<div style="background: #0D9488; color: white; padding: 8px 16px; border-radius: 6px; font-weight: 600; margin-bottom: 24px;">AGENT MODE — uses your existing Claude Code / Codex CLI session.</div>

Mode B · Agent

Runs synthesis + query inside the Claude Code or Codex CLI session that's already open on your machine — no separate API key, no incremental cost beyond your existing agent subscription.

Status

Shipping in v1.1.0-rc8 (#316). Every slash command is prompt- driven — /wiki-sync, /wiki-ingest, /wiki-query, /wiki-reflect, /wiki-update, /wiki-lint. The new agent-delegate synthesize backend closes the last remaining gap: llmwiki synthesize now writes rendered prompts to .llmwiki-pending-prompts/<uuid>.md and returns a placeholder page with a  sentinel. The running agent picks the pending prompts up on the next turn and calls llmwiki synthesize --complete <uuid> to rewrite the page with the actual synthesis.

Zero incremental API cost. Zero bytes of session content leave your laptop. Works when ANTHROPIC_API_KEY is unset.

Setup (no API key)

Just copy the slash commands into your global Claude Code commands dir:

mkdir -p ~/.claude/commands
cp .claude/commands/wiki-*.md ~/.claude/commands/

That's it. Open Claude Code, type /wiki-sync, and it runs.

Daily flow

You: /wiki-sync
Claude: (runs python3 -m llmwiki sync, ingests new pages)

You: /wiki-query when did I last change the convert pipeline?
Claude: (reads wiki/index.md + the relevant source pages, synthesizes)

You: /wiki-graph
Claude: (writes graph/ + opens site/graph.html in your browser)

When to pick this mode

You already pay for Claude Code or Codex CLI — no extra bill.
Exploratory work where you want the model to narrate what it's doing ("I'm now reading wiki/sources/...").
Small-to-medium corpus (< 50 sessions) — serial synthesis isn't a bottleneck.
Local-only privacy — no session content leaves your laptop except through the agent's existing plumbing.

Limitations vs Mode A