autopilot: event-driven autonomous pipeline

[Angelmmiguel]

Overview

Autopilot allows Rover to monitor repository events and act on them autonomously. It builds on top of existing Rover primitives (tasks, workflows, sandboxes) to handle the initial response to routine coding events — filing PRs for new issues, fixing CI failures, posting code reviews, and asking for clarification when needed.

Goals

• Allow Rover to monitor repository events and act autonomously on events that have a clear next step
• Augment the developer by handling triage and routine execution, so they can focus on reviewing, designing, and providing direction
• Track every decision the system makes for observability and debugging
• Support different levels of autonomy (self-driving vs assistant/dry-run mode)

Major Changes

AI-driven pipeline

Events are processed through a 10-step pipeline where AI agents participate in decision-making:

Event → Coordinate → Plan → Workflow → Commit → Resolve → Push → Notify
              ↘ Noop
              ↘ Cleanup

Trace-based tracking

Every pipeline execution is tracked as a trace — a tree of spans linked by actions. The full decision history is inspectable from the TUI or by reading the span/action JSON files directly.

~/.rover/data/projects/<projectId>/
├── spans/<spanId>.json
├── actions/<actionId>.json
└── autopilot/
    ├── state.json
    ├── traces.json
    └── memory/daily/YYYY-MM-DD.md

Configuration

rover autopilot                             # self-driving (default)
rover autopilot --mode assistant            # dry-run
rover autopilot --allow-events maintainers  # default actor filter

Custom instructions via .rover/AUTOPILOT.md and step-specific overrides (.rover/AUTOPILOT.coordinate.md, etc.).

Security model

• Actor filtering: only maintainers by default (--allow-events)
• Mode: assistant mode for safe observation without side effects
• Sandboxing: workflow tasks execute inside containers with isolated filesystems

Related

• Rover autopilot #553 - Autopilot discussion
• Cost Control #523 - Cost Control
• Support multiple sources and enrich task context #474 - Support multiple sources and enrich task context