I keep running into the same debugging problem with coding agents: when an agent behaves differently than expected, the important part is often hidden in the request/response layer. Things I usually want to inspect: - the actual prompt and request body - tool schemas sent to the model - tool calls and tool results - streamed response chunks - token usage, cost, and latency - diffs between turns or providers I found ccglass useful for this because it works as a local dashboard for OpenAI-compatible and coding-agent traffic, including setups like Claude Code, Codex, Qoder, Ollama, LM Studio, OpenRouter, and others. Curious how others here debug this today. Do you mostly rely on provider logs, custom proxy logging, SDK instrumentation, or a local inspection tool?
