Multi-Harness RL Invariants — Cross-Harness Generalization

The Research Finding (FineEnvs / Orchard 2026)

Models trained inside a single agent harness (e.g. OpenHands or Claude Code) experience severe performance drops when deployed to a foreign harness:

• Catastrophic Format Failure: Models hallucinate parameters, emit broken JSON, or fail to parse shell tool outputs when moving to another harness.
• Degraded Resolve Rate: OpenSWE-32B drops 58.8 points (from 62.4% down to 3.6%) when transferring to Kimi-CLI.
• Harness Overfitting: Swapping the harness moves benchmark resolve rates by up to 13 points, while swapping the underlying model moves scores by only 2.5 to 5 points.

The ThumbGate Architecture

ThumbGate acts as the universal Canonical Action Normalizer & Invariant Gateway:

[Claude Code] ──┐
[Antigravity] ──┼─► [Canonical Action Quadruplet] ──► [Cross-Harness Invariant Diode] ──► [Target System]
[Strands]     ──┤    (entity, op, target, params)     (Format, Security, Bounds)
[OpenHands]   ──┘

1. Canonical Action Quadruplets

Regardless of the model's originating harness, every action is normalized into:

• harness: Source harness (claude_code, antigravity, strands, openhands, terminus)
• domain: shell, filesystem, network, generic
• operation: exec, read, write, delete
• target: Command line string, file path, or URL
• payload: Contextual arguments

2. Format & Safety Invariant Gates

• Prevents command injection and forbidden operations (rm -rf /, git push --force origin main, gh pr merge --admin).
• Blocks path traversal (../../../etc/passwd).
• Evaluates before any execution occurs.

3. Transfer Scorecard & RL Rewards

Calculates multi-harness diversity multipliers and invariant pass rates for RL / DPO model training:

node scripts/strands-box-doctor.js --json