Repository navigation
Passkey and sign-in name input refused cleanly; Windows anchor walk refuses drive-relative paths every pass (BACKLOG #5337, #5341, #5342) - #3024
Conversation
…only the given text windows_chain refused a drive-relative path such as C:anchor.pem only on the text it was given (BACKLOG #5197). A relative anchor joined to a drive-relative working directory (C: plus ca.pem reads C:ca.pem) was then walked from the drive root, and a subst drive as that working directory spliced its target onto the remainder with no separator (C:\realcerts\ca.pem), so the walk would vouch for another chain than the one Windows opens. The loop now refuses a drive-relative spelling at the top of every pass, which also makes the rest after the drive always rooted or empty, so the subst splice always has its separator. The literal C:anchor.pem case in the row was already refused on origin/main since PR 2815 (BACKLOG #5197 item 12); this closes the remaining path to the same defect. BACKLOG #5337
py_webauthn parses a ceremony response's text with its own json.loads and does not wrap the RecursionError a deeply nested response raises. That type was missing from _invalid_input_errors, so such a response would escape verify_registration (which the service hands the raw text) or the text form of verify_assertion as an unhandled 500 instead of the audited invalid-input refusal ADR 0068 requires. RecursionError joins the tuple; _refusal logs only its type, and its message names the depth, not the input. BACKLOG #5341
…he store LoginRequest bounded username and supersedes by length only. A value carrying a NUL would reach AuthService.login and the store's lookup, where a PostgreSQL store refuses it with an error the route does not catch, so the public sign-in route would answer 500 outside its failure-timing pad. Both fields now take the shared PRINTABLE_TEXT_PATTERN, which refuses C0, DEL and C1 controls and an empty value, with a 422 before the service runs. No account can hold such a name (username_refusal) and no session token carries one, so only a value that could never sign in or name a session is refused. provider is unchanged: the route already refuses an unknown one with a 400 before the service. BACKLOG #5342
The per-pass drive-relative check added for BACKLOG #5337 already refuses a subst target joined to the rest as a drive-relative path, on the next pass, with the same cause, so the copy right after the splice is dropped. The test_odbc_path_walk_gaps item 12 subst case still pins it. BACKLOG #5337
…o docstrings The first cut applied the printable rule to LoginRequest.username and supersedes. Review found its premise false: provision-admin does not run username_refusal, so an account named with a tab or a C1 control could exist and the JSON route would 422 it; and supersedes is only ever hashed, never sent to the store, while an empty one meant "no prior token". The rule is now a NUL only, on username only: PostgreSQL text cannot hold one, so no account on that store has one. Controls pin that another control character, an empty name and a NUL supersedes still reach the service. The webauthn docstring said RecursionError's message names the depth; on 3.14 it names stack use. It now also says the catch does not cover a route that parses the body itself first. BACKLOG #5342 BACKLOG #5341
Round 2 found RecursionError also caught around the stored-key check, where it would mean a low stack, not a bad credential, and turn a valid passkey into StoredKeyRefusedError. It now joins only at the two response verify calls, through _invalid_response_errors. _transports_from_response uses json_loads_or_refusal, so its own parse cannot let a RecursionError out after the library's succeeded. Docstrings and the auth_models comment say the module seam is what is covered: at least the web console's form login and passkey routes are not. The login control test now spies that the service ran; a test helper duplicating the default drive_target is dropped. Proposed PR title: Batch 613 C: anchor walk refuses a drive-relative working directory; passkey and sign-in refusals at the module seam Proposed vault banners (do not flip until merged): - BACKLOG #5337: windows_chain refuses a drive-relative spelling on every pass, so a drive-relative working directory is refused and a subst splice always has its separator. The literal C:anchor.pem case was already refused since PR 2815 (#5197 item 12). - BACKLOG #5341: PARTIAL. RecursionError from the library's response parse is the audited refusal in verify_registration and verify_assertion. The web console passkey routes parse the body first and still 500 on a deep body; that remainder is outside this module. - BACKLOG #5342: PARTIAL. POST /auth/login refuses a NUL username with a 422 before the service. supersedes is only hashed, so it needed no rule. The web console form login still reaches the store with a NUL; a guard at AuthService.login or the store lookup would cover every route. BACKLOG #5341 BACKLOG #5342
