How do you see what Claude Code is actually doing?

Use --verbose flag (floods terminal with JSON) or the third-party Local Log Viewer tool. Claude Code v2.1.20+ replaced detailed logs with opaque one-liners like 'Read 3 files'.

Why is AI coding agent output hard to debug?

Agents replaced verbose action logs with condensed summaries for cleaner UX. But this hides what files changed, what prompts were used, and how tokens were spent. Debugging failures requires the raw log which is unreadable JSON.

Is there a debugger for Claude Code?

Yes, claude-devtools reconstructs tool calls, token attribution, and subagent trees in a visual UI. The Local Log Viewer (69 HN points) parses Claude Code's log files into a readable format.

How do I check what files Claude Code changed?

With --verbose you get JSON dumps. The Local Log Viewer parses these into a navigable UI showing file paths, diffs, and tool call chains. Claude Code's default output since v2.1.20 hides this detail.

What is the biggest usability pain with AI coding agents?

Opacity. You can see inputs and outputs but not the reasoning and actions in between. When the agent does something unexpected, tracing it requires either verbose JSON or third-party log parsing tools.

← Back to dashboard

clawsmith.com/signal/ai-coding-agent-log-visibility-hidden-output

⚠ IssueWide Opendev_tool_cliLive

AI coding agent terminal output is opaque - verbose mode floods with unreadable JSON

After Claude Code v2.1.20, action logs were replaced with one-liners like 'Read 3 files' and 'Edited 2 files' - no file paths, no diffs, no content. The --verbose flag dumps raw JSON and system prompts into the terminal, which is unreadable. A developer built a Local Log Viewer (69 HN pts) to fix this. A second tool called claude-devtools reconstructs tool calls, subagent trees, and token attribution in a UI. The core frustration: you can't see what your agent is actually doing without custom tooling.

Product Idea from this Signal

A CLI tool that turns AI coding agent terminal output into a navigable trace tree

246 ▲

AI coding agents like Claude Code and Cursor emit dense verbose JSON logs that are nearly impossible to read in real time, making it hard to understand what the agent actually did, which tool calls succeeded or failed, and where tokens were burned. Developers are left guessing why a run went wrong or what the agent changed. This tool parses the raw output stream in real time, reconstructs a structured trace tree of every tool call, file diff, subagent spawn, and token spend, and renders it as a navigable terminal or local-web UI that makes agent behavior immediately legible.

developer-toolsai-agentsobservabilitydebuggingclaude-codecli

Competitive88 leadsView Opportunity →