Let agents inhabit a workspace across separate calls instead of only submitting one-shot execs. Add workspace shell open/read/write/signal/close across the CLI, Python SDK, and MCP server, with persisted shell records, a local PTY-backed mock implementation, and guest-agent support for real Firecracker workspaces. Mark the 2.5.0 roadmap milestone done, refresh docs/examples and the release metadata, and verify with uv lock, UV_CACHE_DIR=.uv-cache make check, and UV_CACHE_DIR=.uv-cache make dist-check.
2.2 KiB
2.2 KiB
Task Workspace GA Roadmap
This roadmap turns the agent-workspace vision into release-sized milestones.
Current baseline is 2.5.0:
- workspace persistence exists and the public surface is now workspace-first
- host crossing currently covers create-time seeding and later sync push
- persistent PTY shell sessions exist alongside one-shot
workspace exec - no export, diff, service, snapshot, reset, or secrets contract exists yet
Locked roadmap decisions:
- no backward compatibility goal for the current
task_*naming - workspace-first naming lands first, before later features
- snapshots are real named snapshots, not only reset-to-baseline
Every milestone below must update CLI, SDK, and MCP together. Each milestone is also expected to update:
README.md- install/first-run docs
docs/public-contract.md- help text and runnable examples
- at least one real Firecracker smoke scenario
Milestones
2.4.0Workspace Contract Pivot - Done2.5.0PTY Shell Sessions - Done2.6.0Structured Export And Baseline Diff2.7.0Service Lifecycle And Typed Readiness2.8.0Named Snapshots And Reset2.9.0Secrets2.10.0Network Policy And Host Port Publication3.0.0Stable Workspace Product3.1.0Secondary Disk Tools
Definition Of Done For The Roadmap
The workspace product is ready to leave beta when:
- the public contract is workspace-first rather than task-first
- an agent can inhabit a sandbox through shell, exec, service, diff, export, snapshot, reset, and explicit host-crossing operations
- the main docs lead with the workspace product, not one-shot VM execution
- the remaining deliberate deferrals are secondary disk tools rather than core workspace features