Claude Code hid file-access data — a new open-source observability tool emerged

Agentic coding tools are increasingly able to read, write, refactor and reason across entire codebases. But the more authority these systems are given, the more consequential their actions become — and the more important it is for developers to understand exactly what those systems are doing.

That dynamic was perhaps best exemplified this past week after Anthropic altered the Claude Code CLI to show less granular detail about which files were being read, searched or modified during a session – a move that prompted a developer to build an open source tool restoring that visibility.

Visibility regresses in Claude Code

In a blog post titled Claude Code is Being Dumbed Down, an author going by the name of Yoshi observed that changes made to version v2.1.20 of Claude Code reduced detailed visibility into file-level operations.

As the author demonstrated, earlier versions of the CLI surfaced each file read individually, along with contextual information such as how many lines were inspected. Here, it shows the tool reading package.json, tsconfig.json and README.md, along with the number of lines inspected in each.

Following the update, that granular output was replaced with a summary line indicating the number of files accessed, with the detailed trace moved behind a keyboard toggle.

This regression had in fact already been noted elsewhere, including community bug reports showing that Claude Code was now hiding detailed file-level usage — and for developers, that kind of detail is more than cosmetic. Seeing which files an agent reads, searches or modifies provides context for its reasoning, helps users steer sessions more precisely, and makes it easier to catch mistakes before they propagate across a codebase.

As the chatter escalated on forums such as Hacker News, Claude Code chief Boris Cherny responded at length, arguing that the change was intended to reduce cognitive overload in the CLI and make the interface easier to follow. The detailed file-level activity had not been removed, he noted, but moved into “verbose mode” that can be enabled to surface the full execution trace. In other words, the underlying data remains available, it’s just no longer shown by default.

While verbose mode restores detailed output, some argued that it doesn’t address their specific need: immediate, lightweight confirmation of which parts of the codebase the agent is exploring during its initial planning phase.

“I don't want verbose mode – I want Claude to tell me what it's reading in the first 3 seconds, so I can switch gears without fear it's going to the wrong part of the codebase,” one user responded. “By saying that my use-case requires verbose mode, you're saying that I need to see massive levels of babysitting-level output … to be able to do this.”

Rebuilding the visibility layer

In response, one developer took it upon themselves to create a new open source project dubbed claude-devtools, a local application designed to surface detailed execution traces from Claude Code sessions in a more structured and readable way.

The tool runs alongside Claude Code and reads the session logs already stored locally in the user’s ~/.claude/ directory. By reading those logs in real time, it reconstructs file reads, writes, searches and tool calls into a structured interface, effectively rebuilding the visibility layer that some users felt had been obscured.

“I built this because I got tired of the Claude Code CLI hiding details from me,” the developer wrote, arguing that while verbose mode restores much of the underlying output, it does so by flooding the terminal with low-level system detail.

“I wanted a middle ground — full observability without the noise,” they continued.

The application presents inline diffs when files are edited, breaks down how tokens are consumed across files, tool outputs and model reasoning, and visualises sub-agent execution paths that can otherwise appear interleaved and difficult to follow in the CLI.

In truth, this dynamic mirrors a broader shift already underway in enterprise AI tooling. As coding agents move deeper into production use, organisations are placing increasing emphasis on metrics, evaluation frameworks and operational visibility. Recent enterprise deployments have introduced dashboards tracking agent performance, task completion rates, failure modes and cost efficiency — reflecting a growing expectation that AI systems must be measurable.

The Claude Code brouhaha is a more developer-driven expression of that same pressure. When agents are given the ability to read and modify code autonomously, transparency becomes foundational to trust. Developers want to see not only what changed, but how and why — and they expect that visibility to be baked in, not toggled on.