[dependabot]

Bumps the python-examples group with 4 updates in the /demos directory: cmcp-runtime, openai, starlette and pytest.
Bumps the python-examples group with 10 updates in the /examples directory:

Package	From	To
cmcp-runtime	0.7.0	0.7.1
starlette	1.8.0	1.10.0
pytest	9.1.1	9.2.0
certifi	2026.10.19	2026.10.23
pycparser	3.11	3.12
pytest	9.1.1	9.2.0
transformers	5.19.0	5.22.0
torch	2.14.1	2.15.0
cmcp-runtime	0.7.0	0.7.1
cedarpy	4.12.1	4.12.2
pydantic	2.14.1	2.14.2
pydantic-core	2.50.1	2.50.2
starlette	1.8.0	1.10.0

Updates cmcp-runtime from 0.7.0 to 0.7.1

Release notes

Sourced from cmcp-runtime's releases.

cmcp 0.7.1

Three security fixes, backported onto 0.7.0 so they ship without the breaking Agent Manifest binding change and the opaque provider removal waiting on main. Full detail in CHANGELOG.

• GHSA-frxc-7f57-6f4r: deny a Cedar Allow that comes with evaluation errors, instead of skipping the errored policy and allowing.
• GHSA-mrcx-xvv4-m7f8: bound the TPM AIA certificate fetch. Each hop resolves the host once, connects only to that checked address, and the hop and byte limits hold.
• GHSA-2p85-666x-xm7x: classify upstream drift against the drifted tool's own catalog entry.

SDK on PyPI as cmcp-runtime 0.7.1.

Changelog

Sourced from cmcp-runtime's changelog.

[0.7.1] - 2026-10-19

Security

• A Cedar Allow reached with evaluation errors was enforced as Allow (GHSA-frxc-7f57-6f4r). Cedar skips a policy whose condition errors, so a forbid reading a context attribute the current phase does not carry could not stop a matching permit. CedarBackend now denies any Allow that comes with evaluation errors, and the error reaches the tool_call or egress_denied audit row. So that a policy written for one phase does not deny in the other, both phases now carry the same context keys: egress adds session_max_sensitivity (the label string), arguments, baa_covered, compliance_domain, server_identity, destination_class and attestation_platform, and ingress adds egress (false), sensitivity_level, injection_events, reset_count and response_size_bytes (0). Without this the shipped BFSi demo denied every response at egress. A policy that reads a key outside that set, or workflow_id when the caller declared none, now denies.
• The TPM AIA certificate fetch was unbounded (GHSA-mrcx-xvv4-m7f8). The chain walk followed redirects to any address and read responses without a limit. It now refuses destinations that do not resolve to global addresses, checks every redirect the same way, follows at most three and ignores responses over 64 KiB. Each hop's host is resolved once and the connection goes to the checked address (the request still names the host, and HTTPS still verifies the certificate against it), so a second DNS answer cannot point the fetch somewhere else. The fetch no longer uses environment proxies. An on-prem CA whose AIA host has a private address falls back to the vendored bundle.
• Upstream drift was classified against the wrong catalog entry (GHSA-2p85-666x-xm7x). The scanner received the server of whichever entry was last in the catalog. Detection and fail-closed denial were not affected; only the classification written to the audit row was.

Commits

• 4a7163a fix(security): cmcp-runtime 0.7.1 backport (Cedar errors, AIA fetch, drift cl...
• See full diff in compare view

Updates openai from 3.31.0 to 3.41.0

Release notes

Sourced from openai's releases.

v3.41.0

3.41.0 (2026-10-28)

Features

• api: accept typed references in Realtime response inputs (#4189) (8ecb487)
• api: add response tool usage models (#4201) (55fc500)
• api: filter agent sessions by metadata (#4200) (77fd116)
• websocket: recover queued binary frames after reconnect (#4191) (f079868)

