sleep-consolidate · diff
git:20261114.da9d227 to git:20261115.c017ea4
6 added, 3 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
---
name: sleep-consolidate
description: >-
Nightly sleep-time memory consolidation ("dreaming") workflow. Audits Vault memory
health, verifies turn journal claims against ground-truth transcripts, distills deep
lessons into local/memory/topics/*.md, keeps local/MEMORY.md as a compact Roadmap
Index, prunes stale facts via hq memory, stages skill proposals, and commits
the private local/ Vault git repo.
---
# Nightly Dreaming & Sleep-Time Memory Consolidation (`/sleep-consolidate`)
> **Architectural Purpose (RFC-0006 & RFC-0014):** Do not treat `local/MEMORY.md` as an append-only dump or blindly trust self-reported 1-line turn summaries. Verify what actually happened against ground-truth transcripts, prune obsolete tasks and contradicted assumptions, distill deep technical/domain knowledge into Topic Wiki pages (`local/memory/topics/<slug>.md`), maintain on-demand Entity Dossiers (`local/memory/entities/<slug>.md`) at 0 `SessionStart` token cost, and keep `local/MEMORY.md` strictly as a concise Roadmap Index ($\le 80$ lines). Default execution tier is **Gemini 4 Carbon (Medium)** (`gemini-4-carbon-medium` / `--tier=deep-medium`).
> [!IMPORTANT]
> **Silent Execution & Memory Operations Invariant:** Automated nightly sleep consolidation (`sleep-consolidate` / `[Auto] Nightly Sleep Consolidation`) runs **100% silently in the background**. Never emit chat messages, checklists, progress updates, or completion banners into the Front Desk conversation. All background audits, transcript verification findings, pruned facts, topic/entity updates, and RSI proposals remain strictly confined to the private Vault (`local/`), ledger, and turn journals. Never mention internal bookkeeping, ledger writes, or file paths (e.g., `project-notes.md`, `facts.jsonl`, `MEMORY.md`, `tasks.json`) in conversational replies. Any critical health issues, distilled insights, or pending approvals roll forward into the next morning's 7:30 AM Morning Brief (`/morning-brief`).
## Four-Stage Dreaming Workflow
+ 0. **Dispatch & Overlay Guard:**
+ - Run `hq dispatch resolve --skill=sleep-consolidate --trigger=core/sleep-consolidate --json`: if `enabled == false`, stop cleanly with zero output; if `has_overlay == true` with non-empty `overlay_rules`, apply those active personal rules for this run.
+
### Stage 1: Baseline Memory Audit, Subagent DAG Reconciliation & Stale Reference Check
1. Run `hq obs reconcile` and `hq memory status` to inspect:
- Ground-truth parent-child subagent attribution in `local/memory/tasks.json` and 24h token/context anomalies (`hq obs tokens --since=24h`).
- Current line and token usage of `local/MEMORY.md` against the **80-line Roadmap Index cap**.
- Existing Knowledge Wiki pages (`local/memory/topics/*.md`), Entity Dossiers (`local/memory/entities/*.md`), active vs. expired bi-temporal facts, and any **Stale References**.
### Stage 2: Episodic Sweep & Ground-Truth Transcript Verification
1. Inspect today's turn journals in `local/memory/journal/YYYY-MM-DD/*.json` and recent entries in `local/memory/interventions.jsonl`.
2. **Ground-Truth Verification Rule:** Whenever a journal entry shows `toolErrorCount > 0`, a user correction/intervention, or an architectural/root-cause/completion claim, run:
```sh
hq memory verify-transcript <conversationId>
```
Cross-check the 0-token ground-truth digest (`errors`, `discrepancyWarnings`, `userInputs`, `filesModified`, `commandsExecuted`) against what the agent claimed before promoting any lesson into long-term memory.
### Stage 3: Proactive Pruning, Topic Wiki, Entity Dossiers & Roadmap Index Update
1. **Prune Stale Facts & Completed Priorities:**
- Soft-expire superseded, contradicted, or completed one-off facts via `hq memory prune-fact <fact-id> [--superseded-by=<new-id>]` (and check off finished goal milestones via `hq goals list`).
