Public Files
Each environment has one R2 bucket, a Worker, and a /files/* route on its domain.
| Environment | Bucket and Worker | Public URL |
|---|---|---|
| Development | opencode-dev-files |
https://dev.opencode.ai/files/ |
| Production | opencode-production-files |
https://opencode.ai/files/ |
The Worker removes /files/ and URL-decodes the remaining path to obtain the R2
object key. For example:
R2 key: releases/1.0.0/opencode-linux-x64.tar.gz
URL: https://opencode.ai/files/releases/1.0.0/opencode-linux-x64.tar.gz
R2 key: videos/demo.mp4
URL: https://opencode.ai/files/videos/demo.mp4
Deployment
Wrangler bundles and deploys the Worker, its R2 binding, and its path route.
wrangler.jsonc defines the dev and production environments. Their routes
use the existing proxied stage hostnames and take precedence over broader website
routes.
bun typecheck
bun run build
bun run deploy --env dev
bun run deploy --env production
The existing buckets are reused. For a fresh environment, create its bucket once
with bunx wrangler r2 bucket create opencode-<stage>-files and add the matching
environment to wrangler.jsonc.
.github/workflows/deploy-files.yml deploys development on every push to dev
and production on every push to v2, using the existing
CLOUDFLARE_API_TOKEN GitHub secret. The token needs Workers Scripts and Workers
Routes write permissions, plus zone read access. Creating buckets and uploading
objects also requires R2 write access.
Uploading
Upload through the R2 dashboard or its S3-compatible API using bucket-scoped
credentials. Store objects without a files/ prefix. The public Worker accepts
only GET, HEAD, and CORS OPTIONS; upload access remains with authenticated
R2 clients. There is no bucket listing or directory index.
Set HTTP metadata when uploading:
Content-Type: the actual media type, such asvideo/mp4,image/png, orapplication/gzip. Missing types default toapplication/octet-stream.Content-Disposition: useattachment; filename="..."when a file should download rather than display inline.Cache-Control: defaults topublic, max-age=3600(one hour). Usepublic, max-age=31536000, immutablefor versioned/content-addressed files, and a short TTL orno-storefor mutable pointers such aslatest.json.
The Worker preserves the object's HTTP metadata, returns ETag and Last-Modified validators, supports conditional GET/HEAD, and streams single byte ranges for video seeking and download resumption, including If-Range. Unsupported or malformed range formats fall back to a full response. Public CORS allows these files to be consumed from other origins.
Release Binaries
The release workflow (.github/workflows/release.yml) uploads release files.
Its publish-cli and publish-desktop jobs run in the release environment,
check out only .github/release/, and never execute repository code:
publish-cli.sh and publish-desktop.sh upload verified, signed build artifacts
to the production bucket for every release channel:
R2 key: bin/0.0.0-dev-123/opencode-linux-x64.tar.gz
URL: https://opencode.ai/files/bin/0.0.0-dev-123/opencode-linux-x64.tar.gz
R2 key: bin/2.0.0/opencode-desktop-mac-arm64.dmg
URL: https://opencode.ai/files/bin/2.0.0/opencode-desktop-mac-arm64.dmg
CLI and Node CLI archives contain the standalone executable at their root.
Linux uses .tar.gz; macOS and Windows use .zip. Desktop uploads keep their
electron-builder filenames and include .app.tar.gz bundles.
The jobs derive S3 credentials from the CLOUDFLARE_API_TOKEN secret and set
the content type, download filename, and immutable cache headers.
They register cli, cli-node (dev builds), and desktop (beta and versioned
releases) with distribution opencode, each only after all of its files have
uploaded. Every metadata.files entry contains the direct CDN url, SHA-256
checksum, and byte size. Desktop metadata.manifests use those CDN URLs with
SHA-512 checksums computed from the final signed files.
Caching
Complete GET responses up to 512 MiB are cached through the Workers Cache API at the serving Cloudflare location, subject to object Cache-Control. Cached complete objects can also satisfy range requests. Cold range requests stream only the requested bytes from R2 and do not populate the cache. Larger objects stream from R2. This cache does not use Tiered Cache.
Query strings are retained in cache keys but do not change the R2 key. Cookies
and authorization do not affect public file contents. Replacing/deleting an R2
object does not purge an already-cached response: use versioned keys, wait for
the TTL, or purge the public URL through Cloudflare. Request Cache-Control: no-cache bypasses the cache for revalidation.
Verification
After uploading an object, check its full response, metadata, and byte ranges:
curl -I https://opencode.ai/files/videos/demo.mp4
curl --fail -H 'Range: bytes=0-31' -D - \
https://opencode.ai/files/videos/demo.mp4 -o /dev/null
Expect 200 for HEAD and 206 with Content-Range for the range request. A
missing key returns 404, unsupported methods return 405, and matching
If-None-Match/If-Modified-Since requests return 304.