Reef Files

Highlights

• Split training data preparation into record-to-record, record-to-task, and task-to-batch stages (#766).
• Add OPD recipes, generated-task training, and AutoEnvScaling examples (#683, [#707], [#755], [#790], [#791]).
• Let recipes choose the training method for each job (#695).
• Support Slime weight publication to vLLM through the disk transport (#710).
• Share sandbox execution through Harbor environments (#803).
• Fix interrupted training recovery, engine startup port conflicts, relative artifact work directories, and OpenRouter web-search streams (#728, [#730], [#723], [#712], [#769]).

Upgrade notes

This release changes configuration and extension interfaces.

• Use the public configuration sections. Remove schema-version; legacy layouts, custom service definitions, and execution.services are no longer accepted (#782).
• Update composite recipe components to use the implementation, config, and runtime recipe shape (#784).
• Update custom training stages to the three-stage pipeline. Configured plugins use a class with from_config (#754, [#766]).
• Update evaluation imports to reef.evaluation and shared contracts to reef.core.evaluation (#785).
• The legacy reef.service.slime_driver entry point and unused runtime APIs have been removed (#804).

See the configuration reference and training pipeline guide before upgrading custom deployments or recipes.

What's Changed

• fix(service): route requests to a checkpointed release's adapter by @MikeZheng777 in https://github.com/Human-Agent-Society/reef/pull/709
• fix(artifact): publish releases with relative work directories by @YihongT in https://github.com/Human-Agent-Society/reef/pull/712
• fix: reject proposals exceeding the requires limit by @jane-jhu in https://github.com/Human-Agent-Society/reef/pull/713
• Say when a grounding document was cut short by @Kaustubh1235 in https://github.com/Human-Agent-Society/reef/pull/714
• chore(deps-dev): bump the python-minor-and-patch group with 2 updates by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/716
• chore(deps): bump the docs-site-minor-and-patch group in /docs/site with 6 updates by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/717
• chore(deps): bump virtualenv from 21.7.12 to 21.7.13 by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/719
• fix(slime): honor training.timeout-s for worker RPCs by @GGQ1996 in https://github.com/Human-Agent-Society/reef/pull/715
• Publish four Guidance-TTT paper solutions and evaluation commands by @Chonghe-Jiang in https://github.com/Human-Agent-Society/reef/pull/720
• Add Yihong Tang to the contributors list by @quao627 in https://github.com/Human-Agent-Society/reef/pull/721
• docs: correct stale behavior claims in active documentation (#603) by @Yi-111-a in https://github.com/Human-Agent-Society/reef/pull/699
• docs: add ASD-STE100-inspired agent writing guidance by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/724
• feat(train): put the trainer behind the InferenceEngine contract and let Slime publish to vLLM over the disk transport by @ruizhang0101 in https://github.com/Human-Agent-Society/reef/pull/710
• refactor(harness): split adapter quirks into AdapterRenderer steps by @yanxz in https://github.com/Human-Agent-Society/reef/pull/725
• fix(inference): retry SGLang and vLLM engine startup on port conflicts by @augustuszzq in https://github.com/Human-Agent-Society/reef/pull/723
• fix(train): resume a run that never saved its optimizer, and keep its lr schedule by @MikeZheng777 in https://github.com/Human-Agent-Society/reef/pull/728
• fix(sglang): start encoder-only engines through encode_server.launch_server by @augustuszzq in https://github.com/Human-Agent-Society/reef/pull/731
• docs(operate): say what x-reef-release-id binds instead of promising a pinned version by @augustuszzq in https://github.com/Human-Agent-Society/reef/pull/732
• fix(slime): list what an interrupted RUNNING job left on disk and document its recovery by @augustuszzq in https://github.com/Human-Agent-Society/reef/pull/730
• feat(record2dataset): share and forward task generation inputs by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/726
• docs: simplify Reefine tutorial around interactive use by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/734
• fix(docker): install Reef's core dependencies in the image by @augustuszzq in https://github.com/Human-Agent-Society/reef/pull/738
• fix(sglang): keep encoder-only engines out of engine control requests by @augustuszzq in https://github.com/Human-Agent-Society/reef/pull/743
• fix(slime): apply the torch.distributed patch before Megatron binds collectives by @ruizhang0101 in https://github.com/Human-Agent-Society/reef/pull/736
• docs(docker): drop the fixed 1e-5 logprob gap expectation by @Yi-111-a in https://github.com/Human-Agent-Society/reef/pull/745
• docs(troubleshooting): correct the ambiguous training job recovery entry by @augustuszzq in https://github.com/Human-Agent-Society/reef/pull/744
• fix(sglang): retry router startup on port conflicts by @augustuszzq in https://github.com/Human-Agent-Society/reef/pull/751
• refactor(plugins): name every configured plugin by class and build it with from_config by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/754
• feat(task-generation): select the task generator in the deployment and train with a built-in recipe by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/755
• chore(deps): bump the github-actions group with 3 updates by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/756
