feat: agentic-ai/opencode local model setup with MLX launch script

[ulises-c]

Summary

Initial setup of OpenCode as a local AI coding agent, with opencode-local — a launcher that auto-discovers MLX models, starts an OpenAI-compatible server, and opens OpenCode. Designed to run alongside Claude Code using local models on Apple Silicon.

Changes

Main: OpenCode initial setup + local model integration (opencode-local)

• agentic-ai/OpenCode/ — new directory with config, install script, and launch tooling
• opencode.json — registers mlx_lm.server (OpenAI-compatible) as a local provider, pre-configured with Qwen 3.5 9B (4bit, MLX)
• bin/opencode-local — launch script that auto-discovers MLX models in ~/.models/, presents a picker, starts the server, waits for it to be ready, then launches OpenCode. Cleans up on exit. Logs to ~/.local/share/opencode-local/server.log
• install.sh — symlinks opencode.json to ~/.config/opencode/ and opencode-local to ~/.local/bin/
• README.md — activation, usage, model recommendations

Minor: Ghostty terminfo fix

• Added term = xterm-256color to macOS and Linux ghostty configs so remote servers that don't recognize xterm-ghostty still render correctly

Housekeeping

• Updated TODO.md with ghostty standardization and OpenCode follow-up notes

[ulises-c]

Memory crash investigation

Reviewed ~/.local/share/opencode-local/server.log from today's sessions. Two distinct issues found:

1. GPU OOM crashes (confirmed)

Two crashes with Qwen3.5-9B-4bit on the 16GB Mac Mini:

libc++abi: terminating due to uncaught exception of type std::runtime_error:
  [METAL] Command buffer execution failed: Insufficient Memory
  (00000008:kIOGPUCommandBufferCallbackErrorOutOfMemory)

Root cause: The KV cache grows unbounded as conversation context accumulates. Each turn appends to the message history, expanding the prefill KV cache until it exhausts Metal's unified memory allocation. Crashes occurred after 4–6 conversational turns.

Fix applied in 86caf28:

• --max-tokens 4096 — caps response generation length
• --prompt-cache-bytes 4294967296 (4GB) — limits total KV cache memory, triggers cache eviction when exceeded

2. GLM-4.7-Flash-5bit model unavailable

The model directory at ~/.models/mlx-community--GLM-4.7-Flash-5bit exists but contains no .safetensors files (likely a partial/corrupted download). The lighter model would be preferable for 16GB machines but needs re-downloading first:

huggingface-cli download mlx-community/GLM-4.7-Flash-5bit \
  --local-dir ~/.models/mlx-community--GLM-4.7-Flash-5bit

Recommendation

Consider documenting these limits in README.md — specifically that Qwen3.5-9B-4bit on 16GB systems may need periodic server restarts for long sessions, and GLM-4.7-Flash-5bit is the safer default for this hardware.

[ulises-c]

Try again with MTP. Do some more fine tuning on KV cache, context length, etc. Merging as is as an initial setup for OpenCode