Knowledge Garden

Knowledge Garden

Audit Basic Memory notes for structural and quality issues, then report actionable, copy-paste-ready findings. Read-only — never writes or modifies notes. Fixes are handed off to /knowledge-maintain.

This skill is the scoped, interactive sibling of the knowledge-gardener agent. It owns the common case — auditing a handful of named notes inline in the main session — and delegates the heavy, graph-wide sweep to the agent so the full audit's hundreds of note reads stay out of the main context window.

Arguments

The user names the notes or topics to audit after the invocation:

1. Decide venue (scoped inline vs delegate)

This is the first and most important step. Classify the request:

Delegate to the agent when the request is graph-wide — any of:

• No arguments are provided.
• The argument is a whole-graph phrase: "my knowledge graph", "the whole graph", "everything", "all notes", "full audit", "--full", "graph health".
• The request implies graph-wide checks the inline path cannot do efficiently: schema drift across a whole type, version drift across ecosystems, cross-project scope leak, orphan detection across the entire graph, tag alignment.

To delegate, launch the read-only auditor and stop: call Agent(subagent_type="knowledge-gardener", description="Full graph health audit", prompt="<the user's original request, verbatim>") — i.e. pass a description summarizing the audit as "Full graph health audit" and a prompt that is the user's original request, verbatim.

Relay the agent's report. Do not also run the inline path. If the Agent call fails or returns no usable report (unknown subagent type, error payload, or empty output — the knowledge-gardener agent may not be installed), say so explicitly: name the failure, do not fabricate a clean result, and offer to run the scoped inline path on a named subset as a fallback. Never report a passing audit you did not actually receive.

Run inline (scoped) when the request names a bounded set of notes or topics (roughly 1–8 targets). Proceed to step 2. If a "scoped" request resolves to more than ~8 notes, stop and recommend delegating to the agent instead — inline audit of a large set bloats context and is exactly what the agent exists for.

2. Resolve each target to a note

For each named target, find the canonical note:

• Plain title (e.g. npm-umzeption) → confirm it exists with read_note directly; if not found, fall back to search_notes.
• Prefixed identifier (brew:ripgrep) → map the prefix to its directory and locate via list_directory(dir_name="<dir>", file_name_glob="*<name>*").
• Topic phrase → search_notes(query="<phrase>", page_size=10), take the matching cluster, and audit those notes.

Prefix-to-directory mapping: npm:→npm/, crate:→crates/, go:→go/, composer:→composer/, pypi:→pypi/, gem:→gems/, brew:→brew/, cask:→casks/, action:→actions/, docker:→docker/, vscode:→vscode/, gh:→gh/, plugin:→plugins/, skill:→plugins/, git:→engineering/git/ (git_builtin notes are conventionally plain-titled like git-replay and also resolve via the plain-title path above).

Exclude schema notes (permalinks under /schema/) — they are structural definitions, not subject content.

3. Audit each note

Load the parsed note and run the audit dimensions:

read_note(identifier="<permalink-or-title>", output_format="json")

Reading as JSON gives the parsed observations and relations arrays — the source of truth for the note body. Then for each note check:

• Schema — schema_validate(identifier="<permalink-or-title>"). Validate the single named note via identifier — never note_type, which batch- validates every note of that type (hundreds of notes) and defeats the scoped purpose. Report error_count / warning_count, plus unmatched_observations and unmatched_relations (silently-absorbed categories/verbs the schema does not declare). If the validator output appears to repeat the same value, trust the read_note JSON (loaded above) as the source of truth for the actual body content rather than inferring a duplication from the validator response alone.
• Structure — required sections present (## Observations with [category]-tagged items, ## Relations with [[wiki-links]]); frontmatter correct (type in snake_case, title/directory matching the prefix convention); ## Relations is the final section (no observation-bearing section after it).
• Relations — verbs are canonical (relates_to, depends_on, etc. — flag related_to, relates to, trailing-colon forms). Use build_context(url="<ecosystem-dir>/<note>", depth=1) (the note's path, e.g. npm/npm-foo — not a bare title) to confirm each [[wiki-link]] target resolves and to surface inbound edges.
• Orphan / isolation — use two passes, because build_context traverses edges only and cannot surface a true zero-link note at all: outbound edges come from the read_note JSON relations array (loaded above); inbound edges come from build_context. A note absent from build_context results has no inbound links; a note with neither outbound nor inbound is a zero-link orphan.
• Fourth-wall quality (rules below, mirrored from the vp-note-quality skill — a skill cannot load another skill's content, so they are inlined here) — flag self-referential content in subject-domain notes: claims like "absent from Raindrop/BM/Readwise", "Connection to the Knowledge Graph" sections, or a lede describing coverage rather than what the subject IS. Exemption: meta-notes whose subject IS the knowledge graph (notes under engineering/agents/*, axioms, tool catalogs, conventions) may reference BM/Raindrop/Readwise freely — apply the rules strictly only to subject-domain notes (packages, people, patterns, concepts, history).

4. Report

Produce a structured report per note. Tier findings and, where possible, give the exact remediation a /knowledge-maintain pass could apply:

## Audit: <note title>
- **Permalink / type:** `<permalink>` (`<type>`)
- **Schema:** PASS / N errors, M warnings (+ any unmatched obs/relations)

### Findings
- **[Critical|Warning|Info] <short title>** — <what + why>.
  *Fix:* <concrete edit_note find_replace target, or "schema question → /schema-evolve">.

(repeat per note)

## Summary
| Note | Schema | Structure | Relations | Orphan | Fourth-wall |
|------|--------|-----------|-----------|--------|-------------|

When the audit set came from a topic phrase (resolved via search_notes in step 2 rather than explicit titles), state in the report that you audited the search-match set, not necessarily the whole cluster — e.g. "Audited N notes matching 'IndieWeb'; the full cluster may contain more." Never present a topic-phrase audit as exhaustive coverage.

Then, if any note has actionable fixes, suggest the handoff: "Run /knowledge-maintain <notes> to apply the structural fixes." Distinguish:

• Note-level edits (missing section, trailing-obs-after-relations, verb drift) → actionable by /knowledge-maintain.
• Schema questions (unmatched observation categories recurring across a type) → defer to /schema-evolve <type>, NOT a note edit.

Edge Cases

• Note not found — report which target failed to resolve; suggest /intel if it looks like an undocumented package/tool.
• Scope creep — a topic phrase resolving to 9+ notes → recommend delegating to the knowledge-gardener agent rather than auditing inline.
• BM unavailable — read-tool failures surface as raw error strings; report and suggest retrying.
• Mid-audit BM failure — if a read_note or schema_validate fails for one note in a multi-note audit, continue the rest, then report which notes completed, which were skipped on failure, and which were not reached. Never present a truncated batch as a full audit.
• Ambiguous "audit X" — if it is unclear whether X names a note or a whole area, prefer the scoped read; only delegate when the request is unambiguously graph-wide.

Guidelines

• Read-only — never write, edit, or delete. Hand fixes to /knowledge-maintain.
• Venue first — always classify scoped vs graph-wide before doing any work.
• Verify before reporting — read the live note (output_format="json") before asserting a structural defect; a flagged issue may be a parse artifact.
• Cite specifics — every finding references the note and the offending line/section, with a copy-paste remediation where possible.
• Don't duplicate the agent — the graph-wide 10-step sweep (schema drift, version drift, scope leak, tag alignment) lives in the agent; delegate to it rather than reimplementing those checks here.