- Gate browser session creation on a successful refresh token. When the
IdP returns no refresh token, SessionAuthBackend would silently reject
every subsequent request and bounce the user back to /oauth/login in a
loop. The callback now bails with a 400 + correlation ID + actionable
hint about offline_access *before* writing browser_sessions or setting
the cookie. Pinned by a new end-to-end unit test.
- Evict orphaned browser_sessions rows in SessionAuthBackend when the
associated refresh token is gone, instead of letting them accumulate
until TTL cleanup. Best-effort; deletion errors stay non-fatal.
- Demote identity-bearing logs in the Flow 2 OAuth callback (user_id,
scopes, audience, expires_at) from INFO to DEBUG so they don't leak
into multi-tenant log aggregation on every provision.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three findings from the latest review on #758 (1 medium, 2 low):
Medium:
- browser_oauth_routes.oauth_logout: move delete_browser_session into a
finally block so an error from delete_refresh_token can no longer leave
an orphan browser_sessions row. The orphan was not exploitable
(SessionAuthBackend rejects sessions without a live refresh token), but
it lingered until the hourly cleanup cron — a correctness gap. New
regression test pins the fix.
Low:
- oauth_callback_nextcloud: drop redundant ``or None`` from
``expected_nonce=nonce``. ``nonce`` is already ``str | None`` and
``secrets.token_urlsafe`` never produces an empty string, so the
coercion was a no-op that could mislead future readers into thinking
empty-string was a valid skip-the-check path.
- storage.RefreshTokenStorage.initialize: fail fast at startup when
SQLite < 3.35, since ``DELETE ... RETURNING`` (used in
``delete_browser_session``) needs that minimum. Ubuntu 20.04 ships
3.31 and would otherwise hit OperationalError on every logout.
Prerequisite also documented in docs/installation.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Flow 2 (oauth_authorize_nextcloud) now generates a nonce, stores it on
the oauth_session row, forwards it to the IdP, and verifies it via
expected_nonce in oauth_callback_nextcloud — closes the last replay-
protection gap (round-3 finding 1).
- _origin_matches_self fails closed when mcp_server_url is missing
instead of allowing the logout, and the diagnostic log is promoted
from warning to error so the misconfiguration is monitorable
(round-3 finding 2). New regression test pins the new behaviour.
- The five user_id-accepting helpers in oauth_tools.py (get_provisioning_status,
provision_nextcloud_access, revoke_nextcloud_access, check_provisioning_status,
check_logged_in) are renamed with leading underscores to make the
trust boundary structural rather than documentary
(round-3 finding 3).
- create_browser_session and delete_browser_session now emit audit_log
rows so session establishment / teardown match the pattern used by
the rest of the security-relevant storage operations
(round-3 nit 5). delete_browser_session selects user_id before delete
so the audit row is attributable.
- oauth_login_callback no longer reflects raw IdP-error text or
exception strings into the HTML failure page; users see a generic
"internal error occurred" message + a correlation ID, with the
detail logged server-side keyed by the same ID (round-3 nit 6).
The XSS regression test is updated to pin the stricter contract.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Blocking:
- AS proxy callback now calls verify_id_token before caching the proxy
code so a tampered IdP response can't smuggle identity claims.
Important:
- Browser OAuth flow generates and verifies an OIDC nonce; new alembic
migration 006 adds the nonce column to oauth_sessions.
- _origin_matches_self logs a warning when CSRF check is bypassed.
- oauth_tools.py uses get_shared_storage instead of fresh handles.
Nits:
- New token_utils.get_oidc_discovery shares the 5-minute cache with
verify_id_token; oauth_login (integrated) and _revoke_refresh_token_at_idp
now use it instead of issuing fresh discovery fetches.
- Drop typing.Optional from oauth_tools.py in favour of X | None.
CI:
- test.yml generates an ephemeral Fernet TOKEN_ENCRYPTION_KEY per run
with openssl, removing the dependency on a missing repo secret.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Six findings from the latest claude-bot review on PR #758:
- JWKS cache had no kid-miss refresh path (Medium): on IdP key
rotation every login failed for up to _OIDC_CACHE_TTL. Evict
and refetch once before raising, per OIDC core §10.1.1.
- _should_use_secure_cookies fell back to nextcloud_host scheme,
but the cookie is issued by the MCP server. Switch to
settings.nextcloud_mcp_server_url so split-scheme deployments
get the right Secure flag.
- _origin_matches_self compared raw netloc strings, which include
the port. Browsers omit default ports per RFC 6454 §6.2; an
mcp_server_url like :443 falsely 403'd every legitimate logout.
Normalise (scheme, host, port) tuples with default ports stripped.
- delete_oauth_session exists in storage.py — drop the stale
"we don't have this method" comment and call it eagerly so
replays can't be processed and the table doesn't accumulate
completed-but-not-yet-expired browser-login rows.
