Add a capability

A capability exposes a named executable runtime through the recorded experiment interface. It does not grant extra host authority; native research agents may already have their own shell tools outside this numerical execution boundary.

1. Add a versioned capability manifest under capabilities/.

2. Define the immutable runtime path, probe command, resource environment, and compatible skill identity.

3. Add or update the corresponding skill under skills/.

4. Make the health probe deterministic and inexpensive.

5. Test workbench execution, evidence-stage source sealing, commissioning, and scientific-command binding.

6. Regenerate public schemas if an action model changed.

When the configured executable is a launcher or interpreter that invokes sibling scientific binaries, declare those binaries and build records through identity_files. Simjecture hashes the declared regular files at discovery and checks them again before every execution. Paths must remain inside the runtime; missing files, directories, symlinks, traversal, and duplicates are rejected.

Capability availability, commissioning, and scientific validity are different states. A healthy runtime may execute workbench code. It may produce scientific evidence only after the applicable workflow gates are satisfied: classic campaigns use instrument claims and prospective contracts, while capability-backed minimal studies use source-bound methods review and the relevant claim/repair commitments. See methods review.

Do not place a problem-specific experiment plan in a general capability skill. Skills may capture stable interface facts and numerical hazards; the agent owns the active scientific design.