v0.1.0

Added

• authlm doctor command for diagnosing Python environment, store selection, metadata file permissions, and credential fingerprint integrity.
• authlm list now supports --json and --format json flags to output credentials in structured JSON format.
• authlm.api now re-exports validate.
• authlm env POSIX shell output format (_format_shell) now includes the export keyword (e.g. export OPENAI_API_KEY='...').
• validate() now accepts an optional metadata_store parameter and persists last_validated_at upon a successful probe.
• authlm connect google --method oauth_browser now prints a pre-flight warning
  when the default Google OAuth client ID is in use, explaining that Google
  requires a user-created Cloud project with the Generative Language API enabled
  and the generative-language.retriever scope registered. Points to
  AUTHLM_GOOGLE_CLIENT_ID and Google's OAuth quickstart docs.
• OAuthConfig now exposes device_code_content_type (default
  application/x-www-form-urlencoded, set to application/json for OpenAI)
  to support providers that expect JSON device-code requests.
• is_default_client_id(provider_id, client_id) helper in _auth_table
  returns whether a client ID matches the hardcoded default for a provider.
• Project scaffolding: pyproject.toml with hatchling build, ruff, mypy (strict),
  and pytest configuration.
• authlm package skeleton with py.typed marker and dynamic version sourced from
  src/authlm/_version.py.
• Core exception hierarchy in authlm.errors (AuthLMError and seven subclasses).
• Core credential types in authlm.credentials (Pydantic models: ApiKeyCredential,
  OAuthCredential) with a discriminated CredentialUnion over the v0.1.0 types.
• parse_credential() deserialization helper in authlm.credentials (uses the
  discriminated union to restore the correct subclass from JSON).
• compute_fingerprint(secret) in authlm.credentials returning a truncated
  SHA-256 digest for non-secret change detection.
  (Additional types AwsCredential and AzureAdCredential are defined in the spec
  for v0.2.0 but are not implemented in this release.)
• Test infrastructure: tests/conftest.py with environment isolation fixtures and a
  smoke test.
• MemoryStore in authlm.stores.memory_store: in-process credential store for
  tests, cleared on process exit.
• EnvStore in authlm.stores.env_store: read-only store that reads API keys
  from environment variables (module-level _ENV_VAR_MAP covers v0.1.0
  providers: openai, anthropic, google, openrouter).
• KeyringStore in authlm.stores.keyring_store: OS keychain-backed credential store
  via the keyring library, with a JSON index file for enumeration (keyring has no
  enumeration API).
• EncryptedFileStore in authlm.stores.encrypted_file_store: Fernet-encrypted
  credential file store, with PBKDF2-HMAC key derivation from a passphrase.
• get_default_store in authlm.stores: auto-selects a CredentialStore from the
  AUTHLM_STORE env var (one of keyring, encrypted_file, env, memory),
  otherwise picks KeyringStore when a real keyring backend is available, and
  falls back to EnvStore with a warning when no keyring is present. Honors the
  AUTHLM_USER_PATH env var for the keyring index and encrypted file locations.
• Provider and ConnectionMethod Protocols in authlm.providers.base, plus the
  OAuthGrant StrEnum enumerating supported OAuth flow types
  (authorization_code_pkce, device_code). Both Protocols are
  @runtime_checkable for isinstance() tests; no implementations are included
  in v0.1.0 yet.
• authlm._auth_table with v0.1.0 provider auth metadata: OAuthConfig and
  AuthTableEntry Pydantic models, an AUTH_TABLE covering openai,
  anthropic, google, and openrouter (with public OAuth client IDs and
  PKCE/device endpoints for the three OAuth-capable providers), and
  get_auth_entry / get_oauth_config lookups. Client IDs are overridable
  per provider via AUTHLM_{OPENAI,ANTHROPIC,GOOGLE}_CLIENT_ID env vars.
• authlm.validation.validate() async probe that GETs each provider's
  validation_url from _auth_table (OpenAI /v1/models, Anthropic
  /v1/models with x-api-key + anthropic-version headers, etc.) and
  returns True on 2xx, False on 401/404, raises AccessDenied on 403
  entitlement denial, and TokenEndpointError on other 4xx. Refuses warned
  subscription methods (claude_pro_oauth_browser,
  claude_pro_oauth_device) unless force=True is passed, and is
  no-op-detectable for providers without a validation_url.
• Connection methods: APIKeyMethod, OAuthPKCEMethod (with loopback HTTP
  server), OAuthDeviceCodeMethod (with polling). All implement the
  ConnectionMethod Protocol.
• 4 built-in providers: OpenAIProvider, AnthropicProvider (with
  warned Claude Pro browser/headless methods),
  GoogleProvider, OpenRouterProvider.
• providers.registry with list_providers, get_provider, get_method.
• Public async API in authlm.api: get_credential,
  get_valid_credential, refresh (handles refresh-token rotation
  and classifies errors per spec §5.3), should_refresh, connect
  (orchestrates method + store + metadata), and validate.
• 5-command CLI in authlm.cli (Click group, entry point authlm.cli:cli):
  connect (interactive method picker with [y/N] warning confirmation for
  warned methods, refuses non-TTY without --method), list (ASCII table of
  stored credentials with backend and last-validated columns), status
  (per-credential metadata block; --validate probes the credential, --force
  allows probing warned methods, --all iterates aliases), disconnect
  (confirmation prompt; --yes to skip), env (exports credential as shell
  env vars in shell / docker / github formats; eval "$(authlm env openai)"
  for shell). The CLI is a thin sync wrapper: each command bridges to the
  async authlm.api functions via asyncio.run() per spec §2.3.
• Per-subcommand --store option mirrors the AUTHLM_STORE env
  var (one of keyring, encrypted_file, env, memory); --metadata-path
  mirrors AUTHLM_METADATA_PATH. Tests use --store=memory for isolation;
  users can use it to override the store on a per-invocation basis. The
  options live on each subcommand rather than the group, to avoid Click
  option-precedence ambiguity.
• authlm.api.connect() now accepts optional on_prompt and open_browser
  keyword-only parameters. When passed, they are propagated to the
  OAuthDeviceCodeMethod / OAuthPKCEMethod via new with_on_prompt() and
  with_open_browser() methods (mirroring the existing
  APIKeyMethod.with_secret_prompt()). Backward compatible: existing
  callers that omit these params see no change.
• authlm.cli is a subpackage (mirrors connection_methods/ and
  providers/); authlm.cli._context provides get_metadata_path()
  (metadata path resolution with chain: explicit → AUTHLM_METADATA_PATH
  → AUTHLM_USER_PATH → get_user_data_path()); authlm.cli._formatters
  provides format_list_table() and format_status_table().
• Dependabot bumps: anthropic 0.112.0 → 0.113.0 (PR #6), astral-sh/setup-uv
  SHA pin updated (PR #5).

