Harbor Trajectories

Interesting agent trajectories from Harbor computer use evaluations.

This repo is deployed to Vercel using a Blob-backed jobs archive so we can keep large Harbor job data without hitting Vercel's serverless function bundle size limit.

How It Works

1. The app runs as a Python function at api/index.py.
2. Viewer UI assets are served from vendored Harbor static files in vendor/.
3. Job data is loaded from one of these sources, in order:
   • JOBS_TAR_URL environment variable (preferred)
   • config/jobs_tar_url.txt fallback
   • local jobs/ folder (mainly local dev fallback)
4. On first request (cold start), the function downloads jobs.tgz from Blob, extracts it to /tmp/harbor-jobs, and serves data from there.

Why This Setup

Vercel serverless functions have an unzipped size limit. Bundling jobs/ (and the full Harbor dependency tree) can exceed it. This setup avoids that by:

1. Excluding heavy local folders from deploy uploads via .vercelignore.
2. Loading data from Vercel Blob at runtime.
3. Vendoring only the Harbor viewer code/assets needed at runtime in vendor/.

Repository Files Used In Deploy

• api/index.py: runtime entrypoint and Blob download/extract logic.
• vercel.json: rewrites all routes to Python function and includes vendor/**.
• .vercelignore: excludes jobs/, local harbor/, and local-only files.
• scripts/upload-jobs-to-blob.sh: uploads new jobs tarball and updates env vars.
• config/jobs_tar_url.txt: Blob URL fallback if env var is missing.

Prerequisites

• Node + npm/bun (for local tooling if needed)
• Python 3.12+
• Vercel CLI (npm i -g vercel)
• Vercel account with project access

One-Time Setup

1. Login to Vercel:

vercel login

2. Link this folder to your Vercel project:

vercel link

3. Create and link a Blob store to the project (only once):

vercel blob store add harbor-trajectories-jobs

When prompted, choose to link it to the project and select environments.

4. Pull env vars locally (script also does this when needed):

vercel env pull .env.vercel.local

5. Upload jobs data and set Blob URL env vars:

bash scripts/upload-jobs-to-blob.sh

This script does all of the following:

1. Tarballs jobs/ to a temporary .tgz

2. Uploads it to Vercel Blob

3. Updates config/jobs_tar_url.txt

4. Sets JOBS_TAR_URL for production, preview, and development

5. Deploy:

vercel deploy --prod

Updating Data Later

Any time jobs/ changes, run:

bash scripts/upload-jobs-to-blob.sh
vercel deploy --prod

Verify Deployment

Check these endpoints:

curl -s https://<your-domain>/api/health
curl -s https://<your-domain>/api/config
curl -s 'https://<your-domain>/api/jobs?page=1&page_size=1'

Expected:

1. /api/health returns {"status":"ok"}
2. /api/config shows jobs dir under /tmp/harbor-jobs/...
3. /api/jobs returns real jobs, not an empty list

Troubleshooting

"No jobs in /var/task/jobs"

Cause: deployment is not using Blob URL (wrong project/domain, missing env var, or old commit).

Fix:

1. Confirm you are opening the correct deployment/project URL.
2. Run bash scripts/upload-jobs-to-blob.sh.
3. Redeploy: vercel deploy --prod.
4. Re-check /api/config.

"Serverless Function exceeded 250 MB"

Cause: large files got included in deploy artifact.

Fix:

1. Ensure .vercelignore is present and includes /jobs and /harbor.
2. Keep jobs/ in Blob only (do not include in function bundle).
3. Redeploy.

Blob commands say token is missing

Fix:

1. Ensure Blob store is linked to the Vercel project.
2. Run vercel env pull .env.vercel.local.
3. Re-run upload script.

Security Note

.env.vercel.local can contain secrets and is ignored via .gitignore. Do not commit it.