fix(auth): match complete WWW-Authenticate parameters by tarunag10 · Pull Request #3012 · modelcontextprotocol/python-sdk

tarunag10 · 2026-06-28T07:45:59Z

Summary

match WWW-Authenticate auth-param names only at parameter boundaries
prevent suffix matches such as error_scope satisfying scope
add regressions for scope and resource_metadata prefixed parameters

Test plan

uv run pytest tests/client/test_auth.py::TestWWWAuthenticate -q
uv run ruff check src/mcp/client/auth/utils.py tests/client/test_auth.py
uv run pyright src/mcp/client/auth/utils.py tests/client/test_auth.py
git diff --check

Related: #2902

cubic-dev-ai

No issues found across 2 files

_{Re-trigger cubic}

tarunag10 · 2026-06-29T14:14:31Z

Pushed a follow-up commit (2e2e819) to make the parser quote-aware, not just boundary-aware.

The extractor now splits WWW-Authenticate auth-params on commas only outside quoted strings, then compares auth-param names exactly before returning a value. This preserves the substring-name fix while also avoiding false matches inside quoted values such as realm="api, scope=decoy".

Additional validation added:

quoted scope decoy before the real scope
quoted-only scope decoy returns None
quoted resource_metadata decoy before the real resource_metadata

Validation run locally:

uv run pytest tests/client/test_auth.py -k "www_auth" -q — 29 passed
uv run ruff format --check src/mcp/client/auth/utils.py tests/client/test_auth.py
uv run ruff check src/mcp/client/auth/utils.py tests/client/test_auth.py
git diff --check

cubic-dev-ai

1 issue found across 2 files (changes from recent commits).

Prompt for AI agents (unresolved issues)


Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/mcp/client/auth/utils.py">

<violation number="1" location="src/mcp/client/auth/utils.py:69">
P2: `extract_field_from_www_auth` fails to match fields in subsequent `WWW-Authenticate` challenges because the scheme prefix is kept in the parsed parameter name (e.g. `"Bearer scope"`), so valid headers with multiple challenges are misparsed.</violation>
</file>

_{Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic}

cubic-dev-ai · 2026-06-29T14:17:05Z

-    match = re.search(pattern, www_auth_header)
+    for param in _iter_www_auth_params(www_auth_header):
+        name, separator, value = param.partition("=")
+        if separator != "=" or name.strip() != field_name:


P2: extract_field_from_www_auth fails to match fields in subsequent WWW-Authenticate challenges because the scheme prefix is kept in the parsed parameter name (e.g. "Bearer scope"), so valid headers with multiple challenges are misparsed.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/mcp/client/auth/utils.py, line 69: <comment>`extract_field_from_www_auth` fails to match fields in subsequent `WWW-Authenticate` challenges because the scheme prefix is kept in the parsed parameter name (e.g. `"Bearer scope"`), so valid headers with multiple challenges are misparsed.</comment> <file context> @@ -26,14 +64,16 @@ def extract_field_from_www_auth(response: Response, field_name: str) -> str | No - match = re.search(pattern, www_auth_header) + for param in _iter_www_auth_params(www_auth_header): + name, separator, value = param.partition("=") + if separator != "=" or name.strip() != field_name: + continue </file context>

tarunag10 · 2026-06-29T14:33:54Z

Pushed one more test-only follow-up (c05bc04) after the GitHub matrix exposed a coverage-only failure.

The failing jobs had all tests passing, then failed the repo-wide 100% coverage gate because the new quote-aware splitter had untested escaped-character branches. I added a regression for an escaped quote inside a quoted auth-param value, which exercises that branch while preserving the real scope extraction.

Validation run locally:

uv run pytest tests/client/test_auth.py -k "www_auth" -q — 30 passed
uv run ruff format --check src/mcp/client/auth/utils.py tests/client/test_auth.py
uv run ruff check src/mcp/client/auth/utils.py tests/client/test_auth.py
git diff --check

tarunag10 · 2026-06-29T14:36:29Z

Pushed another test-only coverage follow-up (32950d8) for the remaining splitter branches.

Added regressions for:

a Bearer challenge with no auth-params
empty comma segments between auth-params
a trailing comma after the final auth-param

Local validation now covers 33 selected auth tests:

uv run pytest tests/client/test_auth.py -k "www_auth" -q — 33 passed
uv run ruff format --check src/mcp/client/auth/utils.py tests/client/test_auth.py
uv run ruff check src/mcp/client/auth/utils.py tests/client/test_auth.py
git diff --check

fede-kamel

I differential-tested this against main and the two sibling PRs for #3009 (#3041, #3050) — full matrix in #3009 (comment).

