auth-go — shareable OAuth 2.0 client library for CLIs

Provider-agnostic Go library for CLIs that authenticate end-users via OAuth 2.0 device flow (RFC 8628) or loopback authorization code + PKCE (RFC 8252), present resource-scoped bearer tokens to data APIs, and (when the auth host and data API live on different origins) exchange tokens via RFC 8693 STS.

No global state, no env-var reads, no implicit URLs. Every endpoint, identifier, and default value is supplied by the embedding CLI through a Config struct.

go get github.com/entireio/auth-go@latest

Subpackages

The internal/oauthhttp package holds shared HTTP body-reading + JSON-decoding helpers (detects HTML responses from captive portals / proxy intercepts and surfaces them as actionable errors instead of unmarshal failures). It is unexported in the Go sense — not importable by other modules — and not part of the public API surface.

Security

Defense-in-depth checks layered on top of server-side validation:

• HTTPS required. Both sts.Client and deviceflow.Client reject http:// BaseURLs unless AllowInsecureHTTP is set, and that opt-in is still limited to loopback (localhost / 127.0.0.1 / ::1) so production misconfigurations fail loudly.
• alg:none JWTs rejected. tokens.ParseClaims decodes the JWT header and refuses the unsigned shape (any case variant of none). Even though claim use is routing-only, this keeps an obvious attack surface closed.
• verification_uri validated. The device-code response field is what your CLI echoes and opens in the user's browser — a malicious AS pointing it at a phishing page would be a credential-harvesting vector. The library rejects non-https (loopback http excepted), embedded user:pass@host userinfo, and control characters in the URI.
• Resource origins validated. tokenmanager.Token requires TokenRequest.Resource to be an origin URL: absolute scheme + host, HTTPS unless loopback HTTP is explicitly enabled, and no userinfo/path/query/fragment. This prevents cache fragmentation and accidental STS requests for surprising resource strings.
• OAuth responses are bounded and strict. Success and error bodies are capped at MaxResponseBytes; oversized responses are rejected, HTML/captive-portal responses get actionable errors, and JSON success bodies with trailing data are refused.
• STS wire shape is explicit. tokenmanager defaults subject_token_type to the RFC 8693 access-token URN and exposes Config.SubjectTokenType for STS endpoints that require the structural JWT URN.

Quick start

import (
    "github.com/entireio/auth-go/deviceflow"
    "github.com/entireio/auth-go/sts"
    "github.com/entireio/auth-go/tokenmanager"
    "github.com/entireio/auth-go/tokenstore"
)

const (
    issuer   = "https://auth.example.com"  // auth host base URL
    clientID = "my-cli"                    // public OAuth client_id
)

store := tokenstore.NewKeyring("my-cli")  // service name = your CLI's name

// One Manager per CLI process. Construct from your CLI's identity.
mgr, err := tokenmanager.New(tokenmanager.Config{
    Issuer:   issuer,
    ClientID: clientID,
    STSPath:  "/oauth/token",  // RFC 8693 endpoint; usually the OAuth token endpoint
    Store:    store,
    Scope:    "cli",
    // Optional: defaults to sts.SubjectTokenTypeAccessToken. If your STS
    // requires the structural JWT URN instead, uncomment:
    // SubjectTokenType: sts.SubjectTokenTypeJWT,
})
if err != nil { /* misconfiguration */ }

Login

dfc, err := deviceflow.New(&deviceflow.Client{
    BaseURL:        issuer,
    ClientID:       clientID,
    Scope:          "cli",
    DeviceCodePath: "/oauth/device/code",
    TokenPath:      "/oauth/token",
})
if err != nil { /* required field missing */ }

dc, err := dfc.StartDeviceAuth(ctx)
// show dc.UserCode + dc.VerificationURI to user, then drive the poll loop:
ts, err := dfc.PollUntil(ctx, dc)
if err != nil { /* surface RFC 8628 §3.5 sentinel as needed */ }
if ts == nil { /* defensive: PollUntil's contract is non-nil on nil err */ }

if err := mgr.SaveCoreToken(*ts); err != nil { /* keyring failed */ }

deviceflow.New (and sts.New) validate required fields at construction time rather than at the first request — misconfiguration becomes a startup error rather than a confusing AS-side rejection mid-flow. Field-bag &deviceflow.Client{...} construction still works for callers that prefer it.

SaveCoreToken takes the full tokens.TokenSet so RefreshToken, absolute ExpiresAt, and Scope survive the round-trip through the keyring — earlier versions silently dropped these.

PollUntil is the helper most embedders want. It honours dc.Interval, applies the RFC 8628 §3.5 +5s bump on slow_down, stops at the dc.ExpiresIn ceiling, and returns terminal sentinels (ErrAccessDenied, ErrExpiredToken, ErrInvalidGrant) unwrapped so callers can errors.Is. Use PollDeviceAuth directly only when you need to render per-tick state in your own UI.

Calling a data API

bearer, err := mgr.TokenForResource(ctx, "https://api.example.com")
if errors.Is(err, tokenmanager.ErrNotLoggedIn) {
    // prompt user to run `mycli login`
}
// bearer is valid for https://api.example.com
req.Header.Set("Authorization", "Bearer "+bearer)

The manager picks the right strategy automatically:

• Same-host (Issuer == resource): hands back the core token verbatim.
• JWT-aud-includes shortcut: same, when the core token's audience already covers the resource (e.g. multi-audience tokens).
• Otherwise: runs an RFC 8693 exchange against Issuer + STSPath, caches the exchanged token by (core, resource, audience, requested_token_type, scope) until expiry.

