Copilot review round 2 (suppressed comments), all three real:
- `require.resolve("<pkg>/package.json")` is governed by the package's
`exports`, and Node keeps no special case for `./package.json` — so a
package declaring an `exports` map without it threw
ERR_PACKAGE_PATH_NOT_EXPORTED even though its bin was installed and
spawnable, a false "not installed" pointing at `npm install` for a
package already on disk. Walk `resolve.paths()` instead, which never
consults `exports`. The docblock claiming the opposite is corrected.
- `shell: true` (needed at all on Windows to start the `.cmd` shims) makes
Node hand `cmd.exe` one space-joined string, so an argument holding a
path with a space splits in two. Every generated path in
pack-and-verify lives under `tmpdir()`, which sits beneath the user
profile. Quote them via `shellArgs()` at all four spawn sites, the
command path included — the hazard predates this PR, so fix the class.
- Document the new module: `scripts/lib/resolve-node-bin.mjs` in AGENTS.md's
scripts/ tree, its tests in the README's `test:scripts` row.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01387r3MotGk52hrkfjzHFsN
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review on #1997:
- A string-form `bin` is npm's shorthand for ONE command, named after the
package, so ignoring `binName` let a typo silently resolve the package's
only executable while bypassing the documented "declares no such bin"
failure. Match it against the manifest's unscoped name instead.
- A declared bin whose file is absent (partial install) was returned
unchecked. `process.execPath <missing>` still spawns and exits 1 with
empty stdout, which `rawProjectListing` records as "no diagnostic
captured" — reproducing the bogus every-file-uncovered report this
helper exists to eliminate. Throw here, where the remedy is actionable.
Also restores `rawProjectListing`'s docblock, which the new `tscEntry`
block had orphaned above itself.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01387r3MotGk52hrkfjzHFsN
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot round 3 (PR #2039): clearing the banner when `selectedServer`
changes is not enough. Start disconnecting A, switch A → B → A, and by
the time the original request rejects both guards match again — the
attempt token is unchanged and the selected server is A once more — so
the stale failure lands on the freshly re-selected A. Bump the token on
every switch, so a round trip invalidates prior disconnects the same way
a one-way switch does.
New test drives exactly that interleaving; it fails without the bump
(the one-way test does not, which is why this needed its own).
Also attaches the before/after TUI screenshots the review asked for —
the review was right that "no UI change" stopped being true once the
banner was added — and corrects the PR description accordingly.
Signed-off-by: cliffhall <cliff@futurescale.com>
Two suppressed review findings.
With an override configured, the wrapper cloned and parsed every
successful application/json response — every tools/call result and every
resource payload on the direct CLI/TUI transports — only to conclude it
was not a metadata document. It now tests the request URL for the RFC
8414 / OIDC well-known paths first, so ordinary traffic costs one string
test and is returned untouched.
The read path also trusted StoredMCPServer's `string` type for the two
override fields, which a hand-edited mcp.json does not honor: the CLI and
TUI read the file directly and then `.trim()` these, so a number would
crash the server load. Guarded with cleanEndpointOverride, mirroring
cleanAuthorizationParams.
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot round 2 (PR #2039) surfaced no new visible comments but six
suppressed ones. Four were actionable.
`clients/tui/src/App.tsx` — a stale-completion race in the new banner:
start disconnecting server A, switch to B, then let A's disconnect
reject, and A's failure appears under B. Rapid same-server retries had
the same shape, and the stale attempt is the one most likely to reject.
Claim a monotonic attempt token before awaiting and publish only if it
is still current and the server has not changed; `handleConnect` bumps
the same counter so a connect retires an in-flight disconnect. The
rejection is still owned when superseded — just not published. The
regression test drives a deferred rejection across a server switch, and
fails without the guard.
`tsconfig.lint.json` — the root type-aware block matches
`test-servers/src/**/*.tsx` but the parser project omitted that
extension, so such a file would have failed with "not found in any of
the provided project(s)" instead of being linted. Added, and AGENTS.md
now states the invariant: the include must stay a superset of the lint
globs.
`clients/web/server/{server,vite-hono-plugin}.ts` — the `open()`
rejection branch was untested, and duplicated. Both were fixed by
removing the duplication: `openBrowser` now lives beside
`printServerBanner` in `web-server-config.ts`, with one implementation
to test. Two copies of a one-line `.catch` is two chances for one to
drift back to a bare `open(url)`. The dev plugin is not booted in tests
— it returns early under VITEST by design — but it calls the same
tested helper.
