Loading ToolBench
High-density context genome compressor and Bayesian ROI Darwin memory for AI agents. Provides surgical targeted context synthesis, symbol search, code skeleton extraction, blast radius analysis, error distillation, and continuous evolution harness.
CookieGli exhibits significant definition quality gaps. While all 9 tools have schemas and descriptions, the descriptions are overly technical and jargon-heavy (e.g., '[01_TASK_BOOST]...BM25 symbols + focus skeleton + 2026 reasoning effort'), making them difficult for LLMs to parse for tool selection. Parameter descriptions are sparse or missing context. No output schemas are documented. Error handling guidance is absent. Tool names use prefixes (cookiegli_*) rather than action verbs, reducing clarity. The server targets code analysis, a specialized domain, but descriptions assume deep familiarity with internal architecture (GenomeEngine, Darwin memory, BM25 ranking) rather than explaining user-facing intent.
[04_IMPACT_ANALYSIS] cookiegli_blast_radius: Forward-to-ingress dependency graph & minimal targeted test suite for modified or specified files.
[01_TASK_BOOST] cookiegli_boost: Primary entrypoint for coding/debugging tasks. Synthesizes Layer 2 dynamic context (<600t) with BM25 symbols + focus skeleton + 2026 reasoning effort.
[07_DARWIN_MEMORY] cookiegli_darwin_record: Knowledge persistence with Bayesian ROI (Laplace-smoothed).
[05_ERROR_DISTILLER] cookiegli_distill_lesson: Distill traceback/panics into Darwin rules with Bayesian ROI.
[02_SYMBOL_BTREE] cookiegli_find_symbols: Sub-millisecond exact B-Tree symbol lookup for known identifier names (classes, functions, methods).
[06_ARCHITECTURE] cookiegli_get_genome: High-density Layer 1 static codebase architecture map (<600t).
Tool names lack action verbs. All tools use 'cookiegli_*' prefix instead of verb_noun pattern (e.g., 'search_symbols', 'get_skeleton'). LLMs cannot infer intent from names alone and must parse jargon-heavy descriptions.
Descriptions are overly technical and assume deep domain knowledge. Examples: '[01_TASK_BOOST]...BM25 symbols + focus skeleton + 2026 reasoning effort' and '[07_DARWIN_MEMORY]...Bayesian ROI (Laplace-smoothed)'. LLMs cannot determine when to select these tools without understanding internal architecture.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
[03_CODE_SKELETON] cookiegli_get_skeleton: Extract folded code skeleton preserving verbatim focus symbol.
[02_SYMBOL_IR] cookiegli_search: Industrial Okapi BM25+ full-text search across codebase symbols using SQLite FTS5 BM25+ ranking.
[01_TASK_CONTEXT] cookiegli_synthesize_context: Synthesizes surgical targeted context for a specific coding or debugging task.
No output schemas documented. Tools return results but LLMs cannot plan downstream calls or extract required fields. Example: cookiegli_search returns results but structure is undocumented.
Parameter descriptions lack actionable constraints. Example: 'max_tokens' has no guidance on valid range (1 - 10000?). 'scope' in cookiegli_darwin_record has no format specification. LLMs cannot validate inputs without explicit constraints.
No error handling guidance. Tools provide no recovery hints (e.g., 'If symbol not found, try cookiegli_search()'). Agents cannot self-correct on failures.