- extract_user_id_from_token's unused ctx param renamed to _ctx
to signal "intentionally unused" at the signature level.
- provisioning_decorator instantiated RefreshTokenStorage per
call. Switch to get_shared_storage() for the lock-protected
process-wide singleton.
Plus pre-push self-review catch: lazy-logging on the unchanged
except arm in session_backend.py.
Adds 5 regression tests:
- JWKS rotation: success on refetch
- JWKS rotation: still-missing-kid surfaces original error
- JWKS rotation: network error during refresh wrapped as
IdTokenVerificationError
- default-port CSRF: explicit :443 in config + portless Origin
- default-port CSRF: portless config + explicit :443 in Origin
- scheme-mismatch CSRF: same host, different scheme rejected
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses all 9 findings from the review on PR #758:
Blocking:
- _revoke_refresh_token_at_idp now reads config from oauth_ctx["config"]
(the production-shaped nested dict). Previously read flat keys, causing
IdP revocation to silently no-op in production. Test fixtures rebuilt
to the realistic nested shape so the bug can't regress unnoticed.
- HTML error responses in oauth_login_callback now wrap IdP-controlled
error_body, str(e), and the attacker-controlled error/error_description
query params in html_escape. New test_browser_oauth_xss.py pins this.
Important:
- New _safe_next_url helper validates the ?next= query param at write
time (oauth_login), in oauth_logout, and on read from the session row
in oauth_login_callback. Blocks https://, // (protocol-relative), and
CRLF/whitespace injection.
- verify_id_token now caches discovery + JWKS (5-min TTL) using the
same pattern as oauth_routes._get_cached_discovery. New caching
regression test pins to one fetch per URL across multiple calls.
- /oauth/logout is now POST-only at the route layer (defeats passive
CSRF via <img src>). oauth_logout also validates Origin/Referer
against the configured mcp_server_url. Logout UI in user_info.html
converted from <a href> to <form method="post">.
- New storage.cleanup_expired_browser_sessions() called from the hourly
cleanup loop in app.py — previously these rows accumulated for users
who never explicitly logged out.
Nits:
- Demoted INFO logs that leaked oauth_config.keys() / client_id /
token-storage state to DEBUG. Operator-relevant outcome lines
(login successful, refresh token stored, logged out) stay INFO.
- verify_id_token algorithms widened to RS256, PS256, ES256 — covers
Azure AD (PS256) and Cognito/some Keycloak realms (ES256). Symmetric
and "none" remain off the allowlist.
- Migrated all Optional[X] usages in auth/storage.py to X | None per
CLAUDE.md.
Breaking change: GET /oauth/logout now returns 405. The in-tree logout
UI was migrated to a POST form; any external bookmark or curl-based
caller that relied on GET will need to switch.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pre-launch hardening for the hosted Astrolabe Cloud offering. Addresses
all five findings raised in #626 (Tim Kaufmann, code review of v0.65.0).
Re-verified against master before fixing.
Finding 3 (LLM-controllable user_id) — drop user_id from the public
signatures of provision_nextcloud_access, revoke_nextcloud_access,
check_provisioning_status, check_logged_in. Tool wrappers now always
derive identity from the verified AccessToken; user_id is no longer
accepted as MCP input. Adds parameterized CI-guard test that locks the
schema.
Finding 2 (predictable session cookie) — replace mcp_session=<user_id>
cookie with a cryptographically random session_id mapped server-side
(new browser_sessions table, alembic 005). Cookie value is opaque,
expires, revocable. SessionAuthBackend looks up user_id via the new
mapping and additionally requires a refresh token to fail closed.
Finding 4 (logout doesn't revoke refresh token) — oauth_logout now
calls the IdP revocation_endpoint (RFC 7009) when advertised, deletes
the stored refresh token regardless, and clears the browser_sessions
row. Cleanup is best-effort: logout always 302s.
Finding 1 (unverified ID token decodes) — verify_id_token helper does
JWKS signature + issuer + audience + exp + nonce checks per OIDC core
3.1.3.7. Used by both OAuth callback handlers (browser + MCP). Removes
the four "verify_signature: False" decodes that previously trusted IdP
claims unconditionally. Drops dead-code _validate_token_audience in
token_broker. Refactors token_utils + provisioning_decorator to read
user_id from the verified AccessToken instead of re-decoding the JWT.
Finding 5 (hardcoded Fernet keys in docker-compose.yml) — replace the
three inline TOKEN_ENCRYPTION_KEY values with required env var
interpolation; document in env.sample.
Test coverage: 4 new unit test modules (signature pinning, browser
sessions, ID-token verification, logout + revoke + session backend).
693 unit tests pass; ruff/format/ty clean.
Migration note: existing browser admin-UI sessions become invalid on
rollout (cookies are looked up against the new browser_sessions table,
which starts empty). Users re-login. MCP API access is unaffected.
Tracked on Astrolabe Cloud POC board card #37.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
