8000 GitHub - idvxlab/VibeTrace-OLD at dev · GitHub
[go: up one dir, main page]

Skip to content

About

No description, website, or topics provided.

Resources

Stars

263 stars

Watchers

0 watching

Forks

 
 

Repository files navigation

VibeTrace: See how your agents think

Version 1.5 · Visualizing Agent Runtime Behavior for Human Intervention in Vibe Coding

A web dashboard for OpenCode that connects to a local OpenCode HTTP server (REST + Server-Sent Events). VibeTrace gives developers a live, layered view of agent execution and supports action-level process intervention — fork, inspect, and steer runs without losing trajectory context.


UI preview

VibeTrace action-flow view


Two core capabilities

1. Layered real-time execution visualization

VibeTrace turns raw OpenCode message streams into a structured, reviewable execution surface:

  • Task-aware session view — Detects when the user switches tasks inside a session and surfaces each completed task as its own tab, so trajectories from different goals are not mixed together.
  • Subtask panels from todos — Within a task, planner todos are materialized into subtask panels. Each panel shows the action-flow trace for that slice of work, making it easy to locate and audit specific steps.
  • Rich action-flow rendering — Orthogonal layout of mapped tool/agent steps, branching forks, contextual tooltips, and click-to-focus linking between the flow, todos, and transcripts. Actions duration toggles between fixed step spacing and horizontally scaled blocks keyed to measured duration. Actions color switches the palette between tokens and tool type lenses. Toolbar timeline / summary changes how subtasks are arranged in the rail; fullscreen is available for the flow view.
  • Per-panel analysis — Each subtask panel can show a trace summary and automated error diagnosis to help you understand what happened and where things went wrong.
  • Cross-linking — Optional connectors from todo rows into a linked card or into the focused action when one is selected.

2. Action-level process intervention

Beyond observation, VibeTrace supports interactive steering grounded in the live trajectory:

  • Fork from any action — Branch the session at a specific tool/agent step, capture a pre-fork panel snapshot, and continue in a new OpenCode session while preserving comparison context.
  • Trajectory-based branching — Fork connectors and ghost trails show how a branched run diverges from the parent path, so you can experiment without losing sight of the original execution.

Installation & Running

VibeTrace runs as three local processes: the OpenCode HTTP server, the memory-worker (Python), and the Vite UI. You start each one from the command line.

Prerequisites: OpenCode CLI, Node.js, and Python 3 (python on Windows, python3 on macOS/Linux).

1. Install OpenCode

Follow the upstream installation guide, then verify:

opencode --version

2. Clone and install dependencies

git clone -b V1.5 https://github.com/idvxlab/VibeTrace.git
cd VibeTrace
npm install

3. Configure environment

Copy the environment template. The defaults work out of the box for a local OpenCode server on port 4096 with no HTTP password:

cp .env.example .env.local

Key settings in .env.local (see .env.example for the full list):

Variable Default Purpose
OPENCODE_PROXY_TARGET http://127.0.0.1:4096 Vite dev proxy → OpenCode
OPENCODE_BASE http://127.0.0.1:4096 memory-worker → OpenCode
VITE_OPENCODE_BASE (empty) Use Vite same-origin proxy (recommended)
VITE_MEMORY_WORKER_BASE (empty) Worker API also proxied through Vite

If you use a different OpenCode port, update OPENCODE_PROXY_TARGET and OPENCODE_BASE to match.

Do not set VITE_OPENCODE_SERVER_PASSWORD / OPENCODE_SERVER_PASSWORD for a normal unsecured opencode serve.

4. Start services (three terminals)

Run every command below from the VibeTrace project root (the directory where you ran npm install). Keep all three processes running while you use VibeTrace.

Terminal 1 — OpenCode (backend API)

opencode serve --port 4096

If your OpenCode install already listens on port 4096 by default, opencode serve alone is fine. Otherwise pass --port 4096 so it matches .env.local.

You should see something like:

opencode server listening on http://127.0.0.1:4096

Terminal 2 — memory-worker (trace ingest & panel analysis)

From the project root:

npm run worker:py

Listens on http://127.0.0.1:8714 by default.

Terminal 3 — Vite UI

From the project root:

npm run dev

Open http://127.0.0.1:5173 in your browser.

Daily use

  1. From the project root, start Terminal 1 (opencode serve --port 4096), then Terminal 2 (npm run worker:py), then Terminal 3 (npm run dev).
  2. Use VibeTrace in the browser. Trace ingest and per-panel analysis run in the background via the worker.

Optional environment variables

Variable When to set
SKILL_WRITE_ROOT Where analyzed skills are written
VITE_TRACE_SESSION_TURN_LIMIT More history turns per ingest (default 5)

The model is chosen in the composer dropdown. If nothing is selected, VibeTrace sends opencode/big-pickle. You do not set a model in .env.local.


Tech stack

React 19 · TypeScript · Vite · Tailwind CSS 4 · d3 · react-tooltip


License

MIT

About

No description, website, or topics provided.

Resources

Stars

263 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0