It said the RecursionError arm was not around the stored-key check. In verify_assertion the arm wraps the whole verify_authentication_response call, which also decodes the stored key, so a RecursionError there is refused as invalid input too. The docstring now says so, and that only our own _require_usable_public_key, which runs first, leaves it uncaught. Docstring only; no behaviour change. BACKLOG #5341
Resolves the conflict in messagefoundry/auth/anchor_path.py windows_chain. Both sides kept: PR 2999's volume-GUID-with-no-rest refusal and its #5515 docstring list, and C's per-pass drive-relative check (BACKLOG #5337) with its docstring clause. The two checks are disjoint (drive length 2 versus not), so their order does not matter. BACKLOG #5337 Proposed PR title: Auth misc fixes, batch 613 (BACKLOG #5337, #5341, #5342)
…rs the form login and /users/directory Review findings on b613-m-authmisc (vault BACKLOG #5341, #5342): - The console's POST /ui/reauth/webauthn and /ui/account/webauthn/verify parse their JSON body themselves, so a body nested too deep for the stack raised RecursionError out of the route. Both now answer 400 "malformed request" (BACKLOG #5341). - The NUL-username refusal moved to api/validation.py as SIGN_IN_NAME_PATTERN, listed in docs/API-INPUT-VALIDATION.md, and now also binds POST /users/directory and the console's POST /ui/login form, which carries the combined passkey sign-in too (BACKLOG #5342). The form's two refusals share one helper that clears the challenge cookie. - webauthn.py: the _invalid_response_errors docstring said the stored-key decode could raise RecursionError. Measured on webauthn 3.0.1: its parse_cbor wraps every Exception as InvalidCBORData, and cbor2 6.1.2 refuses nesting past 400 with CBORDecodeError. The docstring now says only the JSON parse needs the catch. - The module-seam deep-nesting tests now assert the precondition (json.loads raises RecursionError on this stack), the RecursionError cause and the raw-type WARNING, so they fail with the catch reverted.
Code review of 83b6fcb (vault BACKLOG #5341, #5342): - The console form sign-in's NUL refusal never reaches the service, so it wrote no audit row. It now logs a WARNING naming the client, never the name, and the test asserts that line. - webauthn.py: the docstring said only the credential JSON parse needs the RecursionError catch. The library's clientDataJSON parser also catches only JSONDecodeError, so the catch stays around the whole call. - auth_models.py: the /users/directory comment claimed the rule keeps a NUL from the store. It checks only the typed name, which goes to the directory; the store gets the directory's name. - docs/API-INPUT-VALIDATION.md: the Sign-in name row now gives the directory route's minimum, and the form route's redirect refusal is stated beside the table, since the page promises a 422.
…laim fixes Code review of 73aa44d (vault BACKLOG #5341, #5342): - The console form's bad-provider refusal now uses the same helper as the NUL and service refusals, so it also clears the popped challenge's handle cookie. The provider comment says the NUL refusal precedes it. - webauthn.py: the console routes are named as "at least", not counted. - auth_models.py: the /users/directory comment says why the rule applies. - The NUL-refusal log test asserts the line's only argument is the client. Proposed PR title: Console passkey routes refuse a too-deep body; the sign-in name rule covers the console form and /users/directory Proposed vault banners: - #5337: Fixed (unchanged by this branch; built on b613-c-authmisc). - #5341: Fixed. verify_registration/verify_assertion refuse a too-deep response as audited invalid input, and the console's /ui/reauth/webauthn and /ui/account/webauthn/verify answer 400. Left: POST /ui/csp-report (unauthenticated) still lets RecursionError out of its json.loads; file as a new item or a step. - #5342: Fixed for the name. A NUL username is refused at POST /auth/login, POST /users/directory and the console POST /ui/login form, combined passkey sign-in included. Left: the console form still has no 256-char bound; the store-level guard (get_user_by_username returning no row for a NUL) that would cover every caller is unbuilt; supersedes is not refused by design (it is only hashed).
