Dotfiles

• .config/gh/config.yml
• .gnupg/gpg-agent.conf
• .bashrc
• .gitconfig
• .gitconfig-private
• .inputrc
• .vimrc

Prerequisites

• Homebrew
• Ghostty
• Tmux
• GNU stow
• ZSH
• GH Cli
• delta
• Tree-sitter

CLI tools

• ripgrep
• fd
• bat
• eza
• fzf

Once Homebrew is installed the remainder of the prerequisites can be installed with the following command:

brew install git tmux stow zsh gh eza delta zellij ripgrep fd gnu-sed coreutils
cargo install tree-sitter-cli --locked

Installation

Checkout the repository into ~/.dotfiles and use stow to create the symlinks:

git clone git@github.com/marcofranssen/dotfiles.git ~/.dotfiles
cd ~/.dotfiles
stow .

Configure ViM

curl -sS https://raw.githubusercontent.com/marcofranssen/dotfiles/main/install_plugins_vim.sh | bash

Configure git

Prerequisites

curl -sS https://raw.githubusercontent.com/marcofranssen/dotfiles/main/install_gitconfig.sh | bash

Configure gh cli

curl -sS https://raw.githubusercontent.com/marcofranssen/dotfiles/main/install_ghconfig.sh | bash

Configure AI

nono.sh

I'm using nono.sh to sandbox my agents.

Fresh setup, outside any nono sandbox:

brew install nono

# Installs regular files under ~/.config/nono/packages (not in this repo or Stow).
nono pull nolabs-ai/pi
nono pull nolabs-ai/omp
nono pull nolabs-ai/hermes

# Link the custom profiles from this repo.
cd ~/.dotfiles && stow .

nono list --installed

Custom profiles in .config/nono/profiles remain versioned and Stow-managed.

The agent extension and skill symlinks that nono pull creates under ~/.omp and ~/.pi are intentional; they point into the installed package store. After installing or updating packs, restart the agent's nono session.

The Oh My Zsh function in .oh-my-zsh/custom/nono.zsh preserves Stow-managed profile symlinks when running nono profile promote <name>. Load it into an existing shell with:

source ~/.dotfiles/.oh-my-zsh/custom/nono.zsh
nono profile promote omp --diff
nono profile promote omp

For symlinked profiles, it validates the draft, shows a diff, asks for confirmation (unless --yes), and copies the draft to the resolved repository file. The draft is retained. This is a manual apply workflow, not native promotion: it does not provide nono's base-hash conflict checks or atomic replacement. Run it outside the sandbox and avoid concurrent profile edits. Regular profiles and other nono commands pass through to the executable unchanged.

The composed OMP profile extends nolabs-ai/omp, github, and aws. The github overlay owns GitHub credential capture and gh configuration grants. The aws overlay owns host-side SigV4 signing and the sandbox's placeholder credentials; it currently uses the bedrock AWS profile in us-east-1 for STS, Bedrock, and Bedrock Runtime.

When applying drafts from ~/.config/nono/profile-drafts, promote the dependencies before the composed profile, outside the sandbox:

nono profile promote aws
nono profile promote github
nono profile validate --draft omp
nono profile promote omp
nono run --profile omp -- omp

Keep ~/.aws protected. OMP uses placeholder keys for provider availability and client-side signing; the nono supervisor resolves the real AWS profile and re-signs requests. Existing sandbox permissions do not change until restart.

Pi reuses the same github and aws overlays alongside nolabs-ai/pi, while keeping its Pi Lens and SSH known-hosts grants in the Pi profile. After the shared profiles are promoted, apply the Pi draft outside the sandbox:

nono run --profile pi -- pi

Pi

Install the Pi coding agent:

pnpm add -g --ignore-scripts @earendil-works/pi-coding-agent

Start Pi and authenticate each provider that should be available to the parent session or subagents:

pi

Then run /login and complete the flows for OpenAI Codex and Amazon Bedrock. Both providers are intentionally enabled. The default parent model is OpenAI Codex Terra, while specialist subagents can use either provider. After adding a provider or changing enabledModels, restart Pi: the model picker's scoped model list is created when a session starts.

