Issues

• Selecting overlapping suites in one batch overwrites the same log and output identity

  bug
  TIGER-AI-Lab/ClawBench#479 · Filter by author reacher-z opened on Nov 4, 2026

• After the target request is blocked, the browser stays alive until a polling watchdog notices

  bug
  TIGER-AI-Lab/ClawBench#478 · Filter by author reacher-z opened on Nov 4, 2026

• Agent process exit codes are discarded, so CLI crashes are recorded as ordinary agent exits

  bug
  TIGER-AI-Lab/ClawBench#477 · Filter by author reacher-z opened on Nov 4, 2026

• Stage-2 judges get no run date, so relative-date tasks are judged without a temporal anchor

  bug
  TIGER-AI-Lab/ClawBench#476 · Filter by author reacher-z opened on Nov 4, 2026

• clawbench-batch builds the image before discovering that a resumed batch has nothing to do

  bug
  TIGER-AI-Lab/ClawBench#475 · Filter by author reacher-z opened on Nov 4, 2026

• Harbor and Polar leave setup-only provider credentials in the agent's environment

  bug
  TIGER-AI-Lab/ClawBench#474 · Filter by author reacher-z opened on Nov 4, 2026

• Managed browser sessions leak when the runner is killed after the provider creates them

  bug
  TIGER-AI-Lab/ClawBench#473 · Filter by author reacher-z opened on Nov 4, 2026

• The random-click baseline can finish successfully without clicking anything, and is not reproducible

  bug
  TIGER-AI-Lab/ClawBench#472 · Filter by author reacher-z opened on Nov 4, 2026

• The null and random-click baselines still require a working model, an API preflight, and a mailbox

  bug
  TIGER-AI-Lab/ClawBench#471 · Filter by author reacher-z opened on Nov 4, 2026

• Hermes receives a different instruction from every other harness, including the mailbox password

  bug
  TIGER-AI-Lab/ClawBench#470 · Filter by author reacher-z opened on Nov 4, 2026

• Configured temperature and max_tokens are silently ignored by several harnesses while run-meta records them as applied

  bug
  TIGER-AI-Lab/ClawBench#469 · Filter by author reacher-z opened on Nov 4, 2026

• Playwright MCP is launched through an unpinned npx spec while provenance records the pinned global install

  bug
  TIGER-AI-Lab/ClawBench#468 · Filter by author reacher-z opened on Nov 4, 2026

• The HTTP layer records requests but no response status, so site drift looks like agent failure

  bug
  TIGER-AI-Lab/ClawBench#467 · Filter by author reacher-z opened on Nov 4, 2026

• Body-filter tasks can reach the real server when CDP omits postData

  bug
  TIGER-AI-Lab/ClawBench#466 · Filter by author reacher-z opened on Nov 4, 2026

• The CDP capture handler is one-shot: a slow Chrome or a dropped socket leaves the episode unobserved

  bug
  TIGER-AI-Lab/ClawBench#465 · Filter by author reacher-z opened on Nov 4, 2026

• clawbench-analyze's self-report gap is measured from the wrong text

  bug
  TIGER-AI-Lab/ClawBench#464 · Filter by author reacher-z opened on Nov 4, 2026

• A header-only MP4 counts as a present recording

  bug
  TIGER-AI-Lab/ClawBench#463 · Filter by author reacher-z opened on Nov 4, 2026

• Kernel replay finalization spends one deadline on both processing and download

  bug
  TIGER-AI-Lab/ClawBench#462 · Filter by author reacher-z opened on Nov 4, 2026

• clawbench-batch crashes on Windows before starting any job

  bug
  TIGER-AI-Lab/ClawBench#461 · Filter by author reacher-z opened on Nov 4, 2026

• clawbench-reproduce cannot pin the Hugging Face dataset revision it scores

  bug
  TIGER-AI-Lab/ClawBench#460 · Filter by author reacher-z opened on Nov 4, 2026

• clawbench-rescore rejects a judge configured with api_keys

  bug
  TIGER-AI-Lab/ClawBench#459 · Filter by author reacher-z opened on Nov 4, 2026

• ProRL reward reporting silently drops sessions that produced no reward

  bug
  TIGER-AI-Lab/ClawBench#458 · Filter by author reacher-z opened on Nov 4, 2026

• clawbench-prorl-submit --dry-run prints the judge and PurelyMail secrets

  bug
  TIGER-AI-Lab/ClawBench#457 · Filter by author reacher-z opened on Nov 4, 2026

• The runner accepts eval schemas the interceptor cannot execute, and finds out after startup

  bug
  TIGER-AI-Lab/ClawBench#456 · Filter by author reacher-z opened on Nov 4, 2026

• Workspace discovery can adopt an unrelated ancestor's .env and models.yaml

  bug
  TIGER-AI-Lab/ClawBench#455 · Filter by author reacher-z opened on Nov 4, 2026