This skill builds on durable-streams/server-deployment. Read it first for server setup basics.

Run through each section before deploying to production.

Expected:

./durable-streams-server run --config Caddyfile

Fail condition: Importing DurableStreamTestServer from @durable-streams/server in production code.

Fix: Download the Caddy binary from GitHub releases and configure with a Caddyfile.

Expected:

durable_streams {
  data_dir ./data
  max_file_handles 200
}

Fail condition: No data_dir in Caddyfile — server uses in-memory storage and loses data on restart.

Fix: Add data_dir pointing to a persistent directory.

Expected:

const res = await stream({
  url: "https://your-server.com/v1/stream/my-stream",
})

Fail condition: Using http:// URLs in production. Pre-signed URLs and auth tokens are bearer credentials — HTTP exposes them in transit. HTTP/1.1 also limits browsers to ~6 concurrent connections per origin.

Fix: Configure TLS on the Caddy server (Caddy provides automatic HTTPS by default).

Expected:

const handle = await DurableStream.create({
  url: "https://server.com/v1/stream/my-stream",
  contentType: "application/json",
  headers: { "Stream-TTL": "86400" }, // 24 hours
})

Fail condition: Streams created without TTL persist forever, causing unbounded storage growth.

Fix: Set Stream-TTL (seconds) or Stream-Expires-At (ISO timestamp) on stream creation. Use exactly one, not both.

Expected:

headers: { "Stream-TTL": "86400" }
// OR
headers: { "Stream-Expires-At": "2026-04-01T00:00:00Z" }

Fail condition: Providing both Stream-TTL and Stream-Expires-At returns 400 Bad Request.

Fix: Use one or the other. TTL is relative (seconds from creation), Expires-At is absolute.

Expected:

Reads use offset-based URLs that are naturally cacheable:

GET /v1/stream/my-stream?offset=abc123

The server returns Cache-Control and ETag headers automatically for historical reads. CDNs can cache and collapse requests — 10,000 viewers at the same offset become one upstream request.

Fail condition: Overriding or stripping Cache-Control headers at the CDN/proxy layer.

Fix: Allow the server's Cache-Control and ETag headers to pass through to the CDN.

Stream-Cursor prevents CDN cache collisions when the same offset returns different data (e.g., after stream truncation). Ensure your CDN does not strip this header.

Fail condition: CDN strips Stream-Cursor from responses.

Fix: Configure CDN to pass through Stream-Cursor response header.

Expected:

const res = await stream({
  url,
  offset: "-1",
  live: true,
  onError: (error) => {
    if (error.status === 401) return // Stop retrying
    return {} // Retry transient errors
  },
})

Fail condition: No onError handler — permanent errors (401, 403) silently retry forever.

Fix: Add onError handler that stops retrying for non-transient errors.

Wrong:

const res = await stream({ url: "http://api.example.com/v1/stream/my-stream" })

Correct:

const res = await stream({ url: "https://api.example.com/v1/stream/my-stream" })

Pre-signed URLs and auth tokens are bearer credentials. HTTP exposes these in transit. Also, HTTP/1.1 limits browsers to ~6 concurrent connections per origin.

Source: packages/client/src/utils.ts warnIfUsingHttpInBrowser

Wrong:

const handle = await DurableStream.create({
  url: "https://server.com/v1/stream/my-stream",
  contentType: "application/json",
})

Correct:

const handle = await DurableStream.create({
  url: "https://server.com/v1/stream/my-stream",
  contentType: "application/json",
  headers: { "Stream-TTL": "86400" },
})

Without TTL, streams persist forever causing unbounded storage growth.

Source: PROTOCOL.md TTL and Expiry section

Wrong:

headers: {
  "Stream-TTL": "86400",
  "Stream-Expires-At": "2026-04-01T00:00:00Z",
}

Correct:

headers: {
  "Stream-TTL": "86400",  // OR Expires-At, not both
}

The protocol requires exactly one. Providing both returns 400 Bad Request.

Source: PROTOCOL.md TTL and Expiry section

This skill's patterns conflict with writing-data. autoClaim: true is convenient for serverless/ephemeral workers but sacrifices cross-restart coordination. Persistent long-running workers may benefit from explicit epoch management for proper multi-worker coordination.

See also: durable-streams/writing-data/SKILL.md § Common Mistakes

• Using Caddy production server (not dev server)
• data_dir configured for persistence
• HTTPS enabled
• TTL or Expires-At set on all streams
• Not using both TTL and Expires-At together
• CDN passes through Cache-Control, ETag, and Stream-Cursor headers
• onError handler configured for live streams
• max_file_handles tuned for expected stream count

• server-deployment — Initial server setup
• writing-data — Producer configuration for production

Targets durable-streams v0.2.1.