Her · हेर — a detective for your Claude Code sessions

Source: HuggingFace Date: 2026-06-07 URL: https://huggingface.co/blog/build-small-hackathon/her-blog

Summary

Her (हेर) is a hackathon project that post-mortems Claude Code sessions by parsing the .jsonl trace files the CLI writes. It runs Nemotron-Mini-4B-Instruct locally via ZeroGPU, keeping session data in private auto-deleted namespaces, and produces a report view, a journey graph of tool calls sized by token cost, and a conversational “Ask Her” copilot that cites specific turns. The deterministic evaluation engine flags risky operations (production deploys, destructive shell commands) first; the model only supplies natural-language narrative on top.

Implications

• Agentic-engineering patterns. Retroactive session analysis is a gap that grows faster than the agents themselves — as Claude Code and peers run longer autonomous loops, the audit surface compounds. Her is a proof-of-concept for a category (session observability) that will likely graduate from hackathon toy to production tooling.
• Small-model economy. The choice of a 4B parameter local model for the explanation layer — not a frontier API call — is deliberate: privacy-sensitive session data shouldn’t leave the machine. This validates the small-model-as-local-reasoning-layer pattern for developer tooling.
• Supply-chain trust. Surfacing tool provenance (Homebrew, npm, PyPI) from within a session trace points toward a future where agent audits can flag unexpected package installation mid-run — a supply-chain trust capability embedded in the observability layer.