Changed

• OAuthPKCEMethod and OAuthDeviceCodeMethod now log initial prompt/browser URLs at DEBUG level instead of INFO.
• README now documents that OpenAI OAuth methods produce Codex-scoped tokens targeting
  the Codex backend, not the standard OpenAI API.
• Expanded README with status/build/license badges, a "Why AuthLM?" motivation
  section, a feature comparison table (vs llm keys, LiteLLM, provider SDKs),
  installation instructions (from source, since not yet on PyPI), CLI usage
  examples, a credential-stores reference table, and a security section
  linking to SECURITY.md. Corrected the broken uv sync --extra test --all-extras dev command to uv sync --all-extras.
• connect (CLI) refuses to run with no --method on a non-TTY stdin and
  prints a clear error message; this avoids hangs in CI / scripts and matches
  the spec's "explicit over implicit" principle.
• The CLI sets logging.getLogger("authlm").setLevel(logging.WARNING) at
  startup so INFO-level logs from the OAuth methods (e.g. PKCE "Opening
  browser" info) do not pollute eval $(authlm env ...) stdout. The
  device-code URL and user code are routed to stderr via a custom on_prompt
  callback.
• cli group now uses invoke_without_command=True so authlm (no
  subcommand) prints help text and exits 0 instead of Click's default
  MissingCommand (exit 2). Implemented in commit eab5f1f.
• Updated v0.1.0 design spec (.agents/specs/v0.1.0-authlm.md) to reflect
  v0.1.0 reality: plugin system and models.dev integration removed from
  v0.1.0 scope (deferred to v0.2.0), authlm status --backend and
  authlm.set_store() added to v0.1.0 scope, compute_fingerprint wired
  into MetadataEntry for change detection, Field(repr=False) on all
  secret fields, ProviderNotAvailable/AliasCollisionError deferred to
  v0.2.0, ConnectionMethod.validate() removed from Protocol (validation
  goes through validation.validate() directly), roadmap replaced with
  4-version plan (v0.2.0 → v1.0.0).
• Created version spec outlines: .agents/specs/v0.2.0-authlm.md
  (Extensibility & Ecosystem), .agents/specs/v0.3.0-authlm.md
  (Robustness & More Stores), .agents/specs/v1.0.0-authlm.md
  (Stable Release).
• Updated AGENTS.md to reflect v0.1.0 reality (first-party providers
  only, no plugin system, set_store() in stores, --backend in status
  command, Field(repr=False) on secrets, version spec references).
• Updated README.md with honest v0.1.0 scope, roadmap section, and
  removed plugin system / models.dev from the features list.

