[Woodie-W]

Summary

Sprout currently selects mini-swe-agent as its model agent. This adds an optional opencollab-base slot that runs OpenCollab's native Single2 session through the existing worker and ToolAccess, including its multi-tool messages, session state and termination semantics.

• Pin the public OpenCollab SDK revision behind an optional installation extra; preserve the default mini contract and SDK-independent service imports.
• Index closed native turns against actual disk checkpoints, restore SDK state and workspace with the matching history prefix, and route continuations through the existing Run Store and SPROUT search/review jobs.
• Validate reviewer actions against the selected actor's recorded tools. Export actual messages, schemas and provenance so restored prefixes, reviewer actions and SDK-generated submission receipts are excluded from assistant loss.
• Add an explicit review50-v1 compatibility protocol for historical tool names, including native submission and the /workspace alias. It does not distribute or require the review50 dataset.
• Add installation/evaluation/RL documentation and a separate optional-SDK CI job.

The trainer, advantages, loss, optimizer and serving-weight synchronization remain owned by Miles. No replacement trainer, reward shaping, task data, model weights, credentials or host configuration is included.

Validation

Validated at 30b1ad6ce55a:

• Non-slow suite: 4064 passed, 436 skipped, 77 deselected; optional SDK suite: 44 passed. Skips remain explicit infrastructure conditions.
• Isolated PostgreSQL Run Store, branch API and fidelity tests: 576 passed; test database removed afterward.
• All eight upstream gates, including Ruff, strict mypy, package layout checks, runtime tool checks and Go analysis.
• All 44 upstream mini/Harbor interface contracts.
• Wheel build and fresh installation of the wheel plus the publicly pinned SDK, with packaged tool resources checked outside the checkout.
• A real SWE-Smith task in Podman: parent execution, exact snapshot/session restoration, child execution, original official grading, and native message/provenance export. The model endpoint is scripted to verify transport. Both official rewards were zero; this is not policy-learning or GPU-training evidence.

Scope

This targets the current worker/Run Store architecture. It preserves its search stopping and variable-size group contract, so RL needs a compatible upstream Miles client. Legacy Harbor task bundles must have their declared environment setup prepared in their task images. Single2 is supported; OpenCollab team/workflow orchestration is outside this PR.