Loading ToolBench
Powerful, self-contained AI context layer. Combines persistent memory, temporal knowledge graph, and hybrid code search into one local engine. Zero cloud, 100% private.
Synapse has 15 tools with complete schemas and descriptions, but exhibits significant design issues. Two tools (synapse_kg_manage, synapse_knowledge_graph_manage) are duplicates with identical functionality, violating single-responsibility principle. The mega-tool pattern (synapse_kg_manage with 13 enum actions) concentrates too much functionality into one tool, making it hard for LLMs to reason about when to call it. Descriptions are present but generic (194 char baseline met). Parameter descriptions exist but many lack actionable constraints. Error handling and recovery guidance are absent from visible code. Tool naming is inconsistent: 'synapse_kg_manage' vs 'synapse_knowledge_graph_manage' creates confusion. Output schemas are not documented in the provided code.
Proactively predict and return the context most likely needed for the current coding session based on historical patterns.
Dismiss a pattern warning — it won't fire for the specified reason.
Show mistake patterns ranked by how often they've triggered — your personal bug frequency report.
List all recorded mistake patterns, optionally filtered by category.
Teach Synapse about a mistake you just fixed so it never happens again.
Scan a code snippet or file for patterns matching past mistakes.
Add a new failure pattern to the library from your own experience.
Duplicate tools: synapse_kg_manage and synapse_knowledge_graph_manage are identical. LLMs will waste reasoning cycles deciding between them. Consolidate to one canonical name.
Mega-tool anti-pattern: synapse_kg_manage combines 13 distinct actions (add_entity, delete_entity, ingest_markdown, manage_branches, write_diary, etc.) into one tool via enum. This violates single-responsibility and forces LLMs to reason about which action to select. Split into separate tools: create_entity, delete_entity, ingest_markdown, manage_branches, write_diary.
synapse_symbol_query combines 6 actions (callers, definition, implementations, usages, get, rename_preview) into one mega-tool. Split into: get_symbol_definition, find_symbol_callers, find_symbol_usages, find_implementations, preview_rename.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2025-06-18+ | v2 |
Before implementing an architectural decision, get a structured analysis of how this approach has historically failed.
After a decision plays out, record whether the predicted failures occurred to improve future pre-mortem accuracy.
Record outcomes of a session to improve future predictions.
Unified controller for all knowledge graph mutations (add, delete, ingest, branch management, diary write). Use "action" to select mode.
Unified controller for all knowledge graph mutations (add, delete, ingest, branch management, diary write). Use "action" to select mode.
Reads the specific content of a node from the generated PageIndex tree using the node_id.
Generates a hierarchical TOC-style semantic tree (PageIndex) for a given text or markdown file, allowing LLMs to reason about the document structure before reading specific chunks. Returns the tree with node IDs and summaries.
Unified controller for static code intelligence (callers, definitions, usages, implementations). Use "action" to select mode.
Output schemas not documented in visible code. Tools return results but LLMs cannot see what fields to expect. Document return types for all 15 tools (e.g., synapse_kg_manage returns {success: bool, entity_id?: string, error?: string}).
No error handling guidance visible. Tools lack recovery hints (e.g., 'If entity not found, try search_entities() first'). Add actionable error messages and recovery paths.