dream-consolidate

name	dream-consolidate
description	Use when a scheduled or manual Dream Loop pass should refresh ACTIVE.md, strengthen LEARNINGS.md, and drain unresolved inbox signal.

Dream Consolidate

Use during off-hours or explicit maintenance. It refines public memory and resolves leftover uncertainty.

Use This Skill When

• the user asks for memory maintenance
• an automation reviews unresolved inbox evidence
• routes need refresh, promotion, rejection, or archive

Read These Files

• the active memory root's inbox/
• the relevant slices of the active memory root's ACTIVE.md
• the relevant slices of the active memory root's LEARNINGS.md
• the active memory root's ARCHIVE/ when needed for lineage

Main Actions

1. review unresolved inbox evidence and current public memory
2. move misplaced explicit strong signal out of inbox/ immediately
3. refresh or retire stale ACTIVE.md entries
4. strengthen LEARNINGS.md with validated winning routes
5. archive losing, stale, rejected, or superseded routes with source trace
6. stage judgment-heavy proposals before landing them
7. append an audit report with route rationale

Public-Memory Decisions

Refresh ACTIVE.md for hot, temporary, or immediately behavior-changing guidance.

Strengthen LEARNINGS.md for reusable winning routes with clear evidence.

Keep only inferred, ambiguous, weak, or still-competing signal in inbox/.

Auto-land from inbox/ only when the item survived one automation cycle, has no contradictory source or reviewer objection, has source trace, is executable, and has a clear destination.

Archive routes that are no longer hot, replaced, unsupported by evidence, rejected, or still unpromotable after review.

Validation Gate

Before promoting a candidate into LEARNINGS.md, check:

• source evidence: the candidate has a concrete task, correction, audit finding, or repeated failure behind it
• outcome link: the candidate explains how it would improve or stabilize a future run
• blast radius: the candidate is narrow enough to avoid changing unrelated tasks
• rejection condition: the audit names what would make the candidate wrong or stale
• rollback path: the audit preserves enough context to retire or reverse the route later

If the evidence is real but still weak, keep the item in inbox/ or stage it as a proposal. Do not promote it just because it is plausible.

For judgment-heavy changes, produce a staged proposal with accepted edits, rejected candidates, target layer, evidence, reviewer/subagent notes, and rollback notes. Staging is not a public memory layer; it is an audit checkpoint before changing ACTIVE.md or LEARNINGS.md.

Hard Constraints

• do not silently edit AGENTS.md; report any needed AGENTS.md change to the user before or while applying it
• keep policy-like AGENTS.md changes proposal-first and human-reviewed
• never invent learnings without source trace
• never silently delete evidence; archive it
• every promotion, rejection, or archive must be explainable
• age alone is not enough for promotion
• validation-gated promotion beats plausible self-revision
• staged proposals must not become a third daily memory surface
• keep one final winning route when routes compete
• use reviewer or subagent cross-checks for promotion, rejection, archive, or conflict decisions
• use a single-agent fast path only for low-risk cleanup with no meaningful judgment call

Required Output

Report files read/changed, ACTIVE.md and LEARNINGS.md decisions, validation-gate evidence, staged proposals if any, rejected alternatives, reviewer or subagent resolution, archive actions, and remaining gaps.