OpenTask for GitHub

A Manifest V3 browser extension that makes the org's repo skills and common bot directives discoverable right inside GitHub's issue/PR comment box. It is built Chrome-first but deliberately portable to Edge, Firefox, and Safari.

Chrome Web Store listing is under review. In the meantime, install directly from the latest release ZIP.

• Key Features
• Overview
• Installation
• Usage
• Configuration
• Privacy
• Multi-Browser Support
• Development
• Roadmap
• Contributing
• License

Key Features

• 📜 Open Source: Available under the Apache 2.0 License.
• ⚡ Skill Tab-Completion: Type ! in a comment to open a caret-anchored, keyboard-navigable dropdown of the current repo's skills, fuzzy-filtered as you type.
• 🎛️ Quick Prompts Palette: A searchable palette of @opentask directives plus editable templates. Open with Ctrl/Cmd+Shift+P or the ⚡ toolbar button.
• 🔍 Repo-Aware: Resolves owner/repo from the page and fetches .agents/skills/ via the GitHub Contents API, cached per repo.
• 🤝 Non-Intrusive: Never touches GitHub's native @ / # / : completion; insertions fire an input event so draft-autosave and preview stay in sync.
• 🌐 Multi-Browser Ready: One dist/ bundle; only chrome.storage is used, so a port is a manifest tweak, not a rewrite.
• 🔒 Private-Repo Support: Optional fine-grained PAT (Contents: read), stored in extension storage. Supports per-owner tokens for multi-org setups.
• 🚀 One-Click Agent Install: Open any GitHub repo and use the Tasks tab in the repo navigation to install the OpenTask Agent workflow via a pull request. Requires a PAT with Contents: write, Pull requests: write, and Workflows: write.
• 🧩 Skills Install/Uninstall: In the repo's Skills tab (repo nav), a searchable, multi-select list of the Inference Gateway skills registry. Skills matching the repo's top languages are suggested first; applying opens a single PR that adds/removes skill folders under .agents/skills/.
• 🤖 Agents Panel: The Agents tab (repo nav) lets you select A2A agents from the agents registry to include in the workflow. Selected agents are spun up alongside the OpenTask agent at runtime.
• 📝 Free-Text Tasks: The Tasks tab sends a task to the agent. Leave Create a GitHub issue checked to open an @opentask issue, or uncheck it to run the task directly via workflow_dispatch (infer-action's direct-prompt) with no public issue.
• 🏗️ Project Init: The Init tab (repo nav) dispatches the installed workflow to scaffold an AGENTS.md for the repo and open a PR. Optional extras (configured in Settings) include a .githooks/pre-commit hook, a CLAUDE.md → AGENTS.md symlink, and a .claude/skills → .agents/skills symlink.
• 📋 Project Board Tracking: The installed workflow tells the agent to keep an issue's project-board Status in sync (In Progress on start, Done on completion), best-effort and board-agnostic. Board writes need a token with Projects: read and write - enable the GitHub App option, since the default GITHUB_TOKEN can't reach Projects v2; without it the agent skips board updates silently.
• 🤝 GitHub App Bot: Run the agent as your own GitHub App instead of github-actions[bot]. When configured, the workflow mints an installation token via actions/create-github-app-token and checks out + comments as the App, so its comments and commits are attributed to (and verified for) the App.
• 🖥️ Self-Hosted GPU Models: From the extension popup, provision a RunPod GPU that serves a llama.cpp OpenAI-compatible endpoint. Pick a catalog GGUF (or any custom Hugging Face repo:quant), choose a GPU, and deploy; the popup then hands you the repo secrets/variable to route tasks at llamacpp/<model>. See Self-hosted GPU models.
• 🧩 Plugin Support: Optional infer-action plugins extend the agent's capabilities. Toggle them on in Settings and re-install the workflow to bake them in.
• 🔧 Configurable Permissions: Control what the agent may do at runtime: create PRs, create issues, and comment on issues/PRs. Unchecked capabilities stay blocked.
• ⏱️ Configurable Timeout: Set the per-run job timeout for the generated workflow (default 25 minutes).

Overview

On the first !, the extension resolves owner/repo from the page URL and calls the GitHub Contents API (GET /repos/{owner}/{repo}/contents/.agents/skills) from the background service worker, caching the result per repo for 10 minutes. Repos with no skills directory simply show nothing - native completion is untouched.

The quick-prompts palette is a self-contained popup opened by a keyboard shortcut or a ⚡ button injected into the comment toolbar. Both surfaces share the same insertion path, which writes through the native textarea setter so React-controlled composers and GitHub's own draft/preview state stay consistent.

