Autonomous delivery pipeline for AI coding agents. Idea to shipped PR.
Don't manage your AI. Let it manage the work.
/plugin marketplace add https://github.com/jasonnoahchoi/abilities
/plugin install abilities@abilitiesOr from a local clone:
git clone https://github.com/jasonnoahchoi/abilities.git
/plugin marketplace add /path/to/abilities
/plugin install abilities@abilitiesThe plugin's postinstall hook automatically installs the abi CLI (formerly named abilities) via uv tool install bllshttng-abilities (or pip install --user bllshttng-abilities if uv is unavailable). abi --help is on your PATH in a new Claude Code session. The legacy abilities binary still works for one release but prints a deprecation warning on stderr.
The abi CLI powers the autonomous delivery pipeline. Install it one of two ways:
/plugin install abilitiesThe plugin's postinstall hook runs uv tool install bllshttng-abilities (or pip install --user bllshttng-abilities if uv is unavailable). abi --help is on your PATH in a new Claude Code session.
uv tool install bllshttng-abilities
abi --versionRequires Python 3.11+ and uv (https://docs.astral.sh/uv/).
/abilities:ralph "add user authentication with OAuth"
Ralph handles everything: design, planning, TDD, code review (6 parallel agents), and PR creation. It refuses to stop until the feature ships.
Size profiles control how much ceremony ralph adds:
/abilities:ralph -S "fix login redirect" # small: build + PR
/abilities:ralph -M "add OAuth" # medium: + verify, docs, external review
/abilities:ralph -L "rebuild auth system" # large: + research, adversarial, browser, goals
Default is medium. Individual flags override any size: /abilities:ralph -M --adversarial "feature".
423 PRs shipped across real projects. The pipeline tracks cost per task - median $16/PR on Opus, with ~96% cache read ratio from progressive skill loading. Model routing keeps costs down: Opus for code review, Sonnet for execution, Haiku for verification and PR creation.
/abilities:megaralph --council --loop vision.md
Megaralph assembles a think-tank council to debate priorities, generates a roadmap, then continuously executes tasks through ralph until the roadmap is done or budget is exhausted. Each task is automatically sized (-S/-M/-L) based on complexity. Retrospectives run every 3 tasks to catch priority drift.
/abilities:megaralph --status # see progress
/abilities:megaralph --retro # run a retrospective
/abilities:megaralph --continue # resume after interruption
/abilities:think "user auth" # explore the design space (1-on-1)
/abilities:think-tank "user auth" # multi-persona product debate (1-on-many)
/abilities:plan "user auth" # create implementation plan
/abilities:do path/to/plan.md # execute with TDD
/abilities:sigma-review # run parallel review agents
/abilities:create-pr # ship it
/abilities:debug "login fails on mobile Safari"
/abilities:bbb # launch companion in a split pane
bbb reroll # roll for a new buddy (gacha)
bbb current # show your buddy and stats
A reactive companion that watches your session and gives personality-driven commentary via Haiku. Each buddy has randomized stats (debugging, patience, chaos, wisdom, snark) that shape how they react. A high-debugging buddy spots real bugs; a high-snark buddy roasts your code.
Built in honor of Claude Code's /buddy command, removed in v2.1.97 (Apr 8, 2026).
/abilities:tower-watch # mirror transcript in a side pane
/abilities:tower-play # browse past sessions in a TUI
/abilities:copy-this # clipboard
/abilities:copy-this --file out.md # file
| Guide | What you'll learn |
|---|---|
| Getting started | Install, first feature, configuration |
| Ralph pipeline | The autonomous loop - flags, gates, cross-project, resume |
| Think and plan | Design exploration, plan creation, wave-based execution |
| Execution modes | When to use ralph vs do vs operator vs ralphclaw |
| Utilities | Tower, copy-this, debug, feels, code review, distill |
For deep technical details, each skill's SKILL.md is the source of truth (e.g., skills/ralph/SKILL.md).
Auto-merge (opt-in): ralph, ralphclaw, and megaralph can merge PRs automatically after review. See skills/_shared/auto-merge.md for setup, CLI modifiers, conflict resolution behavior, and failure modes.
Skills that work without the full plugin. Install via npx skills add or copy directly:
npx skills add bllshttng/abilities/tdd
# or manually:
cp -r skills/tdd .claude/skills/tddFull autonomous pipeline (portable):
npx skills add bllshttng/abilities/ralphclaw
npx skills add bllshttng/abilities/think
npx skills add bllshttng/abilities/spec
npx skills add bllshttng/abilities/do
npx skills add bllshttng/abilities/create-pr
npx skills add bllshttng/abilities/check-pr| Skill | What it does |
|---|---|
| ralphclaw | Autonomous delivery pipeline (portable orchestrator) |
| think | Design exploration with BDD acceptance criteria |
| spec | Implementation planning (full default, quick for single-file) |
| do | Lightweight plan executor |
| create-pr | PR creation with description |
| check-pr | External review polling and implementation |
| debug | Scientific method bug hunting |
| fix | Autonomous fix loop with auto-revert |
| codemap | Structural codebase analysis via PageRank (+ --db for schema context) |
| distill | Extract reusable skills from conversation patterns |
| ship-docs | Generate architecture and how-to documentation |
| tdd | Test-driven development discipline |
| think-tank | Multi-persona product debate (1-on-many) |
| what-if | Scenario exploration |
| tower-watch | Live session mirror |
| tower-play | Session TUI browser |
| copy-this | Conversation export |
| bbb | Reactive companion with personality stats |
| cache-keepalive | Keep prompt cache warm during idle |
| token-doctor | Diagnose token burn and cache issues |
| feels | Learned behavioral patterns |
| speculate | Run parallel variations of the same task |
| git-worktrees | Create isolated git worktrees for features |
| setup | Interactive settings.yaml wizard |
7 skills remain internal (require the full plugin): ralph, operator, sigma-review, megaspec, cross-project-pipeline, audit, expedition.
Skills are plain markdown and work across Claude Code, Gemini CLI, and Codex CLI. Orchestration hooks are provider-specific (see GEMINI.md and AGENTS.md for Codex/Gemini setup).
| Feature | Claude Code | Gemini CLI | Codex CLI |
|---|---|---|---|
| Skills | Full | Full | Full |
| Ralph pipeline | Full | Full | Full |
| Subagent dispatch | Yes | Sequential fallback | Sequential fallback |
| Hooks (stop, feels) | Yes | Yes | Yes |
| Cost tracking | Yes | Partial | Partial |
/abilities:setup # interactive wizardOr create .abilities/settings.yaml:
project:
vision: "Your project description"
config:
max_iterations: 20
budget_cap: 25
external_reviewer: gemini # gemini | coderabbit | claude | noneRequired: jq, gh (authenticated)
Recommended: Python 3 (metrics, cost tracking)
Optional: Playwright (browser testing), textual (tower TUI)
MIT - Jason Noah Choi