heartbeat

name	heartbeat
description	Autonomous batch orchestrator — fills the pipeline, dispatches all board tasks as parallel sub-agents, triages inbox, consolidates knowledge, archives stale work, logs to agent log, and git syncs. Use when the human says "run a heartbeat", "do housekeeping", or on any scheduled autonomous run.

heartbeat

The autonomous batch orchestrator. Bootstraps context, fills the board pipeline, dispatches ALL in-progress tasks as parallel sub-agents, triages inbox, consolidates knowledge, archives stale work, logs to agent log (inbox/log/YYYY-MM-DD), and git syncs. Adapts to whatever vault structure and conventions the human chose — always read the current AGENTS.md for the actual layout. Never write to the human's daily/ notes — they are pure date anchors whose value is in backlinks.

When to use

• Scheduled autonomous runs (hourly or configured interval)
• "Run a heartbeat", "do housekeeping", "triage my inbox"
• After a burst of work to clean up and consolidate
• This is the unattended mode — no human in the loop

Inputs

Input	Required	Description
scope	no	full (default), tasks, inbox, knowledge, health. Limits which steps run.

Process

STEP 1 — Bootstrap context

Start from [[AGENTS]] and follow its indexes to discover the current workspace:

• The root AGENTS.md — the full operating manual and source of truth for vault layout
• Folder indexes — each folder's AGENTS.md describes its contents and purpose
• The skills index — available skills and triggers
• The board — wherever the human's task board lives
• The knowledge map — current graph entry point

Don't hardcode paths. The human may have reorganized. Let AGENTS.md tell you where things are.

STEP 2 — Fill the pipeline

The board should always have work flowing.

A) Inbox → Board: Read the inbox folder. For any item that is a task, add a card to the board under Todo. Non-task items are triaged in Step 4.

B) Todo → In Progress: If In Progress is light, pull top cards from Todo. The right WIP limit depends on the human's capacity — start with 3, adjust based on throughput patterns.

After this step, In Progress should have cards (unless the entire board is empty).

STEP 3 — Dispatch all tasks (parallel)

This is the heartbeat’s main job. Dispatch ALL In Progress tasks as parallel sub-agents.

Parallelization rules:

• Every unrelated task gets its own sub-agent running concurrently.
• Serialize tasks onto the same thread ONLY when they belong to the same project or have true data dependencies — one sub-agent handles them in sequence so context carries over.
• Maximize throughput. All In Progress tasks should progress simultaneously, not one at a time.

Each sub-agent:

1. Reads the task’s card and any linked notes for context
2. Does the actual work (research, build, organize, process). Uses appropriate skills.
3. Updates the task with a ## Progress entry (most recent at top): what changed, why, how validated, what remains.
4. Moves the card: → Done (append @{YYYY-MM-DD}) or → Blocked (explains why). Each card should link to its progress log.
5. Creates a progress log for the task (in the vault's log location) with what was done, evidence links, and what remains.

Planning (complex tasks only): If a task needs multiple steps or internal parallelism, create a plan in .plans/<task-slug>.md before dispatching. Simple tasks skip planning.

Skip cards that clearly need human input — leave them in Todo with a comment after doing preliminary research.

STEP 4 — Triage inbox

A) Quick capture drops. Check the human's quick capture surface for unchecked items. Process and check them off.

B) Inbox files. Read every file in the inbox (skip AGENTS.md, skip items already on the board).

For each item:

1. Assess what it is and route it per the "What goes where" table in [[AGENTS]]. Use the vault's actual folder layout.
2. Rich content gets priority — fetch full content, enrich, extract insights, connect to graph. Use appropriate skills.
3. Preferences or self-knowledge → meta folder. If a preference changes a convention in AGENTS.md, update the protocol too.
4. Preserve provenance: Add processed_to: "[[destination]]" to the inbox item, then move to the vault's processed location.
5. Update folder indexes in any folder that gained or lost files.

After this step, inbox should contain only AGENTS.md and items <48h old that need human context.

STEP 5 — Knowledge consolidation

Skip if entity folders have fewer than 5 non-index notes total.

1. Split monoliths: Scan for notes covering multiple concepts. Split each idea into its own atomic note, then replace the original with a hub that embeds the atoms. Every split increases the graph's connectable surface area.
2. Extract unattributed atoms: Find quotes, findings, or claims embedded in longer notes without their own note or ^block-id. Give each one a block ID or its own note so it's independently linkable. A quote by a person → its own note linked to that person's people/ entry.
3. Merge duplicates: If 2+ notes cover the same concept, merge. Mark redundant notes with superseded_by.
4. Promote fleeting → permanent (evergreen): Notes with type: fleeting, >7 days old, and inbound links > 0 deserve promotion. Ensure 2+ outbound wiki-links and update existing notes to link back. Also scan human-written notes for evergreen candidates — sentence-like titles, 2+ inbound links, 7+ days old. Suggest promotion to the human; never auto-promote. Add candidates to the board as a card: "Evergreen candidates ready for review."
5. Prune stale fleeting notes: Zero inbound links + >30 days old → propose deletion to human (list them; don't delete autonomously).
6. Strengthen connections: Convert plain-text concept mentions to [[wiki-links]]. Resolve contradictions (create question notes). Bridge notes in the same domain that should be linked.
7. Audit embed composition: Check hub notes and project deliverables — are they embedding atomic notes or rewriting content? Convert paraphrased sections to transclusions (![[atom#^core-claim]]). Higher-level notes should be compositions of embeds with thin connective prose.
8. Elevate meta awareness: Deepen the introspective core by seeding new meta notes or amending existing dimensions. If a convention in AGENTS.md drifted from practice, update the protocol. Update [[Knowledge map]] and .base views if the graph changed.