`core/mcp/remote/node/server.ts` — `closeAbortedStream`'s rejecting-close
path, the entire reason the helper exists, had no test. Exported it (as
this file already does for `requestIdForSendWait` / `mcpParamHeadersOnly`)
and covered both legs, asserting no unhandledRejection fires.
Also corrects AGENTS.md, which said each client lists "both" tsconfig
projects — true for cli/tui/launcher, but web lists four leaf projects.
Signed-off-by: cliffhall <cliff@futurescale.com>
Two review findings on the fetch wrapper:
The content-type test was a substring match for "json", which also
matches `application/x-ndjson` — how the streamable-HTTP transport can
serve its long-lived server-push channel (`isLongLivedStreamResponse`).
Awaiting `.json()` on a clone of an unbounded stream never resolves, so
the MCP connection would hang for as long as an override was configured.
It now matches the media type exactly (`application/json` or a `+json`
suffix), with a regression test driven by a stream that never closes.
The overrides also leave `issuer` as discovery returned it, so pointing
one at an authorization server that advertises a different issuer is
rejected on callback by the RFC 9207 / SEP-2352 mix-up defense. That
check is deliberately not weakened; the limit is now stated on the
module, in the field description, and in the configuration guide.
Signed-off-by: cliffhall <cliff@futurescale.com>
Two review findings on the fetch wrapper:
Reading the body with `response.text()` meant every successful JSON
response was rebuilt while overrides were configured, even when it was
not a metadata document. That dropped native properties a synthesized
`Response` cannot carry (`url`, `redirected`, `type`) from responses the
wrapper has no business altering, and threw outright on a body-less
204/205, whose status Fetch forbids a body on — so an unrelated transport
response could fail. It now parses a clone and returns the caller's own
response unless a document is actually patched.
`new URL` also accepts embedded credentials while Fetch rejects a request
URL carrying them, so `https://user:pass@host/token` passed both the form
and the runtime and then failed deep in the OAuth exchange.
`oauthEndpointUrlError` now rejects a username or password.
Signed-off-by: cliffhall <cliff@futurescale.com>
Review catch (Copilot, PR #2039): folding the rejected disconnect into
`connectError` meant nobody ever saw it. InfoTab renders that value only
when `serverState.status === "error"`, and a failed disconnect leaves the
status "connected" — so the message existed in state and never on screen.
Worse, it fed `connectError ?? inspectorLastError`, where a stale value
would go on masking a later, real connection error.
Give it its own `disconnectError` state and a header banner that renders
regardless of status, cleared on the next connect or disconnect attempt
and on a server switch, so it can neither linger nor be attributed to
the wrong server. The tests now assert the rendered banner rather than
noting that nothing is visible.
Signed-off-by: cliffhall <cliff@futurescale.com>
The overrides are per-server configuration for the *resource* server's
authorization server, but the fetch wrapper that applies them is shared
with the enterprise-managed (EMA) flow, whose IdP OIDC discovery runs
through the same fetch and satisfies the metadata predicate. An
authorization override could therefore redirect the IdP login to the
resource AS, and a token override could send the IdP code or refresh
token there.
OAuthManager.getEndpointOverrides() now returns nothing when the server
is enterprise-managed — the same boundary redirectToExternalAuthorization
draws for the custom authorization parameters (#2018) — with a regression
test, the note surfaced on both form fields, and the rule documented.
Signed-off-by: cliffhall <cliff@futurescale.com>
A floated promise is invisible at review time — the call reads like an
awaited one minus four characters — and the unhandled rejection it
produces surfaces in a different test, in a different file, as a stack
pointing at SDK internals. In #1947 two un-held `callTool` calls made
`npm run ci` unpassable and took a full investigation to attribute; the
fix was two lines. `@typescript-eslint/no-floating-promises` catches the
class outright, so enable it at `error` everywhere.
The rule is type-aware and `tseslint.configs.recommended` is not, so each
scope's config gains a parser project. The four clients name their own
tsconfig projects (both of them — `src` and `__tests__`/`server` live in
different ones). The root scope has no tsconfig at all, so add
`tsconfig.lint.json`: a lint-only project covering `core/**`,
`test-servers/src/**`, and `vitest.shared.mts`, emitting nothing and
gating nothing. It sets `moduleResolution: bundler` deliberately — under
NodeNext `core/`'s extensionless imports fail to resolve, every import
degrades to `any`, and the rule silently stops seeing promises at all.
Measured surface, and the fixes:
- core (2) — `stream.close()` in two Hono `onAbort` listeners. Hono
invokes subscribers with a bare `subscriber()`, so a promise returned
from the listener is floated one layer up; `closeAbortedStream` owns
the rejection where it can actually be owned.
