Silent Failures in My Test Suite: Config Traps in OpenTelemetry MCP
Thirumalaiboobathi · reddit · 2026-08-24
The author discovered three bugs in opentel-mcp v0.12.0 where features appeared configured but produced nothing.
- A sampling policy used a key name (mcp.tool.schemadrift.detected) that existed as a span event but never as a span attribute, causing it to never match without errors.
- Calls to models not in the pricing table lacked a cost attribute, making them look like free calls to numeric thresholding policies, leading to sampled-out real spend.
- Unpriced calls bypassed the budget tracker entirely, disabling protection without warnings.
To prevent recurrence, the author added a test that parses the committed YAML and asserts every referenced attribute is a real exported constant and actually set via span.setAttribute in the source code. This method of verifying example configs against code implementation is highly recommended.
More from coding & agent
- Agent drift and safety studies reveal risks in automated systems — AiJohnAllen · 2026-08-24
- Agent performance degrades on long runs; lies snowball across bots — AiJohnAllen · 2026-08-24
- Cursor Grok Bot source code exposed via runtime maps, reconstructed — banteg · 2026-08-24
- Experiment: Using Grok bot to autonomously sell a business on Flippa — IndraVahan · 2026-08-24
- Open Source Tool Tests MCP Spec Conformance in 60 Seconds — hasmcp · 2026-08-24
- AI Agents Often Fail in Production Due to Non-Model Issues — owenbrooks473 · 2026-08-24