dreams

name	dreams
description	Review brain inbox drafts, preserve durable memory, archive evidence, sync GBrain, and optionally commit. Public workflow defaults to no push.
triggers	["$dreams","dreams","review brain inbox","promote memory drafts","review gbrain drafts"]
mutating	true

Dreams

Use $dreams to run the reviewed memory maintenance cycle for a user's own brain repository. The brain markdown repo is the source of truth; GBrain is the generated index/search layer.

Public defaults:

• brain root: explicit --brain <path> or detected git brain repo
• apply: yes after validation
• sync: yes unless --no-sync
• commit: yes unless --no-commit
• push: no unless --push is explicitly passed

Never index inbox/, archive/, raw/, hook payloads, transcript dumps, review ledgers, .dreams/ state, or source snapshots as normal semantic memory. They are filesystem/git provenance only. Only reviewed curated pages in allowlisted namespaces should enter the generated index.

Public Contract

Draft review and indexing are separate stages:

• Inventory candidate drafts under the configured brain repo's inbox/, including inbox/auto/YYYY-MM-DD/*.md created by stop hooks.
• Read the brain repo rules before classifying drafts: README.md, AGENTS.md, schema.md, RESOLVER.md, and the target directory README.md.
• Classify every draft with a memory handling mode: preserve-full, preserve-substantial, synthesize, source-note, discard, duplicate, or blocked-safety.
• Promote only safe, durable, reviewed material into curated allowlisted pages.
• Preserve important user-authored reasoning when fidelity matters. Do not flatten original theses, decisions, lessons, or distinctive wording into a generic summary.
• Archive reviewed source snapshots for provenance after ledger validation.
• Run redaction, review-ledger, session-owner, session-state, curated-page, and index-allowlist checks before declaring success.
• Sync after writes unless --no-sync or --dry-run.
• Commit after validation unless --no-commit or --dry-run.
• Do not push unless --push is explicitly passed.

Memory Boundary

Only these namespaces are semantic memory by default:

• people/
• companies/
• projects/
• decisions/
• gotchas/
• concepts/
• meetings/
• originals/

Everything else is evidence unless the operator intentionally changes the brain schema. If an indexed result comes from inbox/, archive/, raw/, source snapshots, scripts, reports, or ledgers, treat it as contamination and fix the indexing boundary rather than relying on it.

Safety Rules

• Never promote secrets, credentials, private keys, raw connection strings, unredacted local paths, raw access instructions, or unredactable sensitive data.
• Never copy raw stop-hook metadata such as transcript_path, git_remote, raw transcript text, or local evidence paths into curated pages.
• If a draft is unsafe, mark it blocked-safety and leave it for manual recovery rather than archiving or promoting it.
• If a fact is useful but weakly supported, promote only with low confidence, explicit provenance, and a narrow claim.
• If a draft is duplicate, point at the existing curated target instead of creating a competing page.
• Safety beats recall. A smaller safe memory is better than a searchable leak.

Importance Scorecard

Use importance_score to decide preservation depth. Score each dimension 0-3:

• future_retrieval: how likely future agents/users will need this
• durability: whether it remains useful beyond the current session
• specificity: whether it contains concrete names, commands, dates, paths, or constraints
• authorship: whether it contains user-authored thinking or decisions
• fidelity: whether exact wording/rationale matters

Guidance:

• high total or explicit full-preserve signal -> preserve-full
• concrete durable fact/decision/gotcha -> preserve-substantial
• useful but broad context -> synthesize
• weak provenance but useful source pointer -> source-note
• routine status/log noise -> discard
• already represented in curated memory -> duplicate
• unsafe or unredactable -> blocked-safety

Review Ledger

Every handled draft must produce a review-ledger row before source archiving or finalization. The row is the auditable bridge between unreviewed draft evidence, curated page edits, and later retrieval checks.

Each non-safety row needs:

• decision and preservation_mode
• importance_score
• evidence_quality and concrete quality_reason
• non-empty memory_units
• coverage explaining represented and discarded content
• targets for any curated memory change
• target_actions describing create/update/append/noop work
• sanitized provenance such as source_anchor and source_excerpt_summary

Each memory_units[] item needs:

• summary
• status: durable, source-note, duplicate, discard, blocked-safety, or status-noise
• source_anchor: a short sanitized pointer to the reviewed draft section
• importance_score
• preservation_mode
• target linkage when it contributes to curated memory

For stop-hook sessions, source_anchor should include the relevant segment-N or draft-M:segment-N marker when the draft has multiple capture segments.

