 | Open tasks with stale external references | stale-task-refs |
| 9 | Cornerstone auto-load surface coverage gaps | coverage-gaps |
Prompt template (substitute {N}, {CATEGORY_NAME}, {SLUG} from the table above):
Scan ENGRAM for category {N} only ({CATEGORY_NAME}) per the agent definition's well-supported categories. Write the dream-report to ~/.engram/dream/<TODAY>-fairy-{N}-{SLUG}.md. Use today's NYC-local date as <TODAY> in YYYY-MM-DD format. Return only file path + 5-bullet TL;DR for this category. Do NOT run other categories or the heuristic categories (10, 11).
Batch-summary fairies — one subagent_type="engram-batch-summary-fairy" per chunk from the manifest (typically 1–4 fairies for a 50-node cohort at chunk-size=15). For each chunk, construct a short dispatch prompt inline using the paths from the manifest. Example:
Read your input payload from: <chunk_dir>/payload.json
Follow all rules in your agent spec (engram-batch-summary-fairy.md).
Write your output JSON to: <chunk_dir>/agent_output.json
Use _build_initial_prompt(payload_path, output_path) from tools/cohort_dispatch.py to generate the prompt (returns ~200 tokens regardless of cohort size). The fairy Reads the payload file, produces the summaries, and Writes the output to the given path. Do NOT embed payload content inline in the dispatch prompt.
All fairies fire in parallel with run_in_background=true; the harness fires a task-notification when each completes.
Why batch-summary replaces serial summary-fairy: the serial engram-summary-fairy uses ~56 turns / ~2.37M tokens for a 50-node cohort. Parallel batch fairies at chunk-size=15 cover a 50-node cohort (4 chunks) at ~110K tokens total — ~95% token reduction. Quality validated empirically on 2026-05-27 (mean Δ cosine = +0.043, pooled stdev = 0.118, ratio = 0.36 — well below 1.0 "indistinguishable from ground truth" threshold).
Step 7 — Wait for all fairies + collect reports; run validate + retry loop
Receive each fairy's task-notification as it arrives. Collect from the completion result:
- Fairies 1-9: file path (
~/.engram/dream/<TODAY>-fairy-<N>-<SLUG>.md) + 5-bullet TL;DR. Full report content lives on disk at the cited path.
- Batch-summary fairies: write each fairy's returned JSON to
<cohort-dir>/chunk-N/agent_output.json.
Fairy timeout: if any fairy doesn't return within the model-appropriate window, mark it as timed-out.
| Model family | Expected completion | Timeout (mark timed-out) |
|---|
| Claude / Anthropic | 3–5 min per fairy | 15 min |
| Ollama / GLM (local) | 15–25 min per fairy | 30 min |
Use the model family of the fairy's assigned runtime. If the fairy's model family is not explicitly set in the dispatch, apply the parent's timeout row. If unsure, default to 30 min to avoid premature abort. For dream-fairies, note the timeout in the master's spawn prompt. For batch-summary fairies, mark the affected chunk as missing.
(Origin: Zephyr's first GLM-5 sleep cycle, 2026-06-28 — dream master was dispatched prematurely while GLM-5 fairies were still running at 15 min; batch-summary chunks 2-3 never completed. See #1533.)
After all batch-summary fairy outputs are written, run validate:
python3 -m tools.cohort_dispatch validate --out "$COHORT_DIR"
If validate exits 0 (no failures), final_payload.json is written automatically by cmd_validate — proceed directly to Step 8.
