kg-maintain

name: kg-maintain user-invocable: true description: | Knowledge graph maintenance. Tend the garden — regular, light care keeps it healthy. Not a separate task: woven into every session, every capture.

GARDEN RHYTHM — three modes, applied as needed: Water (routine): after each task, glance at 2–3 recently-touched nodes. Are their gists still accurate? Any note worth adding? Prune (when dense): merge duplicate nodes, shorten verbose gists (→ notes), split oversized nodes, remove stale touches, delete edges to removed concepts. Fertilize (on use): when a node proves valuable, connect it to newly-discovered related nodes. One new edge makes a node far more durable.

AFTER CAPTURE: when you save a node, immediately ask — - Do any adjacent nodes now need updating? - Is this a duplicate of something existing? Merge if so. - Does this node's gist still fit, or did context shift?

REACTIVE TRIGGERS (act immediately, mid-conversation): Uncertainty (spinning wheels, deja vu, about to search) → kg_search first. About to assume something → check KG; if missing, capture the assumption. User correction → update the stale node before continuing. Node just proved useful → add one edge to current context. Gist feels vague after using it → sharpen while context is live.

Archival is reversible — leave archived nodes alone unless content is factually wrong and can't be fixed by updating. Deletion is a last resort, not routine cleanup.

Maintenance Reference (Detailed)

When Invoked Directly (/kg-maintain)

Run a focused maintenance pass in this order:

1. kg_read(cwd) — check health stats: orphan %, avg edges/node, size warning
2. Always: scan all gists against the current kg-capture standard — tighten any that exceed it, regardless of graph size
3. Always: spot-check notes on recently-touched nodes — rewrite any that have grown chaotic or redundant (see "Notes Hygiene" below)
4. If graph is large or has size warning → Prune: merge duplicates, split oversized nodes, remove stale touches
5. After pruning → Fertilize: connect nodes clarified during pruning, add missing edges
6. Water throughout: update any gist that feels stale given what you just read

Announce findings: "Graph health: N nodes, E edges, O% orphans. Running [prune|fertilize|water] pass." Report what changed: nodes merged, gists tightened, edges added.

What a Healthy Graph Looks Like

A healthy graph is a mesh of connections, not isolated facts. Most nodes participate in at least one edge. Health stats show this at a glance:

• Low orphan rate — most nodes connected
• Reasonable edge density — linked but not over-connected
• Mix of levels — user patterns inform project decisions

Maintenance Operations

When auditing the graph with kg_read:

• Disconnected nodes — appear in no edges. Connect them with edges if appropriate. Only delete if truly orphaned AND factually incorrect.
• Duplicates — overlapping gists or IDs. Merge: keep/update richer one, delete the other.
• Outdated knowledge — about removed code or old decisions. Update/improve the node with current state rather than deleting.
• Broken edges — pointing to renamed or removed concepts. Update the edge target, or kg_delete_edge if the relationship no longer exists.
• Archived nodes — automatic process, leave them be. They're memory traces, not clutter.
• Orphaned nodes — invisible in kg_read; searchable via kg_search. Automatically deleted after 365 days without recall. To rescue: read an adjacent archived node — the promotion chain will surface its orphaned neighbors back to archived.
• Verbose gists — if a gist exceeds the current kg-capture standard, tighten it. Move procedure/details to notes. Gist = headline only. See "Gist Hygiene" below.
• Bloated notes — notes that have grown into a changelog or contain contradictions. Rewrite to current truth only. See "Notes Hygiene" below.
• Oversized nodes — single node spanning multiple responsibilities. Split and link. See "Oversized Node Detection" below.

Gist Hygiene

No extra tool needed — gists are visible in the main kg_read output. Scan them all.

Standard: check the current kg-capture skill for the active gist length target. Any gist exceeding that limit needs tightening — this applies to all nodes regardless of graph size or when the node was created. Old nodes do not get a pass on the current standard.

Signals:

• Gist uses "and" to join independent ideas → split the node
• Gist restates what the ID already says → delete the redundant part
• Gist contains a procedure step → move to notes
• Gist longer than the current target → cut; move excess to notes

Action: rewrite to subject + key fact only. Everything else belongs in notes or an edge.

Notes Hygiene

Notes are hidden by default, so they don't inflate the visible graph — but they accumulate silently. A note added in session 1, amended in session 5, and partially updated in session 20 becomes contradictory and bloated. Periodic rewrite is the fix.

When to rewrite notes (not just append):

• Node has been touched 3+ times and notes feel like a log of changes rather than current truth
• Notes contain contradictory statements ("X is required" and "X is optional" both present)
• Notes repeat information already in the gist
• Notes contain a chain of "actually…" corrections — collapse to the final truth
• A node covers two separable concerns — split and redistribute notes

How to rewrite:

1. Read the full notes block as-is
2. Extract the current truth: what invariants hold now? What constraints? What rationale?
3. Discard the history (corrections, "turns out", "actually") — keep only the conclusion
4. Rewrite as a clean set of bullets, each a standalone fact

Notes are not a changelog. They're a compressed memo to a future session that has no other context.

When to do a notes pass:

• During any /kg-maintain invocation: spot-check 3–5 nodes with most touches
• Any time you read a node's notes and feel confused before feeling informed
• After a long debugging session where the same node was updated repeatedly

Oversized Node Detection

When gist + notes together are very large, or a single node spans multiple distinct responsibilities:

Signals:

• A single node covering two separable concepts → split into two nodes linked by an edge
• An edge with a very long notes field → distill to one sentence; create a node for the reasoning if it matters

Actions:

1. Split node — when a node covers two separable concepts, split and link with an edge (relates_to, part_of, depends_on).
2. Promote to CLAUDE.md — if notes contain operational procedures or project conventions, they belong in CLAUDE.md rather than the KG. Remove them from the node once moved.

This pass is independent of graph size — a small graph can have a bloated node.

Operational Safety

Tool result size limit

Claude Code persists tool results over ~50K chars to disk — model sees only a 2KB preview. kg_read output is the full graph as text. If kg_read shows a size warning (>40K chars):

1. Review nodes for staleness, duplicates, disconnected entries
2. Delete or merge low-value nodes
3. KG_MAX_TOKENS default is 5000 per graph level — two graphs ≈ 40K chars, safe under 50K limit

Project renames

When project folder renamed, graph slug changes. Server handles via alias detection. If project graph is unexpectedly empty, check ~/.knowledge-graph/projects/ for old name.

Server restart

Server can be safely restarted (kg-memory restart). Validates PIDs, uses setsid, drains connections, write-through persistence means no data loss.

First Session in a Project

When project graph is empty:

1. Don't document the whole codebase. Capture what surprises you.
2. 2-5 foundational nodes for major components, connected by edges.
3. Add knowledge organically as you work.
4. Quality over quantity — every node earns its place through reuse.