• chore(deps): bump source-map-js from 1.2.1 to 1.2.2 in /docs/site by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/760
• chore(deps-dev): bump tinker from 0.31.0 to 0.33.1 in the python-minor-and-patch group by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/757
• chore(deps): bump the docs-site-minor-and-patch group across 1 directory with 5 updates by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/758
• chore(deps): bump sharp from 0.35.4 to 0.35.5 in /docs/site by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/761
• chore(deps): bump werkzeug from 3.1.8 to 3.1.9 by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/762
• chore(deps): bump multidict from 6.7.1 to 6.9.1 by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/763
• chore(deps): bump mako from 1.4.1 to 1.4.2 by @dependabot[bot] in https://github.com/Human-Agent-Society/reef/pull/764
• refactor(train): share the processor buffer and honor max_staleness for computed-feedback candidates by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/729
• feat(harness): declare locked config keys in adapter descriptors by @akushonkamen in https://github.com/Human-Agent-Society/reef/pull/746
• fix(train): restore the committed marker when a killed weights job saved nothing by @Cross2pro in https://github.com/Human-Agent-Society/reef/pull/750
• feat: add OPD recipe by @jayzou3773 in https://github.com/Human-Agent-Society/reef/pull/683
• feat(opd): score the student on a frozen teacher engine by @jayzou3773 in https://github.com/Human-Agent-Society/reef/pull/707
• refactor: split experience reuse into record-to-record, record-to-task, and task-to-batch stages driven by a training pipeline by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/766
• refactor(record2record): narrow the record stage to forming sets and ending records by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/780
• fix(record2dataset): read the primary verifier reward and correct two Designer prompt lines by @Benjamin-eecs in https://github.com/Human-Agent-Society/reef/pull/694
• refactor: unify commit completion and publication decisions by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/779
• fix(record2record): poll only while judgments run, and rename derivation_pending to background_jobs_running by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/781
• refactor(config)!: make the public layout the only config layout by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/782
• refactor(recipe)!: configure composite components with the recipe section shape by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/784
• refactor: centralize dispatcher worker and reload coordination by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/783
• refactor(evaluation)!: make candidate evaluation a top-level package and move harness episode evaluation into it by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/785
• refactor(harness): share skill frontmatter helpers across adapters by @yanxz in https://github.com/Human-Agent-Society/reef/pull/772
• refactor(tests): organize suites by responsibility and enforce placement rules by @hanfeiyu in https://github.com/Human-Agent-Society/reef/pull/793
• ci: fail a pull request whose base branch is not main by @yanxz in https://github.com/Human-Agent-Society/reef/pull/773
• fix(record2dataset): score agent episodes reliably and let the Designer use the Responses API by @simonucl in https://github.com/Human-Agent-Society/reef/pull/789
• feat(reefine): let a request retire a held entry by id by @akushonkamen in https://github.com/Human-Agent-Society/reef/pull/778
• docs: add Reef citation metadata by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/794
• fix(record2dataset): train the generated-task recipe on multi-turn agent episodes by @simonucl in https://github.com/Human-Agent-Society/reef/pull/790
• feat(record2dataset): add the Coding-Agent based (AutoEnvScaling) record2dataset session generator and its task checks by @simonucl in https://github.com/Human-Agent-Society/reef/pull/791
• docs: remove the OPD row from the recipe catalog by @MikeZheng777 in https://github.com/Human-Agent-Society/reef/pull/796
• refactor(evaluation): simplify selection decisions by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/795
• feat(train): let recipes choose the training method for each job by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/695
• fix(service): translate OpenRouter web search items for Codex streams by @akushonkamen in https://github.com/Human-Agent-Society/reef/pull/769
• refactor!: remove logic left redundant by the recent unifications by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/804
• refactor(sandbox): share execution through Harbor environments by @BobbyZhouZijian in https://github.com/Human-Agent-Society/reef/pull/803

New Contributors

• @GGQ1996 made their first contribution in https://github.com/Human-Agent-Society/reef/pull/715
• @Yi-111-a made their first contribution in https://github.com/Human-Agent-Society/reef/pull/699
• @augustuszzq made their first contribution in https://github.com/Human-Agent-Society/reef/pull/723
• @akushonkamen made their first contribution in https://github.com/Human-Agent-Society/reef/pull/746
• @Cross2pro made their first contribution in https://github.com/Human-Agent-Society/reef/pull/750
• @hanfeiyu made their first contribution in https://github.com/Human-Agent-Society/reef/pull/793

Full Changelog: https://github.com/Human-Agent-Society/reef/compare/v0.1.2...v0.1.3