SkillRadar v0.5.0 — Measurable project-aware routing quality

SkillRadar v0.5.0 — Measurable routing quality

SkillRadar v0.5.0 moves routing quality from intuition to an auditable product contract.

Highlights

Matching v2.1: task-first, project-aware

$skill-router still treats the user's explicit task as the primary signal. v0.5.0 can additionally read a small set of non-secret project metadata — dependency names in package.json, common config filenames, and framework directories — to break close ties.

Project context is deliberately bounded:

• maximum project-context bonus: 6 points;
• if a candidate has zero user-task coverage, project context can add at most 2 points;
• source files, environment-variable values, credentials and candidate third-party scripts are not read for project-context routing;
• route output separates task evidence from project-context evidence.

New route evidence includes:

• context.mode
• context.signals
• match_details.matched_signal_weights
• match_details.project_context_signals
• match_details.project_context_bonus

Complementary Top-3 task coverage

The Top-3 reranker now rewards candidates that cover important task facets not already covered by earlier selections. It still considers skill similarity and source diversity, but no longer treats “three different skills” as sufficient when all three cover the same part of the task.

For the golden multi-capability task:

Build a Next.js React shadcn/ui AI dashboard with streaming, tool calling and App Router

v0.5.0 must return all three core capabilities in the Top-3:

• nextjs
• ai-sdk
• shadcn

Golden Router Quality Benchmark

A new benchmark is part of CI and release gating. The initial v0.5 baseline is:

• 7 / 7 golden routing cases passed;
• pass rate: 100%;
• unsafe D/Blocked results in Top-3: 0;
• average Top-1 match score: 85.1;
• project-context fixture: passed;
• explicit task-dominance fixture: passed.

The benchmark covers AI streaming/tool calling, Next.js App Router, shadcn design systems, React performance, frontend polish, a multi-capability AI dashboard, and Chinese multi-facet spatial rendering.

Results are persisted in data/router-benchmark-latest.json on main so ranking changes can be compared over time.

No hidden auto-tuning

v0.5.0 does not introduce opaque self-learning or automatic production weight changes. Ranking weights remain code-reviewed and benchmark-gated. Future optimization work must improve the benchmark without weakening safety or acceptance thresholds.

Existing v0.4 guarantees retained

• bundled offline-first schema-v2 registry;
• core + design + general discovery pool;
• fail-closed security scanning and retained-candidate provenance gating;
• D / Blocked exclusion from automatic routing;
• C-grade safer-alternative advisory;
• read-only Skill Budget Doctor;
• public Codex marketplace installation and upgrade smoke tests.

Install

codex plugin marketplace add changchangidea-oss/SkillRadar --ref v0.5.0
codex plugin add skillradar@skillradar

For an existing pinned SkillRadar marketplace, remove the old plugin and marketplace first, then add the v0.5.0 ref.