See how your agents think. Steer what happens next.
A visual workspace for human–agent collaboration with OpenCode.
Understand execution, intervene at any action, and turn experience into reusable skills.
Installation · Visualization · Intervention · Skill evolution
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 durationtoggles between fixed step spacing and horizontally scaled blocks keyed to measured duration.Actions colorswitches the palette between tokens and tool type lenses. Toolbartimeline/summarychanges 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.
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.
VibeTrace turns execution traces and feedback into reusable skills:
- Learn from completed tasks — Task-switch detection separates completed work from the next goal, providing task-level evidence for skill evolution.
- Distill and refine skills — The evolution pipeline evaluates trace evidence, distills candidate changes, scores them, and writes accepted changes to local skill files. Skills can be created, revised, merged, or removed as experience accumulates.
- Learn from feedback — Task-level and panel-level feedback can be distilled into reusable guidance, with links back to the source trajectory for review.
Analysis is triggered through the dashboard and uses models configured in OpenCode. Background model availability and quota affect whether analysis and evolution can complete; installing the plugin does not by itself train or change the underlying model.
Requires OpenCode, Node.js 20+, and Python 3.9+. The commands below work on macOS, Windows, and Linux.
First, check your OpenCode version:
opencode --versionnpx --yes @xiaokeyan/opencode-vibetrace@1.0.2 setupnpx --yes @xiaokeyan/opencode-v2-vibetrace@1.0.1 setupThen restart OpenCode (fully quit and reopen Desktop) and open your project as usual.
For V2, run opencode service restart before reopening OpenCode.
VibeTrace opens its dashboard automatically.
OpenCode V1:
npx --yes @xiaokeyan/opencode-vibetrace@1.0.2 uninstallOpenCode V2:
opencode plugin remove @xiaokeyan/opencode-v2-vibetrace@1.0.1Restart OpenCode afterwards. For V2, also run opencode service restart.
Your chat history and VibeTrace data are preserved.
React 19 · TypeScript · Vite · Tailwind CSS 4 · d3 · react-tooltip