OpenCode, an AI coding agent with 161k stars on GitHub, is fundamentally a web-stack tool for piping LLM | bash with a security posture the author describes as "let me bend over for you daddy." Testing with Qwen3.6-27B on an M4 Max revealed fractal poor decision-making.

Prompt cache misses are rampant: OpenCode globs filesystems and re-reads AGENTS.md every SSE turn, prunes context from tool calls on every agent→user transition (discarding results beyond PRUNE_PROTECT = 40_000), and injects the current date into turn-0 system prompts causing midnight cache misses. The author waited 10 minutes for prefill on a 0.5 TB/s memory bandwidth system.

Pruning deletes early reads like specs when interruptions occur, leaving the agent without context. Compaction forces full-session prefill just to generate bullet points, interacting poorly with pruning. System prompts are verbose with "shit opinions" — Beast Mode insists you "CANNOT successfully complete this task without using Google" and "Definitely don't just read the source code." No global prompt modification exists; switching modes triggers cache misses.

The author concludes everyone should stop using OpenCode immediately.