Bug Fixes

• examples: detect incomplete WebSocket responses (#4187) (8559299)
• examples: handle interruptions and disconnects in push-to-talk (#4192) (35fc9f3)
• keep consumer fixtures clean under strict Pyright (#4204) (5012565)

Chores

• api: document upcoming model shutdowns (#4199) (5b43cce)
• api: update image-model documentation (#4198) (fa67521)
• enforce generated-file ownership (#4193) (1b37b9a)
• migrate development type checking to ty (#4202) (45b9531)

v3.40.0

3.40.0 (2026-10-27)

Features

• api: add prompt cache breakpoint TTL options (#4188) (3596d18)

Bug Fixes

• ci: use an explicit browser demo dependency (#4185) (a3f4302)

Chores

• api: record completed generated-code separation (#4190) (d7ab5ee)
• deps-dev: bump griffe from 1.14.0 to 2.3.0 (#4069) (766b849)
• deps: bump actions/create-github-app-token from 3.2.0 to 3.3.0 (#4178) (06dbb5d)
• deps: bump actions/download-artifact from 8.0.1 to 8.0.2 (#4110) (2404d04)
• deps: bump actions/upload-artifact from 7.0.1 to 7.0.2 (#4107) (0adf711)
• deps: bump github/codeql-action/analyze from 4.38.2 to 4.39.0 (#4177) (1e42321)
• deps: bump openai/codex-action from 1.12 to 1.13 (#4070) (904ed3f)
• update generated-code verification and preserve SDK helpers (#4186) (9412200)

... (truncated)

Changelog

Sourced from openai's changelog.

3.41.0 (2026-10-28)

Features

• api: accept typed references in Realtime response inputs (#4189) (8ecb487)
• api: add response tool usage models (#4201) (55fc500)
• api: filter agent sessions by metadata (#4200) (77fd116)
• websocket: recover queued binary frames after reconnect (#4191) (f079868)

Bug Fixes

• examples: detect incomplete WebSocket responses (#4187) (8559299)
• examples: handle interruptions and disconnects in push-to-talk (#4192) (35fc9f3)
• keep consumer fixtures clean under strict Pyright (#4204) (5012565)

Chores

• api: document upcoming model shutdowns (#4199) (5b43cce)
• api: update image-model documentation (#4198) (fa67521)
• enforce generated-file ownership (#4193) (1b37b9a)
• migrate development type checking to ty (#4202) (45b9531)

3.40.0 (2026-10-27)

Features

• api: add prompt cache breakpoint TTL options (#4188) (3596d18)

Bug Fixes

• ci: use an explicit browser demo dependency (#4185) (a3f4302)

Chores

• api: record completed generated-code separation (#4190) (d7ab5ee)
• deps-dev: bump griffe from 1.14.0 to 2.3.0 (#4069) (766b849)
• deps: bump actions/create-github-app-token from 3.2.0 to 3.3.0 (#4178) (06dbb5d)
• deps: bump actions/download-artifact from 8.0.1 to 8.0.2 (#4110) (2404d04)
• deps: bump actions/upload-artifact from 7.0.1 to 7.0.2 (#4107) (0adf711)
• deps: bump github/codeql-action/analyze from 4.38.2 to 4.39.0 (#4177) (1e42321)
• deps: bump openai/codex-action from 1.12 to 1.13 (#4070) (904ed3f)
• update generated-code verification and preserve SDK helpers (#4186) (9412200)

3.39.0 (2026-10-26)

... (truncated)

Commits

• b7d440b release: 3.41.0 (#4195)
• 5012565 fix: keep consumer fixtures clean under strict Pyright (#4204)
• 45b9531 chore: migrate development type checking to ty (#4202)
• 35fc9f3 fix(examples): handle interruptions and disconnects in push-to-talk (#4192)
• 55fc500 feat(api): add response tool usage models (#4201)
• 77fd116 feat(api): filter agent sessions by metadata (#4200)
• 5b43cce chore(api): document upcoming model shutdowns (#4199)
• f079868 feat(websocket): recover queued binary frames after reconnect (#4191)
• 8559299 fix(examples): detect incomplete WebSocket responses (#4187)
• fa67521 chore(api): update image-model documentation (#4198)
• Additional commits viewable in compare view

Updates starlette from 1.7.0 to 1.10.0

Release notes

Sourced from starlette's releases.

