Background
Chimera has chimera/env/ with Local, Docker, Git, Remote, Cloud, PersistentShell abstractions. The "Cloud" backend is largely a placeholder — actual cloud sandbox provider integration (Modal, Daytona, Northflank, e2b, Coder Cloud) was scoped in earlier session memory but never built.
For the ferret sandbox-first CLI and for ProgramBench-style benchmarks (which require isolated containers), cloud sandboxes would let users run agents without setting up Docker locally — and would unlock the "agent runs in cleanroom, no network" path that some benchmarks require.
Scope
Three providers, each as an Environment subclass:
- Modal — Python-native sandboxes, pay-per-second compute.
chimera/env/modal_env.py.
- Daytona — workspace-as-code, supports custom images.
chimera/env/daytona.py.
- Northflank or e2b — pick whichever has cleaner Python SDK.
Each:
- Implements
Environment ABC (provision, exec, read, write, teardown)
- Has its own optional extra:
pip install "chimera-run[modal]", etc.
- Includes mocked unit tests + 1 live-gated integration test
- Doc:
docs/env/{modal,daytona}.md
Acceptance criteria
Estimated effort
Large — ~2 weeks. Could be broken into 3 issues (one per provider).
Why this matters
Cloud sandboxes are how teams scale agent benchmarks (each instance gets a fresh container). chimera ferret without cloud sandboxes is local-only; with them, it's competitive with the existing open-swe-agent workflow.
Background
Chimera has
chimera/env/with Local, Docker, Git, Remote, Cloud, PersistentShell abstractions. The "Cloud" backend is largely a placeholder — actual cloud sandbox provider integration (Modal, Daytona, Northflank, e2b, Coder Cloud) was scoped in earlier session memory but never built.For the
ferretsandbox-first CLI and for ProgramBench-style benchmarks (which require isolated containers), cloud sandboxes would let users run agents without setting up Docker locally — and would unlock the "agent runs in cleanroom, no network" path that some benchmarks require.Scope
Three providers, each as an
Environmentsubclass:chimera/env/modal_env.py.chimera/env/daytona.py.Each:
EnvironmentABC (provision,exec,read,write,teardown)pip install "chimera-run[modal]", etc.docs/env/{modal,daytona}.mdAcceptance criteria
chimera ferret --env modaletc.Estimated effort
Large — ~2 weeks. Could be broken into 3 issues (one per provider).
Why this matters
Cloud sandboxes are how teams scale agent benchmarks (each instance gets a fresh container). chimera ferret without cloud sandboxes is local-only; with them, it's competitive with the existing
open-swe-agentworkflow.