* fix(inspector): release the upstream stream on Express client disconnect (#2394)

* fix(inspector): release the upstream stream on Express client disconnect

The Express adapter proxied responses with a read loop whose only exit was
upstream EOF, and passed no signal into the outbound fetch. A client leaving
during a long-lived SSE response left the loop writing to a dead socket while
the upstream connection stayed open.

Wire an AbortController to res close and thread its signal through the fetch
and the read loop, matching packages/server/src/node-bridge.ts.

* fix(inspector): carry the abort through the proxy's outbound fetch

The Express adapter aborted on client disconnect but the signal stopped at
the Hono proxy, so the outbound fetch and the upstream MCP connection kept
running. Pass the inbound signal into that fetch, which covers every
iteration of the manual redirect loop, and stop forwarding the resulting
abort to Express as an application error.

The existing test aborted only after the first chunk, so it proved reader
cancellation and would have passed without this. Added a case that
disconnects before the upstream sends headers, where cancelling the reader
cannot help, and cleared the race timeout so it stops holding the worker.

* fix(inspector): narrow abort handling to the expected AbortError

* fix(inspector): match the abort by reason, not only by error shape

fetch rejects with signal.reason verbatim, and the reason is not ours to
choose. The Express middleware calls abort() with no argument, so there it
is a DOMException named AbortError. @hono/node-server, which backs the
standalone inspector and the vite dev API, aborts with a plain string
instead, so an instanceof Error test misses every disconnect on that path
and the request still ends in a logged stack and a 500 written to a socket
that has already gone.

Accept the reason by identity as well, which covers both without assuming
who created the signal.

* fix(inspector): preserve stream failure semantics and await buffered proxy reads

Two follow-ups from review.

res.end() moves out of the finally block onto the success path. Ending
there also ended the downstream response when reader.read() threw on a
genuine upstream failure, so the Express error handler received the
error with res.writableEnded already true and could no longer change
the response. Only listener cleanup stays in finally.

proxyResponse(response) gains return await. It awaits
response.arrayBuffer() for non-streaming responses, and returning the
promise unawaited let a later rejection bypass the catch that turns a
client disconnect into a quiet cancellation. With the signal now
propagated, disconnecting after JSON headers arrive but before the body
completes rejects that read.

Regression coverage for both. The stream case destroys the upstream
socket after its first chunk and asserts the client cannot read the
response to completion, which fails if res.end() runs in finally. The
buffered case disconnects after headers and asserts the server still
serves the next request rather than taking it down with an escaped
rejection.

Both assert on observable behaviour rather than on metadata, since the
existing test of this path asserts only that a handler was mounted and
so cannot fail when the behaviour breaks.

* fix(inspector): do not end an aborted response, and stop overclaiming a test

Both Cubic findings on the previous commit are valid; verified each by
running it rather than by reading.

Cancelling the reader resolves a pending read with done, so the loop
broke and res.end() ran on a response the client had already abandoned.
res.end() is now guarded by !signal.aborted, so only a completed
upstream stream ends the downstream response.

The buffered-body test claimed to guard the return await change and did
not. Measured: reverting the await leaves all tests green, so the claim
was false. Driving the Fetch handler directly and asserting a 499, as
suggested, does not distinguish it either, because every layer below
also catches the escaped rejection. Rather than ship a test whose
docstring asserts something it cannot detect, the test now says what it
does prove, that an abort mid buffered-body settles as 499 rather than
rejecting, and says plainly that it is a behaviour lock and not a
regression guard for that change. It also synchronises on the upstream
having sent headers instead of a fixed timeout, so the abort lands
mid-body rather than in the pre-header path.

Coverage, stated honestly: the stream fix is guarded, and reverting it
fails the stream test. The abort guard and the return await are not
covered. Both are correct by inspection and neither produced an
observable difference from outside the proxy in anything I tried. If
there is a shape that distinguishes them, I will add it.

* chore: enter canary prerelease mode

* chore(typescript): version packages (canary)

* fix(client): decode MCP App HTML blobs as UTF-8 (#2632)

* chore(typescript): version packages (canary)

* fix(server): forward cancellation to proxied prompts (#2568)

* fix(server): forward cancellation to proxied prompts

Signed-off-by: Elioooon <[email protected]>

* test(server): require proxied prompt invocation

---------

Signed-off-by: Elioooon <[email protected]>

* chore(typescript): version packages (canary)

* fix(cli): route mcp-use servers env --help to env command help (#2690)

* chore(typescript): version packages (canary)

* fix(cli): reject an inline value on boolean flags instead of discarding it (#2521)

* fix(cli): reject an inline value on boolean flags instead of discarding it

parseArgs split `--flag=value` for every `--` token but the boolean cases
never read the inline value, so `--tunnel=false` parsed as `--tunnel` and
opened a public tunnel. takeValue now records that it consumed the inline
value, and any flag reaching the end of the loop with an unconsumed inline
value is rejected.