Version 1.10.0

📝 Leaner URL-encoded form parsing

Starlette 1.10.0 parses URL-encoded forms with less memory and CPU, no longer drops the last field when a stream ends early, and lets you compare FormData that holds several uploads under one key.

uv pip install starlette==1.10.0

• Forms with many fields parse faster and use less memory (#3724). FormParser builds each field in the parser callbacks instead of queueing every event and replaying it. On CodSpeed, 1000 fields went from 42.8 ms to 34.9 ms and from 63.5 KB to 54.5 KB, and the max_fields limit case went from 6,880 B to 864 B.
• The last field is kept when the request stream ends without an empty chunk (#3599). The parser is now finalized once the stream is exhausted, so custom streams that never yield a final b"" no longer lose their last field.
• FormData with repeated uploads can be compared (#3619). Comparing two FormData objects with several UploadFile values under the same key used to raise TypeError.

⚠️ Behaviour changes

• Multidict equality keeps the order of values under a repeated key (#3619). Keys can still appear in any order, but QueryParams("a=1&a=2") and QueryParams("a=2&a=1") are no longer equal, matching what getlist() returns.
• starlette.formparsers.FormMessage is deprecated (#3724). The parser no longer uses it, and FormParser.messages is gone. Calling FormMessage(...) emits a StarletteDeprecationWarning.

🧭 Routing and URL generation

• Path parameter values that look like a placeholder are kept as they are (#3725). url_path_for("route", category="{id}", id="42") on /{category}/{id} now returns /{id}/42 instead of /42/42. Substitution also got faster, from 549 ns to 520 ns with two parameters and from 1,078 ns to 779 ns with five.
• Route matching does less work per request (#3665, #3722). Match members are compared by identity, cutting up to 4.9% of the instructions per dispatch in the routing benchmarks.
• WSGIMiddleware strips root_path only at a path segment boundary (#3718). A request to /application under a root_path of /app no longer becomes /lication.
• FloatConvertor reports NaN and infinity correctly (#3716). NaN was reported as a negative float.

🛡️ Middleware, responses and config

• Handled exceptions no longer keep requests alive (#3205). Starlette clears the traceback after an exception handler runs, so an exception instance reused across requests, such as a module-level HTTPException, no longer pins per-request objects in memory.
• Middlewares accept http.response.start without headers (#3721). BaseHTTPMiddleware, GZipMiddleware and SessionMiddleware default the omitted field to an empty list instead of raising KeyError.
• FileResponse ignores Range on methods other than GET and HEAD (#3679). A POST endpoint that returns a file now sends the full file, as RFC 9110 requires.
• Config strips one pair of matching outer quotes from env file values (#3720). KEY='say "hi"' now reads as say "hi" instead of say "hi.

Full changelog: 1.9.0...1.10.0

Version 1.9.0

This release streams TestClient response bodies, so client.stream() returns before the application finishes. It also escapes path delimiters when building URLs, matches host names case-insensitively, and keeps every existing Vary header.

Changed

• Stream TestClient response bodies, so client.stream() returns before the application finishes. Closing an unfinished streamed response now cancels the application #3491.

Fixed

• Percent-encode ?, #, \ and a leading // in the request path when building request.url, and \ and a leading // in url_path_for #3671.
• Use the byte length for Content-Length when the response body is a typed or multidimensional memoryview #3622.
• Include HEAD in the Allow header of HTTPEndpoint 405 responses when GET is defined #3585.
• Match host names case-insensitively in Host routing and TrustedHostMiddleware #3542.
• Preserve every existing Vary header in MutableHeaders.add_vary_header(), used by GZipMiddleware and SessionMiddleware #3675.

