Issues

• JobBench keeps a rollout killed at its time limit as a finished run, with no trace in its trajectory and no exit status in its log

  google-research/harness2#121 · Filter by author leo922oel opened on Oct 21, 2026

• Add a Hermes substrate

  google-research/harness2#117 · Filter by author yqm-307 opened on Oct 21, 2026

• JobBench OpenCode rollouts share npm/pip installs, temp files and other rollouts' workspaces

  google-research/harness2#111 · Filter by author leo922oel opened on Oct 20, 2026

• An improver outage is saved as ok, and resume never re-runs the task

  google-research/harness2#109 · Filter by author raiyan22 opened on Oct 20, 2026
  1

• Publish version or tag for conda recipe

  google-research/harness2#106 · Filter by author millsks opened on Oct 20, 2026

• How to use with Antigravity

  google-research/harness2#100 · Filter by author Paxa opened on Oct 20, 2026

• harness2-report drops a mode's improver spend when the mode fails after its stages ran

  google-research/harness2#98 · Filter by author leo922oel opened on Oct 20, 2026

• Proposal: a documented example of running a static check on an evolved harness

  google-research/harness2#96 · Filter by author sairenchristianbuerano opened on Oct 20, 2026
  1

• Judge cannot route to OpenAI-compatible endpoints: provider whitelist plus a Responses-API-only OpenAI path

  google-research/harness2#95 · Filter by author yangbaiwan392-collab opened on Oct 20, 2026
  2

• Can a deployer pin a harness component so it can't be removed?

  google-research/harness2#94 · Filter by author sequimelectronics opened on Oct 20, 2026
  5

• Exploratory Harvey LAB evaluation with structured evidence tooling (c5/c7) & deliverable synthesis alongside harness composition

  google-research/harness2#91 · Filter by author mathmentorAI opened on Oct 19, 2026
  1

• judge_one trusts all_pass/n_passed from the payload: a passing task reads as a failure, and a partial payload raises TypeError in harness2-report

  google-research/harness2#80 · Filter by author Faisal-Fayaz opened on Oct 19, 2026

• LAB runner silently fails when the timeout command is unavailable

  google-research/harness2#77 · Filter by author wangyu-geek opened on Oct 19, 2026

• Harness^2 analyzed by Claude and compared with similar systems

  google-research/harness2#71 · Filter by author neoneye opened on Oct 19, 2026

• Resume and rollout caches silently reuse outputs after task inputs change

  google-research/harness2#65 · Filter by author codebreaker77 opened on Oct 18, 2026

• Add substrates for Pi and Claude code.

  google-research/harness2#64 · Filter by author Mayorc1978 opened on Oct 18, 2026
  8

• Explain when to use Harness2 before benchmark setup

  google-research/harness2#61 · Filter by author aronchick opened on Oct 18, 2026

• Document the local ledger example at its entry point

  google-research/harness2#59 · Filter by author aronchick opened on Oct 18, 2026

• Add an outcome-based getting-started guide

  google-research/harness2#57 · Filter by author aronchick opened on Oct 18, 2026
  1

• Add an inspectable, credential-free ledger example

  google-research/harness2#55 · Filter by author aronchick opened on Oct 18, 2026
  2

• harness2-report undercounts improver spend for tasks re-run after an interrupted invocation

  google-research/harness2#50 · Filter by author leo922oel opened on Oct 18, 2026
  8

• harness2-judge keeps an old grade after a task is re-run if the re-judge fails

  google-research/harness2#44 · Filter by author raiyan22 opened on Oct 18, 2026

• Related-work citation request: Project Peritus’s harness evolution

  google-research/harness2#39 · Filter by author dollspace-gay opened on Oct 18, 2026

• harness2-judge and harness2-report print tracebacks for some invalid arguments and missing settings

  google-research/harness2#38 · Filter by author Yasas-Sri opened on Oct 17, 2026
  1

• Proposal: --bench local, run Harness² on your own folder of tasks

  google-research/harness2#5 · Filter by author kvishalv opened on Oct 16, 2026
  2