- tui (5) — `handleDisconnect` had no catch of its own and its only
caller is a key handler that cannot await, so a failing disconnect was
a live unhandled rejection; it now surfaces the message the way
`handleConnect` does. The rest are `void` at callers that genuinely
cannot await, over callees that already own their failures.
- web (92) — 83 are Storybook `expect(...)`, which is instrumented and
returns a promise, now awaited (`scrollAreaStoryAssertions` is async
for the same reason). Two tests were asserting against a provider call
they never awaited. `open(url)` in both web servers now warns instead
of escaping a synchronous callback.
- cli, launcher — already clean.
Type-aware linting costs time: `clients/web`'s `eslint .` goes from ~8s
to ~19s, inside `validate`. Narrowing the loaded projects is the lever if
that needs reducing later; dropping the rule is not.
Closes#1959
Signed-off-by: cliffhall <cliff@futurescale.com>
Both new #2006 regressions faked the events response with an object cast
`as unknown as Response`, which AGENTS.md rejects — it erases the check
that the mock actually satisfies the API the hook consumes.
Each now builds a real `ReadableStream` inside a real `Response`. `pull`
runs once per chunk the hook's reader consumes, so awaiting a promise
inside it gates delivery exactly the way the hand-rolled reader did,
while the mock stays type-checked against the Response contract. The
chunk counter replaces the read counter as the "the reader has consumed
the previous chunk" signal.
Both still fail against the pre-fix parser, verified.
Raised by Copilot on PR #2036.
Signed-off-by: cliffhall <cliff@futurescale.com>
Adds two optional per-server fields — `oauth.authorizationUrl` and
`oauth.tokenUrl` — that override the endpoints authorization-server
metadata discovery resolved, so a server can be pointed at a development
or staging authorization server without changing what it advertises.
SDK v2 routes both endpoints through the discovered metadata document and
neither through OAuthClientProvider, so the override is applied by
wrapping the client's base fetch and patching the metadata document in
flight (core/auth/endpointOverrides.ts). Wrapping the base fetch — rather
than only the auth fetch — also covers the discovery the SDK runs from
inside the transport on the 401/refresh path.
The fields round-trip through mcp.json, the /api/servers routes, the
runner (CLI/TUI), and the Server Settings → Authorization panel, where a
value that is not an absolute http(s) URL is flagged inline; at runtime
such a value is dropped with a single warning rather than failing the
connection.
Signed-off-by: cliffhall <cliff@futurescale.com>
Upstream #1965 moved the tsc --listFilesOnly machinery into
scripts/lib/tsc-program.mjs (shared with the new verify:dep-lockstep),
still spawning `npx` — which is ENOENT on Windows. Resolve the conflict
by taking upstream's factoring and porting the #1939 fix into the lib:
rawProjectListing and verify-typecheck-coverage's --showConfig pass now
spawn process.execPath with the cached, hard-failing tscEntry resolver
(exported from tsc-program.mjs). Fixture copies in both main tests pick
up the new lib import graph. verify:dep-lockstep now runs on Windows
too.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The `/api/servers/events` reader in `useServers` searched for `\n\n` and
consumed exactly two characters as the separator. SSE permits CRLF line
endings, so a `event: change\r\n\r\n` frame stayed buffered forever:
`isSseDataFrame()` was never reached and no background refresh fired,
leaving the server list stale until a manual refresh or an LF-delimited
event arrived.
Frame boundaries are now located with a CRLF-aware matcher that reports
the separator's real length, and frame lines split on either terminator.
A bare CR is deliberately not accepted as a terminator: it cannot be
distinguished from the first half of a CRLF straddling a chunk boundary,
so a buffer ending in `…\r\n\r` would read as a completed frame and the
`\n` arriving next would open a bogus one.
Two regressions cover it, both verified failing against the old parser.
Each uses a referentially stable `fetchFn` and holds the frame behind a
promise released after the disk write — with an inline `fetchFn` the hook
re-subscribes every render and that re-subscription's own mount refresh
picks the edit up regardless, so the test would pass unfixed.
Closes#2006
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review: the unit test covered the panel's rendering, but nothing
executed `mrtr_empty`, so the handler, its preset-registry branch, and the
showcase config entry could regress with CI green.
Adds a live modern-era test that loads `mrtr-showcase-http.json` through
`loadConfig`/`resolveConfig` — the same artifacts the README's manual
reproduction uses, so the config entry and the registry branch are covered
rather than just the fixture factory — answers the embedded elicitation, and
asserts the terminal result really is an empty `CallToolResult`: that is the
input the Tools panel must render as "completed, returned nothing".
