| name | karena-deploy |
| description | Generate ready-to-paste kArena shell task factory snippets for live-kBench operations. Use when Codex is asked to start, deploy, run, poll, bootstrap, sync, or manage kArena workflows in karena-shell, including crawl_bugs, submit_kcache, poll_kcache, build_kenv_images, crash evals, LLM judge evals, localization, developer patch population, and HuggingFace import/export tasks. |
kArena Deploy
Workflow
Generate code for the karena-shell ptpython REPL. The shell already provides config, db, Path, and the task factories; do not output imports or setup boilerplate.
Read references/task-factories.md when a request mentions a specific operation, pipeline, or missing parameter.
Output Rules
- Output only the minimal ready-to-paste Python snippet unless the user asks for explanation.
- Assign tasks to descriptive variables when starting more than one operation.
- Use
await t only for one-shot operations when the user asks to wait for completion.
- For polling loops, return task assignments and mention cancellation only if useful:
t.cancel().
- Ask only for required missing values that cannot be inferred, especially
judge_id for LLM judge and agent_config_id when the user wants one specific agent configuration.
- Default
status to "fixed" unless the user says open bugs.
- Default
interval to the factory default unless the user gives a frequency.
- Default
workspace_root to Path("workspace").
Common Snippets
Full evaluation after patches exist:
t_crash_issue = issue_crash_evals(config, db, agent_config_id=74)
t_crash_poll = poll_crash_evals(config, db)
t_judge = issue_llm_judge(db, judge_id=6, workspace_root=Path("workspace"))
t_loc = issue_localization(db)
Bootstrap from Syzbot:
t_crawl = crawl_bugs(config, db)
t_submit = submit_kcache(config, db)
t_kcache = poll_kcache(config, db)
t_build = build_kenv_images(config, db)
t_devpatch = populate_developer_patches(config, db)
Bootstrap from HuggingFace:
t_import = import_from_hf(config, db, "org/live-kbench-bugs")
