🧱 Guard and panel

In videos that circulated about Claude Code mods, two safeguards appeared: a collision (two agents in the same file) and one of "blast radius" (the size of the damage a command that deletes can cause).

The idea is great. The kit makes its own version, using only official resources: the hooks from Claude Code. The split is simple: the rule stays in the script, the screen stays in the mod.

How to read the diagram: the amber box is the control point. Every edit and command passes through it before taking effect. If there is no risk, it follows the blue arrow. If there is a risk, the guard neither blocks nor allows it: it sends the question back to you (N2).

"PreToolUse": [
  {
    "matcher": "Edit|Write|MultiEdit|NotebookEdit",
    "hooks": [ { "type": "command",
      "command": "node \"$CLAUDE_PROJECT_DIR/runtime/mods/runtime-guarda/hooks/colisao.mjs\" pre",
      "timeout": 10 } ]
  },
  {
    "matcher": "Bash",
    "hooks": [ { "type": "command",
      "command": "node \"$CLAUDE_PROJECT_DIR/runtime/mods/runtime-guarda/hooks/raio.mjs\"",
      "timeout": 15 } ]
  }
]
…
"PostToolUse": [ … colisao.mjs pos … ]

The pain is real. It happened in one of our projects: one session emptied a file that another was editing. Neither one was wrong on its own. Each thought the file belonged only to them.

The collision resolves this with memory. After each edit, it notes, “this session just changed this file.” Before the next edit, it checks the note and the file’s date.

How to read the diagram: the blue arrow is session A's normal edit, which is recorded. The red arrow is session B arriving within the 30-minute window. You answer the question in the amber box.

Proven in CHANGELOG 0.3.0: with 5 cases (old file, changed externally, same session, another session, nonexistent), the collision only prompts in cases involving another session and external changes. In Claude, an Edit to a file from another session was blocked, and the file remained intact.

O ray comes before a rm or git clean. It counts how many existing files would be deleted, adds up their size, and lists the first paths. Then it asks: "Do you have a backup? Confirm to continue."

The lightning bolt doesn’t delete anything: it only reads the disk. And if the command wouldn’t delete any existing files, it stays quiet. You can test the script on its own, without opening Claude, by sending it the same request Claude would send.

mkdir -p /tmp/teste-raio/x && echo a > /tmp/teste-raio/x/a && cd /tmp/teste-raio
printf '{"cwd":"%s","tool_name":"Bash","tool_input":{"command":"rm -r x"}}' "$PWD" \
  | node /caminho/do/kit/runtime/mods/runtime-guarda/hooks/raio.mjs

a saída traz "permissionDecision":"ask"
e o texto Raio: este comando apaga 1 arquivo(s)

The guard is already enabled in the kit folder. But Sônia has another folder, the one for monthly closings, where agents also run. She wants the same protection there.

Recipe R7 offers two ways. Choose one, not both.

claude --plugin-dir /caminho/do/kit/runtime/mods/runtime-guarda

com --plugin-dir runtime-guarda e o modo que libera tudo,
rm -r pasta foi barrado pelo "Raio" e a pasta ficou intacta;
Edit em arquivo editado por outra sessão foi barrado pela "colisão".

In module 3.3, you launched sessions with claude --bg and tracked it through observar.mjs, in another terminal. The panel brings this list into Claude Code. It’s optional.

It shows each background session with its name, status, and minutes, a button Update and a button Stop in the ones that are running. Only your click applies to a session.

claude --plugin-dir runtime/mods/runtime-painel

claude plugin validate runtime/mods/runtime-painel
claude plugin test runtime/mods/runtime-painel

validate: passa
test: 1 pass

How to read the diagram: is an illustration, not a screenshot. The line r4-revisor · done · 7 it’s the same session as the observar.mjs showed in module 3.3. The red button only appears for sessions that are still running.

What to look for in the table: both read the same session list. Choose whichever feels more comfortable, not based on functionality.

Mods and plugins run with the same Claude Code permissions. A hook can read your files, run commands, and access the network without asking. That's why it protects so well, and why a malicious mod can do so much damage.

The rule in Recipe R7: read the code of any third-party mod before installing it. You don't need to program for this. Ask the agent to read and explain it without running anything.

Read runtime/mods/runtime-guarda/.claude-plugin/plugin.json, runtime/mods/runtime-guarda/hooks/hooks.json, colisao.mjs, and raio.mjs. Explain in plain language: when each hook runs, what it reads, what it writes, and whether it accesses the network. Do not install or run anything.