Signed-off-by: cliffhall <cliff@futurescale.com>
The Tools screen mounts the Results panel only once a call has produced a
result, but the panel's zero-content branch rendered "No results yet" — the
same wording it would use before anything had run. An empty `content` array
with no `structuredContent` is a legal `CallToolResult`, so a call that had
plainly completed reported itself as a call that never happened.
That reads worst after a multi-round-trip (MRTR) sequence, which is how it was
reported: the user answers one or more embedded elicitations, watches the
Protocol tab group the exchange and mark it COMPLETE, and is then told there
are no results yet. Nothing was lost — the result was stored and rendered; the
panel simply had no wording for "completed, returned nothing".
The panel now renders an "Empty result" alert saying the call completed and
returned no content. This is the remaining half of the same gap #1908 closed
for `structuredContent`-only results.
test-servers: new `mrtr_empty` preset (`createMrtrEmptyResultTool`) — one
elicitation round, then a complete result with `content: []` and no
`structuredContent` — added to `mrtr-showcase-http.json` so the shape is
reproducible by hand. README documents it.
Signed-off-by: Cliff Hall <cliff@futurescale.com>
Signed-off-by: cliffhall <cliff@futurescale.com>
`ServerTasksCapability`'s `list`/`cancel` are objects, not booleans, so
the legacy-tasks fixture failed `tsc -b` in CI. A stale incremental
build cache hid it locally.
Signed-off-by: cliffhall <cliff@futurescale.com>
SEP-2663 moved task support from `capabilities.tasks` to the negotiated
`capabilities.extensions["io.modelcontextprotocol/tasks"]`, but the
Connection Info modal's capability grid tested only for the top-level
key. A modern server that advertised the extension in `server/discover`
therefore showed "Tasks ✗" while the very same extension id was listed
under "Server Extensions" two sections below.
Server capability rows now fall back to an extension id for the
capabilities the modern era expresses that way (tasks is the only one
today), gated on the negotiated era so a legacy connection is still
judged purely on its `initialize` capabilities. Keyed the same way
`InspectorClient.isTasksExtensionNegotiated()` gates the Tasks tab, so
the checkmark and the tab can't disagree.
Signed-off-by: cliffhall <cliff@futurescale.com>
Two findings from the Copilot review of #1965, both real.
1. The candidates were flattened to package names before the version
comparison, so a third install holding a different copy failed the guard
even when no program loads that copy beside another — and the diagnostic
named an install that never took part. `crossInstallPackages` now returns
the co-occurrences themselves (name → program → install → entry paths) and
`findSkew` compares per program, across only the installs that met there.
2. Folding a nested copy onto its outermost install lost the path that
identifies which copy was loaded, and the version was then read from the
install's top-level entry — absent or differently versioned, either way
letting a real pair pass. `classifyModulePath` now also returns the
`entryPath` npm keys the lockfile by, `topLevelLockVersions` becomes
`lockVersionsByPath`, and each holder is priced from the entry the program
resolved. A resolved copy with no lockfile entry is now a hard failure
rather than a silently dropped holder.
The failure output prints `in <program>` with each holder's install path
beside its version, so a nested copy is distinguishable from a top-level one.
Signed-off-by: cliffhall <cliff@futurescale.com>
`verify:dep-lockstep` derived its candidate set from the packages the shared
first-party sources named *directly*, so a package whose declarations reach a
program only through another package's `.d.ts` was invisible to it — it could
skew across installs and recreate the two-copy blow-up #1896 was about while
the guard stayed green. `@modelcontextprotocol/sdk` was the live example: never
written in first-party code, yet 16 of its `.d.ts` files land in clients/web's
test program.
Replace the derivation with the measured one: list every client tsconfig
project with `tsc --listFilesOnly`, map each resolved `node_modules` file to
its owning install and package, and keep the packages that reach ONE program
from TWO installs. That is exactly the set that can put two structurally
distinct copies of a type in front of one checker. The rest of the guard is
unchanged — top-level lockfile comparison, deny-by-default, `TOLERATED_SKEW` as
a name allowlist scoped to within-a-major.
The listing machinery moves to `scripts/lib/tsc-program.mjs`, shared with
`verify:typecheck-coverage`, so the two guards can't disagree about what a
program contains. It is deliberately not cached to disk between the two
processes: a stale fingerprint would make a guard measure a program that no
longer exists and pass on a real miss. The second pass costs ~14s.
`TOLERATED_SKEW` is now empty. `jose` and `@modelcontextprotocol/ext-apps` are
root-only since #1970 and cannot skew; `react` (whose types are `@types/react`,
resolved from one install) and `hono` never reach one program from two
installs. Each would be a rationale for a skew that cannot occur.
