remsleep

name	remsleep
description	End-of-day multi-phase consolidation. Cleanup, replay, synthesis, reflection. Surfaces patterns the user wouldn't catch in real-time.

/remsleep

Purpose

Close the day by reviewing what happened, promoting knowledge that's now solid, reconciling hubs, and writing reflection prompts. Multi-phase, dispatches sub-agents for parallel work.

When to use

• End of day. Manually invoked.
• Weekly variant: Sunday or Monday morning (deeper synthesis).

Hard preconditions

• /lint-wiki must have run today. If it hasn't, fail fast and ask the user to run lint first. Curation on a broken vault is wrong.

Arguments

• --weekly (optional) — deeper synthesis mode for end-of-week. Reads a 7-day window instead of 1-day.
• --for=<YYYY-MM-DD> (optional) — run as if today were that date. Useful for catching up on a missed day.

Phases

The skill runs four phases. Phases 1–3 can dispatch sub-agents in parallel; Phase 4 runs after.

Phase 1 — Cleanup

1. Verify /lint-wiki ran today. If not, halt with a clear error.
2. Scan inbox/meetings/ for unprocessed transcripts (files not in inbox/archived/). If any, invoke /ingest-meeting for each.
3. Scan any annotation sources (inbox/<tool>/annotations/YYYY-MM.jsonl) for unresolved annotations. If any, invoke /ingest-source annotations.
4. Enumerate today's /learn output by reading entries dated today in logs/changelog.md (look for [YYYY-MM-DD] enrich | /learn headers). Collect the set of pages, tasks, and memory events touched.

Phase 2 — Replay (sub-agent)

Dispatch a sub-agent to investigate today's activity, in parallel:

• Read today's logs/changelog.md entries.
• Read any meeting summaries in logs/meetings/ dated today.
• Read git activity for any repos under repos/ (last 24 hours).
• Read today's logs/daily/YYYY-MM-DD.md if present.

The sub-agent extracts:

• Decisions made (explicit or implicit).
• Repeated topics.
• Priority shifts.
• Missed action items.
• New context about people / projects / tools.

Return a structured summary to the orchestrator.

Phase 3 — Synthesis (sub-agent)

Dispatch a sub-agent to do the curation work:

1. Review /learn output from Phase 1. For each needs-review page touched today:

   • Is it well-linked? (Search for backlinks.)
   • Is it substantive (not a stub)?
   • Does it overlap with an existing page?
   • Recommend: keep at needs-review / promote to stable / merge with [[other-page]] / archive.

2. Hub reconciliation. Scan the last 7 days of activity for entities (tools, vendors, frameworks, projects) that should be on a hub but aren't. For each gap:

   • Suggest the hub to update.
   • Suggest the entry text.
   • Flag hubs with robin:last_reconciled older than 14 days.

3. Skill candidate detection. Walk Phase 2's output for workflows the user did 3+ times the same way. Flag (do NOT auto-create) as skill candidates with brief description.

4. Long-range link suggestions. Identify pages that should reference each other but don't (e.g., a new decision that references a project — the project page should link the decision).

5. Conflict surfacing. Identify cases where /learn produced a fact that conflicts with an existing fact. Recommend append-with-history (NEVER auto-resolve).

Return a curation report.

Phase 4 — Reflection

Compose reflection prompts and write to brain/about_user/reflections/questions.html. Behavior depends on self_reflection: config and mode:

• self_reflection: false → skip Phase 4 entirely.
• self_reflection: true, daily mode → 1 terse question for tomorrow morning. One paragraph max.
• self_reflection: true, weekly mode → 2–3 deeper questions. Multi-paragraph context for each.

Prompts are targeted — they reflect the day's actual content (decisions, conflicts, tensions noted in Phase 2). Generic prompts are not useful.

Append to the questions file (don't overwrite — keep history).

Compose the report

After all phases:

1. Write logs/remsleep/YYYY-MM-DD.html (Robin v0.2, robin:type=remsleep, robin:date=YYYY-MM-DD). The report includes:

   • Phase 1: cleanup summary.
   • Phase 2: today's replay (decisions, priorities, etc.).
   • Phase 3: curation recommendations (with the user's decision pending for promotions/merges).
   • Phase 4: reflection prompts written (link to questions file).

2. Append to logs/changelog.md:

   ## [YYYY-MM-DD] remsleep | <one-line summary>
   

3. Emit a compact summary in chat:

   • What was cleaned up.
   • Key replay findings.
   • Curation recommendations awaiting confirmation.
   • Reflection prompt(s) written.

Constraints

• Additive only. Never delete. When facts conflict, append with history markers.
• Suggest, don't decide. Promotion candidates surface for user decision (or, if user has explicitly authorized auto-promotion of well-linked pages, act — but never silently delete or merge).
• Provenance preserved. Every change writes a robin:source if the change came from a specific source.
• Multi-agent. Use sub-agents for Phases 2 and 3 to keep the main context light.

Output shape

The chat summary is concise (under 30 lines). The full report is in logs/remsleep/YYYY-MM-DD.html and the chat summary links to it.

Edge cases

• No activity today. Run anyway. Phase 1 may have annotation work to do. Phase 4 can still produce a prompt. Output reflects the quiet day.
• /lint-wiki not run. Halt with a clear instruction: "Run /lint-wiki first."
• Conflicts surface. Surface them. The user resolves; never auto-merge.
• Skill candidate detected but uncertain. Flag with low confidence. The user decides whether to promote to a skill.