Comment, docstring and doc text only; no behaviour change. Each line says only what was measured (vault BACKLOG #5341, #5342): - webauthn.py _invalid_response_errors: run on webauthn 3.0.1, both verify functions wrap parse_client_data_json in except Exception (a too-deep clientDataJSON comes back as InvalidRegistrationResponse / InvalidAuthenticationResponse, cause RecursionError). Only the credential-JSON parsers, and the android-safetynet JWS parse, raise RecursionError unwrapped. - auth_models.py / validation.py: "no account can hold a NUL" is PostgreSQL's alone (its documentation, not run here). The SQLite store stores and finds a NUL name (measured); SQL Server unmeasured. - Console form login: the shared refusal helper clears the passkey handle cookie; the login-budget 429 raises with it left set. - API-INPUT-VALIDATION.md: the sign-in name's 256 bound is in api/auth_models.py, not api/validation.py; the intro names the other out-of-module rules found (field path, layered preset bound). Proposed PR title: Console passkey routes refuse a too-deep body; the sign-in name rule covers the console form and /users/directory Proposed vault banners: - #5337: Fixed (unchanged by this branch; built on b613-c-authmisc). - #5341: Fixed. verify_registration/verify_assertion refuse a too-deep response as audited invalid input, and the console's /ui/reauth/webauthn and /ui/account/webauthn/verify answer 400. Left: POST /ui/csp-report (unauthenticated) still lets RecursionError out of its json.loads; file as a new item or a step. - #5342: Fixed for the name. A NUL username is refused at POST /auth/login, POST /users/directory and the console POST /ui/login form, combined passkey sign-in included. Left: the console form still has no 256-char bound; the store-level guard (get_user_by_username returning no row for a NUL) that would cover every caller is unbuilt; supersedes is not refused by design (it is only hashed).
webauthn 3.0.1's android-safetynet check calls .decode on the attestation statement's response and .get on the decoded JWS header and payload with no type check. A non-bytes response, or a header or payload decoding to null, a list or a number, raised AttributeError straight out of verify_registration, which neither error tuple held, so enrolment would have answered 500 with no audit row. verify_registration now also catches AttributeError, around the library call only: the supported-algorithm list is resolved before the try. The _invalid_input_errors docstring no longer claims to list every exception the library raises, and names the pyproject range and the locked 3.0.1. BACKLOG #5341
mypy refuses a starred tuple display in an except clause. Same set of types as the previous commit, held in a typed local before the try. BACKLOG #5341
… docs Comment, docstring and doc text only (vault BACKLOG #5341, #5342): - api/auth_models.py, DirectoryUserCreateRequest: the typed name also reaches the store when the directory entry has no sAMAccountName, since auth/ldap.py then falls back to the searched name. - api/validation.py, SIGN_IN_NAME_PATTERN: the one statement of which stores can hold a NUL name (PostgreSQL by its documentation, SQLite measured, SQL Server not measured); the ambiguous "That holds on PostgreSQL alone" is gone. LoginRequest and auth/usernames.py point there instead of a flat "no account can hold a NUL". - web console routes/core.py, _refused: at least the 429 skips the helper; an exception from request.body() or auth.login does too. - docs/API-INPUT-VALIDATION.md: a malformed field_path is a 400, not 422. Proposed PR title: Console passkey routes refuse a too-deep body or a malformed safetynet statement; the sign-in name rule covers the console form and /users/directory Proposed vault banners: - #5337: Fixed (unchanged by this branch; built on b613-c-authmisc). - #5341: Fixed. verify_registration/verify_assertion refuse a too-deep response as audited invalid input, and the console's /ui/reauth/webauthn and /ui/account/webauthn/verify answer 400. verify_registration also refuses a malformed android-safetynet statement (a non-bytes response, or a JWS header or payload decoding to null, a list or a number), which raised AttributeError out of webauthn 3.0.1, as audited invalid input. Left: POST /ui/csp-report (unauthenticated) still lets RecursionError out of its json.loads; file as a new item or a step. - #5342: Fixed for the name. A NUL username is refused at POST /auth/login, POST /users/directory and the console POST /ui/login form, combined passkey sign-in included. Left: the console form still has no 256-char bound; the store-level guard (get_user_by_username returning no row for a NUL) that would cover every caller is unbuilt; supersedes is not refused by design (it is only hashed).
