ganglion（ganglion 命令行）

Ganglion — Operator's Manual

Ganglion is a domain-specific execution engine for Bittensor subnet mining. It provides a pipeline framework for orchestrating autonomous mining agents that search for optimal model configurations. It exposes a CLI, an HTTP bridge API, and a Python library. Ganglion is search infrastructure — it doesn't know what a good model looks like, it knows how to search for one.

Quick Reference

# Scaffold a new project
ganglion init ./my-subnet --subnet sn9 --netuid 9

# Check state (local mode)
ganglion status ./my-subnet
ganglion tools ./my-subnet
ganglion agents ./my-subnet
ganglion knowledge ./my-subnet --capability training --max-entries 10
ganglion pipeline ./my-subnet

# Run (local mode)
ganglion run ./my-subnet
ganglion run ./my-subnet --stage plan
ganglion run ./my-subnet --overrides '{"target_metric":"accuracy"}'

# Start HTTP bridge (remote mode)
ganglion serve ./my-subnet --bot-id alpha --port 8899

# Check state (remote mode)
curl -s "$GANGLION_URL/v1/status" | jq .data

Mode Detection

Ganglion supports two modes. Always check which mode applies before running commands.

• Local mode: No GANGLION_URL set, or GANGLION_PROJECT is set. Use ganglion <command> <project_dir> directly.
• Remote mode: GANGLION_URL is set. Use curl against the HTTP bridge.

if [ -n "$GANGLION_PROJECT" ] || [ -z "$GANGLION_URL" ]; then
  echo "local"
else
  echo "remote"
fi

Response Format

All HTTP bridge endpoints (except health probes) return responses in a standard envelope:

• Success: {"data": <payload>} — use jq .data to extract
• Error: {"detail": {"error": {"code": "ERROR_CODE", "message": "..."}}}

Health probes (/healthz, /readyz) return raw JSON without the envelope.

Interactive API docs: $GANGLION_URL/v1/docs (Swagger UI).

Note: Unversioned routes (e.g. /status) still work but are deprecated. Always use /v1/.

How to Run

Prerequisites: Python >= 3.11, LLM_PROVIDER_API_KEY set (used by the LLM runtime).

Install: pip install ganglion

Scaffold a project:

ganglion init ./my-subnet --subnet sn9 --netuid 9

This creates config.py, tools/, agents/, and skill/ in the target directory.

Start in local mode:

export GANGLION_PROJECT=./my-subnet
ganglion status $GANGLION_PROJECT

Start in remote mode:

ganglion serve ./my-subnet --bot-id alpha --port 8899
export GANGLION_URL=http://127.0.0.1:8899

The project directory must contain a config.py that defines subnet_config (SubnetConfig) and pipeline (PipelineDef). See {baseDir}/references/configuration.md for full config details.

Key Features

Observe State

Query the current framework state — registered tools, agents, pipeline definition, knowledge, metrics, and run history. Local mode uses CLI commands; remote mode uses GET endpoints.

Full reference: {baseDir}/references/commands.md

Execute Pipelines

Run the full pipeline or a single stage. The orchestrator executes stages in dependency order, applies retry policies, injects accumulated knowledge into agent prompts, and records outcomes.

# Local
ganglion run ./my-subnet
ganglion run ./my-subnet --stage plan

# Remote
curl -s -X POST "$GANGLION_URL/v1/run/pipeline" -H "Content-Type: application/json" -d '{}' | jq .data
curl -s -X POST "$GANGLION_URL/v1/run/stage/plan" -H "Content-Type: application/json" -d '{}' | jq .data

Mutate at Runtime (Remote Only)

Register new tools, agents, and components; patch the pipeline; swap retry policies; update prompts. All mutations are validated, audited, and reversible.

# Register a tool
curl -s -X POST "$GANGLION_URL/v1/tools" -H "Content-Type: application/json" \
  -d '{"name":"my_tool","code":"<code>","category":"training"}' | jq .data

# Patch pipeline
curl -s -X PATCH "$GANGLION_URL/v1/pipeline" -H "Content-Type: application/json" \
  -d '{"operations":[{"op":"add_stage","stage":{"name":"validate","agent":"Validator","depends_on":["train"]}}]}' | jq .data

Pipeline operations: add_stage, remove_stage, update_stage. See {baseDir}/references/commands.md for all mutation endpoints.

