| Age | Commit message (Collapse) | Author |
|
The page told the operator to mint with POST /get-test-token, which is
gn-auth's dev password grant and does not exist in the deployment being
set up. Issuing a second user's token also had no good answer: the stub
minted exactly one token per process, so a new user meant starting a
throwaway instance, grabbing --token-file and killing it.
So add --mint-only: mint for --subject, print, write --token-file and
exit without listening. Verification only depends on the key (committed,
so shared by every instance) and on iss, which now falls back to
AUTH_SERVER_URL -- a token minted this way verifies against the
long-running stub untouched.
* scripts/gn-auth-stub.py (main): --mint-only; --issuer defaults to
AUTH_SERVER_URL before the port-derived default. Docstring says how a
second user's token is issued.
* docs/www/ai-instruction.md (1. Get a token): the operator's command is
the stub, not /get-test-token; the token lasts whatever --expires-in said
rather than an hour, and there is no early revocation. (4. What is
public): minting is operator-side. Appendix: "Token issuance" rewritten
around the stub -- public-data-only, expiry as the only revocation, sub
as attribution, one token per user -- with gn-auth kept as what replaces
it later; the stub is no longer listed as a dev leftover to delete, but
as the auth server that must never be routed publicly.
* README.md (Tests): mention --mint-only next to the stub's other flags.
|
|
A stub token had a fixed 24h lifetime, and nothing said what an
ungranted one can reach. Both matter for the case that is not a test
run: handing a caller or an agent a long-lived token that may only read
public data. An ungranted stub token already is exactly that -- the
restriction lives in the authorisation lookup genenetwork4 consults for
anything not public, not in the token's scope -- so only the lifetime
was missing.
* scripts/gn-auth-stub.py (duration): new helper, "30d"/"24h"/"60m"/s.
(sign): take expires_in instead of a hard-coded 86400.
(main): --expires-in, defaulting to 24h as before; docstring says what
an ungranted token can reach and why the key's fixity makes a month-long
token safe.
* README.md (Tests): the public-data-only token, what stays visible
(dataset metadata), the --expires-in example and the iss/port coupling.
|
|
A second instance started while one is already listening used to write
its own token to --token-file (and print it) before failing to bind,
leaving a token nothing was serving in the place the first instance's
token was expected -- tests then failed as if the stub had issued it.
Bind first, then write and print.
* scripts/gn-auth-stub.py (main): construct the server before writing
--token-file or printing the token.
|
|
The suites skip rather than fail whatever needs a database, an index or
a token they cannot get, which is right in the build sandbox but leaves
most endpoints unexercised. With a live MariaDB, a full Xapian index
(GN4_FULL_INDEX) and a token, the same commands run the whole suite:
225 Rust tests pass with 0 failures, and the Python suite runs all 94
with no skips -- as against 77 skipped without them.
The token was the only piece with no local stand-in, so add
scripts/gn-auth-stub.py: it serves the JWKS matching the committed test
key, answers /auth/token (so POST /get-test-token works) and answers
both authorisation endpoints with "no privileges", the stranger case
the non-public 403 tests assert. --grant-resource-id grants one
resource id to the fixture token for the owner-side tests.
* scripts/gn-auth-stub.py: new file.
* README.md (Tests): "Against a real database and index", what the
three inputs are, the full run, and what still skips and why.
|