Loading ToolBench
Git-grounded, provider-neutral context continuity for AI coding agents
ContinuityDB demonstrates solid tool design with comprehensive parameter schemas, clear descriptions, and thoughtful risk classification. All 6 tools have descriptions (194-char baseline met). Input schemas are well-structured with type constraints, enums, and bounds. However, output schemas are not documented in the visible code, and error handling guidance is minimal. Tool names follow verb_noun convention (memory_search, memory_capture, handoff_latest). The server implements tool annotations (destructiveHint/readOnlyHint via Risk field), which is current-spec compliant. Main gaps: no documented return types, limited error recovery guidance, and STDIO-only transport caps protocol readiness.
Explicitly save structured continuation state for a named task.
Resume a named task with the latest approved memory and context.
Capture project memory through server-side risk policy. Agents cannot choose tenant, owner, namespace, status, source trust, or approval.
Build a small, cited and token-budgeted context pack for a coding task.
Record bounded, non-destructive feedback on a visible memory. Feedback cannot commit, correct, delete, or change scope.
Search approved engineering memories in the current project and its explicitly linked dependencies.
Output schemas not documented. Tools return results but LLMs cannot plan downstream calls without knowing field names and types. memory_search, memory_context_pack, handoff_latest should document: result structure, pagination fields (if any), and chaining IDs (e.g., memory_id, project_id).
Error handling lacks recovery guidance. No visible error responses that tell LLMs what to do next (e.g., 'Memory not found. Try memory_search() with a broader query.'). Errors should categorize as retryable, user-fixable, or fatal.
memory_feedback description (70 chars) is minimal. Does not explain when to use it vs. other tools, what 'helpful/incorrect/outdated' signals mean, or how feedback affects future searches. Expand to 100-200 chars with context.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2025-06-18+ | v2 |
handoff_latest and handoff_checkpoint names are domain-specific jargon. LLMs may not infer 'handoff' means 'resume task with context' or 'save task state'. Consider rename to resume_task_with_context and save_task_checkpoint for clarity.
memory_capture accepts optional confidence/importance scores (0-1) but no guidance on interpretation. Does 0.5 mean 'neutral' or 'uncertain'? Clarify in parameter descriptions and document how these scores influence retrieval ranking.