Models

Primary
GPT 5.6 Sol

Primary Codex model for implementation, debugging, review, and repository-level reasoning.

Best code output and amazing tool calling. It does exactly what I tell it to, no more and no less. The only real downside is frontend work, but with good prompting and actual human taste it can still produce good results.

Secondary
GPT Codex 5.3 Spark

Secondary Codex model kept for fallback and comparison runs.

Gemini 3.6 Flash

Fast multimodal model reached through Antigravity CLI prompt runs.

GPT 5.6 Luna

Secondary OpenAI model for alternate reasoning and comparison runs.

Harnesses

Primary
Codex CLI

Primary terminal harness for code edits, tool execution, browser checks, and sub-agent orchestration.

After using OpenCode exclusively for months and spending too much time tweaking it or fixing issues, Codex CLI is a breath of fresh air. It does everything I need cleanly, and I genuinely have no complaints. I hope OpenAI models stay SOTA so I can keep using it forever, and I hope they do not enshittify it. If I did not main GPT models, I would probably switch to Pi or Droid CLI, maybe with T3 Code as a GUI, since the few times I have used them they have been pretty neat too.

Secondary
Hermes Agent

Secondary AI assistant for persistent terminal sessions, alternate model judgment, and parallel operator loops.

I use this through Discord as my personal AI assistant. Hermes fans love to say it is miles better than OpenClaw, but honestly it is not; it is as rough as most other AI assistants. It has given me a ton of headaches and breaks with almost every update because they keep adding sloppy PRs.

Droid CLI

Secondary coding-agent harness for terminal work, parallel tasks, and alternate model routing.

Antigravity CLI

Secondary prompt-run harness used with -p workflows to access Gemini 3.6 Flash, not as a coding-agent replacement.

Do not get me wrong, this harness sucks. I have a Google AI Pro plan with a generous Gemini 3.6 Flash quota, and since the only practical way I can use it is through their harness, I resort to agy -p wherever I do not need much intelligence and I do not want to use my Z.ai/GLM quota.

AGENTS.md

Loading raw operating manual...

Hooks

PreToolUse | Bash Safety

Routes rm commands through trash instead of permanent deletion.

PreToolUse | RTK Rewrite

Routes Bash through the local RTK rewrite guard when supported.

PreToolUse | Daily-Pacer Guard

Tracks rolling API usage state for each tool call and keeps calls inside the 5-hour quota window.

UserPromptSubmit | Daily-Pacer Guard

Checks usage state before prompt submission so active request loops pause while quota drains.

SessionStart | herdr State

Initializes the herdr operator state at session start.

MCPs

discord-py-self

Active Discord automation server.

codebase-memory-mcp

Structural code intelligence through a local MCP knowledge graph.

fff

Fast file finder with frecency-backed grep and filename discovery.

markitdown-mcp

Document-to-markdown conversion server.

perplexity-webui

Rendered Perplexity research bridge.

reddit-mcp-buddy

Reddit research server.

gitquarry

GitHub repository search and inspection server.

opensrc-mcp

Source package fetch and cache server.

github

GitHub API server.

kagi-mcp

Kagi search, extract, summarize, and assistant server.

parkour-mcp

Code-profile research and repository navigation server.

ghidra-headless-mcp

Ghidra reverse engineering server for headless binary analysis and decompilation.

semrush

SEO and competitive research data via the Semrush API.

ida-pro-mcp

IDA Pro reverse engineering server for binary analysis and decompilation.

openaiDeveloperDocs

OpenAI developer documentation server for API references and integration guides.

huggingface

Hugging Face Hub model, dataset, and API server.

CLI

rtk

Shell-command rewrite guard for routing noisy Bash calls into supported local tool invocations.

kagi-cli

Runs Kagi search, extraction, summarization, and assistant queries from the terminal.

fff

Fast local file finder with grep, frecency-ranked paths, and fuzzy filename discovery.

agent-browser

Browser automation for rendered checks, screenshots, clicks, forms, and page inspection.

lit

LiteParse document parsing for PDFs, Office files, screenshots, and local images.

gitquarry

GitHub repository discovery, ranking, and source inspection.

opensrc

Dependency source fetching and local cache access.

markit

Markdown conversion for files, URLs, and streams.

egaki

Terminal-driven image generation and editing.

mullgate

Turns Mullvad into authenticated SOCKS5, HTTP, and HTTPS proxies.

rg

Default raw text and filename search command.

jj

Canonical version-control interface for colocated Git repositories.

optimo

Image and video optimization through format-specific compression pipelines.

facts

Atomic, validatable truth statements about a project, managed as .facts files.

clawpatch

Semantic code review with feature mapping, findings persistence, and automated patching.

browserstack

Cloud-based real device and browser testing via CLI.

crabbox

Remote testbox CLI for leasing cloud boxes, syncing diffs, and running test suites off-box.

