AgenticAle is a free, open-source collection of coding skills for AI assistants. Its everyday workflow coordinates implementation, independent review, and focused fixes. A durable workflow carries larger goals across multiple rounds and sessions.
The workflow is designed to lower model cost by using fast, inexpensive, or local models for routine high-volume work, and stronger models for difficult implementation and independent review. Successful delivery, including a requested pull request, includes independent review, acceptance checks, and fixes for blocking findings, so your attention goes to completed work.
Describe the outcome you want and invoke a workflow explicitly. The session where you invoke it coordinates the work; there is no separate coordinator to select.
/work Fix the date filtering bug
/work Investigate why export is slow and recommend a fix
/work --pr Add CSV export to the report table
/review main check the migration commit for regressions
/autonomous 4 Migrate the report pipeline and verify the new behavior
These examples use conceptual slash syntax. Select the skill or use the
invocation supported by your client, as described in the
installation guide.
On OpenCode, add --stats to a /work invocation to include a session
usage summary — duration, per-session token counts, and a per-provider
breakdown — in the final report.
Open the documentation site for getting started, skill guides, installation and routing, and contributing.
Choose the client setup in the installation guide.
The five skills are work, review, autonomous, source-code-lookup, and
pull-request-description. The three workflow skills are explicit-only; the
supporting skills are loaded when their tasks call for them.
With the Copilot CLI, install the published plugin from the AgenticAle marketplace:
copilot plugin marketplace add alefranz/AgenticAle
copilot plugin install agenticale@agenticaleIn the GitHub Copilot app, open these links. Each one asks you to confirm, then opens a pre-filled form:
- Add the marketplace (
alefranz/AgenticAle) - Install the plugin (
agenticale@agenticale)
Or do it by hand: Customize → + Add → Marketplace with alefranz/AgenticAle,
then + Add → Plugin with agenticale@agenticale.
Start a new session and try one that only reads:
/agenticale:review main
You get an independent review back: the blocking findings, the nits, and the checks it ran. Nothing gets changed, committed, or pushed.
See Copilot installation.
See Codex plugin installation or the standalone skill option.
See the standalone skills installation guide.
Quick start without a checkout:
curl -fsSL https://raw.githubusercontent.com/alefranz/AgenticAle/main/scripts/install.sh | sh
(Windows PowerShell: download install.ps1, then
powershell -ExecutionPolicy Bypass -File install.ps1). See
OpenCode installation
for options and the local-checkout flow.
Work completes a bounded task through implementation, independent review, focused fixes for blocking findings, and acceptance checks. When investigation is needed, it gathers evidence first. Autonomous applies the same review gates to long goals and saves a durable next action for a later session. Each workflow reports its checks and remaining limitations.
Before dispatching implementation, the coordinator briefly assesses what the request would touch — which sources other actors can also mutate, which side effects would survive an interruption, and which compatibility surfaces are involved — and inspects only as much as a concrete uncertainty requires. A small question is answered with a direct read; a focused exploration round runs only when a question needs its own context, so no discovery round is mandatory. A blocking finding that needs concurrency, durability, state-machine, or ownership reasoning keeps the fix role but is dispatched on the configured hard route.
Both workflows keep you informed: a concise decision summary precedes each substantive step. When an allowance is exhausted, the stop report says which allowance is spent, what is completed and what is blocked, and recommends a bounded correction, a bounded replan, or a stop that preserves the current work. Where your host offers choices, it offers them — for example one additional fix-and-review cycle — and a granted extension adds exactly the named rounds without resetting counters or granting extra rights; declining ends the stop. Every incomplete stop includes a copy-pasteable continuation prompt. Provider quota exhaustion is separate: Work stops without durable recovery, while Autonomous applies its own cooldown and handoff policy.
Routing preferences and install options are documented in the model routing guide.
Work leaves changes ready for your review unless you ask it to prepare and open a pull request. Autonomous keeps a durable handoff in the target project; commit, branch, push, and pull request actions are separate requests. See the skill guides.
Use Autonomous for a goal that spans multiple rounds or sessions.
The supporting source code lookup and pull request description guides explain when each skill is loaded and what it returns.