Loading ToolBench
DEPRECATED — superseded by the Skill Heaven Claude Code plugin (gaia-research/gaia-skill-heaven). Agent-native discovery and trust interface for the Gaia Skill Tree
Four tools with complete schemas and descriptions. Naming is clear and verb-based (gaia_search, gaia_inspect, summon, gaia_status). Descriptions are detailed (100-300+ chars) and explain WHAT, WHEN, and WHAT-RETURNS. All parameters have types and descriptions. Tool annotations present (readOnlyHint, destructiveHint, idempotentHint, openWorldHint). However, output schemas are not documented in the source, LLMs cannot see what fields to expect from responses. Error handling is delegated to service layer (not visible). Parameter constraints are well-defined (enums, min/max), but no guidance on recovery or next steps if a call fails.
Return an evidence-backed dossier for one generic or Named Skill, including relationships, implementations, trust, sources, and data freshness.
Find generic and Named Skills in the public Gaia Registry. Returns ranked structured results with trust and source freshness metadata.
Report server version, Registry mode, data-contract compatibility, source freshness, counts, and available tools.
Install the best-matching Named Skill from the live Gaia Registry: resolve the current source commit, reuse a bounded commit-addressed payload cache when available, or shallow-clone transiently on a miss; validate the resolved subpath, discard clone scaffolding, then materialize the whole skill directory (SKILL.md plus any reference/, scripts/, and fixtures) into a session-locked temp directory. Recurses into suiteComponents for suite skills. Never writes to your real configuration. Falls through to the next-best candidate on an install failure and reports what was skipped. The structured result includes a printable card and inspect URL, tree-provided trust fields, per-skill cloneSeconds, materializeSeconds, totalSeconds, and cacheState (cold or warm), an honest ranking disclosure, plus the invocation totalSeconds.
Output schemas not documented. LLMs cannot see what fields gaia_search, gaia_inspect, summon, and gaia_status return. This forces agents to guess at response structure and breaks tool chaining when downstream tools need IDs or references.
Error handling not visible in tool definitions. The code delegates to service.search(), service.inspect(), etc., which catch errors and call toolError(). Without seeing error response structure and recovery guidance in the tool description, LLMs cannot plan recovery steps.
summon tool description does not explicitly state it is non-idempotent and has side effects (cloning, materializing to temp directory). The annotations correctly set idempotentHint=false, but the description should warn agents not to retry blindly.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2025-06-18+ | v2 |
gaia_search and gaia_inspect accept 'id' and 'query' parameters that may be opaque internal IDs. No guidance on whether users can pass human-readable names (e.g., skill display names) or must use system IDs. This forces extra lookup calls.
summon's 'limit' parameter (max 5 candidates) is not explained in the description. Why 5? What happens if all 5 fail? Does the tool return partial results or a complete failure? This ambiguity breaks agent planning.