master
Add OSWorld use-computer runner (#2162)
fix(openhands): support non-apt package managers
feat(auth): replace GoTrue sessions with personal API keys (#2189)
Add --reward-artifact to promote numeric JSON artifacts to reward labels. (#2306)
Fix Modal DinD user routing (#2332)
fix(llms): cap Anthropic cache_control breakpoints at 4 (#2072)
Add separate verifier environments (#1655)
[codex] Upload trial locks and use lock task hash (#2140)
refactor: unify environment hashing on environment_content_hash (#2190)
feat: add anonymous usage telemetry (#2263)
Format the whole repo.
fix(ci): update test_adapter_wizard for new adapter class naming (#1476)
feat(trial): scoped log streaming via Trial.add_log_callback (#1966)
Add harbor leaderboard submit command (#1695)
Harbor Hub: upload, download, and streaming run upload (#1479)
feat(hub): add harbor hub job copy and trial copy commands (#2199)
feat: add harbor hub job delete command (#2204)
Add Hub trial downloads and tighten Hub command UX (#2172)
Add leaderboard update command and docs (#2278)
Document job sharing (#1706)
The new Harbor registry (#1285)
Remove job plugins from config files (#2074)
Strip default values from Harbor Hub job and trial configs on upload. (#2294)
Add sharing.
[codex] Omit default config fields when persisting jobs and trials (#2175)
Simplify trial flow (#1672)
Paginated trial downloading for hosted jobs (#2100)
Refactor Harbor task handling and viewer UI (#2023)
Add EC2 sandbox environment (#2031)
Add compileable task evaluation and auto-verification support (#2150)
perf(registry): enumerate --repo tasks via git ls-tree (#1920)
feat: implement git repo skills (#1909)
fix: dedupe duplicate legacy registry tasks (#1354)
added lock to the show command (#2208)
[Ready For Review] Add multi-step task support for sequential agent evaluation (#1234)
Add --verifier-env
[codex] Add resource enforcement policies (#1697)
feat(langsmith): support network_mode=allowlist via sandbox proxy (#2271)
Handle multi-value metrics.
Add resumable multi-step agent sessions (#2144)
fix(openhands-sdk): use uv to pin Python >=3.12 for SDK venv (#2041)
Make HuggingFace datasets an optional harbor[huggingface] extra. (#2174)
test(pass-at-k): add unit tests for pass@k metric utility (#2046)
Add windows support flag to env.
Use RPC for task version resolution (#1736)
add `harbor run --install-only` flag (#1983)
[codex] add repeatable skill inputs (#1674)
Strip canary markers from instruction text before passing to agents (#971)
\#1850 updated default package download timeout from 20s to 120s (#1883)
tasks/client: skip --filter=blob:none for huggingface.co git URLs (#2328)
[codex] Add extra instruction path support (#1682)
Improve task definition viewer UX and share file browser with trials. (#2305)
feat: log public git repo sources in job telemetry (#2320)
rfc: ATIF v1.7: subagent trajectory, Step extra, and context management event improvements (#1486)
Add sidecar artifact collection and verifier collect hooks (#1775)
Fix cancel event ordering (#2227)
Write trial locks and remove invocation from job locks (#2138)
Classify OpenAI cyber refusals and missing models as distinct errors. (#2299)
feat: id refactoring (#2091)
Network mode support for docker environment (#1799)
Align log entry trial identity (#2137)
feat: support include/exclude patterns for agent and verifier log downloads (#1889)
Minor fixes to analysis & and codex.