Her · हेर — a detective for your Claude Code sessions
Her analyzes Claude Code session logs, explains what happened, and flags risky actions like deploys, config changes, and secrets.
Intelligence analysis by GPT-5.4 Mini
Her is a Hugging Face Space that turns Claude Code’s write-only JSONL session logs into plain-English forensic reports. It traces risky turns, token usage, tools, subagents, and MCP servers, and adds an ask-and-explain copilot for digging into specific decisions.
Her is like a magnifying glass for a robot helper’s work log. It reads the helper’s notes, points out the risky parts, and explains why something happened in simple words instead of a pile of messy computer text.
Analysis
What Her does
Her is a detector-style analyzer for Claude Code sessions. A user drops a session .jsonl file into the Space, and the app reconstructs the run in plain English. It highlights risky actions such as deploys, production changes, config edits, and secrets, and ties each one back to the exact turn where it happened.
How it works
The article says the evaluation side is deterministic: the findings do not depend on the language model. The model is used only to write the English narrative and to suggest softer improvements when a named, fixable pattern is detected. If there is nothing meaningful to report, it stays quiet rather than inventing issues.
Her also exposes an “Ask Her” copilot. That lets a user ask questions like why a tool was used, then jump from the answer to the cited turns and the exact tool call. The Space can analyze a single session or combine multiple sessions into a project view.
Trust and implementation
The author says no third-party AI API is called. Nemotron-Mini-4B-Instruct runs on the Space’s own GPU through ZeroGPU, and uploads go into a private, auto-deleted namespace tied to the run. The piece emphasizes that the forensics engine stays deterministic so the numbers do not change when the model changes.
Product details
A small but useful feature is the offline tool database. Her does not just list CLI tools; it tries to identify them using a database of common tools from Homebrew, npm, and PyPI. That helps it distinguish ordinary commands from deploy tools, database clients, or dev servers that deserve extra attention.
The project started as an operator-focused journey graph, then evolved into a simpler report after feedback. It now combines a React frontend, a Gradio server, deterministic analysis, and model-generated prose into a session review tool for Claude Code.
Key points
- Her turns Claude Code `.jsonl` session logs into plain-English forensic reports.
- It flags risky actions like deploys, production changes, config edits, and secrets, and links them to the exact turn.
- The evaluation engine is deterministic; the model is used only for prose and softer suggestions.
- Ask Her lets users query a session and jump to the cited tool call or turn.
- The Space runs Nemotron-Mini-4B-Instruct on its own GPU and says no third-party AI API is used.
If Her works as intended, it could make agent runs much easier to debug and review. That could help people spot wasted tokens, risky commands, and confusing tool use without reading thousands of lines of JSON.
Its usefulness depends on the quality of the trace and the pattern rules behind the report. If the session data is incomplete or the heuristics miss something important, the explanations could look reassuring while still leaving blind spots.
