agent

KO | EN

gitlite — search

#javascript #python #hacktoberfest #go #android #machine-learning #deep-learning #linux #macos #windows #awesome #kubernetes #security #open-source

(★ 9)

A homegrown GitHub automation tool for agentic engineering workflows

agent Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Sudden Agent

Run programmable agents on any [GitHub Workflow](https://docs.github.com/en/actions/concepts/workflows-and-actions/workflows) (issues, pull requests, comments, schedule, dispatch, etc.) to enable automation across your repo.

## What you can automate

You can attach an agent to any workflow and provide a custom prompt to unlock a wide range of automations. For example:

- [Code review](recipes/code-review.md) - Review PRs, respond to comments, and open follow-up issues.
- [Issue assistant](recipes/issue-assistant.md) - Auto-triage issue threads with clarifying questions and duplicate detection.
- [Manual dispatch](recipes/manual-dispatch.md) - Kick off a one-off run with a custom prompt.
- [Security audit](recipes/security-audit.md) - Run periodic code security reviews and file issues.
- [Tag and release](recipes/tag-and-release.md) - Tag and publish releases when versions change.
- [Todo to issues](recipes/todo-to-issue.md) - Create issues for new TODOs introduced on the default branch.
- [Version bump](recipes/version-bump.md) - Suggest version bumps based on PR impact.

Have a useful workflow? [Share your recipe](https://github.com/sudden-network/agent/new/main/recipes?filename=recipe.yml)

## Persistent sessions

Sessions persist per issue and pull request, so the agent picks up where it left off across new comments, edits, and new commits.

This makes iterative work practical: the agent remembers what it already covered, reacts to changes, and stays consistent throughout the process.

## Action inputs

| Input | Required | Description |
| --- | --- | --- |
| `agent` | no | Agent to run (`codex` default). |
| `agent_api_key` | no | Agent API key. Required unless `agent_auth_file` is set. |
| `agent_auth_file` | no | Agent auth file content (agent-specific). |
| `github_token` | no | GitHub token used by the action (defaults to the workflow token). |
| `github_token_actor` | no | Actor login for `github_token` when using a non-workflow token (e.g. `sudden-agent[bot]`). |
| `model` | no | Agent model override (for Codex, append reasoning effort and service tier with `/`, e.g. `gpt-5.6-sol/xhigh/fast`). |
| `prompt` | no | Additional instructions for the agent. |
| `resume` | no | Enable session persistence. Default: `false`. |
| `sudo` | no | Disable sandbox; allow write + network access. Default: `false`. |

- Session persistence requires the `actions: read` permission to download artifacts.
- Artifact retention is controlled by your repo/org settings (see [Workflow Artifacts](https://docs.github.com/en/actions/concepts/workflows-and-actions/workflow-artifacts)).

## Configuring the agent

- Use `prompt` for per-workflow instructions.
- If you want repo-level instructions, add an [AGENTS.md](https://agents.md/) file and run this action after `actions/checkout` so the agent can read it.

### Codex model selection

Set `model` to `<model>[/<reasoning effort>[/<service tier>]]`. For example, use `gpt-5.6-sol/xhigh/fast` for Sol with extra-high reasoning and fast mode.

Current GPT-5.6 choices are:

- `gpt-5.6-sol` for complex work where capability matters most.
- `gpt-5.6-terra` to balance capability and cost.
- `gpt-5.6-luna` for efficient, high-volume work.

See OpenAI's [model guidance](https://developers.openai.com/api/docs/guides/latest-model) for current model and reasoning options. Omit `model` to let Codex choose its default.

## Authentication

Authentication is agent-specific, but this action exposes two generic ways to pass credentials:

- `agent_api_key`: API key auth (recommended for CI).
- `agent_auth_file`: inject an agent-specific auth file into the runner.

Treat `agent_auth_file` like a password (it grants access to the underlying agent account).

For the default agent (`codex`), `agent_auth_file` can be used to inject Codex's `auth.json` (from `~/.codex/auth.json`) so the CLI can use a ChatGPT subscription.

To use Codex with a ChatGPT subscription in GitHub Actions, store a dedicated `CODEX_AUTH_JSON` secret and use a [GitHub App token](github-app/README.md) so the action can save the updated Codex login after each run. See the [GitHub App setup](github-app/README.md) for the workflow setup and bootstrap command.

## Permissions

This action relies on the workflow `GITHUB_TOKEN`.

> Grant only what you need at the job level.
See GitHub documentation for [permissions](https://docs.github.com/en/actions/security-guides/automatic-token-authentication#permissions-for-the-github_token).

If you provide a non-workflow token, also set `github_token_actor` so the agent can recognize its own comments.

Common permissions:
- `issues: write` to post issue comments (including PR conversation comments).
- `pull-requests: write` to comment on PRs and open PRs.
- `contents: write` to push branches/commits.
- `actions: read` to download/list artifacts.

If you want the agent to open PRs, also enable the repo setting:
Settings -> Actions -> Workflow permissions -> "Allow GitHub Actions to create and approve pull requests."

Note: GitHub blocks `GITHUB_TOKEN` from updating workflow files in `.github/workflows/`.

## Security

- The `GITHUB_TOKEN` is held by the action process and is not exposed directly to the agent.
- The action refuses to run unless the triggering `github.actor` has write access to the repo.
- GitHub side effects are constrained by the workflow `permissions` you grant to `GITHUB_TOKEN`.
- By default, `GITHUB_TOKEN` is scoped to the repository running the workflow: it cannot write to other repositories unless you supply a broader token with cross-repo access.
- Agents run in `read-only` sandbox mode: they can read files but cannot write to disk or access the network, even from shell commands.
- In non-`sudo` Codex mode on GitHub-hosted Linux runners, this action prepares Codex sandboxing by enabling unprivileged user namespaces and clearing Ubuntu's AppArmor user namespace gate for the rest of the job.
- Use clear, scoped prompts and least-privilege permissions.
- Agents can make mistakes, so scope triggers carefully and keep permissions minimal before enabling.
- Keep humans in the loop for decisions that affect code, security, or policy.

This action uses generative AI to produce responses and automation. It runs with the workflow’s permissions; review prompts, triggers, and repo access before enabling it, and keep permissions minimal.

## License

This project is licensed under the MIT License — see the [LICENSE](LICENSE) file for details.

agent

agent Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

agent

agent Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README