Loading ToolBench
A static analysis suite and evaluation engine for auditing agent skills and prompts against a 390-criteria scientific matrix. Provides skill auditing, optimization, and GitHub contribution workflows.
SkillGauge has 12 tools with reasonable naming (verb-first: audit*, optimize*, run*, parse*, get*, fork*, commit*, create*, submit*) but critical gaps in schema documentation and parameter descriptions. Most tools lack visible input schema definitions in the source code provided. Descriptions are present but generic (50-150 chars), missing WHEN to use guidance and dependency hints. No output schemas documented. Error handling is absent, no recovery guidance or categorization. Security concern: GitHub token exposed as required parameter (tools 8-12) violates secret-injection pattern. Tools 1-2 are near-duplicates (auditSkill vs auditSkillAsync) suggesting composition issues.
Audits a package of multiple skill files, computing aggregate package score, integrity score (reference validation), redundancy penalty, and individual skill reports.
Audits a single skill markdown file content against a 390-criteria scientific matrix, returning an AuditReport with overall score, tier classification, and dimension-based metrics.
Asynchronous version of auditSkill for use in async contexts.
Commits a file to a GitHub repository branch. Creates or updates the file with base64-encoded content.
Creates a pull request from a source branch to a target branch in a GitHub repository.
Forks a GitHub repository under the authenticated user's account.
Fetches the authenticated GitHub user details using a GitHub API token.
GitHub API token exposed as required parameter in 5 tools (getGitHubUser, forkRepository, commitFile, createPullRequest, submitContribution). Violates secret-injection pattern, credentials must use server-side injection, not tool parameters. Tokens will leak into agent logs and prompt history.
Input schemas not visible in source code for any tool. Cannot verify parameter types, constraints, or enums. Schema documentation is missing, LLMs cannot infer valid parameter ranges or formats. Rubric requires explicit JSON Schema with types and descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 42 | 2026-07-28+ | v2 |
Distills and compresses a skill markdown content using Gemini API or heuristics fallback. Maximizes SkillGauge metrics by reducing verbose prose, eliminating weak words, and optimizing example ratios.
Clean up text using heuristics (fallback mode if no LLM key is provided). Removes filler phrases, replaces weak words with strong directives, and cleans up formatting.
Parses a markdown skill file and extracts YAML frontmatter metadata, headings, list depths, code blocks, and raw text content.
Runs dynamic LLM tests on a skill file content using Gemini API. Executes test scenarios defined in skill metadata and validates outputs against expected/unexpected keywords.
Automates the entire GitHub contribution flow: forks a repository, commits a file to the user's fork, and creates a pull request to the target repository.
auditSkill and auditSkillAsync are near-duplicates with identical input (content string) and purpose. Violates composition pattern, each tool should do exactly one thing. LLMs will waste reasoning cycles deciding between them. Merge into single async-capable tool or remove sync variant.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. E.g., auditPackage returns 'AuditReport' but structure is undocumented, agent cannot know if it includes per-file scores, integrity metrics, or redundancy penalties.
No error handling or recovery guidance. Tools lack categorization of errors as retryable, user-fixable, or fatal. E.g., commitFile could fail due to branch not found, permission denied, or network timeout, agent has no guidance on what to do next.