Senior iOS engineer at Numan. Previously: Babylon Health, Plum.

I live in London, UK, but am originally from Tuscany, in Italy. Find me dancing salsa and bachata after work! 💃

Sometimes I write about AI and software engineering, sometimes I write about life.

A few things I've been building:

• the last harness — an opinionated pi-based agent harness with primary agents, subagents, context management, safety rails, and install/update tooling.
• pi-assistant — use a Raspberry Pi as a voice AI assistant.
• claudio — a Swift & Vapor reimplementation of OpenClaw.

Skill issue?

• ai-agents-skills — the main collection of agent skills, Claude Code plugins, and Codex plugins I maintain.
  • agent-workflow-audit — audits a repo's agent workflow for wasted steps and unclear instructions.
  • github-librarian — researches GitHub code with gh, returning path-first findings with line evidence.
  • ios-agents-skills — skills focused on iOS and Apple platform development.
  • odds-api-io — gives agents access to sports odds via Odds-API.io.
  • pr-comments-triage — checks whether PR review comments are actually valid before fixing anything.
  • prd-interviewer — builds a PRD through a focused, one-question-at-a-time interview.
  • rebase-and-fix-conflicts — rebases onto main and resolves conflicts while preserving branch intent.
  • remove-ai-code-slop — cleans up telltale AI-generated code patterns.
  • sentry-cli — gives agents tools to interface with Sentry's error reporting.
  • starting-from-scratch — reviews a codebase's history and says what should change if you were starting again.
  • tfl-journey-disruption — plans London journeys and warns about TfL disruptions.

I've also started publishing extensions for pi my current agent harness of choice:

• agent-workflow-audit — runs an isolated repo workflow audit subagent and returns a distilled report.
• brrr — sends brrr push notifications when pi is ready for input.
• confirm-destructive — confirms destructive session actions like clear, switch, and fork.
• context-cap — caps effective model context windows to avoid the dumb zone.
• context-inspector — generates a local dashboard showing where the current session context is going.
• dirty-repo-guard — prompts before session actions when the current repo has uncommitted changes.
• gnosis — exposes the gn repo-local knowledge base CLI as an agent tool.
• inline-bash — expands trusted !{command} snippets before the prompt reaches the agent.
• librarian — adds a cached GitHub research scout for path-first code findings.
• minimal-footer — a minimal two-line footer with branch/repo, context/model, and OpenAI Codex usage.
• notify — terminal, desktop, bell, and sound notifications when pi is ready for input.
• openai-fast — adds /fast controls for OpenAI Codex Fast mode.
• oracle — a read-only oracle tool that selects the strongest reasoning model on the current provider/subscription.
• permission-gate — confirms dangerous bash commands like rm -rf, sudo, and chmod 777.
• quiet-tools — renders collapsed built-in tool rows as compact one-line invocations.
• review — adds interactive review flows for local changes, commits, PRs, and selected paths.
• todo — adds a branch-aware todo tool and a /todos viewer.
• triage-comments — checks PR review comments for validity before implementing anything.

If you're curious about how I set my machines up:

• dot — my public dotfiles

• Twitter @diegopetrucci
• Mastodon @diegopetrucci
• LinkedIn diegopetrucci