graphify — High-Quality Full-Vault Re-Run

Background

Promoted from KB-OS-Context §5 + long-standing UPGRADES item. graphify turns a folder of files into a navigable knowledge graph (community detection + interactive HTML + GraphRAG JSON + GRAPH_REPORT.md). It answers a question qmd doesn’t: “how is this connected” (entities, communities, relationships) vs. qmd’s “find it by meaning.” The two are complementary — qmd can even index graphify’s wiki output.

Current state: only two stale partial runs exist (MBR/graphify-out/, Core/Misc/graphify-out/). A high-quality full-vault run has never been done — it was deferred waiting for (a) vault content to stabilize and (b) an API key for full semantic extraction with parallel subagents (per the UPGRADES note, /graphify must run from a main Claude Code session, not a background agent, for the API-key semantic pass).

Since deferral, the vault has stabilized (Context System + Memory v1 shipped; structure stable per STATUS), so this is now actionable.

Goal

One high-quality graphify run over the stabilized unified vault, producing a current “how it connects” layer, with a decision on where the output lives and whether/how it feeds qmd.

Open decisions (resolve at task start)

Scope: whole vault (/mnt/d/FSS/KB) in one graph, or per-business (Core/MBR/…) graphs? A single vault-wide graph shows cross-business connections but is larger; per-business is cleaner but loses cross-links. (Lean: whole-vault with --mode deep.)
Output home + SSOT: the two stale graphify-out/ dirs — delete/replace? Where does the canonical graphify-out/ live (vault root? Core/Misc/?), and is it gitignored (per-device artifact, like .codegraph/ + qmd index — Principle #8) or committed? Recommend gitignored — it’s a regenerable derived artifact, not source truth.
Exclusions: mirror the qmd kb collection’s noise exclusions (Clippings/, archive/, .obsidian/) so the graph isn’t polluted — the same noise that hurt qmd recall will hurt the graph.
qmd handoff: after the run, add graphify-out/’s wiki/report output as a qmd collection (or into kb)? Decide based on whether the report text adds recall value.

Tasks

Resolve the open decisions above (scope, output home, gitignore, exclusions)
Clean up / supersede the two stale partial runs (MBR/graphify-out/, Core/Misc/graphify-out/)
Run /graphify <scope> --mode deep from a main session (API key present) — not a background agent
Verify outputs: interactive HTML opens, GRAPH_REPORT.md reads sensibly, communities are coherent (spot-check a few cross-business links)
Decide + wire the qmd handoff (index the graph output, or not)
Record the run + decisions in KB-OS/logs/; update KB-OS-Context §5 (mark done) and the UPGRADES Graphify section
(If gitignored) add graphify-out/ to .gitignore with a one-line rationale, like .codegraph/

Acceptance criteria

One current, high-quality full-vault graph exists; stale partial runs superseded.
Output home + gitignore decision made and applied (SSOT clean — no duplicate/stale graphify-out/ dirs left alive).
graphify vs qmd division of labour documented in KB-OS-Context (graphify = connections, qmd = meaning).

References

Skill: ~/.claude/skills/graphify/SKILL.md (/graphify <path> --mode deep, --update, --neo4j, etc.)
Deferred note: UPGRADES → “Graphify (Phase 2+)” + “Graphify high-quality re-run”
Complementarity + division of labour: KB-OS-Context §2 map + §5
Precedent for gitignoring a derived index: .codegraph/ + qmd cache (Principle #8)

Claude Response — 2026-07-13

Summary:

Resolved open decisions: whole-vault scope, --mode deep, output stays at vault-root graphify-out/ (already gitignored), exclusions mirror qmd’s kb collection noise rules
Superseded the two stale partial runs — moved (not deleted, permission denied a direct rm -rf) to a session scratchpad backup
Ran the full pipeline from this main session: 541 files detected → AST (4 code files) + 23 parallel semantic-extraction subagents (--mode deep) → merged into 2,103 nodes / 2,671 edges / 329 communities
Mid-run judgment call, confirmed with Talbot via AskUserQuestion: excluded ~85 auto-generated daily MBR/Logs/Ops/*_Rate-Scanner-Ops.md files from semantic extraction (near-identical machine templates, no unique-entity value) — saved ~4/27 subagent chunks
Session hit its Claude usage limit twice mid-run; all subagent output had already been written to disk before each API error, so no extraction work was lost — verified every chunk before merging
Verified outputs: graph.html structurally valid, GRAPH_REPORT.md God Nodes/Surprising Connections read sensibly, spot-checked real cross-business links (MBR Strategy ↔ CRM data flows, Core research ↔ Later/StrategicPlanUpgrades, MBR IT ↔ rate-scanner)
qmd handoff: added graphify-out/GRAPH_REPORT.md as its own qmd collection (graphify), embedded, verified searchable
Recorded the run in KB-OS/logs/2026-07-13_Graphify-Fullvault-Rerun.md; updated KB-OS-Context §5 and UPGRADES (marked done)
gitignore: no change needed — **/graphify-out/ already covers the vault-root output

Next Steps for Talbot:

Open D:\FSS\KB\graphify-out\graph.html in a browser to visually confirm the graph renders and communities look coherent (I verified it’s structurally valid HTML but couldn’t click through it from this environment) 1.
Confirm the two stale partial-run backups (moved to session scratchpad, not deleted) can be permanently deleted — or say if anything in them is still wanted 1.
Optional: review the 304 heuristically-labeled communities (vs. the 25 hand-crafted ones) next time the graph is browsed, and flag any whose label is unhelpful 1.