@bitkyc08/opencodex
Version:
Universal provider proxy for OpenAI Codex & Claude Code — use any LLM with Codex CLI/App/SDK and Claude Code
91 lines (85 loc) • 3.96 kB
text/typescript
import { chmodSync, copyFileSync, existsSync, linkSync, readFileSync, rmSync } from "node:fs";
import { getConfigPath } from "../config";
import { hardenSecretPath } from "../lib/windows-secret-acl";
export interface AlibabaBackupIO {
exists: (path: string) => boolean;
read: (path: string) => Buffer;
copy: (source: string, destination: string) => void;
harden: (path: string) => void;
/** Publish with no-replace semantics: fails with EEXIST if the destination exists. */
publishNoReplace: (temp: string, destination: string) => void;
remove: (path: string) => void;
}
const DEFAULT_IO: AlibabaBackupIO = {
exists: existsSync,
read: path => readFileSync(path),
copy: (source, destination) => copyFileSync(source, destination),
harden: path => {
// POSIX: a failed chmod must not publish a credential-bearing backup with weak permissions.
// Windows keeps its own control via hardenSecretPath below, which is the required check.
if (process.platform === "win32") {
try { chmodSync(path, 0o600); } catch { /* Windows may not support POSIX chmod */ }
} else {
chmodSync(path, 0o600);
}
if (process.platform === "win32") hardenSecretPath(path, { required: true });
},
publishNoReplace: linkSync,
remove: path => rmSync(path, { force: true }),
};
export class AlibabaBackupIntegrityError extends Error {}
/**
* Immutable pre-migration snapshot of `config.json`.
*
* Written to a temp file first and published by `link`, which fails with EEXIST
* rather than replacing. Exclusive *creation* alone is not enough: a crash
* mid-copy can leave a truncated destination that the next run would accept as a
* valid rollback point. Copy-verify-link means a published snapshot is always a
* complete one.
*
* An existing snapshot is NOT required to equal the current config. Demanding
* that has a false positive that would brick a working install: a run that
* created the snapshot and then aborted for an unrelated reason leaves a
* perfectly valid snapshot, the user then edits `config.json` legitimately, and
* every subsequent start would throw out of `startServer` and refuse to boot. A
* safety net must never become the thing that stops the product.
*
* So an existing snapshot is reused as-is — it is the earliest one, which is
* exactly what a rollback point should be. Integrity is carried by publication.
*
* Deliberately not built on `backupConfigBeforeOpenAiTierMigration`: that
* function's stale/rollback classification exists for a repeatable migration and
* would be wrong here.
*/
export function backupConfigBeforeAlibabaRegionMigration(
configPath = getConfigPath(),
io: AlibabaBackupIO = DEFAULT_IO,
): "absent" | "created" | "reused" {
if (!io.exists(configPath)) return "absent";
const backup = `${configPath}.pre-alibaba-region-v1.bak`;
// The earliest snapshot is the most valuable one: it predates every migration
// this config has been through.
if (io.exists(backup)) return "reused";
const source = io.read(configPath);
const temp = `${backup}.${process.pid}.tmp`;
try {
io.copy(configPath, temp);
// The snapshot contains credentials. Harden it before publication so the
// stable backup path is never exposed with inherited permissions or ACLs.
io.harden(temp);
// Verify before publishing: a short copy must never become the snapshot.
if (!io.read(temp).equals(source)) {
throw new AlibabaBackupIntegrityError(`failed to write a complete backup to ${temp}`);
}
io.publishNoReplace(temp, backup);
return "created";
} catch (error) {
// Another process published between our `exists` check and the link. It went
// through the same copy-verify-link sequence, so it is complete by
// construction — nothing else writes this filename.
if ((error as NodeJS.ErrnoException).code === "EEXIST") return "reused";
throw error;
} finally {
io.remove(temp);
}
}