Full changelog: 1.8.0...1.9.0

... (truncated)

Changelog

Sourced from starlette's changelog.

1.10.0 (October 24, 2026)

Changed

• Compare multidicts such as FormData and QueryParams by key while keeping the order of values under a repeated key, so FormData with repeated uploads no longer raises TypeError. QueryParams("a=1&a=2") and QueryParams("a=2&a=1") are no longer equal #3619.
• Build URL-encoded form fields in the parser callbacks, lowering memory and CPU usage for forms with many fields #3724.
• Compare Match members by identity during routing #3665 and #3722.

Deprecated

• Deprecate starlette.formparsers.FormMessage, which is no longer used #3724.

Fixed

• Parse the last URL-encoded form field when the request stream ends without an empty chunk #3599.
• Keep path parameter values that look like another placeholder in url_path_for() #3725.
• Default omitted http.response.start headers in BaseHTTPMiddleware, GZipMiddleware and SessionMiddleware instead of raising KeyError #3721.
• Clear the traceback of handled exceptions, so reused exception instances no longer keep per-request objects alive #3205.
• Strip only one pair of matching outer quotes from Config env file values #3720.
• Strip root_path from PATH_INFO only at a path segment boundary in WSGIMiddleware #3718.
• Report NaN and infinity correctly in FloatConvertor.to_string() #3716.
• Ignore Range on requests other than GET and HEAD in FileResponse #3679.

1.9.0 (October 18, 2026)

Changed

• Stream TestClient response bodies, so client.stream() returns before the application finishes. Closing an unfinished streamed response now cancels the application #3491.

Fixed

• Percent-encode ?, #, \ and a leading // in the request path when building request.url, and \ and a leading // in url_path_for #3671.
• Use the byte length for Content-Length when the response body is a typed or multidimensional memoryview #3622.
• Include HEAD in the Allow header of HTTPEndpoint 405 responses when GET is defined #3585.
• Match host names case-insensitively in Host routing and TrustedHostMiddleware #3542.
• Preserve every existing Vary header in MutableHeaders.add_vary_header(), used by GZipMiddleware and SessionMiddleware #3675.

1.8.0 (October 13, 2026)

This release drops Python 3.10 support and adds HTTP 103 Early Hints.

Added

• Add Request.send_early_hints() for HTTP 103 Early Hints through the ASGI http.response.early_hint extension #3435.

Changed

• Require Python 3.11 or later, dropping support for Python 3.10 #3592.
• Exclude native gRPC media types (application/grpc and application/grpc+...) from GZipMiddleware compression by default #3565.
• Send http.response.start immediately in GZipMiddleware for responses with an existing Content-Encoding, status 206, or an excluded content type, instead of waiting for the first body chunk #3631.

... (truncated)

Commits

• 8aa828c Version 1.10.0 (#3728)
• a85a5c3 Substitute URL path parameters in a single pass (#3725)
• 6b8b612 Finalize URL-encoded form parsing at stream exhaustion (#3599)
• b973efd Build URL-encoded form fields in parser callbacks (#3724)
• d7731e0 Compare FormData with repeated uploads without raising TypeError (#3619)
• fb3e818 Compare remaining Match members by identity in routing (#3722)
• fc23aa1 Compare Match members by identity in Router.app (#3665)
• fe000fb Replace the removed route decorator hint in the host routing docs (#3713)
• 4114432 Default omitted http.response.start headers in middlewares (#3721)
• 1a01a02 Clear exc.traceback after handled exceptions to prevent memory leak (#3205)
• Additional commits viewable in compare view

Updates pytest from 9.1.1 to 9.2.0

Release notes

Sourced from pytest's releases.

9.2.0

pytest 9.2.0 (2026-10-27)

Removals and backward incompatible breaking changes

• #14716: -c now raises a usage error for a path that does not exist, is a directory, or is in a format pytest has no loader for. Such paths were previously either ignored or reported as an unhandled FileNotFoundError.

