Provider Configuration

Terraform Provider Configuration and Authentication

How a provider accepts connection settings and resolves credentials. Poor authentication UX is the first thing every user of a provider hits; a well-designed credential provider chain is what separates a production-grade provider from a demo. The examples use a fictional examplecloud provider and the Plugin Framework.

References (load when needed):

• references/credential-chain.md — complete, compilable credential chain implementation (providers, chain, file profiles, Configure wiring, tests)
• references/case-studies.md — how the AWS provider (aws-sdk-go-base) and smaller providers structure real credential chains

Provider Schema for Authentication

Every authentication attribute must be Optional, never Required — a Required attribute forces users to put credentials in configuration and makes environment-variable and credentials-file resolution impossible. Mark secrets Sensitive so Terraform redacts them in plan output, and state the environment-variable fallback in each description so tfplugindocs publishes the resolution rules.

func (p *examplecloudProvider) Schema(ctx context.Context, req provider.SchemaRequest, resp *provider.SchemaResponse) {
    resp.Schema = schema.Schema{
        Attributes: map[string]schema.Attribute{
            "endpoint": schema.StringAttribute{
                Optional:            true,
                MarkdownDescription: "API endpoint. May also be set via the `EXAMPLECLOUD_ENDPOINT` environment variable.",
            },
            "api_key": schema.StringAttribute{
                Optional:            true,
                MarkdownDescription: "API key. May also be set via the `EXAMPLECLOUD_API_KEY` environment variable, or in a shared credentials file.",
            },
            "api_secret": schema.StringAttribute{
                Optional:            true,
                Sensitive:           true,
                MarkdownDescription: "API secret. May also be set via the `EXAMPLECLOUD_API_SECRET` environment variable, or in a shared credentials file.",
            },
            "profile": schema.StringAttribute{
                Optional:            true,
                MarkdownDescription: "Named profile in the shared credentials file. May also be set via the `EXAMPLECLOUD_PROFILE` environment variable. Defaults to `default`.",
            },
            "skip_credentials_validation": schema.BoolAttribute{
                Optional:            true,
                MarkdownDescription: "Skip the identity check normally performed during provider configuration.",
            },
        },
    }
}

Never add a Default to a credential attribute, and never hardcode a credential anywhere in the provider. Defaults belong in the resolution logic (where environment variables and files can override them), not in the schema.

The Credential Provider Chain

Resolve credentials by consulting an ordered list of sources and taking the first one that produces a complete set. This is the pattern the AWS provider uses via `aws-sdk-go-base`, and it generalizes to any provider. The canonical precedence, highest first:

1. Static configuration — values set directly in the provider block. Explicit always wins.
2. Environment variables — EXAMPLECLOUD_API_KEY, etc. The CI-friendly path.
3. Shared credentials file — named profiles in ~/.examplecloud/credentials, for humans with multiple accounts.
4. Platform identity — instance metadata, workload identity, or OIDC token exchange, where the platform offers it. Credentials nobody has to store.

Two rules make the chain predictable:

• Resolve secrets as a set, not field-by-field. If the environment supplies an API key but no secret, that source offers nothing — fall through to the next source for both values. Mixing an env-var key with a file-profile secret produces authentication failures that are nearly impossible for users to debug.
• Resolve non-secret connection settings field-by-field. endpoint, profile, or insecure can each independently follow config > env > file > default, because a mismatch there is visible and harmless.

The core abstraction is a single-method interface with a sentinel error that distinguishes "this source has nothing to offer" (fall through) from "this source is misconfigured" (surface it):

// ErrNoCredentials signals a source had nothing to offer. The chain falls
// through to the next source. Any other error means the source was
// configured but unusable (e.g. malformed credentials file) and is
// preserved so the final diagnostics can surface it.
var ErrNoCredentials = errors.New("no credentials found")

type Credentials struct {
    APIKey    string
    APISecret string
    Source    string // which provider supplied them, for logging
}

func (c Credentials) Complete() bool {
    return c.APIKey != "" && c.APISecret != ""
}

type Provider interface {
    Retrieve(ctx context.Context) (Credentials, error)
    Name() string
}

A Chain (itself a Provider, so chains compose) walks the providers in order and returns the first complete set of credentials. Every skipped source is recorded into an aggregate ChainError whose Error() lists each source with the reason it was skipped, and whose Is method makes errors.Is(err, ErrNoCredentials) true only when every source fell through cleanly — so Configure can tell "nothing supplied" from "something supplied but broken" with one check. The full implementation — the chain loop, the static, environment, and file providers, and the NewDefaultChain constructor that owns the canonical order — lives in references/credential-chain.md.

Wiring the Chain into Configure

Configure runs once per Terraform operation, before any resource CRUD. The shape:

func (p *examplecloudProvider) Configure(ctx context.Context, req provider.ConfigureRequest, resp *provider.ConfigureResponse) {
    var config examplecloudProviderModel
    resp.Diagnostics.Append(req.Config.Get(ctx, &config)...)
    if resp.Diagnostics.HasError() {
        return
    }

    // 1. Guard against unknown values (e.g. api_key = some_resource.output).
    if config.APIKey.IsUnknown() {
        resp.Diagnostics.AddAttributeError(
            path.Root("api_key"),
            "Unknown API Key",
            "The provider cannot connect because api_key depends on a value known only after apply. "+
                "Set a static value, or use the EXAMPLECLOUD_API_KEY environment variable.",
        )
    }
    // ... repeat for each auth attribute, then:
    if resp.Diagnostics.HasError() {
        return
    }

    // 2. Resolve credentials through the chain.
    chain := credentials.NewDefaultChain(
        config.APIKey.ValueString(),
        config.APISecret.ValueString(),
        credentials.Options{Profile: config.Profile.ValueString()},
    )
    creds, err := chain.Retrieve(ctx)
    if err != nil {
        if errors.Is(err, credentials.ErrNoCredentials) {
            resp.Diagnostics.AddError(
                "No Valid Credential Sources Found",
                "No examplecloud credentials were found. Sources tried, in order:\n\n"+err.Error()+
                    "\n\nSet api_key and api_secret in the provider block, export "+
                    "EXAMPLECLOUD_API_KEY and EXAMPLECLOUD_API_SECRET, or add a profile to "+
                    "~/.examplecloud/credentials. See https://example.com/docs/auth.",
            )
        } else {
            resp.Diagnostics.AddError("Failed to Resolve Credentials", err.Error())
        }
        return
    }
    tflog.Debug(ctx, "resolved credentials", map[string]any{"source": creds.Source})

    // 3. Build the client once; share it with every resource and data source.