Opencode runs Claude processes that kill my machine

[wojteksbt]

Summary

When running opencode from inside the pai-opencode repository directory, after each prompt it spawns a large number of claude processes, quickly saturating CPU and making macOS unresponsive (often requiring a reboot). This does not happen when running opencode from other directories on the same machine.

This looks like something in this repo’s configuration / integration triggers repeated Claude CLI invocations.

Environment

• OS: macOS
• CPU arch (Apple Silicon / Intel): ARM
• macOS version: 26.2
• pai-opencode version/commit: 1.2.1
• opencode version: 1.1.53
• Installation method for opencode: brew
• Related repo:
  • https://github.com/danielmiessler/Personal_AI_Infrastructure

Steps to reproduce

1. Clone and setup this repo:

   git clone https://github.com/Steffen025/pai-opencode
   cd pai-opencode
   bun run .opencode/PAIOpenCodeWizard.ts

2. Start opencode from the repo root:

   opencode

3. Enter any prompt (even a short one like “hi”) and submit.
4. Repeat with 1–2 more prompts.

Expected behavior

• opencode handles each prompt normally without spawning an unbounded number of claude processes.
• CPU usage remains stable/usable.

Actual behavior

• After each prompt submission, many claude processes appear and accumulate.
• CPU quickly spikes toward 100% across cores, the system becomes sluggish/unresponsive, and recovery sometimes requires a reboot.

Frequency

• 100% reproducible when launched inside pai-opencode repo root.
• 0% reproducible when launched from other directories on the same machine.

What I already tried / additional observations

• Setting export OPENCODE_DISABLE_CLAUDE_CODE=1 does not help.
• Renaming ~/.claude does not help.
• After renaming ~/.claude, opencode recreates a .claude folder and adds a project named:
  • -Users-Wojtek-repos-pai-opencode

Notes / suspicion

Because the behavior depends on the working directory, it may be triggered by something in this repo (e.g., config files, hooks, or integration that causes repeated Claude CLI invocations). Any idea what might happen?

[wojteksbt]

I dug into this and I think I found what’s happening.
When opencode runs inside this repo, one of the local plugins (pai-unified) is listening to message stream events. The issue is that it can treat streamed text chunks (message.part.updated) like new user messages.
Each time that happens, it runs the implicit sentiment handler, which calls the local Inference.ts helper.
That helper currently launches claude directly (spawn("claude", ...)).
So instead of one action per prompt, a single prompt/response can trigger many chunk events, and each one may start another Claude process. This plugin currently hard-depends on the claude CLI for side tasks. For an OpenCode plugin, that’s not ideal.
It should be provider-agnostic by default (or at least make Claude opt-in), otherwise this kind of coupling/failure mode is much more likely.