• #8593: The indirect, ids and scope arguments of Metafunc.parametrize() <pytest.Metafunc.parametrize> -- and therefore of @pytest.mark.parametrize <pytest.mark.parametrize ref> -- are now keyword-only.

  Passing them positionally was a common source of confusing errors, e.g. parametrize("arg1", "arg2", [(1, 2)]) made the argvalues land in indirect and failed with indirect fixture '(1, 2)' doesn't exist; such mistakes now fail with an understandable TypeError. The type annotations have always declared these arguments as keyword-only.

• #993: Plugin import failures at startup are now reported and classified instead of escaping as a raw traceback with the accidental exit code 1:

  • a plugin that cannot be found -- via -p, pytest_plugins or PYTEST_PLUGINS -- exits with pytest.ExitCode USAGE_ERROR (4), like a conftest.py that fails to import;
  • a plugin that is found but raises while importing -- including a broken pytest11 entry point or a missing plugin dependency -- exits with pytest.ExitCode INTERNAL_ERROR (3), with the traceback preserved.

  pytest.main now returns these codes instead of raising ImportError. Also fixed an IndexError from pytest's own internals when the plugin raised an exception with no arguments.

Deprecations (removal in next major release)

• #14742: The private _pytest.python.CallSpec2 name has been renamed to CallSpec. The old CallSpec2 alias remains available for now and will be removed in pytest 10 or 11.

• #14755: Schedule yield_fixture decorator for removal in pytest 10 or 11.

• #14774: Similarly to the deprecation of class-scoped fixtures defined as instance methods in pytest 9.1, all higher-scoped fixtures (module, package, and session) defined as instance methods are now deprecated.

  They suffer from the same issue the original deprecation only fixed for class-scoped fixtures, see 10819 and 14011.

• #14808: Passing a non-string value to a string-typed ini option now emits a pytest.PytestRemovedIn10Warning. This affects [tool.pytest.ini_options] in pyproject.toml. In pytest 10 this will raise a TypeError, matching the behavior of the corresponding TOML config path.

• #7323: Passing bytes to TerminalReporter.write_line() is now deprecated, pass a str instead.

New features

• #14743: Full diff: now contains a legend explaining the + & - symbols in diff output.
• #2781: When strict_markers is enabled, marker names used in -m expressions are now validated against registered markers.
• #6881: Added the parametrize_long_str_id_strategy configuration option to control how long str/bytes parameter values are handled in auto-generated test IDs. Explicit IDs set via ids=[...] or pytest.param(..., id=...) are not affected. Available strategies: short (default, values over 100 chars fall back to argname+index), sha256 (SHA-256 hex digest), legacy (keep full value), and disallow (error requesting explicit IDs).

Improvements in existing functionality

• #10745: A RecursionError whose recursion origin cannot be located now shows only the first and last 10 stack frames, matching what pytest already did when locating the origin failed outright. Previously the full traceback -- often around a thousand near-identical frames -- was rendered, which was both slow and unreadable. Pass --full-trace to see every frame.

... (truncated)

Commits

• ae1d0ff Prepare release version 9.2.0
• 143ce31 pre-commit: exclude codespell on generated release announcements
• 532dc5f tox: add attrs dependency to regen, needed in reportingdemo.rst
• a7b2446 Merge pull request #15245 from pytest-dev/claude/project-thread-9jcf91
• 679d09c [pre-commit.ci] pre-commit autoupdate (#15266)
• 41d12cd docs: clarify typing for yield fixtures
• b9022ba Merge pull request #15182 from 1871698/docs/improve-getting-started-description
• 16f87a7 docs: improve getting-started.rst test execution description
• 082ed70 Merge pull request #15255 from pytest-dev/claude/project-thread-eoi3d6
• ae37a01 Merge pull request #15227 from pytest-dev/claude/project-thread-4etye4
• Additional commits viewable in compare view

Updates pytest from 9.1.1 to 9.2.0

Release notes

Sourced from pytest's releases.