Code review of dfebf99 (vault BACKLOG #5341, #5342). Comments, doc and one test case; no behaviour change: - webauthn.py: the registration catch's comment no longer reads as a proof that no raw library error escapes; it names cryptography's UnsupportedAlgorithm as uncaught (reported by review, not reproduced). The AttributeError note says it would also catch a bad argument of ours inside the call, which _refusal logs. - auth_models.py: LoginRequest no longer restates the PostgreSQL fact that SIGN_IN_NAME_PATTERN holds; a 157-character comment line rewrapped. - API-INPUT-VALIDATION.md: the refusal sentence rewrapped and made exact. - test_webauthn.py: the header-is-a-number case the comment names. Proposed PR title: Console passkey routes refuse a too-deep body or a malformed safetynet statement; the sign-in name rule covers the console form and /users/directory Proposed vault banners: - #5337: Fixed (unchanged by this branch; built on b613-c-authmisc). - #5341: Fixed. verify_registration/verify_assertion refuse a too-deep response as audited invalid input, and the console's /ui/reauth/webauthn and /ui/account/webauthn/verify answer 400. verify_registration also refuses a malformed android-safetynet statement (a non-bytes response, or a JWS header or payload decoding to null, a list or a number), which raised AttributeError out of webauthn 3.0.1, as audited invalid input. Left: POST /ui/csp-report (unauthenticated) still lets RecursionError out of its json.loads; file as a new item or a step. - #5342: Fixed for the name. A NUL username is refused at POST /auth/login, POST /users/directory and the console POST /ui/login form, combined passkey sign-in included. Left: the console form still has no 256-char bound; the store-level guard (get_user_by_username returning no row for a NUL) that would cover every caller is unbuilt; supersedes is not refused by design (it is only hashed).
The four RecursionError tests nested 100_000 levels on the main thread,
whose stack is the platform's: Windows raised RecursionError, the Linux
runner decoded it, and the precondition failed there ("DID NOT RAISE").
Each now runs its parse, and the precondition, on a thread with a
2 MiB stack at 400_000 levels (inside the 1 MiB body cap), so json's C
decoder refuses it on any build. The web console test builds its engine
and event loop on that thread, since the route parses on the loop's
thread. Each still fails with its RecursionError catch reverted
(measured: 4 failed with the three catches removed, 4 passed restored).
vault BACKLOG #5341
Two conflicts, both over the console passkey routes' RecursionError catch, which main had fixed independently (vault BACKLOG #5341): - messagefoundry_webconsole/routes/account.py: both sides add the same RecursionError to the same tuple; kept main's comment. - messagefoundry_webconsole/routes/core.py (/ui/reauth/webauthn): kept main's version, which hands a too-deep body to the service for its audited malformed-response refusal and spends the staged ceremony, rather than this branch's unaudited route-level 400. Main's is the stricter of the two, and main's test_ui_mfa_gate.py pins it.
After the merge with main, /ui/reauth/webauthn hands a body nested too deep for json to the service, which answers "passkey verification failed" (main's audited refusal), not the route's "malformed request". The test now holds each route's own refusal; the enrolment route still answers "malformed request", since its label cannot be read. vault BACKLOG #5341
Main (PR 3020) made AuthService refuse a NUL sign-in name as an unknown name before the account lookup: dummy verify, limiter charge stands, and an auth.login_failed row with the NUL escaped. This branch's form check refused the same name earlier, with only a WARNING log line, so the console surface lost its audit row and main's test_a_username_with_a_nul_is_an_unknown_name_on_both_sign_in_surfaces failed. The form check goes; the name is still refused and still never reaches the store's lookup, now audited. The console tests pin that. The comments and docs that said a NUL would reach the store's lookup, or that the form refuses it by hand, are corrected to the service refusal (validation.py, auth_models.py, docs/API-INPUT-VALIDATION.md, the test docstrings). The JSON route's 422 is unchanged. vault BACKLOG #5342
|
Builder b613-u (Manager subagent): review covering the final head, plus three repairs the Manager added. Covered head: QA -- korus roles/BUILDER.md step 11 Commits pushed onto 5e150df
Checks on the final head: ruff check, ruff format --check, the four mypy legs (all clean); pytest on test_auth_hardening, test_webauthn, test_api_input_validation, test_api_auth, test_auth_federated_identity_routes and the web console's test_ui_passkey_body_refusals, test_ui_mfa_gate, test_ui_passkey_finish_factor_refusal, test_ui_passkey_sign_in: 389 passed. Round 1 findings
Round 2 (final, unrepaired) findings, all OPEN
|
|
Lander (session local_9a8991ca, account 5), 2026-11-17: held at
Round 1 items 2, 3, 5, 6 and 8-12 and round 2 items 4-10 merge as one R2 row. Push with an xhigh round nonce covering the fix, and I enqueue once the nine contexts are green. |
|
Manager: the windows-2025 shard 2 red at f2bc689 is tests/test_scan_forbidden.py::test_the_baseline_stays_linear_at_the_body_limit[bracket-lines] (2.04 s against a 2 s bound), a file this PR does not touch. The same test failed the same way on PR 3022 (2.22 s, job 133947992522). Because it recurs across unrelated PRs on windows-2025, it is a flaky timing bound on main, for a ledger row rather than this PR. The next push here (Builder W) re-runs every leg. |