- Remove completed one-off tasks or deleted worktrees from `## Active Priorities` in `local/MEMORY.md`.
2. **Distill Detailed Knowledge into Topic Wiki Pages (`local/memory/topics/<slug>.md`):**
- Put multi-paragraph architectural notes, subsystem invariants, debugging lessons, or domain playbooks into `local/memory/topics/<slug>.md` (beginning with `# <Topic Title>` and `> **Summary:** <1-line description>`).
3. **Reconcile On-Demand Entity & Counterparty Dossiers (`local/memory/entities/<slug>.md`):**
- When verified turns reveal durable collaborator/counterparty roles, handles, spaces, or working-style preferences, upsert them via:
```sh
hq memory add-entity --name="<Name>" --kind="person" --email="..." --space="..." --context="..." --note="..."
```
- Entity dossiers carry **0 bytes of `SessionStart` token overhead** and are retrieved on-demand via `hq memory lookup <query>`.
4. **Update `local/MEMORY.md` ($\le 80$ lines) & Record Verified Atomic Facts:**
- Apply incremental delta edits to `local/MEMORY.md` (`## Active Priorities`, `## Durable Preferences & Facts`, `## Knowledge Map (Topic Wiki)`).
- Add verified atomic facts via `hq memory add-fact --topic="<topic>" --statement="<concise fact>" --source="<convId>" --verified`.
### Stage 4: Verified Recursive Self-Improvement (`hq rsi`) & Private Vault Commit
- 1. **Run the 4-Stage RSI Pipeline (RFC-0010 & RFC-0014 — Never Unilaterally Edit `SKILL.md` or Kernel Code):**
+ 1. **Run the Two-Tier 4-Stage RSI Pipeline (RFC-0010, RFC-0014, & RFC-0019 — Never Unilaterally Edit `SKILL.md`, Overlay Rules, or Kernel Code):**
- Mine 0-token ground-truth friction events (error-recovery chains, `PreToolUse` denials, user interventions, `PRELUDE` draft edit-diffs, and `proactive_noise` idea dismissal rates $>70\%$) and update active canary records:
```sh
hq rsi mine --since=24h
```
- - Synthesize structured ACE delta hypotheses (`hq rsi hypothesize`) and run the 4-Gate Sandboxed Verification Harness + stage verified passes:
+ - Synthesize structured Tier 1 personal hypotheses targeting `<vault>/skills/<slug>/SKILL.md` or `<vault>/memory/rsi/rules/<slug>.md` (`0` dirty root git files; opt-in maintainers pass `--upstream` for Tier 2 upstream worktree PRs) and run the 4-Gate Sandboxed Verification Harness:
```sh
- hq rsi verify <rsi-id> && hq rsi propose <rsi-id>
+ hq rsi hypothesize && hq rsi verify <rsi-id> && hq rsi propose <rsi-id>
```
- When upgrading to a new frontier model, run `hq rsi ablate --model=<new-model>`.
2. **Task Lifecycle Consolidation, Hot/Cold Rotation & Sidebar Archive Sync:**
- Run `hq tasks compact` (automatically rotates eligible completed tasks/subagents to cold storage, archives their threads in Antigravity's sidebar, and drains any queued offline transitions; use `hq archive <slug>` if explicit worktree cleanup is needed).
3. **Verify Budget, Commit Vault & Silent Completion:**
- Run `hq memory compact`, `hq memory status`, and `hq rsi status` to confirm `local/MEMORY.md` is $\le 80$ lines (`[OK]`).
- Commit changes to the private Vault repository (workspace-agnostic across `~/.gemini/antigravity/hq/vault` and `./local`):
```sh
hq memory commit --message="chore(memory): nightly dreaming consolidation YYYY-MM-DD"
```
- **Silent Turn Finish:** End the turn cleanly without emitting chat responses, status banners, or completion summaries into the Front Desk thread.