Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
/**
* advisor/apply.ts — pure resolution + validation for `gbrain advisor --apply`.
*
* Kept separate from the CLI command so the allowlist + injection guard are
* unit-testable without spawning a process. The CLI confirms + executes; this
* module decides WHAT (if anything) is safe to run.
*
* Safety model (#10/C5):
* - Only findings carrying a `dispatch_id` are runnable (the allowlist).
* - The command is a STRUCTURED argv; it is rejected if any token contains a
* shell metacharacter (defense in depth — we never invoke a shell).
* - argv[0] must be 'gbrain' (the advisor never runs arbitrary binaries).
*/
import type { AdvisorReport } from './types.ts';
export type ApplyResolution =
| { ok: true; argv: string[]; display: string }
| { ok: false; error: string; runnable: string[] };
const SHELL_META = /[;&|`$<>(){}\n]/;
export function resolveApplyTarget(report: AdvisorReport, id: string): ApplyResolution {
const runnable = report.findings.filter((f) => f.fix.dispatch_id).map((f) => f.fix.dispatch_id!) as string[];
const finding = report.findings.find((f) => f.fix.dispatch_id === id);
if (!finding) {
return { ok: false, error: `No runnable finding with apply id "${id}".`, runnable };
}
const argv = finding.fix.command_argv;
if (!argv || argv.length === 0) {
return { ok: false, error: `Finding "${id}" has no runnable command.`, runnable };
}
if (argv[0] !== 'gbrain') {
return { ok: false, error: `Refusing to run: fix does not invoke gbrain.`, runnable };
}
if (!argv.every((a) => typeof a === 'string' && !SHELL_META.test(a))) {
return { ok: false, error: `Refusing to run: fix command contains unexpected characters.`, runnable };
}
return { ok: true, argv, display: argv.join(' ') };
}