about summary refs log tree commit diff
diff options
context:
space:
mode:
-rw-r--r--RELEASE_NOTES.md100
1 files changed, 100 insertions, 0 deletions
diff --git a/RELEASE_NOTES.md b/RELEASE_NOTES.md
new file mode 100644
index 0000000..32b013a
--- /dev/null
+++ b/RELEASE_NOTES.md
@@ -0,0 +1,100 @@
+# Release Notes
+
+## Unreleased
+
+Found and fixed while standing up a from-scratch gn-auth instance (fresh
+SQLite database, migrations applied, no prior data) to test a new
+external client (GeneNetwork4) against the OAuth2 token and resource
+authorisation endpoints. The first three bugs below only show up on a
+fresh deployment with no existing JWK/users/sysadmins -- an
+already-running instance with data from before these fixes will not
+have hit them. The fourth is independent of deployment age -- any
+instance, old or new, hits it whenever a client sends a malformed or
+otherwise invalid bearer token to `/auth/resource/authorisation`.
+
+### Fixed
+
+- **`JWTBearerTokenValidator` could reject every token for hours after
+  a fresh start.** Its JWKS key set is cached in memory and only
+  refreshed every `jwt_refresh_frequency_hours` (default 6). On a
+  fresh deployment, no JWK exists until the first token is ever
+  issued (keys are generated lazily, see `jwks.py`), which can happen
+  *after* the validator was already constructed with an empty key
+  set. Every token issued before the next scheduled refresh then
+  failed to validate, even though a usable key existed on disk by the
+  time the token arrived. The validator now also refreshes
+  immediately whenever its cached key set is empty, not only when it
+  is stale. (This fix also corrects a second, smaller issue in the
+  same function: `_last_jwks_update` was never actually updated after
+  a refresh, so once the 6-hour window first elapsed, every later
+  token validation would re-read the JWKS directory from disk.)
+  -- `gn_auth/auth/authentication/oauth2/resource_server.py`
+
+- **`flask init-dev-clients` could not actually authenticate its own
+  dev client.** The command stored the dev client's secret
+  (`yadabadaboo`) as plain text, but client secret verification
+  (`OAuth2Client.check_client_secret`) always runs the stored value
+  through the Argon2 verifier, which raised `InvalidHashError` on a
+  non-hash string. Every grant type requiring `client_secret_post` --
+  i.e. all of them -- failed for the dev client. `create-oauth2-client`
+  already hashed its secret correctly; `init-dev-clients` now does the
+  same. -- `gn_auth/wsgi.py`
+
+- **Make `Resource.created_by` mandatory**; Originally, the fix
+  generated by Claude was to allow a `None` for the `created_by`
+  attribute. This was fundamentally wrong, though it was not Claude's
+  fault, since the typing was wrong too.
+  The fix was updating the typing to make `created_by` a mandatory
+  attribute, and adding some notes to the effect that the
+  "GeneNetwork System" resource (and any other existing resources at
+  the point of setup) should be manually updated to have the
+  system-administrator user as their creator.
+
+- **`POST /auth/resource/authorisation` crashed on an invalid bearer
+  token instead of degrading to the public-only view.** Its exception
+  handling only set the response for an authlib error of
+  `missing_authorization` (no `Authorization` header at all); a
+  present-but-invalid token raises `invalid_token` instead, which fell
+  through without setting the response variable, raising
+  `UnboundLocalError` (a 500). Any token-related failure now gets the
+  same anonymous/public-only fallback. -- `gn_auth/auth/authorisation/resources/views.py`
+
+### Verified by hand against a fresh instance
+
+- `POST /auth/token` (password grant) issues a JWT; `GET
+  /auth/public-jwks` lists the key once one has been generated.
+- `POST /auth/resource/populations/create` links an InbredSet group to
+  a new `inbredset-group` resource; `GET
+  /auth/resource/populations/resource-id/<speciesid>/<inbredsetid>`
+  resolves it back to the same `resource-id`.
+- `POST /auth/resource/authorisation` returns the full role/privilege
+  list for the resource's owner, and just `{"public-read": false}` for
+  an anonymous *or* invalid-token caller on a non-public resource --
+  the cases an external resource server (e.g. GeneNetwork4) needs to
+  distinguish from "authorized".
+- This is also exercised automatically: GeneNetwork4's
+  `tests/auth.rs` runs its full token-verification and
+  resource-authorization flow (including the caching layer) against a
+  real gn-auth instance.
+
+### Noted, not changed here
+
+- The `system-administrator` role's broad `system:*` privileges are
+  special-cased as globally applicable only inside gn-auth's own
+  `@authorised_p`-protected endpoints
+  (`gn_auth/auth/authorisation/checks.py`'s
+  `__system_privileges_in_roles__`). `POST /auth/resource/authorisation`
+  -- the endpoint an external resource server actually calls -- does
+  not use that special case; it reflects only `user_roles` rows tied
+  to the specific resource queried. A sysadmin therefore does not
+  automatically show up with privileges on a resource created after
+  migration `20250729_03` ran (which only granted the role on
+  resources that existed at the time). Worth a decision before any
+  external client relies on "sysadmin" meaning "full access
+  everywhere".
+  *fredm* (on review of this) notes that this code is an artifact of a
+  previous state of affairs that has long been resolved, and thus,
+  this code (especially the `__system_privileges_in_roles__` function)
+  can be removed. This is because at this point, everything (except
+  users), is considered and treated as some form (or other) of a
+  resource.