metadata
title: Agent Failure Atlas
emoji: 🧭
colorFrom: yellow
colorTo: indigo
sdk: static
app_file: index.html
fullWidth: true
pinned: false
license: apache-2.0
short_description: Private, client-side analysis of AI-agent trace failures.
Agent Failure Atlas — Static Space
A browser-only showcase for the Open Agent Failure Atlas.
- Trace files are parsed and analyzed locally in the browser.
- The Space makes no model calls and does not upload user-provided traces.
- It supports Hugging Face Session Trace Simple Format (STS), one-row normalized datasets, and OpenAI-style message JSONL.
- The complete Python CLI, native adapters, release tooling, and tests remain in the main repository.
Public artifacts
- GitHub
- Benchmark Dataset
- DeltaStore Trace Explorer
- DeltaStore Replay Traces Dataset
- DeltaStore browser app
Deploy
Create a Hugging Face Space using the Static SDK, then upload the contents of this directory to the root of the Space repository. The YAML front matter above configures sdk: static and app_file: index.html.
No build step or server dependency is required.
Local verification
python -m http.server 8000
# Open http://localhost:8000
node --test tests/core.test.mjs
Scope
The browser implementation mirrors the six deterministic detector families used by the Python MVP:
- scope and unauthorized-tool violations
- approval bypass
- prompt injection in tool output
- runaway repeated actions
- silent or unhandled tool failure
- secret exposure
The bundled 20-session result is a synthetic software smoke benchmark. A perfect fixture score does not establish real-world detector accuracy or model safety.