Curupira

Curupira (by Caipora Labs) is the product name. The PyPI project, primary console script, and Python import are curupira. The short command curu is the same entry point.

Curupira runs automations on your machine. It takes a GitHub issue or pull request, or a local cron occurrence, and hands it to a coding-agent CLI you already have.

Each automation in the settings TOML watches one source (issues, pull requests, or a cron schedule) and carries its own prompt. All automations share one discovery, scheduling, and execution pipeline: run drains currently available tasks, while run --watch polls every automation continuously.

Requirements

• A supported platform wheel from PyPI (the CLI is a Rust binary; Python 3.11+ is used for packaging and optional official trigger extras)
• gh installed and authenticated (gh auth login) when using GitHub automations
• Only the CLIs used by the configured profiles need to be installed: claude, codex, the Cursor CLI (agent), gemini, GitHub Copilot CLI (copilot), kilo, opencode, pi, or qwen

Installation

Install the Rust CLI from npm or PyPI. Both ship the same executable and use ~/.curupira/ for configuration and state. The npm install does not need Python.

npm install -g @caipora-labs/curupira
# or, without a global install:
npx @caipora-labs/curupira --help

# Python
uv tool install curupira
pip install curupira

Official triggers (curupira-plugins) and curupira-sdk are optional. Install them when you need forge task sources:

pip install 'curupira[extensions]'  # installs curupira-plugins (+ curupira-sdk)
pip install 'curupira[github]'      # same package; forge-named extras are aliases

To pin a specific version, use uv tool install "curupira==X.Y.Z".

The project is caipora-labs/curupira. curupira --version and the short alias curu --version report the installed version. Task-source extensions (except built-in cron) are separate processes; coding agents are built into the Rust CLI. See Process extensions.

Documentation

See the full guide in Portuguese for installation, automation configuration, providers, and operational commands. Newcomers can start with the automations reference (triggers, minimal TOML examples, first-run commands, and common errors).

Configuration

The default settings file is ~/.curupira/settings.toml. Download the example configuration directly to that location, then adjust repositories, paths, queries, and prompts:

mkdir -p ~/.curupira
curl -fsSL https://raw.githubusercontent.com/caipora-labs/curupira/main/packages/curupira-cli/curupira.example.toml \
  -o ~/.curupira/settings.toml

The ~/.curupira directory is created automatically when the default file is first loaded. Pass --config path/to/settings.toml to use a different file; relative workspace, state, and automation paths are resolved from that file's directory.

[settings]
max_active_tasks = 1
workspace_dir = "~/.curupira/workspaces"
state_db_path = "~/.curupira/state.sqlite3"
log_level = "INFO"
# Optional OTLP/HTTP trace endpoint; omit it to disable telemetry.
# otlp_endpoint = "http://localhost:4318/v1/traces"

[settings.polling]
poll_interval_seconds = 30
batch_size = 100
cron_poll_interval_seconds = 1

[repositories.api]
remote = "https://github.com/acme/api.git"
# setup_script = "scripts/bootstrap.sh"

[agents.defaults]
profile = "opencode-default"
timezone = "UTC"

[agents.profiles.opencode-default]
provider = "opencode"

[automations.resolve-ready-issues]
trigger_type = "plugins-py/github-issues"
repository = "api"
repo = "acme/api"
labels = ["agent-ready"]
linked_pull_request = false
prompt = "Resolve issue ${issue_number}: ${issue_title}\n\n${issue_body}"

Automations

For a newcomer-oriented map of every built-in trigger (minimal TOML, placeholders, and first-run commands), see the automations reference.

[automations.<name>] is a keyed map; the map key is the automation ID and is carried into every task identity. repository selects a [repositories.<alias>] checkout with a full Git remote URL (and optional path / setup_script). trigger_type selects the source:

• "plugins-py/github-issues" / "plugins-py/github-pull-requests" — discover matching GitHub issues and pull requests; see the GitHub task source guide
• "plugins-py/github-project-items" — discover issues and pull requests in a GitHub Project board column (needs the read:project scope)
• "cron" — produces occurrences from schedule instead of querying a forge

Every automation requires repository and prompt. GitHub triggers also require forge repo identity; cron requires schedule. Optional profile selects a named CLI profile; otherwise the default profile applies. Relative paths resolve from the TOML directory. Different repository aliases cannot share one workspace path. Automations keep file order, and one-shot selection follows that order.

Providers and native options

Curupira supports these coding-agent CLIs, selected with provider in a profile:

• Claude Code (claude)
• Codex (codex)
• Cursor (cursor)
• Kilo CLI (kilo)
• Gemini CLI (gemini)
• GitHub Copilot CLI (copilot)
• OpenCode (opencode)
• pi (pi)
• Qwen Code (qwen)

How model, effort, agent, and permission options map to each CLI's native arguments is documented on the Providers page.

Prompts and placeholders

Placeholders use ${name} syntax. The host provides ${repository} and ${automation}. Every other name is a key from the trigger result, substituted as a string. A missing name is left unchanged. For dynamic prompts, Jinja syntax ({{ name }}, {% if %}, {% for %}) works next to ${name}; the keys of the GitHub triggers are listed in the GitHub task source guide. Cron prompts receive only the host variables.

