Roadmap

1.2 — Sandboxed reviews — delivered

Run each pull request review in an ephemeral Dev Container.
Let the model explore the checked-out repository through read-only tools instead of sending the source code in the prompt.
Anchor review comments to changed lines and classify findings by severity.

If a repository has no Dev Container configuration, run its review in a default container based on debian:stable-slim.
Use /workspace as the workspace path for the default container and as the fallback when workspaceFolder is not specified.

Introduce a type-safe model output for the action Steep should take, such as Review or Implementation.
Route review requests (for example, “@steep, review this PR”) to the existing review flow.
Recognize implementation requests (for example, “@steep, rename method XXX to YYY”) and establish their dispatch path. Execution of implementation requests is planned for 2.0.

Use the typed action from 1.3 to route explicit implementation requests to an implementation workflow.
For a request on an issue, clone the repository from its default branch and implement the requested features on a new branch.
For a request on an existing pull request, clone its branch and implement the requested modifications there.
When the agent's iteration loop ends, the orchestration layer automatically pushes the resulting changes: it opens a pull request for issue-based work, or updates the existing pull request branch. The agent does not receive or invoke a push tool.
Keep implementation work isolated in a sandbox; do not modify the default branch directly.

Read applicable AGENTS.md files from the repository and use their project-specific instructions when carrying out implementation requests.
Allow configuring a dedicated model for pull request reviews and another for implementation requests.
Route each action type to its configured model, with a clear fallback when only one model is configured.

Automatically discover project errors reported by Sentry, without waiting for a user comment.
Investigate the errors in an isolated sandbox, implement candidate fixes, and validate them before reporting the results.