Knowledge Store

Cross-run strategic memory that compounds over time. Records patterns (what worked) and antipatterns (what failed), then automatically injects relevant history into agent prompts. Knowledge is queried by capability and filtered by bot_id for multi-bot setups.

# Local
ganglion knowledge ./my-subnet --bot-id alpha --capability training

# Remote
curl -s "$GANGLION_URL/v1/knowledge?capability=training&max_entries=10" | jq

Rollback

Undo any mutation. Every mutation is recorded in an audit log with rollback data.

curl -s -X POST "$GANGLION_URL/v1/rollback/last" | jq
curl -s -X POST "$GANGLION_URL/v1/rollback/0" | jq    # undo ALL mutations

Multi-Bot Workflows

Multiple OpenClaw sessions share a knowledge pool via --bot-id. Each bot's discoveries flow into the shared pool. Cooperation emerges from shared knowledge, not explicit coordination.

# Two local sessions
ganglion run ./my-subnet --bot-id alpha
ganglion run ./my-subnet --bot-id beta

# Two remote servers
ganglion serve ./my-subnet --bot-id alpha --port 8899
ganglion serve ./my-subnet --bot-id beta  --port 8900

MCP Integration

Ganglion is a bidirectional MCP system: it can consume external MCP servers (client mode) and expose its own tools as an MCP server (server mode).

MCP Client Mode — Consuming External Tools

Connect to external MCP servers to add tools to the agent's repertoire. Tools from MCP servers appear as regular Ganglion tools with a configurable prefix.

# Static: add to config.py
# from ganglion.mcp.config import MCPClientConfig
# mcp_clients = [MCPClientConfig(name="weather", transport="stdio", command=["python", "-m", "weather_server"])]

# Dynamic: add at runtime via API
curl -s -X POST "$GANGLION_URL/v1/mcp/servers" -H "Content-Type: application/json" \
  -d '{"name":"weather","transport":"stdio","command":["python","-m","weather_server"]}' | jq .data

# Check connected MCP servers
curl -s "$GANGLION_URL/v1/mcp" | jq .data

# Disconnect / Reconnect
curl -s -X DELETE "$GANGLION_URL/v1/mcp/servers/weather" | jq .data
curl -s -X POST "$GANGLION_URL/v1/mcp/servers/weather/reconnect" | jq .data

MCPClientConfig options: name, transport (stdio|sse), command (for stdio), url (for sse), env, tool_prefix, category, timeout (default 30s).

MCP Server Mode — Exposing Ganglion Tools

Run Ganglion as an MCP server so external agents (Claude Code, Claude Desktop, OpenClaw) can call Ganglion tools directly.

# stdio transport (Claude Desktop / Claude Code)
ganglion mcp-serve ./my-subnet --transport stdio

# SSE transport (HTTP-based clients)
ganglion mcp-serve ./my-subnet --transport sse --mcp-port 8900

# Multi-role with access control
ganglion mcp-serve ./my-subnet --roles ./roles.json

Claude Code integration: This repo includes .mcp.json which auto-configures Claude Code to connect to Ganglion's MCP server via stdio. Claude Code will see all Ganglion tools natively in its tool palette.

Exposed MCP Tools (31 total)

Observation (11) — read-only state queries: ganglion_get_status, ganglion_get_pipeline, ganglion_get_tools, ganglion_get_agents, ganglion_get_runs, ganglion_get_metrics, ganglion_get_leaderboard, ganglion_get_knowledge, ganglion_get_source, ganglion_get_components, ganglion_get_mcp_status

Mutation (6) — write operations: ganglion_write_tool, ganglion_write_agent, ganglion_write_component, ganglion_write_prompt, ganglion_patch_pipeline, ganglion_swap_policy

Execution (3) — run operations: ganglion_run_pipeline, ganglion_run_stage, ganglion_run_experiment

Admin (5) — rollback and MCP management: ganglion_rollback_last, ganglion_rollback_to, ganglion_connect_mcp, ganglion_disconnect_mcp, ganglion_reconnect_mcp

Compute (6) — infrastructure tools: compute_status, compute_jobs, compute_job_detail, compute_routes, write_dockerfile, build_image

Multi-Role MCP Serving

Run one process with multiple MCP server instances, each with different access levels and auth tokens.