9.2.0

pytest 9.2.0 (2026-10-27)

Removals and backward incompatible breaking changes

• #14716: -c now raises a usage error for a path that does not exist, is a directory, or is in a format pytest has no loader for. Such paths were previously either ignored or reported as an unhandled FileNotFoundError.

• #8593: The indirect, ids and scope arguments of Metafunc.parametrize() <pytest.Metafunc.parametrize> -- and therefore of @pytest.mark.parametrize <pytest.mark.parametrize ref> -- are now keyword-only.

  Passing them positionally was a common source of confusing errors, e.g. parametrize("arg1", "arg2", [(1, 2)]) made the argvalues land in indirect and failed with indirect fixture '(1, 2)' doesn't exist; such mistakes now fail with an understandable TypeError. The type annotations have always declared these arguments as keyword-only.

• #993: Plugin import failures at startup are now reported and classified instead of escaping as a raw traceback with the accidental exit code 1:

  • a plugin that cannot be found -- via -p, pytest_plugins or PYTEST_PLUGINS -- exits with pytest.ExitCode USAGE_ERROR (4), like a conftest.py that fails to import;
  • a plugin that is found but raises while importing -- including a broken pytest11 entry point or a missing plugin dependency -- exits with pytest.ExitCode INTERNAL_ERROR (3), with the traceback preserved.

  pytest.main now returns these codes instead of raising ImportError. Also fixed an IndexError from pytest's own internals when the plugin raised an exception with no arguments.

Deprecations (removal in next major release)

• #14742: The private _pytest.python.CallSpec2 name has been renamed to CallSpec. The old CallSpec2 alias remains available for now and will be removed in pytest 10 or 11.

• #14755: Schedule yield_fixture decorator for removal in pytest 10 or 11.

• #14774: Similarly to the deprecation of class-scoped fixtures defined as instance methods in pytest 9.1, all higher-scoped fixtures (module, package, and session) defined as instance methods are now deprecated.

  They suffer from the same issue the original deprecation only fixed for class-scoped fixtures, see 10819 and 14011.

• #14808: Passing a non-string value to a string-typed ini option now emits a pytest.PytestRemovedIn10Warning. This affects [tool.pytest.ini_options] in pyproject.toml. In pytest 10 this will raise a TypeError, matching the behavior of the corresponding TOML config path.

• #7323: Passing bytes to TerminalReporter.write_line() is now deprecated, pass a str instead.

New features

• #14743: Full diff: now contains a legend explaining the + & - symbols in diff output.
• #2781: When strict_markers is enabled, marker names used in -m expressions are now validated against registered markers.
• #6881: Added the parametrize_long_str_id_strategy configuration option to control how long str/bytes parameter values are handled in auto-generated test IDs. Explicit IDs set via ids=[...] or pytest.param(..., id=...) are not affected. Available strategies: short (default, values over 100 chars fall back to argname+index), sha256 (SHA-256 hex digest), legacy (keep full value), and disallow (error requesting explicit IDs).

Improvements in existing functionality

• #10745: A RecursionError whose recursion origin cannot be located now shows only the first and last 10 stack frames, matching what pytest already did when locating the origin failed outright. Previously the full traceback -- often around a thousand near-identical frames -- was rendered, which was both slow and unreadable. Pass --full-trace to see every frame.

... (truncated)

Commits

• ae1d0ff Prepare release version 9.2.0
• 143ce31 pre-commit: exclude codespell on generated release announcements
• 532dc5f tox: add attrs dependency to regen, needed in reportingdemo.rst
• a7b2446 Merge pull request #15245 from pytest-dev/claude/project-thread-9jcf91
• 679d09c [pre-commit.ci] pre-commit autoupdate (#15266)
• 41d12cd docs: clarify typing for yield fixtures
• b9022ba Merge pull request #15182 from 1871698/docs/improve-getting-started-description
• 16f87a7 docs: improve getting-started.rst test execution description
• 082ed70 Merge pull request #15255 from pytest-dev/claude/project-thread-eoi3d6
• ae37a01 Merge pull request #15227 from pytest-dev/claude/project-thread-4etye4
• Additional commits viewable in compare view

Updates cmcp-runtime from 0.7.0 to 0.7.1

Release notes

Sourced from cmcp-runtime's releases.

cmcp 0.7.1

Three security fixes, backported onto 0.7.0 so they ship without the breaking Agent Manifest binding change and the opaque provider removal waiting on main. Full detail in CHANGELOG.

