Optional Integrations

Integrations add capabilities on top of the core loop, declared in fabric.yaml (Configuration) — and they are never load-bearing: with every integration off, every command behaves exactly as the core docs describe. The design rule: integrations add capabilities; they are never load-bearing.

One page per integration (deep dives)

Integration	Page	One-line
judgment	Judgment — the decision-model tier	typed calibrated judging: Jev cloud / Laya on-device / Tev1+Nimble via ollama / generic — pros, cons, measured numbers
graphify	Code call-graph	AST provenance, stale-claim detection after refactors, code navigation
embeddings	Semantic re-rank	offline cosine fusion for query ranking (measured value conditions)
obsidian	Two-way vault	human edits → evidence; generated notes → your vault
MCP	wf-mcp server	the 0-token core surface as native tools for any MCP client
team sync	Team Sync + the freshness cycle	corpus sharing via git; scheduled upstream drift capture

The quick matrix

Integration	When active	Cost	Ship it when…
judgment	mining near-miss refinement; wf verify-effects second opinion; eval gates; retrieval fusion	on-device (0) or ~$0.0004/call (Jev)	mining gets real traffic; self-preference risk matters; fusion rerank wanted → page
graphify	doc→code provenance; AST stale-claim detection; code-navigation block	0 tokens (AST)	docs reference code symbols; refactors rename code → page
embeddings	semantic re-rank on top-40	~5 ms/query offline (fastembed)	corpus > ~2k claims; word divergence across projects → page
obsidian	two-way: harvest human edits as evidence; mirror notes via REST API	0 tokens	humans curate the wiki by hand → page
MCP	core surface as native tools	0 tokens	your harness speaks MCP; multi-client setups → page

All off by default; each page carries pros/cons, measured examples, and the exact setup — plus the degrade contract (server down / route off ⇒ deterministic behavior stands).

The loudness contract (gate notifications)

Pending human decisions — the things integrations like judgment propose — reach you where you stand:

• Terminal banner + bell (the no-setup default): every command that creates a pending decision prints human decisions pending — promotions: 1 (wf gate); wf status carries a persistent ⚠ Gate: pending … → wf gate line.
• macOS notification center (default on darwin) and webhooks (WF_GATE_WEBHOOK_URL) — adapter config below.
• Payload: selection facts (what, why, evidence), never hidden reasoning. Fire-and-forget: notification failure never fails the emitting command; the durable record is always registry/pending-gate.md + the queues.

notify:                  # optional; absent = terminal (default) + macOS bump
  - kind: webhook
    url_env: WF_GATE_WEBHOOK_URL
  - kind: macos
  - kind: terminal
# notify: []             # silence everything — the manifest stays the record

Agent harnesses

Wiki Fabric installs into any agent harness — see the matrix in Getting Started. wf harness install detects what you use and writes the always-on block + skills in each tool's native format; wf harness status shows the current state.

Session capture (wf capture chat <slug>) covers four harnesses: Claude Code, opencode, OpenAI Codex CLI, and Gemini CLI. Formats were built from the harnesses' own storage schemas (verified against google-gemini/gemini-cli's chatRecordingTypes.ts and claude-mem's transcript schema). Design note vs claude-mem's approach: they capture via hooks (SessionStart/UserPromptSubmit/PreToolUse events) — real-time but plugin-dependent; wiki-fabric reads the session stores the harnesses write anyway — zero install requirements, sha256 anti-loop, and thread-index frontmatter on every capture.

Enabling

wf integrations                    # show what's active and what each changes
wf install --with-graphify         # enable at install time
wf update --with-graphify          # enable on an existing fabric