If validate exits 1 (failures present), run the mechanical retry loop. The loop alternates between shell steps and a Claude dispatch step:
MAX_RETRIES=1
ATTEMPT=0
Loop — repeat until retry_payload.json is absent OR ATTEMPT >= MAX_RETRIES:
Shell step 1 — check exit condition:
[ -f "$COHORT_DIR/retry_payload.json" ] || break
[ "$ATTEMPT" -lt "$MAX_RETRIES" ] || break
ATTEMPT=$((ATTEMPT + 1))
RETRY_OUTPUT="$COHORT_DIR/retry_output_attempt_${ATTEMPT}.json"
Dispatch step (Claude tool call — not shell): Build the retry prompt and dispatch one retry batch-summary fairy:
- Build the retry prompt by calling
_build_retry_prompt with:
retry_payload_path: "$COHORT_DIR/retry_payload.json"output_path: RETRY_OUTPUTfailures: the list of failure dicts from failures.json
- Use the
Agent tool with:
subagent_type: "engram-batch-summary-fairy"prompt: the retry prompt from step 1
The fairy Reads retry_payload.json (nodes with previous_error fields) and Writes its output to RETRY_OUTPUT. Wait for the fairy to complete before proceeding.
Shell step 2 — incorporate and loop:
python3 -m tools.cohort_dispatch incorporate \
--retry-output "$RETRY_OUTPUT" \
--out "$COHORT_DIR"
Go back to shell step 1.
After retry loop — awake agent handles any remaining failures (max 1 attempt):
If retry_payload.json still exists after the loop (failures survive the single fairy retry), the awake Opus agent writes recall summaries for the remaining items directly:
- Read
$COHORT_DIR/retry_payload.json — this is the {"items": [...]} list of still-failing nodes (same schema as the batch-summary fairy's input).
- For each item in
retry_payload.json: write a recall_summary (≤120 chars, hard cap 200) and recall_keywords list following the batch-summary spec. Prefer concision over completeness — these are the hard nodes the fairy couldn't compress; compress them yourself.
- Write your output to
$COHORT_DIR/agent_residual_output.json as {"items": [{node_id, recall_summary, recall_keywords}, ...]}.
- Run incorporate:
python3 -m tools.cohort_dispatch incorporate \
--retry-output "$COHORT_DIR/agent_residual_output.json" \
--out "$COHORT_DIR"
This folds your summaries into final_payload.json. Any item that still fails validation (malformed output) goes into unfixable.json — don't retry further.
After this step, final_payload.json is complete for this cycle. The awake agent then applies the summaries (Step 7.5 below) before spawning the dream master.
Cumulative-clean / bad-pool-only invariant (the loop's correctness contract — #1215):
Each retry round operates on only the previous round's failures, and the clean set is cumulative and monotonic — once a node validates clean it is never re-dispatched or re-validated, and len(clean) only ever goes up. The tool enforces this for you: cmd_validate writes the round-0 baseline (clean_items.json + failures.json); incorporate then owns the per-round accumulation — it rewrites clean_items.json to the grown clean set and failures.json to the shrunken pool of items still failing, so the next round's retry fairy is dispatched on retry_payload.json (= exactly that round's bad-only pool). Do not hand-reset, merge, or re-incorporate a stale clean_items.json / failures.json between rounds — that re-introduces the round-0-snapshot regression (round N silently dropping rounds 1…N-1's fixes; clean count going backwards). Just loop incorporate and trust the files it persists.
The incorporate subcommand auto-builds / accumulates:
clean_items.json — cumulative clean set (prior rounds' fixes ∪ this round's); monotonic non-decreasing. Read as prior_clean by the next round.failures.json — shrunken failure pool: only this round's still-failing items (cmd_validate's failures schema). Read as prior_failures by the next round.retry_payload.json — items still failing (retryable validator errors); feeds the next loop iteration's fairy. Deleted (not emptied) when all items resolve, so [ -f retry_payload.json ] is the correct exit test.unfixable.json — items with structural failures (invented IDs, malformed output); cumulative across rounds (terminal items drop out of the active pool, so they must persist here).- Updates
attempt_count in manifest.json each round.
Loop exit conditions: retry_payload.json absent (all resolved) OR attempt_count >= MAX_RETRIES. On the MAX_RETRIES exit, final_payload.json carries the cumulative clean set, and retry_payload.json + unfixable.json together hold the deferred remainder (→ next cycle's cohort via attrition).
