OpenTrace：把雜亂日誌轉化為下一步最佳行動

What I built

OpenTrace is a local-first debugging assistant for developers. A user pastes
an error, stack trace, or service log and receives a structured report with a
summary, severity, evidence, likely cause, safe next actions, and questions
that would improve the diagnosis.

Live demo: https://opentrace-ai-debugger.onrender.com

Source code: https://github.com/samainakhatun115-cpu/opentrace-ai-debugger

Why this problem

Logs are often technically detailed but operationally unclear. The final error
message may hide the first useful signal, and a generic chatbot can present a
guess as a fact. OpenTrace makes the distinction visible: evidence is shown
separately from hypotheses, and every response includes safe next steps rather
than destructive commands.

How open-source AI is used

When Ollama is available, OpenTrace sends the pasted log to a local
Ollama-compatible open-source model and requests a strict JSON report. The
server does not store logs or request history. The model name and endpoint are
configurable through environment variables.

If Ollama is unavailable, OpenTrace uses a deliberately limited heuristic
fallback and labels the result Demo Mode. It never presents the fallback
as model output.

Architecture

flowchart LR
  A[Browser] --> B[Node HTTP server]
  B --> C{Ollama available?}
  C -->|yes| D[Open-source local model]
  C -->|no| E[Clearly labelled demo heuristic]
  D --> F[Structured JSON report]
  E --> F
  F --> A

Technical decisions

Node.js standard library: keeps the project easy to inspect and run.
Ollama integration: supports local inference without requiring a hosted AI key.
Strict JSON output: gives the UI stable fields and makes uncertainty easier to communicate.
No persistence: logs can contain secrets, so OpenTrace does not retain them by default.
Safe fallback: the app remains demonstrable without hiding an unavailable model dependency.

What I learned

The most important product decision was not to make the assistant execute
commands. A debugging helper should help a developer decide what to inspect
next, while the developer remains responsible for validating production changes.
Making evidence and limitations visible is more useful than producing a
confident-looking answer.

Run it locally

cd OpenTrace
node server.js

Open http://localhost:3000. For model-backed analysis, install Ollama and run:

ollama pull llama3.2:3b

Limitations and next steps

The demo fallback is not a replacement for model inference.
The tool does not inspect infrastructure or execute commands.
A future version could add redaction before analysis and adapters for more local model runtimes.

Acknowledgements

Built as a new project for the Hacktoberfest Open-Source AI Challenge: Week 4.
AI assistance was used during development and drafting.