Claude Code sessions already leave behind the raw material for accountability: a .jsonl trace with every turn, tool call, and token. The problem is that most teams do not actually use that evidence. In practice, the file is a write-only artifact — dense enough that even a 4,000-line log can hide the moment an agent reached for production, why the context budget disappeared, or which subagent quietly consumed the run.

That is the gap Her is trying to close. Hugging Face describes Her — from the Marathi word for detective — as a detective for Claude Code sessions, built to drop a session file onto the page and investigate what happened. Instead of forcing operators to sift through JSON, it reconstructs the session in plain English, then maps the narrative back to the exact turn where a decision occurred. The tool is explicit about the kinds of events it treats as important: deploys, config and production changes, and secrets.

That matters because observability without attribution is not much use in deployment governance. If a release picked up a risky config change, the question is rarely just what changed; it is who or what introduced it, in which turn, and through which tool path. Her’s value proposition is that it takes the session trace and makes those links readable enough to audit. It shows the turn sequence, the tool usage, the tokens consumed, and the production decisions that emerged from the trace. For teams running Claude Code in production-adjacent workflows, that is the difference between a log you store and a record you can actually review.

The governance angle is straightforward. When a system can point to the precise moment a deploy command was invoked, or identify that secrets were handled in a risky way, post-hoc review becomes much more concrete. Instead of arguing about the broad shape of an interaction, operators can inspect the turn-level sequence and see how the model or agent moved from reasoning to action. That makes Her useful as an audit aid, but also as a risk flagger: not every session needs review, yet the ones that include production changes, config edits, or secret exposure deserve a closer look.

Hugging Face also positions Her as something more than a passive viewer. There is a built-in copilot called Ask Her, which lets users probe the trace once it has been reconstructed. That matters operationally because a narrative is only useful if teams can interrogate it. A reviewer might want to ask why a particular tool was used, whether a token spike lined up with a specific subagent, or which step preceded a risky action. Ask Her gives the trace a second pass, which is especially helpful when the first read surfaces a gap that needs follow-up.

For teams trying to operationalize this, the cleanest approach is to treat Her as a governance checkpoint rather than a curiosity. Session traces can be dropped into the page after a deploy, during an incident review, or as part of a rollout sign-off. The plain-English reconstruction then becomes a review artifact alongside the usual logs, change records, and CI/CD history. Used that way, Her can help a team spot whether an agent made an unexpected production move, whether a config change was introduced through the right path, and whether a session drifted into risky territory before it reached users.

That said, the tool’s usefulness depends on the quality and completeness of the underlying session data. A readable narrative is not the same thing as truth independent of the trace. If a session file is incomplete, or if the surrounding workflow is already poorly governed, Her can only reflect what it can see. The blog is careful on that point: it suggests rather than asserts, and it stays silent when there is nothing worth saying. That restraint is a feature, but it also means teams should not mistake a clean narrative for an end-to-end control system.

There is also an integration cost to consider. A traceability layer only helps if it is wired into the places where decisions already matter: CI/CD gates, post-deploy reviews, incident retrospectives, and policy checks around secrets handling and config management. In other words, the artifact needs a home in the operational workflow, not just a link in a notebook. That is where Her becomes more than a visualization layer. It becomes a control point that can make Claude Code usage more inspectable without asking humans to decode raw JSON every time.

For product teams shipping Claude Code-powered features, the immediate takeaway is not that every session needs a detective. It is that the session trace is now a governance surface, and tools like Her make that surface readable enough to use. If your deployment strategy depends on confidence in agent-driven changes, then the question shifts from whether traces exist to whether you can review them quickly enough to matter. Her’s pitch is that you can — by translating opaque JSONL dumps into a narrative tied to exact turns, risky moves, and production decisions.

AI News

Her turns Claude Code traces into an auditable story

Get the brief in your inbox.