athal7's dotfiles

Manages ~ on macOS via chezmoi. Covers shell, editor, AI tooling, calendar automation, and a library of AI agent skills.

Quick start

sh -c "$(curl -fsLS get.chezmoi.io)" -- init --apply athal7

You'll be prompted for a few required values (name, email, code directory, GitHub token, calendar names). Optional integrations — Slack, Figma, Elasticsearch, Linear, PagerDuty, ICS feeds, etc. — can be added by editing ~/.config/chezmoi/chezmoi.toml after init. See the commented sections in .chezmoi.toml.tmpl for the full list.

Machine-specific config

Per-machine values that don't belong in version control (secrets manifest, calendar configuration, per-org overrides) live in ~/.local/share/chezmoi/.chezmoidata/local.yaml. Copy local.yaml.example from this repo's root and fill in your values.

orgs.<org> keys cover per-GitHub-org behaviors used by the agent skills:

• issues: linear routes issue management to Linear instead of GitHub Issues
• automated_review declares an embedded code reviewer (e.g. GitHub Copilot review) — the review skill consumes prior automated findings instead of duplicating them locally

Note: the example file lives at the repo root rather than under .chezmoidata/. Files inside .chezmoidata/ are merged into chezmoi data at runtime, which would leak placeholder values into the live config.

What's configured

• Shell — zsh (dot_zshrc.tmpl, dot_zshenv.tmpl, dot_zprofile.tmpl)
• Editor — Neovim (dot_config/nvim/)
• Git — config, aliases, hooks (dot_config/git/)
• Terminal — Ghostty (dot_config/ghostty/)
• AI tooling — OpenCode config, MCPs, plugins, agent instructions (dot_config/opencode/)
• Packages — brew, cask, mise, GitHub releases (.chezmoidata/packages.yaml)
• Calendar automation — sync, lunch guard, family scheduler (dot_local/bin/, Library/LaunchAgents/)
• Homebridge — Google Nest via HomeKit (dot_homebridge/)
• macOS services — LaunchAgents for background processes (Library/LaunchAgents/)
• Agent skills — see skills/

Local model setup (LM Studio)

LM Studio serves the local model used by OpenCode for agentic coding. Ollama remains installed for the minutes skill and cloud-proxied models, but qwen3-family models leak XML tool-calls through Ollama's OpenAI-compatible endpoint (see OpenCode #26162, #24316, #4428). LM Studio is the most-reported working setup.

One-time bootstrap after chezmoi apply installs the cask:

# 1. Launch LM Studio.app once — sets up ~/.lmstudio/ and accepts quarantine
open -a "LM Studio"
# (close the app after first launch)

# 2. Install the lms CLI on PATH
npx lmstudio install-cli

# 3. Verify
lms --help

# 4. In LM Studio GUI → Developer tab, enable "Run server on login"
#    (makes the server auto-start without keeping the GUI open)

# 5. Download Qwen 3.5 27B Unsloth UD-IQ3_XXS (~11.5 GB)
#    LM Studio's `lms get` can't target a specific quant — download via curl.
mkdir -p ~/.lmstudio/models/unsloth/Qwen3.5-27B-GGUF
curl -fL -o ~/.lmstudio/models/unsloth/Qwen3.5-27B-GGUF/Qwen3.5-27B-UD-IQ3_XXS.gguf \
  "https://huggingface.co/unsloth/Qwen3.5-27B-GGUF/resolve/main/Qwen3.5-27B-UD-IQ3_XXS.gguf?download=true"

# 6. Load with a stable identifier matching opencode.json
lms load unsloth/Qwen3.5-27B-GGUF --identifier qwen3.5-27b --context-length 32768 --ttl 999999

# 7. Verify the server is up
curl http://localhost:1234/v1/models | jq

After step 4, the server auto-starts on each login. Reference the model in OpenCode as lmstudio/qwen3.5-27b.

To benchmark: bench-opencode -m lmstudio/qwen3.5-27b drives the OpenCode HTTP API through 6 realistic agent scenarios. See dot_local/bin/executable_bench-opencode.

Picking a model

Glukhov maintains an empirical comparison of local models for OpenCode agentic coding, with measured error rates on real coding tasks: https://www.glukhov.org/ai-devtools/opencode/llms-comparison/. Top recommendations as of May 2026:

• Qwen 3.5 27B Unsloth UD-Q3_XXS — 5.0% error rate (clear winner)
• Qwen3-Coder-Next UD-IQ4_XS — fastest, 8.8% error rate
• Gemma 4 26B A4B IQ4_XS — 6.3% error rate

Glukhov tests on llama.cpp, but LM Studio can serve the same Unsloth GGUFs and ships patched chat templates. Quantization source matters — Unsloth quants have patched templates that Bartowski quants do not, which materially affects tool-call quality.

Agent Skills

Agent Skills-compatible skills deployed to ~/.agents/skills/. Works with OpenCode and any compatible agent.

Skills use a capability-based composition system — workflow skills declare what they requires, and integration skills self-register via provides. The capability harness discovers providers by scanning installed skills — no manifest file needed. This lets workflow skills stay tool-agnostic: swap providers by installing a different integration skill. See agentskills/agentskills#311 for the spec proposal.

Mechanical workflow enforcement

Skills tell the agent what to do. Without mechanical enforcement, every skill is a polite suggestion the agent skim-selects from. High-stakes actions — git commit, git push, any gh write — are gated through opencode's permission config in opencode.json: patterns marked "ask" prompt for approval before the action runs, and the user picks once, always, or reject per prompt. Common reads like gh pr list and gh issue view are explicitly "allow" to avoid prompt fatigue.

Installing individual skills

With the GitHub CLI:

gh skill install athal7/dotfiles commit
gh skill install athal7/dotfiles review

With chezmoi — declare skills in .chezmoidata/packages.yaml and use a run_onchange_ script to install and update them weekly. See my sync script as a reference:

# .chezmoidata/packages.yaml
packages:
  skills:
    - repo: athal7/dotfiles
      skill: commit
    - repo: athal7/dotfiles
      skill: review