[rollout] feat: add best-effort file artifacts for workflows/graders

[JoyboyBrian]

What

• Add ctx.artifacts_dir to AgentWorkflowContext and GraderContext so user code can write rollout artifacts (logs, traces, generated files) to a known directory.
• Add osmosis_ai/rollout/utils/file_artifacts.py with HARBOR_ARTIFACTS_DIR (sandbox /logs/artifacts/) and default_artifact_root() (~/.osmosis/artifacts/).
• Wire artifact collection through HarborBackend (agent runner + grader runner) and LocalBackend, landing files on the host under ~/.osmosis/artifacts/<rollout_id>/artifacts/.
• Update the rollout scaffold template and docs/rollout-sdk.md with an Artifacts section and usage example.
• Add unit tests covering context wiring, Harbor backend collection, and local backend collection.

Why

Rollouts often produce files (traces, screenshots, large/binary outputs) that don't belong in the sample payload. This gives workflows and graders a single convention — write under ctx.artifacts_dir — and the backends collect the files to the host after each rollout. Artifact handling is best-effort and never affects rewards or rollout status.

How to Test

• uv run pytest tests/unit/rollout/test_context.py tests/unit/rollout/test_harbor_backend.py tests/unit/rollout/test_local_backend.py
• uv run ruff check . and uv run ruff format --check .
• uv run pyright osmosis_ai/

Checklist

• PR title follows [module] type: description format
• Appropriate labels added (e.g. enhancement, bug, breaking)
• ruff check . and ruff format --check . pass
• pyright osmosis_ai/ passes
• pytest passes (new tests added if applicable)
• Public API changes are documented
• No secrets or credentials included

---

Summary by cubic

Add best-effort file artifact support for rollouts. Workflows and graders can write to ctx.artifacts_dir, and Harbor/Local backends collect files to ~/.osmosis/artifacts/<rollout_id>/artifacts/ without affecting rollout status or rewards.

• New Features
  • Added artifacts_dir to AgentWorkflowContext and GraderContext.
  • Harbor: runners set artifacts_dir to /logs/artifacts; backend moves artifacts to host before cleanup, copies when trials are kept, and keeps the trial dir if relocation fails (logs, doesn’t fail rollout).
  • Local: creates per-rollout dir at ~/.osmosis/artifacts/<id>/artifacts/logs/artifacts; passes it to contexts and degrades to None if creation fails.
  • New osmosis_ai/rollout/utils/file_artifacts.py with HARBOR_ARTIFACTS_DIR and default_artifact_root().
  • Updated scaffold and docs/rollout-sdk.md with an Artifacts section and example.
  • Added unit tests for context wiring, Harbor relocation/failure behavior, and Local creation/degradation.

^{Written for commit f4d564f. Summary will update on new commits.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[cubic-dev-ai]

4 issues found across 11 files

Confidence score: 2/5

• osmosis_ai/rollout/backend/local/backend.py and osmosis_ai/rollout/backend/harbor/backend.py build artifact paths from rollout_id without sanitization, so traversal/absolute segments could write, copy, or delete files outside the intended artifacts root; merging as-is carries a concrete host-path safety risk — validate/normalize rollout_id and enforce containment within the artifact root before any filesystem operation.
• osmosis_ai/templates/_scaffolds/rollout/main.py.tpl assumes artifacts_dir is always present in the grader example, but GraderContext.artifacts_dir can be None; copied scaffold code may raise runtime errors in environments where artifact directory creation fails — add a None guard in the template example before using the path.
• docs/rollout-sdk.md shows grade(...) -> None while the ABC uses -> Any, which can mislead implementers and static typing expectations; this is low risk but increases integration friction — align the doc signature to -> Any before or shortly after merge.

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[cubic-dev-ai]

No issues found across 11 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Re-trigger cubic}