The stowed configuration lives at .pi/agent/settings.json (linked to ~/.pi/agent/settings.json). Use /model to change the parent model in a running session. A default provider is not a provider lock: every authenticated model listed in enabledModels remains selectable, and subagents use their fully-qualified configured model independently of the parent model.

Model routing

The configuration uses a cost-aware three-tier model mix:

• Luna for cheap, bounded reconnaissance and generic delegation.
• Terra for normal substantial work, including source-backed research.
• Sol for adversarial reviews where deeper reasoning is worth the cost.
• Claude Sonnet/Opus where their larger context or implementation/advisory strengths are more valuable than minimizing Bedrock usage.

subagents.defaultModel is Luna at low thinking. It is a safe fallback for future or package-provided agents that do not define a model or thinking level; it does not override the explicit mappings below.

Subagent	Primary model and thinking	Fallbacks	Why
scout	OpenAI Codex Luna, low	Bedrock Luna, then Claude Haiku 4.5	Fast, inexpensive codebase reconnaissance and compact handoffs.
delegate	OpenAI Codex Luna, low	Bedrock Luna, then Claude Sonnet 4.6	Routine, well-scoped tasks and summaries should use the lowest capable tier.
researcher	OpenAI Codex Terra, medium	Bedrock Terra, then Claude Sonnet 4.6 / Sonnet 5	Terra is the normal source-evaluation and synthesis tier; Sonnet fallbacks preserve access to much larger context when needed.
worker	Claude Sonnet 5, high	Claude Sonnet 4.6	The primary implementation lane favors Sonnet's coding behavior and 1M-token context. High thinking reduces costly implementation/rework cycles.
reviewer	OpenAI Codex Sol, high	Bedrock Sol, then Claude Opus 4.8 / Sonnet 5	Reviews are adversarial correctness/security work, so deeper reasoning is deliberate. Use a one-off Terra override for inexpensive routine reviews.
oracle (advisor)	Claude Opus 4.8, high	Claude Sonnet 5, then Claude Fable 5	Rare high-stakes architecture and decision-consistency checks optimize for judgment rather than cost.

Fallbacks preserve the same model across providers first: Codex Luna, Terra, or Sol retries through its Bedrock equivalent before moving to a different Claude model. The later fallback models are used only for retryable provider/model failures, such as rate limits or temporary availability errors; they are not a task-based quality escalation mechanism. Use a per-run model override when a specific task needs a different tier.

For example, run a bounded implementation with Terra without changing the persistent worker mapping:

/run worker[model=openai-codex/gpt-5.6-terra:high] "Implement the approved plan"

Installed Pi packages

The packages list in .pi/agent/settings.json installs the following extensions and skill sources when Pi starts. Package changes also require a Pi restart.

Package	Purpose
npm:pi-subagents	Adds the multi-model subagent fleet, including the scout, delegate, researcher, worker, reviewer, and oracle roles used above.
npm:pi-web-access	Provides web research tools such as web_search, fetch_content, and get_search_content, enabling source-backed researcher runs.
git:github.com/apmantza/pi-lens	Adds code-intelligence tools: symbol/module discovery, targeted reads, LSP diagnostics, and optional AST/LSP navigation tools.
npm:@piotr-oles/pi-reflag	Provides reflag, a search/command utility used for efficient repository inspection. brew install ripgrep fd
git:github.com/cli/cli	Supplies GitHub CLI-related agent skills and usage guidance. The gh executable itself is installed separately through Homebrew.

Packages can expose tools, commands, agents, or skills. A skill's guidance is loaded only when the task matches it; package installation alone does not force a subagent or tool to run. Keep package names in settings.json rather than installing them ad hoc so the Pi environment is reproducible across machines.