Session Owners

For stop-hook session_id groups, use one session-owner subagent per session group when the host supports native subagents. This is the no-loss path for large agent sessions.

• Every session-owner spawn must request reasoning effort: xhigh. If the host uses multi_agent_v1.spawn_agent, pass reasoning_effort: "xhigh" and omit model unless the operator explicitly requested a model change.
• Spawn session owners as a bounded rolling queue: at most 6 session-owner subagents may be live at once. Start up to six, then wait for a report to be written and accepted before opening the next owner.
• Never spawn all pending owner tasks at once.
• Each owner must write the required JSON report to the assigned report_path. A chat-only summary is not accepted coverage.
• Ledger rows for session drafts must include session_owner_report with session_id, agent_id, report_path, report_sha256, and covered_review_units.
• validate-session-owner-reports.rb and validate-review-ledger.rb must both accept the binding before finalization.

For huge sessions, segment deterministically into segment-1..segment-N. Non-durable segment content still needs coverage so important content does not disappear silently.

Proposal And Evidence Gates

Before applying curated changes, create proposal mirrors under the session's proposals/curated/ tree for every curated target. The proposal mirrors are the reviewable full-file output for changed pages.

After applying, require:

• proposal mirrors match current curated files
• each changed target has a timeline/source entry
• indexed_at is empty or pending before sync and stamped after successful sync
• archive/dreams-session-audits/<session-id>/session-owner-audit.json exists for session-owner runs
• gbrain evidence index-ledger --brain-repo "$BRAIN" --source <source_id> --session "$SESSION" --include-archive-excerpts updates the source_evidence lookup layer for curated rows only
• at least one changed curated target is discoverable with get_evidence({target_slug}) or the CLI equivalent

Evidence indexing must not make inbox/ or archive/ semantic memory. It is a target-slug provenance lookup for reviewed curated claims.

Helper Scripts

Use the shipped helpers from this skill directory with explicit paths:

ruby skills/dreams/scripts/init-session-state.rb --brain "$BRAIN" --no-push
ruby skills/dreams/scripts/inventory.rb --brain "$BRAIN" --session "$SESSION" --copy-sources
ruby skills/dreams/scripts/prepare-session-owner-tasks.rb --session "$SESSION" --inventory "$SESSION/inventory.json"
ruby skills/dreams/scripts/validate-session-owner-reports.rb --session "$SESSION" --inventory "$SESSION/inventory.json"
ruby skills/dreams/scripts/validate-review-ledger.rb --brain "$BRAIN" --session "$SESSION" --require-proposal-artifacts
ruby skills/dreams/scripts/archive-reviewed-sources.rb --brain "$BRAIN" --session "$SESSION"
ruby skills/dreams/scripts/export-session-audit.rb --brain "$BRAIN" --session "$SESSION"
ruby skills/dreams/scripts/redact-staged-artifacts.rb --brain "$BRAIN" --apply
ruby skills/dreams/scripts/sync-temporal-metadata.rb --brain "$BRAIN"
ruby skills/dreams/scripts/stamp-indexed-at.rb --brain "$BRAIN" --timestamp "$INDEXED_AT"
ruby skills/dreams/scripts/check-index-allowlist.rb
ruby skills/dreams/scripts/finalize-session-state.rb --brain "$BRAIN" --session "$SESSION"
ruby skills/dreams/scripts/validate-session-state.rb --brain "$BRAIN" --session "$SESSION"

For public installations, --push is a separate explicit operator decision. Readiness, passing tests, or a completed local commit do not authorize remote publication.

Verification

Before reporting completion, run the relevant local gates:

ruby skills/dreams/scripts/test-inventory.rb
ruby skills/dreams/scripts/test-finalize-session-state.rb
ruby skills/dreams/scripts/test-validate-review-ledger.rb
ruby skills/dreams/scripts/test-validate-session-state.rb
ruby skills/dreams/scripts/test-session-owner-workflow.rb
ruby skills/dreams/scripts/test-check-index-allowlist.rb
ruby skills/dreams/scripts/test-redact-staged-artifacts.rb

If the run changed public package contents, also run the repository's public Dreams/hook tests and package/privacy checks. Do not claim the public surface is ready until the tests and privacy scan are fresh.