SWE-agent cannot be run by the standard as shipped: it needs root in the kits and the internet at start #37

Priority

extreme

Components

• app/nullius (backend)
• app/web (routes)
• templates and static files
• docs and the site
• catalogue (benchmarks, agents, capabilities)
• the standard (ports, kits, relay, runs)
• CI, workflows and skills

Today

SWE-bench Verified's page says "SWE-agent, the authors' own agent, runs only on Harbor." On the standard its reference agent is mini-swe-agent. SWE-agent itself was looked at for import at its latest release, v1.1.0 (0f3acaf), with the import procedure. It has no ACP mode; it has a one-task command that works in a container that exists already (sweagent run --env.deployment.type=local --env.repo.type=preexisting). As shipped it cannot be run by the standard, for two reasons in its own code, each tried:

• It copies its tool bundles to /root/tools (sweagent/tools/tools.py:251, the path is written in). The conformance kits run an agent in the base image as the user of the machine, without root: mkdir: cannot create directory '/root': Permission denied. An agent that cannot pass its kits has no proven port, and no benchmark is matched to it.
• Its default tools install a package when the agent starts (tools/edit_anthropic/install.sh: pip install 'tree-sitter==0.21.3', pip install 'tree-sitter-languages'), and a failed install stops it (check="raise"). In a SWE-bench instance's container the agent has no way out but its model: ERROR: No matching distribution found for tree-sitter==0.21.3.

Its dependencies are also unpinned (pyproject.toml lists names only, no lock file), so an install of the same tag can differ from day to day.

Wanted

SWE-agent offered on the standard as its vendor ships it, and SWE-bench Verified's reference agent when it runs in the instance's own container. Nothing of the agent is patched to get there; what it needs is of the standard, and is the owner's to decide:

• An agent's own setup step in the unit's container before it starts, with the way out a benchmark's preparation has, gone again before the first model call. This is what Harbor's agent setup phase does.
• A way for an agent to pass its kits that does not depend on writing under /root as the machine's user: the kits run in a container where it is root, or the agent is proven on an own-image fixture.

Acceptance

• The owner has said whether an agent may have a setup step with a way out, and how an agent that needs root passes its kits.
• SWE-agent is in the agents catalogue at a pinned release, its dependencies pinned.
• It passes the kits of the ports it offers, and its model calls are in the relay's record.
• It runs a SWE-bench Verified instance in the instance's own container.
• The benchmark's limitation no longer says it runs only on Harbor.

Dependencies and related

The preparation step and Box.offline (Terminal-Bench's tmux) are the pieces an agent's setup step would reuse.