Credit first: this is the only implementation of the three that correctly handles RFC 9110 escaped quotes inside quoted values — Bearer scope="say \"hi, there" parses intact here and truncates at the comma everywhere else, including main.

But I found a regression on multi-challenge headers. _iter_www_auth_params takes everything after the first space as the auth-params of a single challenge, so when the Bearer challenge is not the first one, the field is no longer found:

extract_field_from_www_auth(r('Basic realm="r", Bearer scope="files"'), "scope")
# main -> "files";  this branch -> None

extract_field_from_www_auth(r('DPoP algs="ES256", Bearer scope="files"'), "scope")
# main -> "files";  this branch -> None

The token for the second challenge comes through as Bearer scope="files", whose name (Bearer scope) never equals scope. The DPoP + Bearer co-advertisement is a real pattern (RFC 9449 §7), so this would break scope discovery against DPoP-capable servers.

A possible fix within this design: when a token contains a space before its = (i.e. it looks like Scheme param=…), treat it as the start of a new challenge and strip the scheme prefix before comparing. Alternatively, the escape-aware parsing here could be rebased as a follow-up on top of the minimal boundary fix in #3050, which keeps multi-challenge behavior intact.

fede-kamel · 2026-07-10T20:22:07Z

Follow-up with end-to-end confirmation of the multi-challenge regression, since my first comment used constructed header strings.

I stood up a raw-socket HTTP server returning 401 with two separate WWW-Authenticate headers (which RFC 9110 permits and reverse proxies commonly produce) and fetched it with real httpx:

server sends:  WWW-Authenticate: Basic realm="legacy"
               WWW-Authenticate: Bearer resource_metadata="https://rs.example/prm"

httpx presents:  'Basic realm="legacy", Bearer resource_metadata="https://rs.example/prm"'

extract_field_from_www_auth(resp, "resource_metadata")
#   main → "https://rs.example/prm"
#   this branch → None

Two implications beyond my earlier comment:

Because httpx joins repeated headers with ", ", the regression triggers even for servers that never put two challenges in one header — separate Basic + Bearer headers are enough.
The affected caller isn't just scope discovery: extract_resource_metadata_from_www_auth is how the client locates the RFC 9728 protected-resource metadata, so a None here breaks the OAuth discovery flow at its first step for any server whose Bearer challenge isn't the first header.

Same live scenario passes on main and on the #3050 approach, so this is specific to the first-space challenge parsing here.

fix(auth): match complete WWW-Authenticate params

b8b6001

tarunag10 mentioned this pull request Jun 28, 2026

Auth — 2026 client-side requirements #2902

Closed

cubic-dev-ai Bot reviewed Jun 28, 2026

View reviewed changes

fix(auth): parse WWW-Authenticate params quote-aware

2e2e819

cubic-dev-ai Bot reviewed Jun 29, 2026

View reviewed changes

test(auth): cover escaped WWW-Authenticate values

c05bc04

test(auth): cover WWW-Authenticate parser edge branches

32950d8

This was referenced Jul 10, 2026

WWW-Authenticate parsing matches a field name as a substring of another auth-param #3009

Open

fix: match complete www-authenticate auth params #3050

Open

fede-kamel reviewed Jul 10, 2026

View reviewed changes

fede-kamel mentioned this pull request Jul 10, 2026

Fix substring matches in WWW-Authenticate parsing #3041

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(auth): match complete WWW-Authenticate parameters#3012

fix(auth): match complete WWW-Authenticate parameters#3012
tarunag10 wants to merge 4 commits into
modelcontextprotocol:mainfrom
tarunag10:codex/www-auth-param-boundary

tarunag10 commented Jun 28, 2026

Uh oh!

cubic-dev-ai Bot left a comment

Uh oh!

tarunag10 commented Jun 29, 2026

Uh oh!

cubic-dev-ai Bot left a comment

Uh oh!

cubic-dev-ai Bot Jun 29, 2026 •

edited

Loading

Uh oh!

tarunag10 commented Jun 29, 2026

Uh oh!

tarunag10 commented Jun 29, 2026

Uh oh!

fede-kamel left a comment

Uh oh!

fede-kamel commented Jul 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

tarunag10 commented Jun 28, 2026

Summary

Test plan

Uh oh!

cubic-dev-ai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

tarunag10 commented Jun 29, 2026

Uh oh!

cubic-dev-ai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

cubic-dev-ai Bot Jun 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tarunag10 commented Jun 29, 2026

Uh oh!

tarunag10 commented Jun 29, 2026

Uh oh!

fede-kamel left a comment

Choose a reason for hiding this comment

Uh oh!

fede-kamel commented Jul 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

cubic-dev-ai Bot Jun 29, 2026 •

edited

Loading