Loading ToolBench
A local-first engineering memory layer and MCP server for AI coding tools
traz has 14 well-named tools with complete input schemas and descriptions. Naming follows verb_noun convention (traz_recent, traz_search, traz_add). Descriptions are detailed (100-250 chars typical) and explain WHEN to use each tool. However, output schemas are not documented, LLMs cannot predict response structure for downstream chaining. Parameter descriptions are present but lack explicit constraints (enums, ranges, formats). Error handling guidance is absent, tools do not indicate what to do on failure. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Composition is strong: tools chain well (search → show → diff), and destructive operations are properly separated.
Add a new engineering event to the traz timeline. Use this to record bug fixes, refactors, decisions, or any context worth preserving.
Create a checkpoint: a snapshot of the current state that can be restored later. Useful before context window fills up or before major refactors.
Compress the timeline by removing low-value events and consolidating similar ones. Experimental tool.
Get a structured summary of recent activity, optimized for AI agent consumption. Provide a 'query' to retrieve only relevant context via RAG. Use format='dense' for ~50-70% token savings, and max_tokens to enforce a budget. If you are starting a fresh chat to reset context bloat, use this to instantly regain your bearing without reading old conversation history.
Delete a specific event by ID. Experimental tool.
Show the diff/patch for a specific event.
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination) for downstream tool chaining. E.g., traz_search returns results but response format is opaque.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. traz_delete and traz_rollback are destructive but unmarked; traz_recent is read-only but unmarked. Agents cannot infer safety without annotations.
Parameter constraints not formalized. 'format' param accepts 'markdown', 'dense', 'toon' but no enum constraint. 'limit' accepts 1-100 but no min/max. 'threshold' is 0.0-1.0 but undeclared. LLMs guess valid values.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | <=2025-11-25 | v2 |
Get a quick summary of what happened in the last 24 hours instead of reading the full timeline.
Get recent engineering events from the traz local timeline. If you are starting a fresh chat to reset context bloat, use this to read the latest 'checkpoint' event. Supports token-optimized output: pass format='dense' to get ~50-70% fewer tokens.
Rollback to a previous checkpoint or event. Removes all events after the specified event ID.
Compress a range of events into a single summary event. Useful for reducing noise and token usage.
Search the traz engineering timeline for events matching a keyword. Searches across titles, summaries, tools, types, and filenames.
Show details of a specific event by ID.
Get statistics about the traz timeline: total event count, breakdown by tool, and other metadata.
Get the full timeline of events in chronological order. Experimental tool.
No error handling guidance. Tools do not document what happens on failure (e.g., event_id not found in traz_show, invalid threshold in traz_compress). Agents have no recovery path.
traz_stats has minimal description ('Get statistics...') and empty input schema. Unclear what statistics are returned or when to call it vs traz_timeline.