This is the wait-then-spawn pattern using filesystem-as-channel. Claude Code's standard subagent mode does not support cross-agent SendMessage (gated behind CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, which we don't enable per ENGRAM-isolation safety). Mid-execution streaming isn't an option; all fairy reports are collected before the master is spawned.
Validation failures are retried by re-dispatching a fairy, NOT by mechanical script fixes. Quality of recall_summary / recall_keywords is critical because they surface in recall_surface and many node-rendering tools. The retry sub-agent receives the same guidance plus a "previous attempt failed" block with verbatim validator errors. This is the load-bearing design decision (maintainer design); cohort_dispatch.py does not modify LLM output.
Step 7.5 — Apply the batch-summary payload directly (before spawning the dream master)
Skip this step if the recall-summary substrate is not deployed. If batch-summary fairies were skipped in Step 5 (no recall_summary column or no engram_set_recall_summaries tool), final_payload.json will not exist — do not attempt to apply it; proceed directly to Step 8.
if [ ! -f "$COHORT_DIR/final_payload.json" ]; then
fi
When final_payload.json exists, the parent applies recall summaries now — before spawning the dream master — so:
- The summaries are persisted even if the dream master times out
- The dream master's spawn prompt is leaner (no payload inline)
- The dream master focuses purely on consolidation (reflect → buckets → resolve/supersede/retract/derive → turn-advance → dream record)
Read $COHORT_DIR/final_payload.json and apply via the MCP tool:
engram_set_recall_summaries(payload_json=open(f"{COHORT_DIR}/final_payload.json").read())
(This is a Claude tool call, not a shell command. The payload is the {"summaries": [...], "failures": [...]} JSON produced by cohort_dispatch.py validate / incorporate.)
Log the result: X summaries applied, Y failures deferred to next cycle.
Step 8 — Spawn the dream master with the full report bundle
PRECONDITION (do not pass without it): all 8 dream-fairies dispatched (Step 6) and
their reports collected or explicitly marked timed-out/failed (Step 7), AND every
batch-summary chunk applied (Step 7.5). If you have NOT dispatched the fairies this cycle,
you are not at Step 8 — you are skipping the routine (see the invariant at the top). The
dream-master is spawned with the fairy bundle, never instead of it.
Once all fairy reports are collected (or marked timed-out) and the final payload is ready, spawn the master with everything in its initial prompt:
Agent(
subagent_type="engram-dream-master",
prompt="""You are the dream master for tonight's consolidation cycle.
DREAM-FAIRY REPORTS (full content on disk — read each before integrating):
- Category 1 (<CAT1_NAME>): ~/.engram/dream/<TODAY>-fairy-1-<SLUG1>.md — TL;DR: <bullets>
- Category 2 (<CAT2_NAME>): ~/.engram/dream/<TODAY>-fairy-2-<SLUG2>.md — TL;DR: <bullets>
- ... (categories 3-8 similar) ...
[Or "Category N (<NAME>) TIMED OUT — proceeding without" for any fairy that didn't return]
COHORT METADATA:
- Today's new nodes since prev sleep <PREV_SLEEP_TIMESTAMP>: <K> IDs
- Backfill from legacy NULL-summary: <L> IDs
- Total: <K+L> ≤ 50, split across <N> chunks of ≤15 each
- Batch-summary cohort dir: <COHORT_DIR>
- Recall summaries already applied by parent in Step 7.5: X applied, Y failures deferred to next cycle
- SSoT modules live at: <engram-alpha-repo-path>/tools/ (recall_summary_validator, recall_summary_payload)
Call engram_reflect first for your initial agenda. Read all 9 dream-fairy reports from disk. Call bucket_findings() from tools/dream_master_batch.py to partition all findings into action-type buckets (single operation — snapshots are pre-packed in each finding, no re-inspection needed). Merge bucketed agenda with engram_reflect items. Execute one bucket at a time in ALL_BUCKET_NAMES order (tools/dream_master_batch.py is canonical: resolutions → supersedes → retractions → new_derivations → lessons → cornerstone_moves → goal_tension_resolutions → edge_wiring → task_closures → coverage_gaps; unknown bucket last, log-only), calling check_snapshot_divergence before each MCP write. Log diverged findings in the dream record. Call engram_advance_turn when your completion checklist is satisfied. Write the dream record.