• GHSA-frxc-7f57-6f4r: deny a Cedar Allow that comes with evaluation errors, instead of skipping the errored policy and allowing.
• GHSA-mrcx-xvv4-m7f8: bound the TPM AIA certificate fetch. Each hop resolves the host once, connects only to that checked address, and the hop and byte limits hold.
• GHSA-2p85-666x-xm7x: classify upstream drift against the drifted tool's own catalog entry.

SDK on PyPI as cmcp-runtime 0.7.1.

Changelog

Sourced from cmcp-runtime's changelog.

[0.7.1] - 2026-10-19

Security

• A Cedar Allow reached with evaluation errors was enforced as Allow (GHSA-frxc-7f57-6f4r). Cedar skips a policy whose condition errors, so a forbid reading a context attribute the current phase does not carry could not stop a matching permit. CedarBackend now denies any Allow that comes with evaluation errors, and the error reaches the tool_call or egress_denied audit row. So that a policy written for one phase does not deny in the other, both phases now carry the same context keys: egress adds session_max_sensitivity (the label string), arguments, baa_covered, compliance_domain, server_identity, destination_class and attestation_platform, and ingress adds egress (false), sensitivity_level, injection_events, reset_count and response_size_bytes (0). Without this the shipped BFSi demo denied every response at egress. A policy that reads a key outside that set, or workflow_id when the caller declared none, now denies.
• The TPM AIA certificate fetch was unbounded (GHSA-mrcx-xvv4-m7f8). The chain walk followed redirects to any address and read responses without a limit. It now refuses destinations that do not resolve to global addresses, checks every redirect the same way, follows at most three and ignores responses over 64 KiB. Each hop's host is resolved once and the connection goes to the checked address (the request still names the host, and HTTPS still verifies the certificate against it), so a second DNS answer cannot point the fetch somewhere else. The fetch no longer uses environment proxies. An on-prem CA whose AIA host has a private address falls back to the vendored bundle.
• Upstream drift was classified against the wrong catalog entry (GHSA-2p85-666x-xm7x). The scanner received the server of whichever entry was last in the catalog. Detection and fail-closed denial were not affected; only the classification written to the audit row was.

Commits

• 4a7163a fix(security): cmcp-runtime 0.7.1 backport (Cedar errors, AIA fetch, drift cl...
• See full diff in compare view

Updates starlette from 1.7.0 to 1.10.0

Release notes

Sourced from starlette's releases.

Version 1.10.0

📝 Leaner URL-encoded form parsing

Starlette 1.10.0 parses URL-encoded forms with less memory and CPU, no longer drops the last field when a stream ends early, and lets you compare FormData that holds several uploads under one key.

uv pip install starlette==1.10.0

• Forms with many fields parse faster and use less memory (#3724). FormParser builds each field in the parser callbacks instead of queueing every event and replaying it. On CodSpeed, 1000 fields went from 42.8 ms to 34.9 ms and from 63.5 KB to 54.5 KB, and the max_fields limit case went from 6,880 B to 864 B.
• The last field is kept when the request stream ends without an empty chunk (#3599). The parser is now finalized once the stream is exhausted, so custom streams that never yield a final b"" no longer lose their last field.
• FormData with repeated uploads can be compared (#3619). Comparing two FormData objects with several UploadFile values under the same key used to raise TypeError.