// roles.json
[
  {"name": "admin",    "categories": null,                        "token": "admin-xyz",    "port": 8901},
  {"name": "worker",   "categories": ["observation","execution"], "token": "worker-abc",   "port": 8902},
  {"name": "observer", "categories": ["observation"],             "token": "observer-def", "port": 8903}
]

Roles filter tools by category. null categories = access to all tools. Each role gets a separate port and bearer token for SSE transport. At most one role can use stdio transport.

Per-Bot Usage Tracking

When running with --roles, a shared UsageTracker records per-bot tool calls (tool name, success/failure, timestamp, duration) to .ganglion/usage.db. Query via /usage endpoint on any SSE role.

Connecting Other Agents to Ganglion's MCP Server

OpenClaw and other LLM agents can start Ganglion's MCP server for themselves or for other agents to connect to. Use SSE transport for generic MCP clients that don't support stdio.

# Start Ganglion MCP server with SSE transport (any MCP client can connect)
ganglion mcp-serve ./my-subnet --transport sse --mcp-port 8900

# SSE endpoints exposed:
#   GET  http://127.0.0.1:8900/sse         — SSE stream (tool list + results)
#   POST http://127.0.0.1:8900/messages     — send tool calls
#   GET  http://127.0.0.1:8900/usage        — usage stats (if tracking enabled)
#   GET  http://127.0.0.1:8900/usage?bot_id=alpha — per-bot stats

# With auth (multi-role), include bearer token:
#   curl -H "Authorization: Bearer worker-abc" http://127.0.0.1:8902/sse

For OpenClaw: OpenClaw reads skills and invokes commands via bash/curl — it doesn't connect to MCP natively. To give OpenClaw access to Ganglion tools, use the CLI (local mode) or HTTP bridge (remote mode) documented above. OpenClaw can also *start* an MCP server so that other MCP-capable agents (Claude Desktop, Cursor, etc.) can connect:

# OpenClaw starts the MCP server for other agents
ganglion mcp-serve ./my-subnet --transport sse --mcp-port 8900

For generic MCP clients (Cursor, Windsurf, custom): Point your client's MCP config at the SSE endpoint:

• Server URL: http://127.0.0.1:8900/sse
• Messages endpoint: http://127.0.0.1:8900/messages
• Transport: SSE
• Auth: Bearer token (if using roles)

Common Workflows

See {baseDir}/examples/common-workflows.md for full step-by-step guides.

1. First run: ganglion init → edit config.py → ganglion run
2. Iterative mining: check status → review knowledge → run pipeline → check metrics → repeat
3. Dynamic mutation: observe tools/agents → register new tool via API → patch pipeline → run
4. Multi-bot setup: start multiple servers with different --bot-id values on the same project
5. MCP integration: connect external tool servers → tools appear in registry → agents can use them

When Things Go Wrong

Full troubleshooting: {baseDir}/references/troubleshooting.md

Retry Policies

Four built-in policies control how stages retry on failure:

• NoRetry — single attempt
• FixedRetry — retry N times (default: 3)
• EscalatingRetry — increase temperature per attempt, optional stall detection
• ModelEscalationRetry — climb a model cost ladder (cheap → expensive)

Three presets: SN50_PRESET (escalating + stall detection), SIMPLE_PRESET (fixed), AGGRESSIVE_PRESET (model escalation).

Additional Resources

• Full CLI & API reference: {baseDir}/references/commands.md
• Configuration guide: {baseDir}/references/configuration.md
• Operational procedures: {baseDir}/references/operations.md
• Troubleshooting: {baseDir}/references/troubleshooting.md
• Workflow examples: {baseDir}/examples/common-workflows.md
• Sample API requests: {baseDir}/examples/sample-requests.md
• Health check script: {baseDir}/scripts/healthcheck.sh