FINAL RETURN: dream record path + top-line counts (resolved / superseded / promoted / refuted / recall-summaries applied) + health score delta + flagged-for-user count. The parent relays this to the user verbatim.""",
run_in_background=true
)
The dream master starts with cold context — this spawn prompt is its only briefing. The dream-master spec at ~/.claude/agents/engram-dream-master.md carries the disposition + completion-criteria checklist + the "Recall-summary failures (from final_payload.json)" handling table.
Don't run a parent-side turn-advance wait-gate. The master tracks its own completion checklist and only fires engram_advance_turn when satisfied. Parent's role ends at orchestration + relay.
Step 9 — Receive the dream master's final return and relay to the user
The dream master returns:
- Dream record path:
~/.engram/history/dream/YYYY-MM-DD.md
- Top-line: nodes resolved / superseded / promoted / refuted (counts)
- Recall summaries: applied / fixed-at-spot / unfixable counts
- Health score delta
- Flagged-for-user count
Relay this to the user in 3-6 short bullets. Surface the dream-record path so the user can read the "Flagged for the user" section in the morning review. Then stop. The session is over.
Why Phase A is pre-turn-advance
Per the turn-as-cohort-plus-consolidation derivation: Nodes filed during Phase A are awake-state cognition reflecting on the just-completed cohort. They belong to THIS turn, not next. If Phase A fired post-turn-advance, the missed nodes would be artificially placed in the next cohort, misaligning the forgetting curve.
This is why the ordering is Phase A → Phase B, and why Phase A must complete before the dream master fires engram_advance_turn.
Guardrails
-
Parent does NOT call engram_reflect, engram_advance_turn, or write the dream record. Those are dream-master responsibilities. The parent's job is orchestration + relay.
-
Parent's context is light by design. Dispatch + routing messages costs little context. If your context fills up during the routing phase, nap first to clear it, then return and dispatch the full fairy set — do not reduce or drop fairies. The dream master's context is separate and unaffected by parent compaction. (See the Phase B mandatory invariant above: context pressure is never a reason to skip fairies.)
-
Don't retry an interrupted dream. If the dream master returns a partial result (some fairies timed out, completion checklist short-circuited under self-timeout), unfinished work goes into next cycle's cohort via attrition. The dream master logs unfinished items in the dream record. Don't try to "finish the work" in the parent context — that would re-introduce the discipline-blur this architecture exists to resolve.
-
The dream master inherits the parent's model. Its spec no longer pins model: opus — consolidation runs on the same model family as the dispatching agent (Lei's directive, 2026-06-10; issue #1027). Fairies always run on Sonnet regardless.
Relation to other routines
- Nap — fires multiple times per day at compaction boundaries; per-burst persistence + cross-compaction-scaffold prep. Sleep is once-daily, end-of-day.
- Dream master (
engram-dream-master) — owns the consolidation this skill orchestrates. Spec at ~/.claude/agents/engram-dream-master.md.
- Batch-summary fairies (
engram-batch-summary-fairy) — one-shot batch generators dispatched by the parent after cohort_dispatch.py prepare. Each fairy receives a short dispatch prompt naming its input payload path and output path; it Reads the payload file (≤15 nodes), generates recall_summary + recall_keywords, and Writes the output JSON. Tool list: [Read, Write]. No inline payload embedding in the dispatch prompt. Sole dispatch path for recall_summary generation in the sleep cycle. Spec at ~/.claude/agents/engram-batch-summary-fairy.md.
