Loading ToolBench
Universal agentic connector exposing StepStitch diagnostic and reproduction capabilities over the Model Context Protocol. Provides read-only and draft-only operations for trace analysis, replayability scoring, privacy posture assessment, and Playwright reproduction generation.
StepStitch MCP server has well-structured tool definitions with clear naming (verb_noun pattern) and comprehensive descriptions (avg 150+ chars). All 13 tools have input schemas with typed parameters and descriptions. However, output schemas are not documented in the source code provided, and error handling guidance is absent. Tools are read-only (safe) except for two write operations. Naming is consistent and descriptive (e.g., 'get_trace_summary', 'generate_playwright_repro'), but parameter descriptions could be more prescriptive about constraints and expected formats. No tool annotations (readOnlyHint/destructiveHint) are visible in the schema definitions.
Build sanitized ServiceNow + Salesforce + Genesys DRAFTS. Sends nothing.
Build the named financial-services support DRAFT pack. Sends nothing.
The smallest failing path compiled to Playwright (drops unrelated-route detours). Structural fields only.
Deterministic Playwright reproduction code (text only; never run against production).
The Safe Agent Packet: summary + replayability + privacy posture + diagnostic + Playwright repro composed into one call. Structural fields only; a convenience wrapper over five already agent-safe reads, not new capability.
Signed, independently-verifiable evidence bundle (scrub report + replayability + verdict + SDK build) with a tamper-evident hash. Structural fields only.
Output schemas not documented. Tools return complex objects (summaries, scores, repro code, attestation bundles) but no schema definitions are visible in source. LLMs cannot plan downstream operations or extract specific fields without documented return types.
No error handling guidance. Tool descriptions do not explain what happens on failure (invalid trace_id, permission denied, service unavailable) or what the LLM should do next. Agents cannot recover from errors without explicit recovery paths.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2026-07-28+ | v2 |
Sanitized frontend/API diagnostic summary and recommended next step.
Per-step fragility ranking (selector brittleness + templated routes), worst-first — predicts what will break. Structural fields only.
Per-trace scrub report and the never-captured list.
Reproducibility score, grade (A-F), and warnings.
Sanitized, structure-derived summary of a trace (no raw values).
List recent traces (admin, audited). Returns no trace bodies.
Match this trace against the verified-fix corpus by structure — surfaces 'you've fixed this shape before' with the prior fix ref. Structural fields only.
No tool annotations. Write operations (create_export_preview, create_fs_export_preview) lack destructiveHint annotation. Read-only tools lack readOnlyHint. Agents cannot distinguish safe from unsafe operations without explicit annotations in the schema.
Parameter descriptions lack constraint details. 'trace_id' and 'project_id' parameters have minimal descriptions. No guidance on format (UUID? alphanumeric?), length, or what happens if invalid. 'limit' has max=200 in schema but no explanation of why or what happens if exceeded.
Composition risk: 'get_agent_packet' is a convenience wrapper over five existing tools. While documented as such, this creates redundancy and decision overhead for LLMs. Consider whether agents benefit from the wrapper or should compose the five calls themselves.