[weiddeng]

Summary

Add a minimal Harbor 0.17.0 adapter that installs a checksum-verified Linux S-Code runtime bundle and executes one complete S-Code agent run inside a Harbor task container.

This provides a reproducible way to evaluate S-Code on Harbor-managed tasks, such as TBLite and Terminal-Bench.

Scope

• Add a Harbor BaseInstalledAgent adapter.
• Build and verify a pinned S-Code runtime bundle.
• Use isolated S-Code state for every trial.
• Configure the model endpoint and credential through environment handles.
• Preserve raw events, stderr, final output, and provenance under Harbor agent logs.
• Add focused adapter, bundle-validation, and smoke tests.

The run requires no human input. S-Code uses Full permission mode inside the disposable task container, while Harbor remains the outer isolation and network boundary.

Initial constraints

• Linux x86-64 with glibc.
• One complete agent run per Harbor trial.
• No retry, resume, persistent memory, or ATIF conversion.
• The model endpoint must be reachable from the task container.

Validation

• Adapter and bundle-validation unit tests.
• S-Code bundle self-tests.
• Deterministic Harbor fixture smoke test.

[weiddeng]

Replaced by #129 so the scope can be aligned before implementation. No adapter code was included in this draft.