maestro/docs/tools/requestpackage.md
oss-sync 77ee3bc426
Some checks are pending
CI / build-and-test (push) Waiting to run
sync: update from private repo (ddadfd71)
2026-07-08 23:35:00 +00:00

54 lines
2.4 KiB
Markdown

# RequestPackage
Request a Python wheel that is not already available, so an approver can install
it into **this workspace** and let the run continue.
## When to use it
You tried to `import somelib` (in `Bash` running `python3`, or `RunUserScript`)
and it failed with `ModuleNotFoundError`, and the library is **not** in the
preinstalled set. Instead of retrying `pip install` (which is blocked in the
sandbox), call:
```
RequestPackage({ name: "httpx", reason: "need an async HTTP client for the API calls" })
RequestPackage({ name: "pandas==2.2.2", reason: "pinned for reproducibility" })
```
- `name` — the package. Either a bare name (`httpx`) or an exact pin
(`httpx==0.27.0`). **Only `name==version` pins are accepted** — ranges
(`>=`, `~=`) and URLs/extras are rejected (the overlay must be reproducible).
- `reason` — a concrete justification. It is shown to the approver.
## What happens
- **Interactive runs (a user is watching the chat):** the movement pauses and an
Approve / Deny card appears in the chat. When a task-write user (owner / admin /
space editor) approves, the wheel is installed into this workspace's private
overlay and the run resumes automatically — now `import` works. If denied, you
continue without it.
- **Non-interactive runs (subtasks / scheduled):** the request is recorded (the
user finds it later) and you proceed **without** the package. Don't block on it —
finish with what you have, or `complete({status:"needs_user_input"})` if you
genuinely can't proceed.
## Rules and limits
- Wheels only, from a fixed package index. No source builds, no arbitrary index.
- A package that would shadow the standard library or a preinstalled package
(e.g. `os`, `numpy`, `pandas`) is rejected — those are already importable.
- If the package is already installed in this workspace, the tool tells you to
just `import` it (no approval needed).
- The install is per-workspace. Other workspaces do not see it.
- You cannot approve your own request — approval is a human/operator action.
## After approval
The run re-enters the same step. Re-run your Python; the import now succeeds. If
you call `RequestPackage` again for the same package, it reports "already
installed".
Related: preinstalled packages are listed in the error you get from a blocked
`pip install`. Workspace-wide package management (for operators) lives in
**Settings → the workspace's Python packages panel**.