[BobbyZhouZijian]

Motivation

Versioned deployments still expose HTTP process commands and training bridge wiring in YAML. This implements #414's contributor-directed boundary: users configure components, Reef assembles native inference/training, and recipes consume independently deployed method services through endpoints. The broader RFC remains proposed, so this PR remains a draft.

Changes

• Remove service, services and execution.services from v2 configuration. HTTP settings and canonical flags use reef.*, including --reef.port and --reef.token. CLI and YAML use the same selected-component parser with CLI precedence.
• Assemble provider/local inference or a declared serving runtime automatically. Weight Recipes select a training deployment definition; Slime remains the compatibility default. Its integration owns the driver, Ray bridge and training-owned inference connection.
• Select training.backend through a built-in name, installed reef.training_backends entry point, or dotted TrainingDeployment class. Each definition owns process preparation and HTTP runtime configuration. Backend selection and service assembly share reef/service/deploy/training.py; extension contracts live in reef/train/deployment.py. InProcessTrainingDeployment starts only HTTP and uses its RuntimeFactory to validate options and create a TrainingRuntime. Missing backends fail without falling back to Slime.
• Keep Slime native validation, model binding, Ray roles, command/environment and readiness inside its integration. Move shared reference/argv utilities below the service layer to preserve package dependency direction.
• Organize deployment code by ownership: CLI syntax in cli.py, the ServiceConfig value and service_config_from_mapping() conversion in service_config.py, DeploymentConfig defaults, selected schemas and public layout in deployment_config.py. Keep YAML loading and interpolation in config_utils.py, move process validation to execution.py, and consolidate provider/local inference plus model preparation in inference.py. Remove forwarding-only argument modules and update imports and the codebase guide.
• Weight Recipes use the generic RuntimeTrainingBackend; the old SlimeTrainingBackend import remains an alias. CPU contracts prove an installed in-process backend works without importing Slime/Ray/SGLang/Torch/MLX.
• Keep PRM and user-model services entirely outside Reef. OpenClawRL declares PRM endpoint/tokenizer fields; its example Compose owns auxiliary startup, health checks and GPU assignments. Reef has no Recipe process-registration hook and neither probes nor stops these services. The example allocates GPU 1 to PRM, GPU 2 to the user model, and GPUs 3–7 to Reef/Slime (four actor plus one rollout GPU).
• Migrate all maintained deployment YAML, tutorial/profile imports, packaging smoke configuration, examples, CLI help and documentation. Move repeated Slime CUDA/NCCL environment defaults into its integration.

Compatibility and operational impact

This intentionally revises the initial v2 layout from #417: move service fields to reef, use --reef.* instead of --service.*, and remove explicit process lists. Method-specific services are independently deployed. OpenClawRL uses recipe.config.prm-url and recipe.config.prm-tokenizer-path; it exposes no auxiliary launch settings. In the example, use docker compose down to stop all three containers; shutting down Reef alone leaves the auxiliary services running. External Ray clusters must exclude PRM/user-model GPUs. Unversioned legacy files retain explicit stacks and their existing override/command semantics.

Managed Slime launches ignore ambient SLIME_ARGS_FILE and use training.options. Existing Ray clusters remain externally owned. Slime's default environment sets CUDA_DEVICE_MAX_CONNECTIONS=1 and NCCL_NVLS_ENABLE=0, honoring operator overrides; this preserves the former TTTD/CORAL/OpenClawRL YAML defaults and also applies them to SAO. Migrated examples preserve their separate HTTP/driver readiness deadlines. Experiment metadata now names RuntimeTrainingBackend and the actual runtime class instead of labeling all weight training as Slime. Launch and app-assembly entrypoints in reef.service.deploy remain available. The Python configuration type and converter are now ServiceConfig and service_config_from_mapping(); callers using ServiceSettings or service_settings_from_config() must update those names. Internal imports from deploy.config, settings, components, layout, options, arguments and provider must use their new owning modules; repository callers are updated. No HTTP wire, persistence or numerical algorithm changes.

Verification

• Final full suite: NO_PROXY='*' no_proxy='*' GIT_CONFIG_GLOBAL=/dev/null GIT_CONFIG_SYSTEM=/dev/null .venv/bin/python -m pytest tests -n 4 --dist loadfile --cov=reef --cov-report=term -q --tb=short: 3341 passed, 48 skipped, 88.24% coverage.
• Focused endpoint, lifecycle, OpenClawRL and example configuration suites: 76 passed; training startup suite: 25 passed. These cover CLI > YAML, rejecting auxiliary launch options, CPU-only discovery, and real external HTTP process survival after Reef shutdown/startup failure.
• REEF_TEST_RAY=1 ... pytest tests/reef_service/test_openclawrl_gpu_pool.py -k openclawrl: 8 passed, covering local/external Ray and OpenClawRL's five logical model GPUs. No GPU kernels were run.
• pre-commit run --all-files: passed; after the final test-fixture update, its focused hooks also passed. Mypy with lint CI dependencies: 264 source files passed.
• Docs on Node 22: npm run check:docs, npm run lint, npm run build passed; dependencies were already installed with npm ci earlier in this PR.
• OpenClawRL docker compose ... config --quiet and bash -n .../run.sh: passed. The three-container GPU stack was not started locally.
• Before module reorganization, an installed CPU wheel outside the checkout passed 165 tests, including in-process HTTP startup/shutdown via an installed backend entry point; current package CI repeats this coverage.
• GitHub CI has been triggered for 81cc9c9; its results are separate from the completed local checks above.

This supplies the deployment extension boundary for #325, without merging its MLX numerical implementation. Real MLX/Metal and Slime/Megatron/SGLang training runs were not performed. CPU tests verify configuration, runtime construction and process lifecycle, not numerical or GPU acceptance.

AI assistance

OpenAI Codex implemented the contributor-directed change, added tests, updated documentation and ran verification. The human contributor remains responsible for reviewing and understanding the implementation and results.

Checklist

• The change is focused and contains no unrelated cleanup.
• Tests cover behavior changes, or this pull request does not change behavior.
• Public interface changes include contract tests, or no public interface changes are present.
• Affected user and developer documentation is updated, or no documentation update is required.
• Root README changes update both languages and README.i18n.yaml, or do not affect README content.
• An accepted RFC is linked, or this change does not require an RFC.
• Relevant pre-commit and test checks pass.
• Non-trivial AI assistance is disclosed above, or none was used.

Review and merge

Draft follow-up to #414 and merged #417. RFC acceptance and GPU validation remain outstanding.

[github-actions]

Merge Oncall: @BobbyZhouZijian

The merge oncall will route reviews and keep this pull request moving toward a clear decision.