The Tasks, Skills, Agents, and Init tabs are injected into GitHub's repo navigation bar by the content script. Each opens a popover panel that communicates with the background service worker to install workflows, manage skills, select agents, or scaffold project files.

Installation

From a release ZIP (recommended until the store listing is approved):

1. Download the browser-extension.zip from the latest release.
2. Open chrome://extensions (or edge://extensions).
3. Enable Developer mode.
4. Load unpacked and select the dist/ folder inside the extracted ZIP.
5. Open any GitHub issue/PR, focus the comment box, and type !.

Manual (unpacked) for development or self-building:

bun install
bun run build      # outputs dist/

1. Open chrome://extensions (or edge://extensions).
2. Enable Developer mode.
3. Load unpacked and select the dist/ folder.
4. Open any GitHub issue/PR, focus the comment box, and type !.

Usage

• Skills: type ! at the start of a word to open the dropdown. Arrow keys navigate, Tab / Enter inserts /, Esc closes.
• Quick prompts: press Ctrl/Cmd+Shift+P (or click the ⚡ toolbar button) to open the palette, filter, and Enter to insert the selected template at the caret.
• Install the agent: navigate to any GitHub repo and click the Tasks tab in the repo navigation bar. Pick a model and click Install to open a PR that adds the OpenTask Agent workflow.
• Send a task: in the Tasks tab, type a prompt and choose whether to create a GitHub issue or dispatch the workflow directly.
• Manage skills: the Skills tab shows the skills registry, filtered by the repo's languages. Check skills to install and uncheck to remove, then click Apply to open a PR.
• Select agents: the Agents tab lists available A2A agents from the registry. Check the ones you want included in the workflow, then re-install to bake them in.
• Init a project: the Init tab dispatches the workflow to generate an AGENTS.md and open a PR. Configure extras (githooks, symlinks) in Settings.
• Refine an issue: on issue pages, a Refine button appears in the header (if enabled in Settings). Click it to have the agent rewrite the issue's description in place. Auto-refine on new issue creation is also configurable.

Configuration

Right-click the extension → Options (or the Details page → Extension options):

Accounts

Manage per-owner tokens and bot configurations. Each account pairs a GitHub owner (user or org) with a PAT and an optional GitHub App bot. The account whose owner matches the repo's owner is used automatically.

• Owner: your GitHub username or an organization. Populated from GitHub once you enter a token.
• Personal access token: required to install the OpenTask Agent workflow, send tasks, and list skills in private repos. Use a fine-grained token with Contents: write, Pull requests: write, Workflows: write, Issues: write, and Actions: write. Stored in this browser's extension storage.
• Custom bot (GitHub App): when enabled, the generated workflow authenticates as your GitHub App instead of github-actions[bot]. Create an App via the provided link, then enter its Client ID and the name of the repo secret holding its private key. The App needs Contents: write, Issues: write, Pull requests: write, Actions: write, and Workflows: write.

Quick prompts

A JSON array of { id, label, description, insert } objects shown in the palette. Editable, with a Reset to defaults button.

Install models

A JSON array of { model, keyInput, secret } objects offered in the Tasks tab's model dropdown. The first entry is the default. keyInput is the infer-action provider-key input (e.g. anthropic-api-key) and secret is the repo secret it reads. Add custom models here.

Self-hosted GPU models (RunPod)

Instead of a hosted provider, you can run a model yourself on an on-demand GPU and point tasks at it. OpenTask provisions a RunPod pod running llama.cpp's server (image ghcr.io/ggml-org/llama.cpp:server-cuda), which exposes an OpenAI-compatible endpoint the Inference Gateway reaches via the llamacpp provider.

1. Add your RunPod key. Options → Orchestrator → RunPod API key (from RunPod settings). Stored in this browser's extension storage. The GPU section in the popup only appears once a key is set.

2. Deploy. Open the extension popup:

• Pick a catalog GGUF (Ornith 1.0 9B, Llama 3.1 8B, Qwen 2.5 7B, Mistral 7B, Phi-4, Gemma 2 9B) or type any Hugging Face ref in custom HF repo:quant (e.g. unsloth/Qwen2.5-32B-Instruct-GGUF:Q4_K_M). Ornith 1.0 9B is the default - it's a small model tuned for agentic tool-calling, unlike general chat GGUFs.
• Choose a GPU (RTX 4090 → A100 80GB). The pod is pinned to hosts with CUDA ≥ 12.8 (the image's requirement) and secured with a generated --api-key bearer token.
• Hit Deploy. Status goes provisioning → running; the model's GGUF downloads and loads on the pod first (watch the pod's logs for server is listening).

