Connect

mn:connect — Discover Hidden Links

Invocation marker (both runtimes): begin your reply with the exact line 🧠 mn:connect (mnemo) → running — the user-visible confirmation that this skill actually loaded. Emit it once per invocation, before any other output.

Portable paths

Resolve <mnemo-root> once to the absolute plugin root before reading bundled files or running bundled scripts. In Claude Code, use ${CLAUDE_PLUGIN_ROOT}; in Codex, derive it from this loaded SKILL.md path (skill directory → skills/ → plugin root). Replace <mnemo-root> with that quoted absolute path in every command — never execute the placeholder literally and never hunt versioned cache directories.

When another mnemo skill must run, use the runtime-native path: Claude Code invokes mn:<skill> through its Skill tool; Codex reads <mnemo-root>/skills/<skill>/SKILL.md completely and follows it with the prepared input. For user-facing explicit syntax, render /mn:<skill> in Claude Code and $mnemo:<skill> in Codex.

Analyze a note and discover connections to other notes in the vault that aren't linked yet.

Prerequisites & config

Obsidian must be open. Config at ~/.mnemo/config.json — reads vault, links_section, taxonomy, and taxonomy_roles. Before any mapped-hub operation, require exactly the five semantic roles fact, insight, source, session, and moc; require every target to exist; and require session → session plus moc → moc. The deterministic legacy Zettelkasten fallback is allowed; any other missing/invalid map stops mapped routing and offers the runtime-native setup skill instead of guessing. Schema in <mnemo-root>/references/config-schema.md.

Workflow

Step 1: Identify Target Note

Accept a note name from the explicit invocation, for example /mn:connect "Atom — LongCat-Flash-Prover" in Claude Code or $mnemo:connect Atom — LongCat-Flash-Prover in Codex.

If no argument, ask: "Which note should I analyze for connections?"

Step 2: Read the Note

python3 "<mnemo-root>/scripts/safe-read.py" read <<< '{"file":"{note_name}","vault":"{vault}"}'

Extract:

• Key concepts, technologies, names mentioned in text
• Existing [[wikilinks]]
• Existing links in {links_section} section

Step 3: Search for Related Notes + Backlinks (single grep + backlinks, parallel)

Even better than parallel obsidian search calls: one literal filesystem scan for all concepts. The helper treats each concept as data, not regex or shell syntax. It takes ~50ms for any number of concepts vs ~180ms per obsidian search.

# Run these TWO commands in parallel (single message, two Bash tool uses):
# 1. One literal scan — much faster than N separate Obsidian searches
python3 "<mnemo-root>/scripts/safe-read.py" grep-concepts <<< '{"vault":"{vault}","concepts":["{concept_1}","{concept_2}","{concept_N}"]}'

# 2. Backlinks check
python3 "<mnemo-root>/scripts/safe-read.py" backlinks <<< '{"file":"{note_name}","vault":"{vault}"}'

Collect matching note paths from the scan output. Exclude the target note itself. Backlinks output → notes already connected (exclude from suggestions).

Why not obsidian search per concept: each CLI call is ~180ms. One literal scan = ~50ms total for any N. On 7 concepts: 1.26s → 50ms (25x faster).

Step 4: Generate Suggestions

Compare: notes found by search MINUS notes already linked (wikilinks + backlinks).

For each suggestion, explain WHY it's relevant (shared concept, shared tag, complementary topics).

Tension-nodes (high value, ТРИЗ): if a found note makes an OPPOSITE claim on the same question — that's NOT a duplicate, it's a #contradiction link worth surfacing. Two notes disagreeing on the same benchmark is a synthesis point, not noise.

Step 5: Present (DO NOT auto-apply)

🔗 Connection suggestions for "{note_name}"

Already connected: {N} notes
New suggestions: {N}

1. [[Atom — SCOPE beats TextGrad]]
   Why: Both discuss agentic RL stability — HisPO and SCOPE solve similar problems

2. [[MOC — Agent Self-Correction]]
   Why: Note mentions trial→verify→reflect cycle, this MOC covers the same pattern
   Action: Add to MOC? (currently not listed there)

3. [[Session — ANT-14 TextGrad research]]
   Why: Both evaluate RL approaches for agent improvement

4. [[Atom — Full attention wins on this benchmark]]  ⚡ #contradiction
   Why: opposite conclusion on the same GPU-efficiency question — tension to resolve

Apply these? (y/N, or pick numbers: 1,3)

Links are applied with their one-line context (the «Why»), not as bare [[wikilinks]] (Luhmann: state why you linked).

