code-task

Fixes issues and implements features in real repositories with verified red-green test loops.

6.8k|540|Updated May 6, 2026
One-click install
npx skills add https://github.com/opensquilla/opensquilla --skill code-task
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: code-task
Source: https://github.com/opensquilla/opensquilla/tree/main/src/opensquilla/skills/bundled/code-task
Command: npx skills add https://github.com/opensquilla/opensquilla --skill code-task

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Making code changes to a real repository by hand in a chat session leaves no proof the change actually works. This Skill clones the target repo into a disposable working directory, runs an OpenSquilla agent to implement the fix or feature on a task branch, and independently verifies the result with a red-green-regression test loop before reporting a structured outcome.

Core Features & Use Cases

  • Repository issue fixing: Resolve a GitHub issue by number (via the gh CLI) or a free-text task description against any repo named by path or git URL.
  • Independent verification: Acceptance tests are proven red on the base commit and green on the change, plus a regression run against the existing suite, with internal retries on failure.
  • Multiple verification modes: red-green for existing repos, scratch for self-contained testable code written from scratch, and build for packaging a new app into an installer (.dmg, .exe, .AppImage).
  • Use Case: A user says "fix issue 412 in github.com/acme/widgets" — the Skill clones the repo, implements the fix on a task branch, proves the acceptance test flips from fail to pass, and reports the branch, diffstat, assumptions, and cost.

Quick Start

Ask the agent to fix issue 412 in github.com/acme/widgets using the code-task solve command and wait for the verified result report.

Frequently Asked Questions about code-task

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I fix a GitHub issue automatically with an AI agent?

Run code-task solve with --repo pointing at the repository URL and --issue N for the issue number. The GitHub CLI (gh) must be installed and authenticated; the issue body and comments are pulled in as the task specification.

How does red-green verification work for automated code changes?

The agent writes acceptance tests, the runner proves they fail on the base commit (red) and pass on the change (green), then runs the existing suite to catch regressions. The result state is verified only when all three stages succeed.

Can I use code-task without a repository to write code from scratch?

Yes, use --verification-mode scratch with --task and no --repo. It scaffolds a throwaway git repo, writes the code plus pytest coverage, and reruns the declared acceptance command, returning verification_kind=scratch.

Does code-task run in Docker or a sandbox?

No, version 1 runs directly on the host with no Docker isolation, so the target repository must be trusted. A Docker-isolated backend for untrusted repositories is planned but not yet available.

What should I do when a code-task run reports failed?

Do not relaunch the same task; the runner already retried internally up to max_attempts. Read final_failure_reason and the agent_stdout.log in the run directory, then surface the failure details to the user.

Why is the source repo empty while code-task is running?

code-task clones the source into an isolated run directory and does all work there, so the source repo stays untouched until a run finishes and verifies. Monitor progress via the status.json file in the printed run directory instead.