Skills

benchmark-or-eval-graphPersonal- Build graphs from local or remote data using a scripted workflow.
ce-simplify-codePersonal- Simplify recently changed code for clarity, reuse, quality, and efficiency.
chart-visualizationPersonal- Generate charts from structured data with AntV.
data-visualizationPersonal- Choose truthful visual forms instead of defaulting to bar charts.
ui-unslopPersonal- Unslop frontend UI after building or reviewing a page or layout.
crabboxPersonal- Operate Crabbox leases, machine inventory, and configured template boxes.
batch-grill-mePersonal- A relentless interview that asks every frontier question at once, round by round.
dft-writingPersonal- Generates long-form prose through the Deft Writing API with proxy support.
grill-with-docsFavourite- Stress-tests decisions against the project glossary and captures resolved language.
grill-meFavourite- Runs focused design interviews until assumptions and tradeoffs are explicit.
compact-handoffFavourite- Preserves task state across context compaction and resume boundaries.
repository-test-designPersonal- Designs tests that catch real defects instead of shallow snapshot churn.
optimoPersonal- Optimizes images and videos through local media pipelines.
namingPersonal- Structured naming system for private operator use.
thermo-nuclear-code-quality-reviewFavourite- Runs strict maintainability reviews.
mermaid-diagramsPersonal- Produces useful GitHub-compatible diagrams for systems, flows, and architecture.
language-selectionPersonal- Chooses implementation languages from constraints instead of taste alone.
create-cliFavourite- Shapes command-line interfaces, flags, help text, errors, and output contracts.
dynamic-workflowsFavourite- Orchestrates large tasks into supervised workflows with approval gates and subagents.
animated-faviconsPersonal- Build animated favicons with canvas, badges, and progress indicators.
apple-designFavourite- Apple's fluid interface design and physical motion, translated for the web.
grill-with-factsPersonal- Stress-tests a plan through design grilling and records decisions in .facts files.
make-interfaces-feel-betterFavourite- Polished UI, animations, hover states, and micro-interactions.
emil-design-engFavourite- UI polish, component design, and invisible details for great-feeling software.
changelogFavourite- Generates changelog entries from git history following Keep a Changelog conventions.
bugbashFavourite- Systematically explores software projects to find bugs with full reproduction evidence.
bughuntFavourite- Autonomous pipeline to discover, triage, fix, and audit bugs using parallel subagents.
commit-helperFavourite- Generates clear commit messages from git diffs following Conventional Commits.
ideateFavourite- Timeboxed brainstorming via propose-and-critique subagent pairs.
release-helperFavourite- Automates git tag creation and GitHub release workflows.
quality-codeFavourite- Type safety, real tests, and picking the right abstractions.
review-animationsFavourite- Review animation code against a high craft bar.
write-better-error-messagesFavourite- Review and rewrite error messages and error-handling UX.
codebase-memory-mcpPersonal- Structural code intelligence through a local MCP knowledge graph.
domain-modelingFavourite- Build and sharpen a project's domain model and ubiquitous language.
codebase-designFavourite- Shared vocabulary for designing deep modules with clear interfaces.
grillingFavourite- Interview the user relentlessly about a plan or design before building.
idea-generatorPersonal- Generate concrete, codebase-grounded improvement ideas as an issue-style report.
teachFavourite- Teach a new skill or concept within the workspace.
transitions-devFavourite- Production-ready CSS transitions for web apps.
timeboxed-iteratingFavourite- Run a task iteratively via subagents for a user-specified duration.
animation-vocabularyFavourite- Reverse-lookup glossary for web animation and motion effect terms.
cli-reviewFavourite- Run Greptile CLI review for the current local branch.
greploopFavourite- Iteratively improve a PR until Greptile scores 5/5.
improve-animationsFavourite- Survey animation code and produce prioritized audit plans.
resolving-merge-conflictsFavourite- Resolve in-progress git merge and rebase conflicts.
wayfinderFavourite- Plan huge work as a shared map of decision tickets.
better-accessibilityFavourite- Accessibility engineering for focus, keyboard, ARIA, forms and screen readers.
better-colorsFavourite- OKLCH color space conversion, palettes, and contrast for web.
better-interfaceFavourite- Cross-discipline interface review coordinating accessibility, layout, writing, typography, color and UI.
better-layoutFavourite- Layout structure from grouping and alignment to adaptive breakpoints.
better-typographyFavourite- Web typography from font selection to spacing and accessibility.
better-uiFavourite- Design engineering principles for polished interfaces.
better-writingFavourite- UX writing and interface copy from button labels to error messages and empty states.
effectPersonal- Opinionated guide for building production TypeScript applications with Effect v4.
effect-program-designPersonal- Design and review Effect service modules using the deep module pattern.
improve-codebase-architectureFavourite- Scan a codebase for deepening opportunities and grill through them.
find-fontsPersonal- Scout fonts from a design brief and return a license-checked shortlist.
transitions-polishFavourite- Polish and refine existing motion against the transitions.dev motion-token scale.
declankifyFavourite- Rewrite or review text against a banned AI-speak list and prose structure rules.
pick-ui-libraryFavourite- Pick the right library for a frontend task from a curated, opinionated list.
kagi-account-configPersonal- Configure Kagi CLI credentials, profiles, lenses, and account settings.
kagi-aiPersonal- Use Kagi's paid AI allowance for reasoning, coding, and research.
kagi-assistantPersonal- Use Kagi Assistant for conversational answers and thread continuity.
kagi-monitoringPersonal- Automate repeated Kagi queries with batch search, watches, and notifications.
kagi-usagePersonal- Search the web, extract pages, translate, or check news with the kagi CLI.
oraclePersonal- Oracle CLI browser workflow with direct model-picker recovery.
bikeshedPersonal- Run two independent subagents through a structured debate on the same prompt.

Code Review

Codex Code Review

OpenAI Codex as a GitHub Action for automated PR review and code suggestions.

Greptile

AI code review with full codebase context, line-by-line analysis, and PR scoring.

CodeRabbit

AI-first pull request reviewer with context-aware feedback and real-time chat.

Workspace

VPS

Active remote compute anchor for this operator environment.

herdr

Agent multiplexer that runs Claude Code, Codex, Copilot CLI, and 15+ agents in one terminal.

tmux

Workspace manager enabling persistent background sessions and multi-pane splits.

Tailscale

Zero-config mesh VPN based on WireGuard for secure remote workspace access.

Tabby

Terminal client used as one of the active remote-work surfaces.

Termius

SSH/session manager used for remote access and connection organization.

Socket Firewall

Package-install firewall for supply-chain risk checks.