benchflow 0.6.5

What's Changed

• Accept provider total token usage in review gate by @bingran-you in #840
• Add Prime-RL SFT train wrapper by @bingran-you in #842
• fix(integration): unblock expanded review gate by @bingran-you in #841
• Resolve Prime-RL SFT configs from checkout by @bingran-you in #843
• fix(train): support convert from existing Prime-SFT JSONL by @bingran-you in #845
• Add reproducible eval and SFT artifact CLI hooks by @bingran-you in #844
• Harden Prime SFT tool-call validation by @bingran-you in #848
• Redact trajectory secrets before serialization (fix prime-sft convert) by @bingran-you in #849
• Repair Prime SFT conversion from results JSONL by @bingran-you in #850
• Fix OpenRouter and Prime SFT results artifacts by @bingran-you in #851
• Strip OpenHands chat input field before LiteLLM upstream calls by @bingran-you in #852
• Add Prime-RL SFT exposure controls by @bingran-you in #853
• Package local Prime SFT JSONL for Prime-RL by @bingran-you in #854
• Guard Prime-RL SFT reproduction semantics by @bingran-you in #855
• Add Prime-RL custom trainer compatibility controls by @bingran-you in #856
• Add Mobile300 Prime-RL reproduction profile by @bingran-you in #857
• Fix Mobile300 Prime-RL chat-template profile by @bingran-you in #859
• Add Mobile300 Prime-RL tail-window staging by @bingran-you in #860
• Avoid stale Prime-RL local dataset cache by @bingran-you in #861
• Match Mobile300 custom trainer microstep semantics by @bingran-you in #862
• Sync Mobile300 Prime-RL checkpoint step by @bingran-you in #863
• Add Mobile300 Prime-RL token-suffix staging by @bingran-you in #864
• Add Prime-RL sample-mean SFT loss shim by @bingran-you in #865
• Fix Prime-RL custom trainer SFT staging by @bingran-you in #866
• Stub flash-attn imports for Prime-RL SDPA SFT by @bingran-you in #867
• Preserve row-wise Prime-RL SFT exposure by @bingran-you in #869
• Preserve private Prime-SFT tool argument tokens by @bingran-you in #870
• feat(agents): autoload agent plugin packages via benchflow.agents entry points by @Yiminnn in #873
• fix: litellm-proxy health + ACP shims survive parallel/deepseek runs by @Yiminnn in #871
• feat(agents): namespace shorthand + manifest auto-load (#876) by @Yiminnn in #877
• Add native source adapters for MCP benchmarks by @bingran-you in #878
• fix(acp): derive model-via-env from registration data, not agent names by @Yiminnn in #879
• fix(acp): avoid empty agent log placeholders by @bingran-you in #832
• fix(diagnostics): type diagnostic and usage sources by @bingran-you in #858
• fix(litellm): bridge /v1/responses→chat for chat-only OpenAI upstreams by @Yiminnn in #868
• fix(diagnostics): classify core error markers case-insensitively by @bingran-you in #880
• fix(litellm): honor Gemini provider base URL by @bingran-you in #881
• Materialize Toolathlon credential-backed tasks by @bingran-you in #885
• Registry-driven Claude subscription gate + zero-signal heuristic exemption (Anthropic OAuth unlock) by @Yiminnn in #886
• fix(harvey): map provider env to the OpenAI-compatible adapter by @Yiminnn in #888
• Provision Toolathlon service runtime sidecars by @bingran-you in #889
• Fix Toolathlon k8s Daytona runtime by @bingran-you in #890
• Redact sharded worker payload artifacts by @bingran-you in #891
• Stabilize Toolathlon Daytona runtimes by @bingran-you in #892
• Fix Toolathlon Daytona direct runtime by @bingran-you in #893
• Fix Toolathlon Daytona blocker retries by @bingran-you in #894
• Document Python 3.12 CLI install requirement by @bingran-you in #899
• Use noncanonical PTY mode for Daytona ACP by @bingran-you in #896
• Add reusable task runtime primitive by @bingran-you in #902
• Add optional TRL GRPO integration by @bingran-you in #903
• Add HF task snapshot provenance support by @bingran-you in #904
• Add paired eval lift reporter by @bingran-you in #901
• Document BenchFlow GRPO pipeline by @bingran-you in #905
• Skip TRL schema test without transformers by @bingran-you in #906
• Complete TRL rollout artifacts and selective task snapshots by @bingran-you in #907
• Remove BENCHFLOW_SKILL_NUDGE prompt injection by @bingran-you in #908
• Use native data-agent datasets in GRPO runbook by @bingran-you in #909
• fix(acp): restore Gemini on Daytona by @bingran-you in #912
• fix(sandbox): prevent Daytona heartbeat sleep leak by @bingran-you in #910
• Fix OpenHands Azure reasoning effort propagation by @bingran-you in #911
• chore: release v0.6.5 by @bingran-you in #913

Full Changelog: v0.6.4...v0.6.5