mirror of
https://github.com/stack-auth/stack.git
synced 2026-06-30 21:01:54 +08:00
0e1d98b5d0
## Problem
`hexclave dev` re-execs itself through `npx <pkg>@latest`
(`maybeReexecToLatest`) so users always get the latest dashboard without
reinstalling. But when that npx run **fails**, it exits nonzero and we
did `process.exit(result.code)` — killing `hexclave dev` even though a
perfectly good CLI was already installed locally.
This surfaced on **Replit**, where a user running `hexclave dev` got:
```
npm notice Access denied: Your download has been blocked by the Socket Security Policy.
Reason: AI-detected potential malware.
npm error code ECOMPROMISED
npm error Lock compromised
```
Diagnosis: the user's `pnpm` is aliased to `sfw pnpm` (**Socket
Firewall**), and Replit enforces an org-level Socket Security Policy.
Socket **blocks the `@hexclave/cli@latest` download** (false-positive
"AI-detected malware", almost certainly the bundled minified dashboard).
The interrupted download breaks npx's reify while it holds its cache
lock, which npm reports as `Lock compromised`. The lock message is a
*downstream symptom*; the real failure is the blocked download.
The same class of failure (blocked download, npm error, lock contention,
offline) all share one shape: **npx exits nonzero before our CLI ever
runs**, and today that takes down `hexclave dev`.
## Fix
Use a **startup-marker handshake** to distinguish the two cases:
- The parent passes a temp marker path to the npx child via
`STACK_CLI_REEXEC_MARKER`.
- The re-exec'd child touches the marker the instant it starts
(`signalReexecStartedIfChild`).
- After the child exits (`decidePostReexec`):
- exited 0, or nonzero **with** the marker present → our CLI ran;
**propagate** the exit code (real command result).
- nonzero **without** the marker, or npx not spawnable → our CLI never
ran; **fall back** to the installed CLI instead of failing `hexclave
dev`.
The marker only needs file create/exists (robust on sandboxed/networked
filesystems). If the marker can't be created, we preserve the old
always-propagate behavior, so there's no spurious-fallback risk.
`decidePostReexec` and `signalReexecStartedIfChild` are pure and
unit-tested.
## Verification
- `pnpm test run src/lib/self-update.test.ts` → 23 passing (added cases
for the fallback decision and the marker handshake).
- `pnpm typecheck` / `pnpm lint` → clean.
- End-to-end: forced a real npx failure (unreachable registry → npx
exits nonzero before our CLI runs) and confirmed `hexclave dev` now logs
`Auto-update skipped: …; continuing with the installed CLI.` and
proceeds into the installed dev path instead of dying.
## Notes / follow-ups (not in this PR)
- Workaround for affected users today: `STACK_CLI_NO_AUTO_UPDATE=1` (or
`--no-auto-update`) skips the npx download entirely.
- Worth reporting the Socket false-positive to socket.dev to allowlist
`@hexclave/cli`, and reconsidering shipping a ~165 MB minified dashboard
inside the npm tarball (it's what trips AI-malware heuristics and slows
cold installs).
<!-- This is an auto-generated description by cubic. -->
---
## Summary by cubic
Prevent `hexclave dev` from exiting when `npx @latest` auto-update fails
by falling back to the installed CLI. Signal-based aborts now propagate
as a nonzero exit (never NaN), so Ctrl-C doesn’t relaunch dev.
- **Bug Fixes**
- Fall back to the installed CLI when `npx <pkg>@latest` fails before
the CLI starts (firewalls, offline, npm lock errors). Do not fall back
when the child is killed by a signal; propagate 128+signum with a safe
nonzero fallback if the signum is unknown.
- Startup-marker handshake: parent sets `REEXEC_MARKER_ENV`
(`HEXCLAVE_CLI_REEXEC_MARKER`); child touches it on start;
`decidePostReexec` chooses propagate vs fallback. If the marker can’t be
created, keep the old always-propagate behavior.
- Scrub the marker env from user commands using the exported constant,
covering both Windows and POSIX spawn paths.
<sup>Written for commit
|
||
|---|---|---|
| .. | ||
| cli | ||
| dashboard-ui-components | ||
| js | ||
| next | ||
| react | ||
| sc | ||
| shared | ||
| shared-backend | ||
| tanstack-start | ||
| template | ||
| ui | ||