Closes#1965
Signed-off-by: cliffhall <cliff@futurescale.com>
Extends the previous commit's fix rather than changing it. `waitUntil` now
settles, but the same window is open wherever a key is written without having
just polled: `press()` reaches its `stdin.write` on a turn that may still have
React's passive-effect flush queued (straight after a plain `tick`, or after an
earlier key committed a render), and `mount()`'s absorbing keypress lands one
tick after the mount commit. Both now `await settleInputHandlers()` before
writing, so "frame visible" implies "input handlers armed" at every dispatch
point, not only after a frame wait.
Adds a `test helpers` case that pins the contract: a `setImmediate` sentinel
stands in for the pending effect flush and must have run by the time
`waitUntil` resolves. Verified it fails when the yield is removed, so a future
regression surfaces here instead of as a differently-named flaky OAuth test
under coverage instrumentation.
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review: the comments offered "a nullable number the user cleared" as a
way to get an explicit `null` into this field. Clearing the box reports
`undefined` — `toNumericValue("")` says so, and an existing test pins it — so
the only way `null` arrives is from parent state: a value received from the
server, restored from a deep link, or written by a caller for a nullable
schema. Say that in all three places instead.
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review: a nullable number schema really does produce `null` for both
the raw value and the schema default, but the props declared only
`number | undefined` and the call site cast the null away. The runtime already
handled it — the initializer's `?? ""` exists for exactly that — so the types
were the part that was wrong, and a cast hiding it is how a future null bug
goes unseen by the compiler.
Widen both to `number | null | undefined` and drop the narrowing casts, keeping
the distinction the initializer encodes: `undefined` takes the default, `null`
is a value and renders as an empty box.
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review: seeding with `value ?? defaultValue` also substitutes the
default when the raw value is an explicit `null` — a nullable number field the
user cleared, or one the server sent — where `resolveValue` previously kept the
null and the box rendered empty. A nullable schema with a non-null default
would have opened showing that default over a real value.
Test the `undefined` case explicitly, as `SchemaJsonField` already does, and
pin it: the new test fails with `'30'` where `''` belongs.
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review: the code and test comments still said a defaulted array or
object argument *was* uneditable. It would be, but not through today's seeding
path — `collectSchemaDefaults` assigns `fieldSchema.default` itself, so the
substitution returns the same reference and the re-sync never fires. It needs a
structurally-equal but distinct value object, which parsed wire/deep-link
values and rebuilt nested-object defaults produce, and which these tests
construct. The number field has no such protection and did revert in the app.
State the boundary in all three places, and stop claiming the distinct-object
seed is how the Tools tab seeds the field, which it is not.
Signed-off-by: cliffhall <cliff@futurescale.com>
A field whose schema declares a `default` could not be edited. Unsendable
draft text reports `undefined` by design — an inspector must not report a value
it cannot send — and `resolveValue` turned that straight back into the default.
That is a real change from the previous value, so the draft/value re-sync
rewrote the box to the default and the keystroke was lost. Almost every edit
passes through an unsendable state, which left a defaulted array or object
argument uneditable in the JSON editor, and a defaulted number field unable to
be emptied.
Hand the two draft-holding fields the raw value and the schema default
separately, and compare only against the raw one. The default now seeds what
the field opens with, which is what a `default` means, rather than being
re-imposed while the user types. Every other field keeps resolving through
`resolveValue` unchanged.
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review: seeding `{ b: null }` / `{ cfg: null }` meant the schema's
`default: null` was never resolved, so a regression that dropped null defaults
would have left both #1885 tests green while claiming to pin that path. Render
the harness unseeded instead, which is also how the reporter's form arrived.
Signed-off-by: cliffhall <cliff@futurescale.com>
Three issues report one defect: the schema form's JSON fallback re-escaping
the text being typed. #1853 saw it while typing an array, #1856 on the
Backspace that first makes a valid array invalid, and #1885 on a nullable
parameter whose default `null` had to be edited in place.
The mechanism — unparseable draft text stored back as the field's value, which
the next controlled render re-JSON.stringify'd — was removed by the draft/value
split in SchemaJsonField and the nullable-union collapse (#1928, #2007, #2020).
Those fixes were made for differently-framed issues, so nothing pinned the
flows these three reports actually describe.
Verified against the pre-fix commit: four of the five new tests fail there with
the reported corruption, including '"\"\"a"' for the caret-outside-the-quotes
case and the compounding backslash cascade for the typed array.
Signed-off-by: cliffhall <cliff@futurescale.com>
