view-trajectory

name

view-trajectory

description

Open a Harbor ATIF agent trajectory (.json) in the harbor-trajectory-viewer web UI to inspect the step timeline, clustered tool calls, metrics, subagents, and code diffs. Can also generate an annotations sidecar that adds one-line summaries for each tool-call cluster and translations of every message into a target language. Use when the user asks to visualize, view, open, summarize, or translate a Harbor trajectory / rollout.

View a Harbor ATIF trajectory

@tengxiaohtx/harbor-trajectory-viewer renders a Harbor ATIF trajectory (*.trajectory.json) as a web UI: a step Timeline (user / agent messages, reasoning, tool calls grouped into collapsible clusters, and system / context-management steps), and a separate Diffs tab showing every file change as a unified or split code diff. The sidebar surfaces session metadata, token/cost metrics, and any embedded subagent trajectories you can drill into.

The hbtv command is provided by the package. Prefer npx -y @tengxiaohtx/harbor-trajectory-viewer ….

1. Just view a trajectory

npx -y @tengxiaohtx/harbor-trajectory-viewer "<path-to>.trajectory.json"

Add --no-open for headless, -p <port> to choose a port. Run with no path to open drag-and-drop upload mode.

Locating the file

Harbor writes ATIF trajectories (named trajectory.json or *.trajectory.json) into a job's logs directory. To open the newest one under the current directory:

FILE=$(ls -t **/*trajectory*.json 2>/dev/null | head -1)
npx -y @tengxiaohtx/harbor-trajectory-viewer "$FILE"

If the user names a specific file, use that directly.

2. Add summaries and/or translations (annotations)

The viewer can overlay a sidecar annotations file (<file>.hbtv.json) that adds, per stable id:

• a one-line summary for each cluster of consecutive tool calls (shown in the cluster header), and
• a translation of each user/agent message into a target language (shown beneath the original text).

Workflow — you (the agent) fill in the annotations:

1. Generate the scaffold. This writes a .hbtv.json next to the trajectory with every text unit (with its original text) and every tool-call cluster (with the tool names) already enumerated and correctly keyed:

   npx -y @tengxiaohtx/harbor-trajectory-viewer extract "<file>.json" --lang "<target language>"
   

   Use the language the user asked for (e.g. 简体中文, English, 日本語); omit --lang if they only want summaries.

2. Read the scaffold (<file>.json.hbtv.json). It looks like:

   {
     "version": 1,
     "targetLang": "简体中文",
     "units": [
       { "type": "text", "id": "step-1", "role": "user", "original": "…", "translation": "" },
       { "type": "cluster", "id":        
     
   
   

Notes

• npx fetches the published package, which bundles the built HTML — no build step.
• From a local clone, run pnpm run build first, then node bin/cli.js ….
• Everything is served from localhost; nothing is uploaded.
• ATIF spec: https://github.com/harbor-framework/harbor/blob/main/rfcs/0001-trajectory-format.md