Prometheus Wiki Loop

Three small tools that turn any AI coding session — Claude Code, OpenCode, Codex, Claude Desktop, Kimi Desktop, Mavis/MiniMax Desktop, or a bare shell — into a contribution to a shared, human-readable, version-controlled knowledge base.

The Karpathy pattern, in shell. Plain markdown wiki, LLM-compiled, TF-IDF searchable, no vector database. Every session closes by writing what was learned into the wiki. The next session opens with the wiki already primed.

What's in the box

File	What it does	When it runs
bin/kbd-close	Universal session-close hook. Reads a summary (file, stdin JSON, or fallback), enriches it with active KBD phase context, calls pk ingest to compile into the wiki via the openai-proxy, appends a learning-log JSONL entry, files a skill-update candidate. Always exits 0.	End of every session (Stop hook, PreCompact hook, notify, or manual)
bin/kbd-open	Session-prime companion. Detects active KBD phase from .kbd-orchestrator/current-waypoint.json, reads position reminder + phase goals, runs pk focus on the phase, surfaces pending skill-update candidates + today's learning log. Writes a snapshot to ~/.prometheus/last-open-snapshot.txt.	Start of every session (SessionStart hook, or manual)
bin/wiki-loop-mcp	Zero-dependency Node stdio MCP server. Exposes 7 tools (add_to_wiki, prime_context, search_wiki, focus_wiki, list_recent_learnings, list_wiki_entries, list_pending_skill_updates) so any MCP-capable chat tool can contribute to the wiki by natural-language prompt.	Whenever the chat tool calls a tool

All three scripts:

• Are MIT licensed, < 500 lines each, zero non-stdlib dependencies
• Source ~/.prometheus/.env automatically so the LLM endpoint (openai-proxy, OpenAI, Anthropic, Groq, etc.) is configurable per environment
• Always exit 0 — a hook failure never breaks the calling tool

Install

# 1. Install the wrappers to ~/.local/bin (must be on PATH)
./scripts/install.sh

# 2. Seed the env file with your LLM endpoint
cat > ~/.prometheus/.env <<'EOF'
CLOUD_LLM_URL=https://api.openai.com/v1          # or http://localhost:8181/v1 for openai-proxy
LOCAL_LLM_URL=https://api.openai.com/v1
CLOUD_LLM_API_KEY=sk-...
PK_COMPILE_MODEL=gpt-4o-mini
PK_LINT_MODEL=gpt-4o-mini
PK_FOCUS_MODEL=gpt-4o-mini
EOF

# 3. Make sure pk is installed (from the prometheus-knowledge crate)
#    and the prometheus-knowledge MCP server is running on port 8942.
#    See: https://github.com/Prometheus-AGS/prometheus-knowledge

# 4. (Optional) Register the MCP server in your chat tool
#    Claude Desktop, Mavis/MiniMax, Kimi Code CLI, Codex:
{
  "mcpServers": {
    "wiki-loop": {
      "command": "/Users/gqadonis/.local/bin/wiki-loop-mcp"
    }
  }
}

That's it. From this point:

• Every Claude Code session begins with last-open-snapshot.txt showing your active KBD phase + focused wiki hits + pending skill-updates
• Every Claude Code session ends (or compacts) with the work written into the wiki
• Codex, OpenCode, Kimi Code, Mavis/MiniMax Desktop — same wiki, different trigger surface
• Chat surfaces (Claude Desktop, Kimi Desktop chat, Codex chat) — say "save this conversation to the wiki" and the agent calls add_to_wiki automatically

The full architecture

┌──────────────────────────────────────────────────────────────────┐
│                  AI Tools (any of these)                          │
│                                                                   │
│   Claude Code    Codex    OpenCode    Kimi Code                   │
│       ↓           ↓          ↓           ↓                        │
│   Stop/PreCompact  notify    plugin      /kbd-close skill         │
│       ↓           ↓          ↓           ↓                        │
│   ┌──────────────────────────────────────────────────┐           │
│   │         ~/.local/bin/kbd-close (universal)        │           │
│   └──────────────────────────────────────────────────┘           │
│                            ↓                                     │
│   ┌──────────────────────────────────────────────────┐           │
│   │   Detect KBD phase + enrich + pk ingest + log    │           │
│   └──────────────────────────────────────────────────┘           │
│                            ↓                                     │
│   ┌──────────────────────────────────────────────────┐           │
│   │  ~/.prometheus/knowledge/shared/wiki/*.md        │           │
│   │  ~/.prometheus/learning-log/YYYY-MM-DD.jsonl     │           │
│   │  ~/.prometheus/skill-updates/                    │           │
│   └──────────────────────────────────────────────────┘           │
│                                                                   │
│   Chat surfaces (Claude Desktop, Kimi Desktop, Codex chat)        │
│       ↓  "save this to the wiki"                                  │
│   ┌──────────────────────────────────────────────────┐           │
│   │       ~/.local/bin/wiki-loop-mcp (MCP server)    │           │
│   │   7 tools: add_to_wiki, prime_context, ...       │           │
│   └──────────────────────────────────────────────────┘           │
└──────────────────────────────────────────────────────────────────┘

What it depends on

• bash (for kbd-close / kbd-open)
• node ≥ 18 (for wiki-loop-mcp, zero deps)
• python3 (for the JSON extraction helpers in kbd-close)
• pk CLI (from prometheus-knowledge) — compiled to ~/.prometheus/bin/pk or on PATH
• prometheus-knowledge MCP server running on :8942 (or override with PK_BIN env var)
• An LLM endpoint reachable by the chosen CLOUD_LLM_URL (default: OpenAI; works equally with the local openai-proxy on :8181 or any OpenAI-compatible service)

The related projects

This toolkit is one piece of the Prometheus Fabric, an open-source multi-repository platform for sovereign agentic AI built on BossFang (librefang). Related crates:

• prometheus-knowledge — the Rust pk CLI and pk-cherry MCP server this toolkit calls
• surreal-memory-server — the optional graph-memory + TaskStreams backend
• prometheus-skill-pack — the 280+ skill manifests and 4-layer PMPO pipeline that uses this toolkit for cross-session compounding
• prometheus-entity-management — the React entity graph that consumes the wiki as a knowledge plane

License

MIT — see LICENSE.