3. Wire it into the repo. Once running, the popup shows copy-to-clipboard rows - add them under the repo's Settings → Secrets and variables → Actions:

Kind	Name	Value
Secret	LLAMACPP_API_URL	the pod endpoint + /v1
Secret	LLAMACPP_API_KEY	the generated bearer token (masked; Copy copies the real value)
Variable	DEFAULT_MODEL	llamacpp/<repo:quant> (the model the gateway registers)

DEFAULT_MODEL is the model used for issue-triggered runs. For Run task dispatches, the running llama.cpp model also appears in the Tasks-tab model dropdown. Then re-install the workflow so it wires the LLAMACPP_* secrets onto infer-action.

Each redeploy is a new pod with a new URL and token - update LLAMACPP_API_URL and LLAMACPP_API_KEY again from the popup. Deprovision from the popup terminates the pod so it stops billing.

Permissions

What the OpenTask agent may do while a task runs. These widen infer-action's read-only baseline; unchecked capabilities stay blocked. Re-install the workflow after changing these.

• Create pull requests (commit & push)
• Create GitHub issues
• Comment on issues & pull requests

Issue refinement

Let the OpenTask agent rewrite an issue's description in place. Refine edits the issue body via gh issue edit, so the installed workflow needs Create GitHub issues permission above - re-install after enabling.

• Show a Refine button on issue pages (default on)
• Auto-refine issues you create on GitHub (default off)

Project init

What the Init button (in a repo's nav) asks the agent to scaffold. It always generates an AGENTS.md and opens a PR; these add optional extras. Requires the OpenTask Agent workflow to be installed on the repo.

• Add a .githooks/pre-commit hook
• Symlink CLAUDE.md → AGENTS.md
• Symlink .claude/skills → .agents/skills

Workflow

• Timeout (minutes): per-run job timeout for the generated workflow (default 25). Applies to newly installed workflows; re-run Install on a repo to update an existing one.

Plugins

Optional infer-action plugins the installed workflow pre-installs to extend the agent. All off by default; check the ones you want. Re-install the workflow after changing these.

Theme

Choose how the options page and toolbar popup are displayed: System default, Light, or Dark.

Privacy

Everything the extension stores (your optional token, quick prompts, and a short per-repo skill cache) stays in this browser's local storage - nothing is synced or sent to any server. The only network call is a single GitHub Contents API request to list a repo's skills; there is no backend and no telemetry. See PRIVACY.md for the full data-flow breakdown and how to delete stored data; per-permission justifications for the store listing live in docs/store/privacy-declarations.md.

Multi-Browser Support

The same dist/ is the whole extension, and the only privileged API used is chrome.storage (present on Chrome/Edge/Firefox). Per-browser notes:

Browser	What's needed
Chrome, Edge	Works as-is (background.service_worker). Chrome Web Store and Edge Add-ons use the same dist/ ZIP.
Firefox 109+	Build with task build:firefox (applies manifest.firefox.json overrides: background.scripts + browser_specific_settings.gecko.id).
Safari 16.4+	Build with task build:safari, then wrap with xcrun safari-web-extension-converter on macOS. See docs/store/safari-listing.md for the full packaging and App Store release guide.

If the API surface ever grows beyond chrome.storage, drop in Mozilla's single-file webextension-polyfill and alias browser → chrome.

Development

Activate the Flox environment, then install the repository's pre-commit hook before making changes:

flox activate
task hooks:install  # required first-time setup
task setup          # installs the hook and Bun dependencies
task check          # typecheck, test, and build

Layout: src/content.ts is the imperative controller (DOM detection, caret math, insertion, keyboard, repo-nav injection); src/ui/* is the React view (skill menu, palette, Tasks/Skills/Agents/Init panels); src/background.ts is the service worker that fetches and caches skills, manages workflows, and handles the skills/agents registries.

To produce a store ZIP:

task package:chrome    # builds, zips dist/ to browser-extension.zip, writes SHA-256
task package:edge      # same ZIP, named edge-extension.zip for Edge Add-ons
task package:firefox   # builds with Firefox manifest, zips to firefox-extension.zip, writes SHA-256
task build:safari      # builds with Safari manifest; see docs/store/safari-listing.md for Xcode conversion

The ZIPs are also built and attached automatically to every GitHub Release by the release workflow.

Roadmap

• The newer React/ProseMirror composer on some 2024+ issue pages (v1 targets the classic <textarea>).
• Skill descriptions in the dropdown (would cost one API call per skill; names only for now, one call per repo).
• An org-wide skill catalog (v1 is per-repo).
• A packaged icon set, a build-time watch mode, and CI.

Contributing

Found a bug or have a feature in mind? You're more than welcome to open issues or submit pull requests for any fixes, improvements, or new ideas. Read the Repository Guidelines before contributing; they cover project structure, testing, style, commits, and pull-request expectations.

License

This project is licensed under the Apache 2.0 License.