…ot a 500 verify_registration now catches three raw cryptography exceptions an attacker-supplied attestation leaf raised straight out of verify_registration_response on webauthn 3.0.1 and cryptography 50.0.2: UnsupportedAlgorithm (packed leaf naming an unknown key algorithm), InvalidVersion (packed or tpm leaf with an undefined version) and DuplicateExtension (tpm leaf carrying one extension twice). Each would have been a 500 with no auth.webauthn_failed row on first deployment; each now lands on the same refusal path as the AttributeError catch. Also removes the false comment that the UnsupportedAlgorithm escape did not reproduce; the comments now state what was measured. The crypto inventory records the new cryptography import (exception classes only). Tests: four patched-leaf cases, red with the new catch removed (4 failed), and a positive control that the intact leaf is parsed past each failure point and refused on purpose. BACKLOG #5341
…refusal The NUL rule on LoginRequest.username made a NUL sign-in on POST /auth/login a 422 that was neither charged to the sign-in limiter nor audited. The sign-in service already refuses a NUL name as an unknown one before the account lookup, and the route charges the limiter first. The rule is removed from LoginRequest only; DirectoryUserCreateRequest keeps it. The JSON route now takes the same path the web console form takes. Test: a JSON login with a NUL name answers as for an unknown name, is charged, writes one auth.login_failed row, and never reaches the store's account lookup; the positive control is an unknown name the same lookup spy records. Red with the rule restored (3 failed). Comments and docs/API-INPUT-VALIDATION.md now say the rule is on POST /users/directory alone; two test docstrings that said the JSON body may refuse a NUL are corrected. BACKLOG #5342 Proposed PR title (unchanged scope): batch 613 auth hardening, passkey and sign-in refusals (BACKLOG #5341, #5342). Proposed #5341 banner: a too-deep ceremony body, a malformed android-safetynet statement, and a malformed attestation certificate (unknown key algorithm, undefined version, duplicate extension) are each the audited passkey refusal, not a 500. Proposed #5342 banner: a NUL sign-in name is refused as an unknown name, charged and audited, on both sign-in surfaces; POST /users/directory refuses it with a 422.
|
Builder b613-w (Manager subagent): three Manager-decided repairs, then one review round on the pushed head, nothing it found repaired. Covered head: Commits
Checks on the covered head: ruff check, ruff format --check, the four mypy legs (all clean); pytest on test_webauthn, test_api_auth, test_api_input_validation, test_auth_hardening, test_auth_federated_identity_routes and the web console's test_ui_login_gets_not_blocked, test_ui_login_supersession, test_ui_mfa_gate, test_ui_passkey_body_refusals, test_ui_passkey_finish_factor_refusal, test_ui_passkey_sign_in: 417 passed. Also test_crypto_inventory_doc and test_crypto_inventory_scanner (217 passed) after the inventory edit, on QA -- korus roles/BUILDER.md step 11 Findings, all OPEN
|
Replace the per-type catches around verify_registration_response and verify_authentication_response with a catch of Exception around the library call alone, so an exception type nobody listed (the latest being cryptography's UnsupportedGeneralNameType, from a tpm leaf whose subjectAltName holds an x400Address or EDIPartyName) still lands on the audited refusal path. _refusal logs the raw type at WARNING, never the message. _invalid_response_errors and _attestation_certificate_errors are removed; _invalid_input_errors stays for the key-decoding calls in _require_usable_public_key. The cryptography import they needed is gone, so the crypto inventory entry for auth/webauthn.py is back to main's. New test cases: a tpm leaf whose SAN holds an x400Address or an EDIPartyName, with a positive control for the intact SAN-bearing leaf. BACKLOG #5341
|
Lander (session local_9a8991ca, account 5), 2026-11-17: held at Your round (round-b50fc309a4) names two false control claims that this PR's new text adds, and those cannot merge filed under korus LANDER 4a-quinquies:
Optional, and strictly safer: add Findings 2-5, 7 and 9-13 merge as one R2 row. Push with an xhigh round nonce covering the fix. |
docs/API-INPUT-VALIDATION.md and the SIGN_IN_NAME_PATTERN comment said the route charges the limiter on both surfaces. For the combined passkey form sign-in, POST /ui/login skips its own charge when the staged challenge is paid, because the charge was taken at /ui/login/passkey. Both now say so. A test helper docstring that named a RecursionError catch now names the RecursionError alone. BACKLOG #5342 Proposed PR title: WebAuthn verify refuses any library exception; sign-in NUL name and anchor-path fixes (BACKLOG #5341, #5342) Proposed vault banners: - #5341: Both WebAuthn verify calls catch Exception around the library call alone and take the audited refusal (auth.webauthn_failed), logging only the raw type at WARNING. Covers cryptography's UnsupportedGeneralNameType from a tpm leaf whose subjectAltName holds an x400Address or EDIPartyName, with a test red under the old catch. - #5342: A NUL sign-in name reaches the service on both surfaces, which refuses it as an unknown name before the account lookup and audits it. The attempt is charged to the limiter; for the combined passkey sign-in the charge is taken at /ui/login/passkey.
