You’re running a parent agent with seven subagents, all quietly grinding through a browser-automation task. A single page navigation races mid-flight, a CDP (Chrome DevTools Protocol) call hiccups — and the entire process tree dies. Parent, all seven children, every pending tool call, gone. That’s not a hypothetical. It’s a freshly filed P1 crash in the oh-my-pi agent. And it has a quieter cousin in Claude Code that can silently fork your Windows config.
The issues we found
1. oh-my-pi: a browser stealth patch crashes the whole process (issue #5296). OMP’s patched puppeteer-core@25.3.0 stealth layer calls a debug helper named debugError in its frame-acquisition catch paths. That helper is not defined unless the puppeteer:error debug channel is explicitly enabled — which it never is by default. So when a routine, recoverable CDP error hits during world re-acquisition, the logger itself throws a TypeError instead of the real error. That secondary throw escapes as an unhandledRejection, trips OMP’s postmortem handler, and calls process.exit(1). Because all sibling agents share the process, one browser tab takes down the lot. The reporter documented a comfyui parent plus seven subagents killed in a single shot.
2. Claude Code: Windows path casing silently splits your config (issue #76994). ~/.claude.json keys per-project state by raw directory path, compared byte-for-byte. On Windows — case-insensitive — the CLI launches with an uppercase drive letter (C:/...) while VSCode-spawned sessions lowercase it (c:/...). The same folder becomes two different keys. Result: MCP servers you registered from the CLI never load in VSCode, and vice versa, with zero error. The bug is labeled has repro and the reporter lost weeks to intermittent “where did my MCP server go” confusion.
Are you affected?
Run these quick checks:
# oh-my-pi: did a browser task kill the whole process tree?
grep -i "unhandled_rejection" ~/.omp/logs/*.jsonl | tail -5
grep -i "debugError is not a function" ~/.omp/logs/*.jsonl | tail -5
# Claude Code / VSCode: do you have case-variant duplicate keys?
grep -ioE "\"c?:/users/[^\"]+\"" ~/.claude.json | sort -u | uniq -dIf you see fatal unhandled_rejection entries after browser use, or duplicate path spellings in ~/.claude.json, you’re hit.
The fix
- oh-my-pi: Until upstream patches the stealth layer, avoid long-running multi-agent browser sweeps on the stock build, or pin a patched
FrameManager.jsthat routes catch handlers through the safedebugCatchErrorhelper (the maintainer already proposed the one-line swap). Restart affected sessions — there is no live in-place recovery once the process exits. - Claude Code: Canonicalize your working directory casing across clients. Launch VSCode and CLI from the same path spelling, or register MCP servers at user scope (
--scope user) instead of project scope so the key collision can’t hide them.
Why it happened
Both are defensive-coding misses, not model failures. OMP’s patch reimplemented frame acquisition without reusing Puppeteer’s existing safe logger, so the disabled channel returns nothing and the catch block calls a non-existent function. Claude Code never canonicalizes paths on a case-insensitive filesystem, so two legitimate spellings of one folder become two permanent, divergent configs.
FAQ
Q1: Is the oh-my-pi crash a security exploit? No — it’s an internal crash from a missing function reference, not an external attack. But it is a reliability emergency: one transient browser error can wipe real work.
Q2: Will restarting Claude Code fix the Windows MCP split? Restarting won’t merge the duplicate keys — they’re already written. You must delete the stray case-variant entry (or switch to user-scope servers) so both clients read the same record.
Q3: Can I prevent the oh-my-pi process kill today?
Partially. Set NODE_DEBUG=puppeteer:error to define the helper (trade-off: very noisy logs), or apply the proposed patch locally. The real fix ships with the stealth layer update.
Before you buy any coding agent subscription — compare prices and find exclusive deals at aiFiesta.