| name | kenv-invoker-writer |
| description | Write or update live-kBench kEnv agent integrations. Use when Codex needs to create a kenv-invoker.py, adapt an agent to the kEnv protocol, add a docker/kenv-*.Dockerfile, debug kEnv container inputs or outputs, or ensure an agent writes patch.txt, traj.json, and log.txt for live-kBench evaluation. |
kEnv Invoker Writer
Workflow
First inspect the repo state before editing:
rg -n "kenv-invoker|KENV_|/KBDr|patch.txt|traj.json|log.txt" .
sed -n '1,260p' docs/kEnv.md
Use references/kenv-protocol.md for the protocol contract and existing examples.
Implementation Rules
- Create or update the agent invoker at
<agent-dir>/kenv-invoker.py.
- Create or update
docker/kenv-<agent-name>.Dockerfile only when the agent image does not already exist.
- Extend
kenv:latest or the repo's current kEnv base image pattern.
- Treat
/KBDr/agent-config as the agent behavior config and /KBDr/model-config as the model/API config.
- Run
/KBDr/initialize before invoking the agent.
- Run the agent against the preexisting Linux checkout at
/linux.
- Build the task text from
/root/report.txt and /root/reproducer.txt.
- Preserve
KENV_STATEFUL_EDIT=true support by letting the agent call /KBDr/run_kernel; do not hide or delete that tool.
- Always write these outputs under
/KBDr/output/: patch.txt, traj.json, and log.txt.
- Derive
patch.txt from the agent's native prediction artifact when available; otherwise use git diff from /linux.
Checks
After editing, run the smallest relevant validation:
python3 -m py_compile <agent-dir>/kenv-invoker.py
docker build -f docker/kenv-<agent-name>.Dockerfile -t kenv-<agent-name>:latest .
For a container smoke test, mount the required kEnv inputs listed in the reference and confirm /KBDr/output/patch.txt, /KBDr/output/traj.json, and /KBDr/output/log.txt exist after the run.