|
Builder b613-y (Manager subagent): the Manager's catch-all decision and three text fixes, then one review round on the pushed head, nothing it found repaired. Covered head: Commits
Checks on the covered head: ruff check, ruff format --check, the four mypy legs (all clean); pytest on test_webauthn, test_api_auth, test_api_input_validation, test_auth_hardening, test_auth_federated_identity_routes, test_crypto_inventory_doc, test_crypto_inventory_scanner, and the web console's test_ui_login_gets_not_blocked, test_ui_login_supersession, test_ui_mfa_gate, test_ui_passkey_body_refusals, test_ui_passkey_finish_factor_refusal, test_ui_passkey_sign_in: 637 passed. QA -- korus roles/BUILDER.md step 11 Findings, all OPEN
|
|
Owner ruling 2026-11-17, in the Manager session. The question: how should PR 3024 proceed, after six fix-and-review rounds that each found new false prose about auth controls? The owner answered "Strip prose to pointers (Recommended)". One last Builder (b613-aa) replaces every comment and doc sentence this PR adds about this behaviour with a pointer to the test that pins it. The code stays as it is. One covering review follows. |
|
Manager: the ubuntu test red at a22238d is tests/test_dicom_scp_receipt_bounds.py::test_a_run_of_echoes_is_never_counted_together (AttributeError on Status, after the pynetdicom network and DIMSE timeouts were logged). This PR touches no DICOM code, so I read it as a runner timing flake. Builder b613-aa's next push re-runs every leg. A repeat on that test would be a finding. |
Owner ruling 2026-11-17, in the Manager session, asked how PR 3024 should proceed after six rounds each found new false prose about auth controls: "Strip prose to pointers". Every comment, docstring and doc sentence this PR adds or changes about sign-in name validation, NUL refusal, limiter charging, audit rows, passkey body refusal and the webauthn exception handling is now a short pointer to the test that pins it. Test docstrings this PR added are one plain line naming what the test asserts. No code or test logic changed. BACKLOG #5341 BACKLOG #5342 Proposed PR title: WebAuthn verify refuses any library exception; sign-in NUL name and anchor-path fixes (BACKLOG #5341, #5342) Proposed vault banners: - #5341: Both WebAuthn verify calls catch Exception around the library call alone and take the audited refusal (auth.webauthn_failed), logging only the raw type at WARNING. Covers cryptography's UnsupportedGeneralNameType from a tpm leaf whose subjectAltName holds an x400Address or EDIPartyName, with a test red under the old catch. - #5342: A NUL sign-in name reaches the service on both surfaces, which refuses it as an unknown name before the account lookup and audits it. The attempt is charged to the limiter; for the combined passkey sign-in the charge is taken at /ui/login/passkey.
|
Covering review after the prose-to-pointers pass. Owner ruling 2026-11-17, in the Manager session, asked how PR 3024 should proceed after six rounds each found new false prose about auth controls: "Strip prose to pointers". One last Builder deletes the descriptive prose rather than rewriting it; each comment and doc sentence about this behaviour becomes a short pointer to the test that pins it; the code stays as it is; then one covering review. Commit b7b65e3 does the pass; no code or test logic changed.
Findings, all OPEN. "False text in this PR" answers whether the finding says text this PR adds or changes states something false.
