⚒ FORGE — mission control for Claude Code agents

A localhost terminal-style mission control: every Claude Code session on this PC streams onto a live board — their build tasks as a tiered node pipeline (pending · in-progress ▸ · verified ✓ · cyan ◇ compliance rules), a SELECTED NODE inspector, a LIVE LOG, a rules ticker, LINE: FREE/BUSY/WAITING status, and an elapsed run timer. Plus two meta-skills: /forge-plan runs whole builds on the board; /forge-build forges new skills/agents.

Everything is local: server binds 127.0.0.1, data stays in data/, zero npm dependencies (Node 18+).

Quickstart

node server\server.js        (or double-click start-forge.cmd)

Open http://localhost:4000. Hooks are installed machine-wide (~/.claude/settings.json, 14 events), so every Claude Code session appears immediately. Then, in any Claude Code session:

/forge-plan a pomodoro web app

…and watch the board fill with a task pipeline that executes node by node — wire in + verify — exactly the mission-control view from the video.

The board (v3 — mission control)

• Task pipeline from Claude Code's native task system (TaskCreate/TaskUpdate + TaskCreated/TaskCompleted hooks): dense numbered nodes, status borders, duration badges, [RULE] … tasks as cyan ◇ nodes + one glowing COMPLIANCE n/m billboard, real addBlockedBy DAG edges — and the circuit lights up green along completed chains. Defaults to the newest run; ALL shows per-session swimlanes.
• Run header: run title derived from the /forge-plan goal, with a ▾ dropdown to switch runs; segmented progress bar (one segment per task — click = jump to node); bare HH:MM:SS run timer; LINE: FREE|BUSY|WAITING; events/min sparkline.
• NOW strip: the active task + elapsed, updating live under the top bar.
• SELECTED NODE: click any node — status history, brief, dependencies, failures, and the tool events captured while it ran; EXPORT ⬇ downloads a markdown run report.
• LIVE LOG: narrative feed (→ TASK 07 — wire in …, coalesced tool chatter, verb-colored accents) with search/filters/pause; DETAIL switches to raw events; click any row for full payload JSON.
• REPLAY: scrub the timeline rail to watch the run re-light from any buffered moment; ● LIVE snaps back.
• More: FOLLOW (auto-scroll to the active node), board filters (ALL·ACTIVE·FAILED·RULES, dim-don't-hide), keyboard (/ search · f follow · arrows walk nodes · Esc), STATS (tool usage, failures, task durations), FORGED registry, FX toggle (CRT flavor; honors prefers-reduced-motion), boot splash (skippable).

The skills

Skill	What it does
/forge-plan <goal>	Decomposes the goal into 15–40 native tasks (+ [RULE] invariants), then executes node-by-node: in_progress → work → verify by running it → completed. The board lights up live.
/forge-build <capability>	Forges a new skill/subagent/hook automation: scaffolds correct frontmatter, validates (hooks/validate_artifact.js), registers (hooks/register_capability.js).

Both live in .claude/skills/ here and are installed user-wide in ~/.claude/skills/ so they work from any project.

How it works

any Claude Code session on this PC
  └─ 14 hooks (exec form → node.exe, no shell)          ~/.claude/settings.json
       └─ hooks/send_event.js     truncate → POST, silent, always exit 0
            └─ http://127.0.0.1:4000/events
                 server/server.js     ring buffer + data/events.jsonl (rotated >50MB)
                 ├─ GET /             the mission-control dashboard
                 ├─ GET /stream       Server-Sent Events (live)
                 └─ /api/events · /api/registry · /api/transcript · /api/health

Hooked events: SessionStart/End, UserPromptSubmit, Pre/PostToolUse, PostToolUseFailure, Notification, PermissionRequest, SubagentStart/Stop, Stop, PreCompact, TaskCreated, TaskCompleted. A hook adds ~50–100 ms; if the server is down everything no-ops silently.

Transport note: the dashboard polls GET /api/sync?after=<id> once per second (paged ≤80 events per response) instead of holding an SSE stream — some network-inspection software (AV) buffers browsers' streaming response bodies indefinitely and silently eats responses larger than ~250 KB, which kills EventSource/fetch-streaming while small completed responses always pass. /stream (SSE) remains for curl/tools, and /probe is a built-in transport diagnostic page.

Privacy & config

• Prompts/tool text truncated to ~1,200 chars per field; all storage is local (data/events.jsonl, data/registry.json — don't hand-edit the registry, use register_capability.js).
• Transcript viewer only reads files under ~/.claude.
• Port: FORGE_PORT=4001 node server\server.js + set FORGE_SERVER_URL user-wide.
• Multiple dashboard tabs share the browser's 6-connection limit — close old tabs if one won't connect.
• Known limitation: cross-session shared task lists aren't merged; each session's tasks render in its own band.

Tests

node --test server/server.test.js

Transcript access tests cover POSIX and Windows path rules independently of the host filesystem. HTTP tests use a temporary server, home directory, and synthetic files. They do not read your Claude sessions or write to the repository's data.

Uninstall / pause

Delete the "hooks" block from ~/.claude/settings.json (takes effect immediately — hooks hot-reload). Stop the server; delete data/ to drop history.