Fulcrum Cockpit

Control-plane dashboard and native PI tools for the PI coding agent.

Install

# From this repo
pi install ./agent-integration/pi/cockpit

# Or use the install script
bash agent-integration/pi/install.sh

What you get

Dashboard widget

Live-updating widget in PI's TUI showing:

• Server status (stopped / starting / up / error)
• Running agent count + roles
• Blocked agent count + reasons
• WIP count
• Workspace/project context

Footer status line

Always-visible at the bottom of PI:

● FULCRUM  2 run  WIP:3  :4721

Slash commands

Command	Description
/fulcrum-setup	Confirm workspace IDs (computed from project path, no files written)
/fulcrum-status	Show workspace status: running agents, blockers, WIP count
/fulcrum-start	Start the Fulcrum monitor server
/fulcrum-monitor	Open monitor in browser (http://localhost:4721)
/fulcrum-tasks [status]	List tasks, optionally filtered by status
/fulcrum-create <title>	Create a new task
/fulcrum-run <task_id> <role>	Start an agent run
/fulcrum-complete <run_id> [summary]	Mark a run as completed
/fulcrum-block <run_id> <reason>	Mark a run as blocked
/fulcrum-recall <query>	Search project memories
/fulcrum-workspaces	List all workspaces
/cos <goal>	Inject Chief-of-Staff world-state context into the conversation

Native LLM tools (11 tools, no MCP overhead)

Tool	Description
fulcrum_list_tasks	List tasks in current workspace
fulcrum_create_task	Create a new task
fulcrum_update_task	Update task status or add a blocker note
fulcrum_recall_memory	Semantic search over project memory
fulcrum_write_memory	Persist a note to project memory
fulcrum_start_run	Register a new agent run (call at task start)
fulcrum_heartbeat	Send a heartbeat during long operations
fulcrum_complete_run	Mark a run as completed
fulcrum_block_run	Mark a run as blocked, escalate to CoS
fulcrum_workspace_status	Get full workspace snapshot
fulcrum_build_cos_context	Build Chief-of-Staff world-state context

Policy hook

Every tool call in PI is checked against the Fulcrum policy engine. Currently enforced:

• Team invocation guard: only chief_of_staff may invoke teams

Auto-start

On PI session start, the cockpit:

1. Computes workspace_id and project_id deterministically from the project directory path (sha256[:12] of abs path) — no file needed
2. Starts the Fulcrum monitor server (fulcrum serve monitor) if not already running
3. Polls workspace status every 5 s
4. Injects a Fulcrum-first system-prompt nudge before each agent turn so PI prefers native fulcrum_workspace_status / fulcrum_recall_memory before filesystem search or architecture decisions

Configuration

No project-local config files. IDs are computed from the directory path automatically.

Environment variable overrides:

• FULCRUM_WORKSPACE_ID — override computed workspace_id
• FULCRUM_PROJECT_ID — override computed project_id
• FULCRUM_PORT — override monitor port (default 4721)

Monitor URLs

After fulcrum serve monitor (or auto-start):

Endpoint	Description
http://localhost:4721/status	Server health
http://localhost:4721/agents	All agent runs
http://localhost:4721/board	Task counts by status
http://localhost:4721/tasks	Task list
http://localhost:4721/workspaces	Workspaces
http://localhost:4721/metrics	Workspace metrics
http://localhost:4721/memory-trace	Memory entries

Multi-runtime sharing

The same Fulcrum monitor server is shared by all three runtimes:

Runtime	Integration
PI	This cockpit (native tools + hook)
Claude Code	agent-integration/claude/ (MCP + hook)
Gemini CLI	agent-integration/gemini/ (MCP + hook)