Gear v0.1.2

Fix DSH startup and bootstrap evidence recovery:

• Keep the dormant DSH row install-safe and avoid initializing model clients during startup (#44).
• Preserve the original bootstrap evaluation reference across repeated incomplete evidence repairs, so recovery can finish without rerunning valid cells (#45).
• Update the package version and bundle assertion to 0.1.2.

The fixes were validated on dev before promotion to main in #50. Full regression testing passed with 1192 tests passing and 29 skipped, alongside platform sandbox, CPU training, and Hitch integration checks.

Published to npm as rsi-gear@0.1.2 with the latest tag.

Gear v0.1.1

Gear 0.1.1 adds reusable benchmark resources and user-defined refinement objectives while preserving existing experiment identities and evidence.

Highlights

• Share benchmark inputs through Hitch resource selections, preflight checks and durable retention. Materialize legacy task subsets with bounded clone/copy policies and explicit storage audit/quarantine cleanup.
• Recover interrupted projection publication and quarantine restoration without losing ownership proof or clearing referenced historical data.
• Define raw metrics and frozen weighted objectives with explicit constraints. Keep original verifier measurements separate from derived objective scores, and recover compatible cached measurements without additional rollouts.
• Batch shared evidence provenance verification and pin the DSH rc.2 carrier's compatible HMR dependency.
• Include updated bilingual guides, trajectory replays and reproducible storage validation evidence. Preserve the existing js-yaml/esbuild security updates on main.

Validation

All six release PR and main CI jobs passed, including the complete regression suite, Linux/macOS sandbox and real DSH runtime checks, CPU training, Hitch integration, package contents and the external search API consumer. The release preserves the 15 new storage recovery regression tests.

Main CI: https://github.com/rsi-gear/gear/actions/runs/35953308825

Trusted npm publication succeeded with signed provenance. The published package and this release use squash merge commit b69233b8f4ba60e360fd6b20d3de58b8dd1161ba; the npm latest tag resolves to 0.1.1.

Compatibility

Node.js 22.19+ within the 22.x line, or Node.js 24+, is required. Generic resource dataset v2 execution requires Hitch 0.2.14. Existing legacy datasets, historical absolute references and sealed scoring behavior remain supported. Model training remains experimental.

Release PR: #42
Implementation PRs: #26, #40, #41

Available on npm: rsi-gear@0.1.1.

npm install --global rsi-gear@0.1.1

Gear v0.1.0

Gear's first release brings evaluation-guided harness refinement to compatible agent environments through the bundled Refine Skill.

• Skill-first control plane with native DSH and harness-neutral adapters, isolated candidate workspaces, and Hitch evaluation integration.
• Durable refinement recovery, champion baseline reuse, verifier and resource-use evidence, and reusable experience memory.
• Extensible search policies and scoped candidate promotion with regression guards.
• Experimental Slime model-training bridge with CPU validation and recovery support.
• Bilingual guides, installation instructions, and reproducible harness and search examples.

Node.js 22.19+ within the 22.x line, or Node.js 24+, is required. Model training remains experimental.

Release PR: #23

Validation: 1,089 regression tests passed (29 skipped in the main test job); dedicated Linux/macOS sandbox, real DSH runtime, CPU training, and Hitch rerun contract jobs passed. Package metadata, type checking, build, package contents, and an external search API consumer were verified.

Available on npm: rsi-gear@0.1.0.

npm install --global rsi-gear@0.1.0