re-ensure

re-ensure Skill

Explicit, on-demand bring-up of the reverse-engineering MCP backends. The RE MCPs are bridge proxies — each mcp__<x>__* call forwards over a local socket to a plugin/server that must be running:

MCP bridge	backend port(s)	headless launch?
mcp__ida-pro-mcp	GUI :13337 (preferred) or headless idalib :8745	yes (idalib_server)
mcp__ghidra-mcp	GhidraMCP headless :8089	yes (true no-GUI)
mcp__binary-ninja-mcp	BN GUI plugin :9009	NO — Personal license has no headless; open the GUI manually

If the backend isn't running the bridge connects to nothing and the call fails ("not reachable" / "connection refused" / list_instances → []).

You usually don't need this skill

A PreToolUse hook (~/.claude/hooks/re-autolaunch/) already auto-launches the right backend the moment you invoke a matching RE tool and its port is down. This skill is for explicit control: "make sure Ghidra is up before I start", or bringing up a backend ahead of a batch so the first real call doesn't have to wait.

Mechanism (the launcher)

C:/Users/Prekzursil/.claude/hooks/re-autolaunch/autolaunch.py wraps the re-harness (C:/Users/Prekzursil/.local/opt/re-harness/re_launch.py):

# Probe all backends (no launch) — fast, safe, prints SUCCESS/FAILED + free RAM.
python C:/Users/Prekzursil/.claude/hooks/re-autolaunch/autolaunch.py probe

# Cold-start ONE backend on-demand (detached; idempotent; prefers existing).
python C:/Users/Prekzursil/.claude/hooks/re-autolaunch/autolaunch.py ensure ghidra
python C:/Users/Prekzursil/.claude/hooks/re-autolaunch/autolaunch.py ensure ida
python C:/Users/Prekzursil/.claude/hooks/re-autolaunch/autolaunch.py ensure bn    # -> manual hint (no headless)

# Bring up every LAUNCHABLE backend (IDA + Ghidra). Heavier — only when asked.
python C:/Users/Prekzursil/.claude/hooks/re-autolaunch/autolaunch.py ensure all

ensure kicks a detached re_launch.py (which itself spawns the heavy server detached, polls it ready, registers it) and briefly waits for the port. It prints a terminal-state marker per backend:

• SUCCESS:re-ensure <x> already-up … / … now listening …
• STARTING:re-ensure <x> cold-start launched … — analysis still running in the background
• MANUAL:re-ensure binary-ninja-mcp … — open the BN GUI yourself
• FAILED:re-ensure … — launcher couldn't start

Poll until ready (cold-start can take 1–3 min)

A backend's port binds quickly, but full MCP readiness (idalib auto-analysis / Ghidra program open) can take minutes. After STARTING, poll probe until UP rather than firing a doomed call. Because this can exceed 60s, run the poll in the background (longrun / run_in_background), per the >60s rule — never a blocking foreground sleep loop:

# bounded poll: stop as soon as ghidra:8089 listens (or give up after ~3 min)
for i in $(seq 1 36); do
  if python -c "import socket,sys; s=socket.socket(); sys.exit(0 if s.connect_ex(('127.0.0.1',8089))==0 else 1)"; then
    echo "ghidra UP"; break
  fi
  sleep 5
done

Inspect a stuck launch via the logs: ~/.claude/hooks/re-autolaunch/state/<tool>-spawn.log and the re-harness ~/.local/opt/re-harness/state/<tool>.log.

Safety / RAM

• On-demand & single-backend — IDA and Ghidra are heavy; ensure <tool> starts only that one. Use ensure all only when you really want both.
• Never duplicates the live GUI IDA — if :13337 is up, IDA is treated as up and no headless server is started.
• OOM guard — the auto-launch hook refuses to launch when free RAM is below min_free_ram_mb (default 2048) and tells you to free memory or launch manually. ensure is explicit, so it still launches; check probe (it prints free_ram_mb) first if RAM is tight.
• Never touches port 13337 / the GUI beyond probing; teardown stays the re-harness's job (re_launch.py <tool> --teardown, registry-gated).
• Binary Ninja cannot be headless-launched on a Personal license — ensure bn reports the manual step (open GUI, load binary, start the MCP plugin on :9009).

Configuration

~/.claude/hooks/re-autolaunch/backends.json — ports per backend, the pre-loaded binary (default: the BC libnative-lib.so for 15.4.1), min_free_ram_mb, brief_wait_seconds, and deny_while_starting / deny_when_unlaunchable (set false for inject-only, never-block behavior). Disable everything with RE_AUTOLAUNCH=off or a .disabled file in that directory.

Verify

python C:/Users/Prekzursil/.claude/hooks/re-autolaunch/test_autolaunch.py   # unit/decision/fail-open tests
python C:/Users/Prekzursil/.claude/hooks/re-autolaunch/autolaunch.py probe  # live backend status