By default, tokenmanager sends subject_token_type=urn:ietf:params:oauth:token-type:access_token for the stored core bearer. This is the role-based RFC 8693 token type and works for STS endpoints that treat the device-flow result as an OAuth access token even when its wire format is JWT. If your STS requires the structural JWT token type, configure SubjectTokenType: sts.SubjectTokenTypeJWT.

Logout

if err := mgr.DeleteCoreToken(); err != nil { /* keyring failed */ }

Deletes the keyring entry first; only clears the in-memory exchange cache on success, so a failed delete doesn't leave the CLI thinking it's logged out while the keyring still holds the token.

Design principles

• No globals, no env-var reads, no implicit URLs. Everything ships through Config. The library should compile and run identically inside any CLI.
• Provider-agnostic. deviceflow.Client and sts.Client are field-bag structs; neither knows about your provider's endpoint paths or token-type URIs. Pass them in.
• Bearer-presenter, not bearer-validator. This library is for CLIs that receive tokens from an auth server and present them to a resource server. JWT signature verification is intentionally not done — the resource server validates. tokens.ParseClaims is documented as unverified and used only for routing decisions.
• Per-CLI keyring isolation. Each CLI passes a unique service name to tokenstore.NewKeyring. OS keyrings key by (service, account), so consumers naturally get separate credential stores.
• Caller controls the wire shape. Default values for RFC 8693 fields are explicit and overridable. tokenmanager defaults requested_token_type to the standard access-token URN and subject_token_type to access_token; callers can set RequestedTokenType, SubjectTokenType, scope, audience, resource, and package-specific Extra fields as needed.

Embedding checklist

1. Pick a stable service name for tokenstore.NewKeyring(...). Don't change it later — renaming orphans every existing user's stored credentials.
2. Pick a client_id that the auth server recognises.
3. Decide your STSPath: typically the OAuth token endpoint per RFC 8693 convention, or a dedicated path if your auth server exposes one.
4. Decide whether the STS expects subject_token_type=access_token (the tokenmanager default) or jwt (SubjectTokenType: sts.SubjectTokenTypeJWT).
5. Construct the tokenmanager.Manager once at startup; pass it to your data-API call sites.
6. For multi-environment users (regions, staging), key the keyring by issuer URL — Manager.Issuer() returns the configured value.

Security

The library is the client side of OAuth 2.0 device flow + RFC 8693 token exchange. It receives bearers from an authorization server and presents them to data APIs. The threat model is:

• The auth server is trusted. This library does not verify JWT signatures — that's the data API's job. tokens.ParseClaims is documented as unverified and is used for routing decisions only.
• The data API is trusted. The library hands it a bearer; what the API does with it is out of scope.
• The transport is trusted only when TLS-protected. Both deviceflow.Client and sts.Client reject http:// BaseURLs unless AllowInsecureHTTP is set, and even then only loopback hosts are accepted. The Transport field is for observability and proxies, not for disabling TLS verification.
• OS keyrings are trusted within the user's session. The tokenstore.Keyring impl keys credentials by (service, account); pick a stable, unique service name per CLI and treat anything the keyring returns as untrusted bytes (the impl already rejects JSON-shaped junk and empty access tokens).

Defenses in depth that the library applies regardless:

• Rejects alg:none JWTs and any alg containing non-alphanumeric characters (closes whitespace / zero-width-space bypass attempts).
• Rejects absolute Path / DeviceCodePath / TokenPath values (defeats redirect-via-config attacks against url.ResolveReference).
• Validates TokenRequest.Resource as an origin URL and normalises it for same-host / JWT-audience shortcut comparisons and cache keys.
• Normalises tokenmanager.Config.Issuer so cosmetic differences (trailing slash, host case, default port) can't split keyring state across two effective issuers.
• Sanitises server-supplied error_description and non-JSON error body text (strips control chars, caps length) before wrapping into Go errors — so a hostile AS can't paint terminals or balloon logs.
• Rejects oversized OAuth responses and JSON success responses with trailing data after the first JSON value.
• Defaults STS subject_token_type to access_token, with Config.SubjectTokenType available when an STS expects jwt.
• Refuses to cache exchanged tokens with non-positive expires_in (forces a fresh exchange instead of treating unknown-lifetime bearers as "valid forever").
• Caps cache entries' lifetime at 1h when ExpiresAt is unset.

If you find a security issue, please email the maintainers privately rather than opening a public issue — coordinated disclosure gives a window to ship a fix before the report becomes searchable.

Non-goals

• OIDC discovery / ID tokens. This library is OAuth 2.0 only. If you need OIDC /.well-known/openid-configuration + ID-token verification, layer coreos/go-oidc on top.
• Server-side OIDC. If you're building an issuer, look at zitadel/oidc's op package.

Status

Used in production by entireio/cli. Issues and PRs welcome.

Releasing

Every user-visible change must land with a CHANGELOG.md entry in the same PR, under the top ## Unreleased heading. Tagging a release is then a two-step move:

1. Rename ## Unreleased to ## vX.Y.Z — YYYY-MM-DD, commit, and merge through a regular PR.
2. Tag that merge commit with git tag -a vX.Y.Z -m "..." and git push --tags.

Each ## Unreleased bullet must describe behaviour that changed since the last tag — not since some earlier point. When a CHANGELOG entry survives a release without being moved under a version header, the next release inherits stale claims (e.g. v0.3.4's draft initially still listed v0.3.2's :access_token default as "new").

Version bumps follow 0.x SemVer with one local convention: while pre-1.0, breaking changes may ship as patch bumps when the surface is narrow and well-documented (v0.3.2 changed the default subject_token_type as a patch).

License

MIT — see LICENSE.