Repository navigation
Decouple SWE environment infrastructure from agent harnesses #1249
Description
The SWE environment infrastructure (Apptainer containers, tool execution, eval harnesses) is currently coupled to specific agent harnesses in swe_agents. This means these environments cannot easily be used with other agents and requires config duplication.
As part of the broader sandbox infrastructure work, we should refactor swe_agents to separate the environment initialization and verification from the agent loop, so any agent can use SWE-style sandboxed environments.
Metadata
Metadata
Assignees
Type
Projects
- Status
Activity
feat(swe_env): decouple SWE environment infra from agent harnesses (N…
feat(swe_env): all 6 family harnesses + relocated parser + lifecycle/…
feat(swe_env): model-endpoint egress, reaper wiring, verifier config …
test(swe_env): end-to-end verify() HTTP wire-ownership contract (NVID…
feat(swe_agents): additive SELF_DRIVING swe_env adapter (NVIDIA-NeMo#…
feat(mini_swe_agent_2): opt-in verifier-POST eval (cross-agent swe_en…
feat(swe_agents): delete legacy two-container eval path; verifier pat…
swe_env: make get_harness self-registering + HF-name tolerant (NVIDIA…
swe_env: strip apptainer; NVIDIA-NeMo#1694 owns it — docker-only prov…
anyswe: wire apptainer backend (provider + local-sif + writable /test…
anyswe: trim PR scope — drop openclaw + orphaned scripts/fixtures, fi…
Converge #1677 swe_env decoupling into anyswe + verify on SWE-bench V…