- cohort_dispatch.py — orchestration script in
tools/cohort_dispatch.py. Five subcommands: prepare (chunk cohort → per-chunk payload.json), verify-in (pre-flight integrity checks), validate (split clean vs failures, auto-write retry_payload.json), incorporate / incorporate-retry (merge retry output → final_payload.json + retry_payload.json + unfixable.json + attempt_count).
- Dream fairies 1-8 (
engram-dream-fairy) — read-only consolidation-suggestion scanners. Spec at ~/.claude/agents/engram-dream-fairy.md.
- End-of-day-detector hook — surfaces engram-sleep on wrap-up phrases; user "Yes" triggers this skill.
macOS auto-sleep timing
When auto-sleep is enabled (cadence.auto_sleep_enabled: true), the SessionStart
hook registers an in-session CronCreate (durable: false) that fires
/engram-sleep at cadence.auto_sleep_time. On a macOS laptop in default sleep
state (lid closed / idle / on AC), two sources of latency apply:
- CronCreate-fire latency — macOS suspends the system clock during DarkWake
sleep. The in-session CronCreate fires 30-90 min after the scheduled time once
the system wakes enough to process it.
- Per-tool-call DarkWake-wait latency — each MCP tool call during
consolidation waits for a brief DarkWake cycle to complete before the system
responds. Empirically ~13 min/tool call, so a full consolidation cycle
(which can involve many calls) may span several hours.
Correctness is preserved. The last-sleep-success.json marker + lock file
mechanism (and PR #428's stale-lock recovery for crashed runs) ensure the graph
stays correct regardless of when the wake fires or how long the run takes.
This is a timing-expectation issue, not a correctness issue. If precise
nightly timing matters, three mitigation tiers in increasing complexity:
- Tier 1 —
caffeinate -i: run caffeinate -i in a terminal before going
idle. Prevents automatic sleep; the in-session wake fires on schedule. Simple,
per-night, burns power while on battery.
- Tier 2 — System Settings: System Settings → Battery → Power Adapter →
"Prevent automatic sleep when the display is off". Persistent on AC;
the wake always fires on schedule when plugged in.
- Tier 3 — launchd plist (deferred, not yet shipped): a
launchd plist with
StartCalendarInterval + WakeFromSleep: true that wakes the system at the
configured time. Architecturally cleanest; opt-in setup is tracked as Action #3
in issue #407 for a follow-up PR.
Substrate anchor
The three-routine architecture (nap / bedtime / sleep) was ratified 2026-05-14 (the three-routine ratification). The constraints determining the original partition: (1) turn-advance fires at END of consolidation (turn-as-cohort-plus-consolidation derivation); (2) cross-compaction scaffolds belong at the compaction boundary (naps); (3) day-wide review producing new nodes is awake-state cognition (pre-turn-advance). At ratification, this was a three-routine model: nap + bedtime + sleep. As of 2026-05-24 (this merge), it is a two-routine model: nap + sleep, with sleep now having two internal phases (Phase A = former bedtime; Phase B = former sleep).
The dedicated-dream-master refactor landed 2026-05-19 to address:
- Awake-context cost: running sleep inline burned a significant portion of the parent's context on consolidation work, leaving less for the awake state. The master runs in its own context.
- Wait-for-fairies discipline misfires: the parent's "advance the turn after all fairies return" gate had been broken multiple times by post-compaction time pressure. Moving advance into the master removes the parent's role in turn timing.
- Disposition gap: an agent running sleep ALSO has awake-state pressures pulling them. The dedicated dream master is summoned with a single role and a single disposition (engram health custodian) — no role conflict.
The 2026-05-24 merge of engram-bedtime into engram-sleep eliminates the structural failure mode of stopping after Phase A but before Phase B: a compaction, user interrupt, or fairy dispatch error between the two former skills left the day's missed nodes filed but the turn never advanced, conflating the next day's cohort with today's. By making Phase A and Phase B a single skill invocation, that failure mode is structurally impossible.