STEP 6 — Archive stale Done cards

Check the Done lane for cards with @{date} older than 7 days:

1. Move card's linked files to the vault's archive location (preserve folder structure).
2. Remove the card from the board.
3. Cards younger than 7 days stay visible so the human can review recent work.
4. If a Done card has no date, add today's date — it will be archived next cycle.

STEP 7 — Vault health audit (quick)

Run a lightweight [[introspect]] with depth: quick. Record issues as board cards in Blocked lane — don't fix inline. The human promotes them to Todo when ready.

STEP 8 — Fractal review cadence

Check if any [[fractal-review]] is due. The human's reflective practice runs on cascading timescales — the heartbeat's job is to detect when a review is due and prepare the surface.

1. Weekly: If today is Sunday (or >7 days since last review/weekly tagged note), and fragments exist from the past week → run fractal-review with cadence: weekly. If the human is absent, create the prep note and leave a board card: "Weekly review ready."
2. Monthly: If today is the 1st (or >30 days since last review/monthly), and weekly reviews exist from the past month → run fractal-review with cadence: monthly.
3. Yearly: If today is in January (or >365 days since last review/yearly), and monthly reviews exist → run fractal-review with cadence: yearly.

The heartbeat never writes the review — it creates the prep note and prompts the human. Reviews are created-by: human.

STEP 9 — Skill opportunities

Scan recent agent logs and task progress logs for repeated workflows (2+ occurrences). If a pattern isn't already a skill, suggest one: name + one-line description + what it automates.

STEP 10 — Agent log

Append a summary to today's agent log (inbox/log/YYYY-MM-DD). Cover: pipeline movement, tasks dispatched/shipped/blocked, inbox triage count, knowledge changes, health flags, review prep (if any), meta observations (new seeds, amended dimensions, protocol updates), and anything that needs human input. Keep it scannable — the human should understand what happened in 10 seconds. Never write to the human's daily note (daily/YYYY-MM-DD) — it's a pure date anchor.

STEP 11 — Git sync

Last step of every cycle.

1. git add -A
2. git status --porcelain — if empty, skip. Log "No changes to sync."
3. git commit -m "heartbeat: <one-line summary>" — e.g. heartbeat: dispatched 3 tasks, triaged 2 bookmarks, consolidated 1 note
4. git push — if push fails (conflict), log the error in the agent log. Do NOT force push. Next cycle retries. NOTE: If git is not set up, do git init and if remote is not set up, do git remote, skip git push.

Outputs

• Updated board with cards moved between lanes
• Task progress logs (one per completed/blocked task)
• Clean inbox
• Updated knowledge graph — consolidated, promoted, and better-linked notes
• Updated [[Knowledge map]] with new clusters
• Updated folder indexes where files moved
• Stale Done cards archived
• Meta/ growth — new seeds or amended dimensions from this cycle's observations
• Agent log entry (inbox/log/YYYY-MM-DD) summarizing the cycle
• Git commit + optionally push

Decision authority

• YOU DECIDE (act, don't ask): prioritization, task ordering, triage routing, note promotion, link strengthening, cluster naming, how to partition work across sub-agents, what to archive, agent log content, small protocol updates
• ESCALATE TO HUMAN (Blocked lane): large irreversible impact, creative/subjective input needed, credentials or external access required, structural changes to vault layout or note type system

Default: ACT. The human isn't watching. Make the best call. Note it in the agent log. The human will course-correct on their next visit. If you notice a preference pattern, update AGENTS.md — small convention tweaks don't need permission, structural changes do.

Conventions

• Heartbeat orchestrates; sub-agents execute. Don't run tasks inline in the heartbeat thread.
• Follow the current [[AGENTS]] conventions — they're the human's agreed patterns, not fixed rules.
• Archive, don't delete. Task logs are the audit trail — one per completed task.
• Knowledge writes: new note + update 1–3 existing notes to link back (distributed write). All agent-created notes must include created-by: ai. Never modify the body of created-by: human notes — only update their frontmatter properties. To connect human notes, create proxy connection docs (created-by: ai).
• Update AGENTS.md indexes when files move. Git sync is the last step every cycle.
• Protocol evolution: If this cycle reveals a convention in AGENTS.md doesn't match reality, update it.