Loading ToolBench
Local-first context engine for coding agents. Provides hybrid retrieval (semantic + BM25 + graph), memory management (episodic/semantic/procedural), conversation handoff, and skill routing for AI-powered coding assistants.
Server provides 20 tools with consistent naming (verb_noun pattern) and reasonable descriptions (avg ~120 chars). All tools have input schemas with typed parameters and descriptions. However, output schemas are not documented in the provided source, parameter descriptions lack actionable constraints (ranges, formats, enums), and error handling guidance is absent. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Composition is reasonable but some tools conflate concerns (e.g., memory_inject mixes retrieval and injection).
Score repository context completeness across multiple dimensions
Estimate token count for text using model-specific tokenizer
Explain context retrieval strategy and relevant evidence for a question
Get model-specific token budget information
Create a no-compression history handoff for conversation continuation across sessions
Invalidate cached context state for confirmed stale data
Delete a memory entry by key
Output schemas not documented. Tools return complex objects (e.g., retrieve_context, memory_search, route_skills) but no schema documentation visible in source. LLMs cannot plan downstream calls or extract fields reliably.
Parameter descriptions lack actionable constraints. 'max_tokens' has no range (1 - 8192?). 'target_ratio' has no bounds (0 - 1?). 'limit' parameters undefined. LLMs will pass arbitrary values, risking API failures.
No tool annotations despite clear risk classifications. Tools marked DESTRUCTIVE (memory_delete, memory_evict) or WRITE lack readOnlyHint/destructiveHint/idempotentHint annotations. Agents cannot distinguish safe from risky operations.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | <=2025-11-25 | v2 |
Perform LRU eviction on memory store to manage storage
Inject relevant memory entries into the current prompt context
List recent memory entries across all tiers
Save information to episodic, semantic, or procedural memory tiers
Search across episodic, semantic, and procedural memory tiers
Get database statistics for memory storage
Prepare and assemble context for a query with token budgeting and optimization
Refresh workspace state for changed paths
Generate a map of repository structure and top-level directories
Restore a raw handoff payload to resume a previous conversation
Hybrid retrieval combining semantic search, BM25 lexical matching, and dependency graph analysis to retrieve relevant code context
Route a task to appropriate skills based on semantic matching and scoring
Submit feedback on skill execution with outcome verification
No error handling guidance. Tools provide no recovery hints (e.g., 'workspace_id not found, call list_workspaces() first'). Agents cannot self-correct on failures.
Enum constraints missing. 'memory_type' accepts ['episodic','semantic','procedural','all'] but no enum in schema visible. 'outcome' in skill_feedback has enum but others lack formal constraints.