Fixed

• authlm status --validate now passes metadata_store to validate(), ensuring metadata is created when missing.
• OAuthPKCEMethod loopback server now returns HTTP 404 for auxiliary paths (such as /favicon.ico) without setting an error state.
• redact_body() in _oauth_helpers.py now recursively redacts secrets within nested JSON lists and top-level arrays.
• OAuthPKCEMethod loopback server error propagation now unblocks main thread event-loop waiters immediately on user denial or state mismatch, and sets allow_reuse_address only on POSIX systems (sys.platform != "win32").
• EncryptedFileStore now caches PBKDF2 Fernet key derivation to improve performance on repeated store operations.
• KeyringStore methods get(), set(), and delete() now catch generic Exception backend errors (e.g. Linux D-Bus / SecretStorage failures) and wrap them in SecretStoreError, and _index_write() now writes atomically via a temporary file.
• Windows DACL user SID resolution in _restrict_permissions now resolves process token user SID for domain/cloud-joined accounts.
• Provider connection_methods() (OpenAIProvider, AnthropicProvider, GoogleProvider) now lazy-manage httpx.AsyncClient instances instead of eagerly instantiating them, eliminating socket handle leaks.
• EnvStore.get() now returns None when alias != "default" instead of raising CredentialNotFound, aligning it with the CredentialStore protocol contract.
• connect CLI error messages correctly distinguish between unknown methods and warned methods, listing available method IDs when an unknown one is provided.
• validate() now skips probing for OpenAI's chatgpt_oauth_browser and chatgpt_oauth_device methods by raising AuthLMError, avoiding inevitable 403 errors since those tokens do not have entitlement for the standard validation endpoint.
• status CLI now explicitly notes when fingerprint tamper detection is skipped due to a missing metadata entry, instead of failing silently.
• chatgpt_oauth_device now implements the official two-stage Codex device-code flow, which is not RFC 8628 compliant. It correctly polls for an authorization-code/PKCE tuple using device_auth_id rather than polling the token endpoint directly.
• chatgpt_oauth_browser now pins the OAuth redirect URI to http://localhost:1455/auth/callback to match the single registered URI for the Codex public client ID. Previously, it sent http://127.0.0.1:1455/callback which was rejected by OpenAI.
• validate() now sends Google API keys as ?key= query parameters instead of
  Authorization: Bearer headers. Google's Generative Language API rejects Bearer
  auth for API keys with 401; the key works correctly when sent as a query parameter.
• OpenAI OAuth PKCE authorize URL now includes codex_cli_simplified_flow,
  originator, and id_token_add_organizations parameters. Without these,
  OpenAI's authorize endpoint returned "Invalid authorize request" before the
  sign-in screen appeared.
• OpenAI OAuth device-code flow now sends a JSON body to the device-code
  endpoint instead of form-encoded. OpenAI's /api/accounts/deviceauth/usercode
  endpoint returns 400 for form-encoded bodies with "Input should be a valid
  dictionary or object to extract fields from".
• EncryptedFileStore file permissions now work on Windows: replaced the
  POSIX-only os.chmod(0o600) with a platform-aware _restrict_permissions
  that uses pywin32 (SetNamedSecurityInfo with
  PROTECTED_DACL_SECURITY_INFORMATION) on Windows to strip inherited ACLs
  and grant Read+Write only to the current user. pywin32>=307 is a
  platform-marked core dependency (sys_platform == 'win32').
• Keyring store and encrypted file store now wrap raw backend errors
  (keyring.errors.*, file I/O) in SecretStoreError so consumers can
  catch all secret-store failures with a single AuthLMError subclass.
• get_default_store now emits a WARNING-level log when the encrypted
  file passphrase is sourced from AUTHLM_PASSPHRASE (visible to child
  processes and /proc/<pid>/environ on Linux); prefer an interactive
  prompt.
• validation.validate() now runs the 4xx TokenEndpointError body
  through redact_body in authlm.connection_methods._oauth_helpers
  before including it in the exception message. Bearer <token> substrings
  and JSON-string fields named access_token / refresh_token /
  id_token / client_secret are replaced with [REDACTED], and the
  body is truncated to 200 chars. Prevents credential leakage when OAuth
  providers echo tokens in error bodies (e.g. invalid_token: <token>).
  The state (CSRF) parameter is intentionally not redacted — it is
  not a credential and is useful for debugging.

Security

• Fixed a bug where redact_body() failed to redact secrets inside nested JSON lists and top-level arrays.
• Added "id_token", "token", and "key" to JSON body redaction keys (_REDACT_DICT_KEYS) in _oauth_helpers.py.
• Added "key", "api_key", "secret", "access_token", "refresh_token", "id_token", "client_secret" to the URL query-parameter redaction set (_REDACTED_PARAMS) in _oauth_helpers.py so sensitive parameters passed in query strings are redacted in log output and exception messages.

Removed

• authlm.plugins — Plugin loader (deferred to v0.2.0)
• authlm.hookspecs — pluggy hookspecs (deferred to v0.2.0)
• models_dev module — models.dev integration (deferred to v0.2.0)