Public portfolio hub for applied AI engineering, agent tooling, reusable skills, OSS contributions, and sanitized production engineering case studies by blockedby.
This repository is an index, not a product monorepo. The active work lives in the linked public repositories; this lab provides a concise map of the tools, operating model, and supporting evidence.
- Isolated Pi/Pipi environments and coding-agent runtime integration
- Multi-model delegation across Pi, Codex, Claude, Luna, and Terra
- Reusable Agent Skills for review, browser work, planning, and verification
- Deterministic evidence collection with bounded LLM-assisted interpretation
- Dependency-aware GitHub planning and conflict-safe parallel work
- Backend, Linux, browser-runtime, and operational automation
| Project | What it demonstrates |
|---|---|
blockedby/my-pi-setup |
Pipi, my current isolated Pi environment: pinned runtime, separate state, subagent profiles, parallel workflows, background terminals, browser/Codex integration, and reusable skills. Maintained as a public fork of davis7dotsh/my-pi-setup. |
blockedby/plan-gh-backlog |
Agent Skill and Python CLI for validating roadmap structure, planning dependency-ready batches, and safely publishing idempotent GitHub issue backlogs. |
blockedby/gpt5.6-reviewer |
Evidence-driven code-review contracts, routing, closure semantics, CLI validation, and a reusable review skill. |
blockedby/pi-codex |
Deterministic web search/fetch plus bounded Codex summarization, patch handling, and delegated coding tools for Pi. |
blockedby/browser-chrome-skill |
Portable Chrome DevTools skill with disposable headless sessions and an explicit persistent-browser boundary. |
Pipi composes the working environment around these capabilities: scoped implementation and audit agents, eight-way workflow fan-out, long-running background terminals, fd/rg discovery, ask-user interactions, browser MCP servers, and Codex/Claude/Pi delegation. The individual repositories keep each tool or skill inspectable and independently reusable.
The public skills/general collection contains the reusable engineering layer behind these workflows:
| Skill | Purpose |
|---|---|
backend-quality |
Backend, API, storage, validation, auth, idempotency, and data-safety evidence. |
frontend-quality |
Frontend implementation and UI-quality checks. |
devops-quality |
Configuration, CI, container, deployment, and runtime readiness. |
visual-composition |
Product-quality hierarchy, responsive composition, complete states, and interaction polish. |
completion-verification |
Fresh acceptance evidence before readiness or completion claims. |
git-branching |
Safe PR-first branches, worktrees, rebases, and target synchronization. |
browser-chrome |
Controlled persistent and disposable Chrome automation. |
explanatory-html-pages |
Self-contained technical explainers with readable diagrams. |
modern-skill-revising |
Focused context and instruction design for modern models. |
Standalone skills add specialized workflows: code-review provides evidence-driven review and closure contracts, while plan-gh-backlog validates and publishes dependency-aware GitHub backlogs.
| Area | Public target | Notes |
|---|---|---|
| Voice tooling | blockedby/kwispr |
Actively developed fork of MaksBoi/kwispr: Wayland/KDE dictation with cloud, OpenRouter, and local/offline STT, plus a Rust inference runtime and native desktop integration. |
| Systems automation | blockedby/vibe-practicum-vpn |
Public-safe Linux/KDE VPN and routing tooling, container test labs, guarded operations, and redacted diagnostics. |
| LLM SDK work | blockedby/go-openrouter |
Fork/contribution work around a Go OpenRouter SDK, with upstream attribution. |
| Linux automation | blockedby/linux-kubuntu-tweaks |
Personal Linux/Kubuntu automation and recovery runbook; environment-specific rather than a packaged product. |
| Browser runtime | Upstream Obscura commit | Merged Rust/CDP contribution attributed to blockedby: promise settling, DOM event and navigation side effects, click-submit parity, and end-to-end regression coverage. |
Start with case-studies/ for the reading guide.
- Obscura CDP click and navigation parity — upstream contribution connecting promise evaluation, DOM events, JavaScript navigation, and real click-submit regression coverage.
- Go OpenRouter SDK work — fork status, LLM API feature areas, and exact commit evidence.
- Sanitized private-production workflow — workflow patterns without private code, credentials, logs, domains, or user data.
- Linux/Kubuntu automation runbook — practical systems debugging and recovery automation.
- AI-assisted PR loop
- Task-package template
- Verification checklist
- Pinned repositories checklist
- Profile consistency check
I use agents as engineering tools, not as unchecked decision makers:
- define the problem, constraints, and acceptance criteria;
- split work into bounded slices with explicit ownership and dependencies;
- route exploration, implementation, and audit to the appropriate tool or model;
- keep deterministic evidence separate from model summaries or opinions;
- verify with tests, builds, static checks, browser evidence, or explicit manual checks;
- record what changed, why it changed, and what remains risky.
The submodules/ directory contains optional references to selected public repositories. They are convenient review links, not vendored source or ownership claims over upstream projects.
git clone --recurse-submodules https://github.com/blockedby/agentic-engineering-lab.gitOnly public repositories belong here. Private repositories must not be published or added as submodules.
- No secrets, tokens, cookies, credentials, raw logs, chat IDs, webhook URLs, private endpoints, or real user data.
- No private repositories are published or linked as public evidence.
- Forks and contributions are described as fork work, contribution work, or my changes in a fork.
- Upstream projects and maintainers are attributed clearly.