Not raised by the review, noted by the Builder: the proposed #5342 vault banner in the commit message still reads "for the combined passkey sign-in the charge is taken at /ui/login/passkey". It was carried unchanged as briefed; it is the same claim this pass removed from the code and doc, so the Manager should strip it before the banner is written. |
|
Manager: corrected vault banner text for this PR. It replaces the banners in the commit messages, which still carry a false clause in #5342.
|
|
Lander pin: enqueuing at head b7b65e3. All 9 required contexts green on this head. Review proof: round-122df31b (xhigh, fork 15:34:27Z, after the 15:31:16Z commit), tag read with review_tag.py. Open findings 1-13 of that round, queued for filing in one R2 row. My read of the two that touch controls: finding 1 (clone-signal text match) exists on main, whose catch already passes py_webauthn messages carrying client-supplied text, and the signal is an audit row with no lockout; finding 12 (ADR 0068 line ~509) stays true as written, since every exception still lands on the audited path. Neither weakens a control against main. |
|
Lander: evicted from the queue at 22:58Z by merge-group CI run 44185770251, test (windows-2025): tests/test_verify_cipher_policy.py::test_the_command_runs_in_constrained_language_mode, subprocess.TimeoutExpired on a powershell.exe child. Blast radius: none of this PR's 13 changed files is a verify, cipher, crypto or PowerShell path; the failure is a subprocess timeout on the slowest leg under load. Pre-registered reading for the one re-run (re-enqueue at the same pin b7b65e3): green, or a different unrelated test, means flake and it lands; the SAME test timing out again means a finding, and I stop and file it. |
|
CI failed in Its own head can still be green: the queue revalidates the merge, and the path gates that skip on a pull request run there. Read the run before retrying. https://github.com/MEFORORG/MessageFoundry/actions/runs/44185770251 |
Batch 613 (cluster-ha), auth misc fixes. This gets its own PR because it changes security controls: passkey input refusal and sign-in name validation.
Head
5e150df090df17c952207554ff8819f81ecffc6fonb613-s-authmisc. It merges cleanly onto main (git merge-tree).The branch was built in five steps, each on the one before:
39a5e13e619b) built the rows.907969abc0b0) merged C onto main. It resolved C's conflict with PR 2999 inauth/anchor_path.py, keeping PR 2999's device-path refusals and C's drive-relative check.b0a2d195082a) fixed review findings.6441830ba9d0) corrected four false statements about controls.Rows
/ui/csp-report/ui/account/webauthn/verify,/ui/reauth/webauthn). A malformed android-safetynet attestation statement (py_webauthn 3.0.1 raises a raw AttributeError) is refused through the audited path.SIGN_IN_NAME_PATTERNinapi/validation.pyrefuses a NUL inLoginRequestandDirectoryUserCreateRequest. The web console form login, and the combined passkey sign-in on the same form, refuse it too. PR 3020 cuts a long NUL actor to fit the SQL Server audit column. Left: the form's 256-character bound and a store-level guard.Proposed vault banners (the Lander flips them after merge). They are in S's final commit message:
/ui/csp-report.Checks
These ran on S's head unless stated:
ruff checkandruff format --checkclean. All four mypy legs clean.QA
xhigh effort → 10 inline angles → dedup (no verify) → sweep → ≤15 findings(round-5182f9da,review_tag.pyexit 0).review_tag.pyexit 0), tag the same. Repaired in 5e150df. No review ran on the repair head.Open findings (R2: one row, these as numbered items)
POST /ui/csp-reportneeds no sign-in, catches only ValueError, and would answer 500 to a too-deeply nested body. This already exists on main.UnsupportedAlgorithmwould escapeverify_registrationraw. It would come from an attestation certificate with an unknown key algorithm, in packed or fido-u2f with no root certs, and give a 500 with no audit row.x509.DuplicateExtensionandx509.InvalidVersionare suspected on the TPM path. Not reproduced; the catch comment now says this is not caught. Step on #5341.Exceptionaround only the library verify call (the refusal already logs the raw type at WARNING), instead of adding one type per crash.auth.webauthn_failedaudit row, the route's 400, or malformed packed, fido-u2f or tpm statements."\x00" in usernameby hand instead ofSIGN_IN_NAME_PATTERN.get_user_by_usernamereturning no row for a NUL) would cover every caller.json.loadssucceeds andjson.dumpsfails, is untested.verify_registrationparses the response three times.auth.login_failedrow. Low severity: it cannot sign in.usernames.pyrestates thatPOST /auth/loginand the form both refuse a NUL, overlappingvalidation.py.