Retry transient failures when downloading artifacts

[dan-manges]

rwx artifacts download made a single unretried request to storage, so a connection reset partway through surfaced as a hard failure:

Error: unable to download artifact: HTTP request failed: Get "https://mint-storage-...s3.us-east-2.amazonaws.com/org...
read tcp 192.168.86.51:53581->3.5.89.33:443: read: connection reset by peer

DownloadArtifact now retries with bounded exponential backoff, following the same shape as ListRuns: a downloadArtifactOnce helper returns a retryable flag, and the caller loops on retry.NewBackoff() (5 attempts, 1s→5s, ~12s worst case).

Retried:

• transient transport errors via retry.IsTransient — resets, refused, timeouts, EOF
• transient failures reading the response body, so a reset partway through a large artifact restarts instead of failing
• 429 and 5xx from storage (S3 SlowDown)

Not retried: 400/403/404, so an expired presigned URL or missing key still surfaces immediately. The same presigned URL is reused across the short retry window.

Exhausted retries are wrapped with errors.ErrNetworkTransient, matching retry.RoundTripper, so telemetry buckets them as network_transient_error rather than unknown.

This also covers the parallel multi-artifact download path in service_artifacts.go, which goes through the same client method. Retries are silent under the existing "Downloading artifact..." spinner.

Still buffered, no resume

Artifacts are read fully into memory and a retry restarts from byte zero — there's no Range-based resume. A reset late in a multi-GB download still re-transfers the whole thing.

.rwx/sandbox.yml

Added, since the repo had no sandbox config. It mirrors ci.yml (code → tool-versions → go/go-deps, plus the lsp-* chain so internal/lsp's embedded bundle exists) with a sandbox task running rwx-sandbox.

Verification

• rwx sandbox exec -- go test ./internal/... ./cmd/... — all packages pass
• golangci-lint run ./internal/api/... — 0 issues
• New TestAPIClient_DownloadArtifact_Retries covers reset-then-success, exhaustion after 5 attempts with the transient sentinel, 503-then-success, and 403 not retried. Resets are simulated by hijacking the connection with SetLinger(0).

🤖 Generated with Claude Code