Discovery, concurrency, and cron semantics

GitHub discovery compiles typed filters into Search queries and runs them through GraphQL with a token from gh auth token; see the GitHub task source guide for filters, deduplication, scopes, and prompt fields. Polls fetch up to batch_size items (default 100, up to 1000). When a full cycle finds nothing, the shared poller waits poll_interval_seconds (default 30s); consecutive empty cycles double the wait up to 5 minutes, and any discovery resets it. Each automation deduplicates its own items, so two automations may process the same forge item with different prompts.

Other task sources can be added as trigger plugins registered under the curupira.triggers entry-point group; see the plugins guide.

max_active_tasks bounds concurrently running coding agents (default 1). By default, each task runs in a new worktree beside its base checkout, so tasks for the same repository can run concurrently without sharing edits. The worktree branch is created from the fetched remote default branch and is not pushed. Set checkout = "main" to use the shared checkout instead (this means the shared checkout, not a branch named main, and restores the previous exclusive behavior). path continues to select the base checkout. With checkout = "main", the agent runs on the shared checkout exactly as it is: Curupira does not fetch, pull, or switch branches there. Checkouts are created on demand with git clone <remote> under workspace_dir/<repository-alias> (or the alias path).

An optional setup_script on the repository alias is a repository-relative executable path (no absolute paths or ..). It runs directly, with the checkout root as its working directory, only when the base checkout has just been cloned. It does not run for an existing checkout or in the task worktree. Use checkout = "main" when the agent must run where setup wrote files. A nonzero setup exit prevents the agent from starting; the newly cloned checkout is removed, while an existing checkout is preserved. validate checks the path syntax but does not require the script to exist. run --dry-run does not fetch, clone, create a worktree, or run setup.

Each cron automation coalesces overdue ticks into a single pending occurrence; the same automation never runs concurrently with itself. schedule is a five-field cron expression, timezone is IANA (defaulting to agents.defaults.timezone), and start_date/end_date form an optional inclusive window interpreted in that timezone. Without start_date, the window starts when the automation is first recorded.

OpenTelemetry

Set settings.otlp_endpoint to an OTLP/HTTP trace endpoint (for example, http://localhost:4318/v1/traces) to export one span for each dispatched issue, pull request, or cron occurrence. Each span includes the repository, type, identifier, and result (success or failure); failures also include error.message. The endpoint must accept OTLP over HTTP/protobuf. If the field is omitted, no telemetry is exported.

State files

Running sessions and cron schedule state live in state_db_path (default ~/.curupira/state.sqlite3). The per-user dispatch lock is stored in ~/.curupira/dispatch.lock, and the dedicated log directory is ~/.curupira/logs. Only one run or tui process can dispatch at a time; a second process exits with an error rather than running tasks in parallel. Session records are removed when the agent process ends; run and run --watch resume saved sessions after a restart. If the file exists but is not a compatible database, the application exits with an error instead of deleting it — delete or move the file yourself to start fresh.

Log file

The run and run --watch commands append records to ~/.curupira/logs/curupira.log; restarting the process does not erase existing content. Each task records its start and completion time, repository, type, and identifier. If a task fails, the record includes the error. Set [settings].log_level (or pass --log-level) to DEBUG for more internal detail. Uncaught crashes write separate reports under ~/.curupira/logs/crashes/.

Usage

Validate configuration without calling external CLIs or writing state:

curupira validate

Drain currently available tasks through the shared scheduler:

curupira run
curupira run --size 5

Preview one selected task without reserving, persisting, cloning, or executing:

curupira run --dry-run

Poll all automations with the shared bounded scheduler until interrupted:

curupira run --watch

Run the same continuous scheduler inside an interactive Textual dashboard:

curupira tui

The short alias curu accepts the same subcommands (curu validate, curu run, curu run --watch, curu tui).

validate exits 0 when the configuration is valid and 2 on configuration errors. run, run --watch, and tui exit 1 when any executed task failed, otherwise 0. run --dry-run never reserves or persists cron occurrences and does not perform checkout, worktree, or setup operations.

run --watch runs every CLI non-interactively so concurrent workers never contend for the terminal UI. tui is the interactive alternative. Editing the configuration file while either is running hot-reloads settings after in-flight tasks finish. Transient gh failures are retried with backoff; authentication, configuration, output-format, and agent-task failures are not retried automatically.

Public interface

Curupira is CLI-first. The only supported programmatic surface is curupira.__version__; all other modules are internal implementation details that may change without notice.

Development and validation

uv sync --dev
uv run --no-sync prek install
uv run --no-sync pytest
uv run --no-sync ruff check .
uv run --no-sync ruff format --check .
uv run --no-sync pyrefly check
uv build
uv run --no-sync twine check dist/*

uv sync installs the pure-Python package. uv run --no-sync prek install registers Git pre-commit hooks; Ruff and Pyrefly run through uv so they use the locked project environment. uv build produces a py3-none-any wheel and sdist with no compiler toolchain required.

To inspect branch coverage locally, run uv run --no-sync pytest --cov --cov-report=term-missing; the configured minimum is 85%.

See CONTRIBUTING.md for environment setup and the exact verification commands, and CHANGELOG.md for release notes.

License

MIT — see LICENSE.