⚠️ Behaviour changes

• Multidict equality keeps the order of values under a repeated key (#3619). Keys can still appear in any order, but QueryParams("a=1&a=2") and QueryParams("a=2&a=1") are no longer equal, matching what getlist() returns.
• starlette.formparsers.FormMessage is deprecated (#3724). The parser no longer uses it, and FormParser.messages is gone. Calling FormMessage(...) emits a StarletteDeprecationWarning.

🧭 Routing and URL generation

• Path parameter values that look like a placeholder are kept as they are (#3725). url_path_for("route", category="{id}", id="42") on /{category}/{id} now returns /{id}/42 instead of /42/42. Substitution also got faster, from 549 ns to 520 ns with two parameters and from 1,078 ns to 779 ns with five.
• Route matching does less work per request (#3665, #3722). Match members are compared by identity, cutting up to 4.9% of the instructions per dispatch in the routing benchmarks.
• WSGIMiddleware strips root_path only at a path segment boundary (#3718). A request to /application under a root_path of /app no longer becomes /lication.
• FloatConvertor reports NaN and infinity correctly (#3716). NaN was reported as a negative float.

🛡️ Middleware, responses and config

• Handled exceptions no longer keep requests alive (#3205). Starlette clears the traceback after an exception handler runs, so an exception instance reused across requests, such as a module-level HTTPException, no longer pins per-request objects in memory.
• Middlewares accept http.response.start without headers (#3721). BaseHTTPMiddleware, GZipMiddleware and SessionMiddleware default the omitted field to an empty list instead of raising KeyError.
• FileResponse ignores Range on methods other than GET and HEAD (#3679). A POST endpoint that returns a file now sends the full file, as RFC 9110 requires.
• Config strips one pair of matching outer quotes from env file values (#3720). KEY='say "hi"' now reads as say "hi" instead of say "hi.

Full changelog: 1.9.0...1.10.0

Version 1.9.0

This release streams TestClient response bodies, so client.stream() returns before the application finishes. It also escapes path delimiters when building URLs, matches host names case-insensitively, and keeps every existing Vary header.

Changed

• Stream TestClient response bodies, so client.stream() returns before the application finishes. Closing an unfinished streamed response now cancels the application #3491.

Fixed

• Percent-encode ?, #, \ and a leading // in the request path when building request.url, and \ and a leading // in url_path_for #3671.
• Use the byte length for Content-Length when the response body is a typed or multidimensional memoryview #3622.
• Include HEAD in the Allow header of HTTPEndpoint 405 responses when GET is defined #3585.
• Match host names case-insensitively in Host routing and TrustedHostMiddleware #3542.
• Preserve every existing Vary header in MutableHeaders.add_vary_header(), used by GZipMiddleware and SessionMiddleware #3675.

Full changelog: 1.8.0...1.9.0

... (truncated)

Changelog

Sourced from starlette's changelog.

1.10.0 (October 24, 2026)

Changed

• Compare multidicts such as FormData and QueryParams by key while keeping the order of values under a repeated key, so FormData with repeated uploads no longer raises TypeError. QueryParams("a=1&a=2") and QueryParams("a=2&a=1") are no longer equal #3619.
• Build URL-encoded form fields in the parser callbacks, lowering memory and CPU usage for forms with many fields #3724.
• Compare Match members by identity during routing

[carloshvp]

Reviewed commit e8df6944bd1ce35da307fda6be644e781d61f4f8.

Two upgraded requirement sets are unsatisfiable. The presenter job combines agent-manifest 0.15.0 with cmcp-runtime 0.7.1, whose published metadata requires agent-manifest >=0.12,<0.13. The sovereign connector job combines cedarpy >=4.12.2 with ca2a-runtime 0.4.0, which requires cedarpy ==4.12.1.

Choose a jointly compatible set of released packages and regenerate each affected lock. The demos lock's agent-manifest downgrade from 0.15.0 to 0.12.0 also needs an explicit compatibility/security review rather than implicit resolver acceptance. Preserve platform-specific pins and rerun the affected example/demo jobs.