Screen Code And Tool Agents

Coding and ops agents often read untrusted issues, logs, PR comments, command output, and package metadata. Screen that text before it can steer execution.

Boundary rule

Call POST /v1/parse on tool output before the agent runs shell commands, writes files, opens credentials, calls APIs, or commits code.

const toolResult = await runReadOnlyDiagnostic();
const screen = await fetch("https://parsethis.ai/v1/parse", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.PARSE_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    prompt: toolResult.stdout,
    metadata: {
      source_kind: "tool_output",
      trust_level: "external",
      intended_action: "execute",
      tool_permissions: ["shell", "filesystem", "network"],
    },
  }),
});

const decision = await screen.json();
if (decision.suggested_action !== "allow") {
  throw new Error(`Unsafe tool result: ${decision.suggested_action}`);
}

Pilot check

Start in shadow mode, then enable blocking only for high-risk tool boundaries. Track false positives on normal logs and issue bodies separately from malicious fixture recall.