dream-state

Dream State Skill

Automated architecture maintenance — scan every memory file, validate structural integrity, produce a diagnostic report. Dream observes and measures; it never modifies.

When to Use

• User asks to "dream" or run "architecture maintenance"
• After file reorganizations that may break connections
• Before major meditation sessions (establish baseline)
• After domain learning to validate new connections
• When architecture health is uncertain before skill selection

6-Phase Protocol

Phase	Action	Output
1. Discovery	Scan .github/ for all memory files	File inventory
2. Validation	Parse frontmatter, verify trifecta completeness	Issue list
3. Repair	Auto-fix via consolidation mappings (renames only)	Repair log
4. GK Sync	Check %OneDrive%/AI-Memory/ for cross-platform content	Sync status
5. Reporting	Generate health report with statistics	Markdown report
6. Display	Show results, open report in VS Code	Notification

Health Metrics

Metric	Healthy	Concern
Broken connections	0	Any
Missing frontmatter	0	Any
Incomplete trifectas	0	Any high-use skills
Brand violations	0 (outside exceptions)	Any
Version drift	All match	Any mismatch
Review queue size	0–5 pending	>10 pending (semantic debt accumulating)
Review queue age	Oldest <30d	Oldest >60d (stale reviews indicate blocked curation)

Review-Queue Backpressure (FM10)

The review queue (.github/quality/review-queue.jsonl, defined in FM5) is a leading indicator of semantic debt. When the queue grows faster than it's resolved, the architecture is accumulating unreviewed changes.

Signal	Threshold	Health Impact	Action
Queue size growing	>10 pending entries	needs-attention	Prioritize review session; schedule meditation
High-priority entries stale	Any high priority entry >14d unresolved	needs-attention	Escalate: these are heir-feedback or drift-detected entries
Queue empty	0 pending	healthy	Normal state; curation is keeping pace
Oldest entry age	>60 days	needs-attention	Something is blocking reviews; investigate

Per-volatility weighting: High-volatility files (instructions, skills with frequent edits) generate more queue pressure than stable reference docs. Weight queue entries by the source file's edit frequency when computing backpressure score.

Integration: Dream protocol reads review-queue.jsonl during Phase 2 (architecture scan) and includes queue metrics in the dream report's health assessment.

Ritual Hierarchy

Meditation is the foundational ritual. Dream is a diagnostic that runs subordinate to it.

	Dream	Meditation
Role	Diagnostic tool	Foundational ritual
Mode	Automated, unconscious	Interactive, conscious
Creates files?	Reports only	Memory files + insights
Makes decisions?	Never	Yes
Activation	/dream, or chained after meditation	User says "meditate"
Analogy	Blood test	Doctor's visit

When Dream Chains After Meditation

Dream triggers automatically after meditation when:

Trigger	Condition
Pattern: file reorg	Meditation touched 3+ architecture files
Pattern: trifecta concern	Meditation created or modified skills/instructions
Pattern: staleness	Last dream was >7 days ago
Random	~1 in 5 meditations (keeps architecture fresh)

Connection Format

Skills connect via frontmatter applyTo patterns and semantic search:

---
name: "skill-name"
description: "What this skill does"
applyTo: '**/*pattern*,**/*files*'
---

Valid relations: applies-to, extends, requires, contradicts, supersedes

Report Template

# Dream Report — YYYY-MM-DD

## Summary
- **Files Scanned**: N
- **Trifecta Completeness**: N/N
- **Broken Links**: 0
- **Brand Violations**: 0
- **Status**: HEALTHY / ATTENTION REQUIRED

## Issues Found
| File | Issue | Severity |
|------|-------|----------|

## Recommendations
Prioritized list of manual fixes needed.

Interpreting Results

Status	Meaning	Action
HEALTHY	All checks pass, no broken connections	None — architecture is sound
ATTENTION REQUIRED	One or more issues detected	Review issues table, prioritize by severity

Severity levels: critical (broken trifecta), warning (missing frontmatter field), info (cosmetic).

Escalation: When Findings Need Human Judgment

Dream is a diagnostic — it never modifies files. When findings require judgment (skill consolidation, content quality decisions, contradiction resolution), the natural next step is meditation, which reads the dream report and decides what to fix.

Finding type	Resolution path
Missing frontmatter, broken link with mapping	Author the fix in the next meditation; mechanical
Two skills with overlapping applyTo	Meditation decides merge / split / keep; semantic
Skill flagged as shallow	Meditation enriches or deprecates; requires domain context
Trifecta where instruction contradicts skill	Meditation reconciles intent; requires understanding
Version drift master/heir	Meditation picks canonical; project-specific

The handoff is the dream report file. Meditation's Resolve step explicitly evaluates whether a dream should chain after it; the inverse — meditation reading an existing dream report — happens naturally when a meditation session opens with "check the architecture first."

This pattern replaces the former lucid-dream ritual (retired v8.4.0). The escalation concept lives here because dream is the producer of the findings; meditation is the consumer that decides.

Snapshot Schema (.github/quality/dream-report.json)

dream-cli.cjs writes a single structural JSON snapshot every run. Heirs and graders depend on this shape; preserve it across versions.

Field	Type	Purpose
schemaVersion	number	Currently 1. Bump only with a heir migration plan.
timestamp	ISO string	Last dream run; consumed by proactive-awareness and health-pulse.
workspace	string	Absolute path of the dreamed workspace.
inventory.skills / .instructions / .prompts / .agents / .total	numbers	File counts under .github/.
trifectaIssues[]	{type, skill, detail}	Workflow skills missing matching .instructions.md or .prompt.md.
brokenRefs[]	{sourceFile, line, target}	Markdown / config references that no longer resolve.
health	"healthy" | "good" | "needs-attention"	Derived bucket: 0 / ≤5 / >5 issues.

Stability rule: meditation, dream-creativity-score.cjs, and heir consumers (health-pulse, proactive-awareness.instructions.md) all read these field names directly. Adding fields is safe; renaming or removing requires a fleet-wide migration.

Chronicle Pruning

dream-cli.cjs keeps the newest 50 dream-report-YYYY-MM-DD*.md chronicles in .github/episodic/ after each run; older ones are deleted in place. Rolling architecture history lives in the JSON snapshot and brain-qa outputs — chronicles are only useful as recent narrative context for meditation.