Bare booleans, `--flag value`, and `--flag=value` on value-taking flags are
unaffected.

* chore(cli): bump the changeset to minor for the rejected inline value

Rejecting `--tunnel=false`, which earlier versions accepted, changes
behaviour for existing invocations, so a patch bump understates it.

* test(cli): fold the boolean-flag check into the existing parser suite

Per review: drop the separate test file and keep a single case in the
existing parseArgs suite. The bare-form behaviour of every boolean flag is
already covered there, so the remaining cases were redundant.

---------

Co-authored-by: Aishwary Dongre <[email protected]>

* chore(typescript): version packages (canary)

* fix(agent): stop dispatching tools in runToolLoopNonStreaming when signal is aborted (#2667)

* fix(agent): stop dispatching tools in runToolLoopNonStreaming when signal is aborted

* test(agent): add abort signal coverage for OpenAIResponsesDriver.runToolLoopNonStreaming

* test(agent): use afterEach restoreAllMocks for fetch spy cleanup

* chore(typescript): version packages (canary)

* chore(deps-dev): bump vitest from 4.1.10 to 4.1.11 in /libraries/typescript/packages/server/examples/sentry (#2627)

* chore(deps-dev): bump vitest

Bumps [vitest](https://github.com/vitest-dev/vitest/tree/HEAD/packages/vitest) from 4.1.10 to 4.1.11.
- [Release notes](https://github.com/vitest-dev/vitest/releases)
- [Changelog](https://github.com/vitest-dev/vitest/blob/main/docs/releases.md)
- [Commits](https://github.com/vitest-dev/vitest/commits/v4.1.11/packages/vitest)

---
updated-dependencies:
- dependency-name: vitest
  dependency-version: 4.1.11
  dependency-type: direct:development
...

Signed-off-by: dependabot[bot] <[email protected]>

* chore: add changeset for dependabot updates

---------

Signed-off-by: dependabot[bot] <[email protected]>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* chore(typescript): version packages (canary)

* fix(agent): wrap Gemini array tool results in an object (#2675)

* fix(agent): wrap Gemini array tool results in an object

AI assistance: OpenAI Codex; research support from Instinct and Jev.

* test(agent): trim Gemini tool-result cases to the reviewed table

Signed-off-by: Charan Rathore <[email protected]>

---------

Signed-off-by: Charan Rathore <[email protected]>

* chore(typescript): version packages (canary)

* feat(client): expose OAuth UserInfo in useMcp (#2688)

* feat(client): expose OAuth UserInfo in useMcp

* fix(client): address UserInfo review feedback

* chore(typescript): version packages (canary)

* chore(changelog): update changelogs for v20.3.14 and v2.7.2 releases

- Added details for v20.3.14 including fixes for Express client disconnection and enhancements to the OAuth proxy.
- Documented new features and fixes in v2.7.2, highlighting the exposure of OAuth UserInfo and stricter CLI flag handling.
- Updated references to the Inspector changelog in the TypeScript changelog.

* fix: stop a false vitest release and harden two canary tests (#2710)

The Sentry example's vitest bump cannot apply under the workspace override and would publish a false mcp-use note. Keep the proxy stream test and pre-aborted OpenAI fetch from failing for the wrong reason, and document UserInfo proxying.

* chore(typescript): version packages (canary)

---------

Signed-off-by: Elioooon <[email protected]>
Signed-off-by: dependabot[bot] <[email protected]>
Signed-off-by: Charan Rathore <[email protected]>
Co-authored-by: Mohammed Alkindi <[email protected]>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Steven Huang <[email protected]>
Co-authored-by: Elioooon <[email protected]>
Co-authored-by: Rohith Reddy Nama <[email protected]>
Co-authored-by: aishwary-dongre <[email protected]>
Co-authored-by: Aishwary Dongre <[email protected]>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Charan Rathore <[email protected]>
Co-authored-by: Enrico Toniato <[email protected]>