Step 5.5: Auto-apply mode (the review --full chain only)

Fires only when the invoking payload carries an explicit auto-apply directive — the review --full chain sends one after reading review.full.autoConnect: true from ~/.mnemo/config.json (see <mnemo-root>/references/config-schema.md). A standalone /mn:connect / $mnemo:connect — anything a user typed, or any invocation without that directive — never auto-applies; it renders Step 5 and waits. This is the same shape as the other --full writes: the user opting into --full and flipping the config flag is the consent, exactly as review.lint.autoStampReviewed gates the one write health makes.

When the directive is present, skip the Step 5 confirmation prompt and go straight to Step 6 for the ranked, deduplicated, backlink-excluded suggestions this run produced — then report every link written (tension links included and clearly marked) so nothing is applied invisibly. The suggestions are still the same high-quality set (max 5-7, meaningful-not-generic, orphans excluded); auto-apply removes the keystroke, not the quality bar. If the config flag is off, or the directive is absent, this step does not run — fall through to Step 5.

Step 6: Apply on Confirmation

If the user confirms (Step 5) — or auto-apply mode is active (Step 5.5):

1. Read the target again, choose one unique stable anchor in {links_section}, then add new links with a one-line context (Luhmann — state why connected): - [[Name]] — {why, ≤10 words}, not a bare [[Name]]. Tension links get the marker: - [[Name]] #contradiction — {what disagrees}.
2. Apply the confirmed block through the bundled writer:

python3 "<mnemo-root>/scripts/vault-write.py" <<< '{"action":"insert","vault":"{vault}","note":"{target note}","anchor":"{unique anchor copied from safe-read}","position":"after","content":"{one JSON-escaped Markdown block containing only confirmed contextual links}"}'

3. If a mapped hub suggestion was confirmed, repeat the same exact-anchor insert for the note reached through taxonomy_roles.moc.
4. Verify with the helper's backlinks action.

If an anchor is missing/non-unique or a concurrent edit wins, re-read and retry only the confirmed changes. A dynamic name or explanation can contain quotes, backticks, or $(); keep it JSON data and never use inline Obsidian CLI content.

Advanced modes (optional)

• Radius-2 (Scrapbox/Cosense-style): two notes sharing ≥2 of the target's links are likely related even without shared text. If text search yields <3 results or user asks --deep, run:

python3 "<mnemo-root>/scripts/safe-read.py" shared-targets <<< '{"note_path":"{note_path}","vault":"{vault}"}'

• KJ-Canvas (bottom-up hub, 川喜田 affinity): for --canvas {topic} — gather notes mapped to the fact role, drop them onto an Obsidian Canvas without categories, and let the user group spatially so structure emerges. Then name clusters → new insight notes / mapped-hub revision. In the default taxonomy those are Atoms, Molecules, and a MOC.

Gotchas

Common failures in <mnemo-root>/references/gotchas.md. Tool-routing rationale in <mnemo-root>/references/tool-routing.md. Skill-specific rules:

• Max 5-7 suggestions — don't overwhelm. If you find more, rank and present top-7.
• Don't suggest links to orphan notes — they need their own fixing first (run /mn:health if interested).
• Ghost notes are normal — [[Technology]] pointing to a non-existent note enables entity discovery. Not "unresolved" in the bad sense.
• Avoid generic connections — "both mention Claude" is noise. A connection is meaningful if it shares a concept, approach, or unresolved question.
• Never auto-apply on the standalone path — a user-typed /mn:connect / $mnemo:connect, or any invocation without the explicit review --full auto-apply directive, always asks before writing new wikilinks (Step 5). The only exception is Step 5.5: the review --full chain with review.full.autoConnect: true — there the user opted into --full and flipped the config flag, so the links apply without a per-suggestion y and every write is reported. Default config keeps the flag off, so the default connect behavior is unchanged.