Round 16 reported no inline comments and three suppressed findings, all
valid:
- `getSkillsExtension` treated a non-object extension value (`false`,
`"skills"`) as a declaration. SEP-2133 declares an extension as an
object of sub-options, so a primitive is not one — now rejected,
matching `appElicitation.ts`.
- The `onGetSkill` prop doc claimed the two responses "must agree" and
that a difference means a broken server, contradicting the
implementation, which reports a difference as a finding rather than a
fault. Reworded to what the code does.
- No real-transport test exercised the fixture's `skills/*` handlers.
Added `src/test/integration/mcp/inspectorClient-skills.test.ts`,
parameterized over both protocol eras: capability advertisement, the
paged `skills/list` walk (direct and through `ManagedSkillsState`),
`skills/get`, the `-32602` for an unknown URI, `resources/read` of a
skill file, and delegation of an ordinary resource through the
fixture's `resources/read` wrapper.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
Round 15 generated no new inline comments; all five suppressed findings
were real and unaddressed.
- SkillsScreen: claim the attempt BEFORE issuing the request, for both the
SKILL.md preview and `skills/get`. Recording it only on settle left a
window where an older request that happened to resolve first was still
considered current and published while a newer one was in flight —
`verifyRow` already claimed its row up front, these two did not. Tests
answer the OLDER request first and assert nothing is published.
- SkillsScreen: "No skills listed", not "No skills". SEP-2640 lets a server
return an empty or partial catalog and says an empty result is not proof
it has none — an unlisted skill is still fetchable by URI — so the old
copy had the tool asserting something the protocol explicitly does not.
- SkillsScreen: each row's Verify button gets an `aria-label` carrying its
URI. Every row's visible text is "Verify" and the URI cell is not
programmatically associated with the button, so a screen-reader user had
no way to tell which file each control checked.
- inspectorClient: attribute a rejected `skills/get` envelope with
`markResponseRejected`, so the Protocol tab stops rendering it as a clean
success while the screen shows an error. In the client rather than a
store because `skills/get` has none. Gated on `isClientDecodeRejection`,
with a test that a transport failure is NOT attributed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
The review finally quoted the normative text, which is what I asked for
across five rounds before declining: SEP-2640's `skills/list` section says
"In protocol versions 2026-07-28 and later, the result also carries …
`ttlMs` and `cacheScope`." So the era-aware validation is in.
- skillsSchemas: `ModernListSkillsResultSchema` = the page plus the base
list envelope, with field shapes mirroring `ModernResultEnvelopeSchema`
in `listSalvage.ts` so this repo's two statements of a modern envelope
cannot drift. `InspectorClient.listSkills` picks it from the negotiated
era; the legacy shape stays permissive, because those are 2026-era
attributes a legacy server has no business sending.
Nothing else catches this: `skills/*` is consumer-owned, so it is absent
from the SDK's cacheable-method registry and a modern server answering
`{ skills: [] }` was reaching the conformance UI as a clean list.
- skills.ts: count `frontmatter.description` (and `name`) length in Unicode
CODE POINTS. `String.length` counts UTF-16 code units, so 600 non-BMP
characters measured as 1200 and a perfectly valid description was
reported `malformed-description` — a conforming server failed by an
off-by-encoding, which is the direction this module works hardest to
avoid. The Agent Skills reference validator uses Python `len()`.
- #2248 narrowed: the `skills/list` half is settled here; what remains is
whether `skills/get` carries the same attributes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
All three findings were the same underlying inconsistency: URI comparisons
on raw strings, in a flow that elsewhere deliberately treats normalized
equivalents as the same resource. One `skillUriIdentity` helper now backs
every comparison, so they cannot disagree.
- skills.ts: `manifest-missing-self` compares identities, so a manifest
listing the RFC-equivalent `skill://demo/%53KILL.md` is recognized as the
entry's own file — it is fetchable as that file, so reporting it missing
was the tool disagreeing with itself.
- skills.ts: `skillEntriesMatch` normalizes the entry URI and every manifest
URI before comparing, so a server that canonicalizes an escape between the
listing and the fetch is not reported as a changed snapshot.
- SkillsScreen: the `wrongUri` check uses the same identity, so a
canonicalizing server is not accused of answering for a different skill —
the read path already accepts exactly that equivalence.
- skills.ts: apply the Agent Skills name grammar to the RAW value. Trimming
first let `" demo "` pass, and whitespace is not in the grammar — so an
entry could report "Conforms" with a name that can never equal its URI
path segment. The trimmed copy now only distinguishes absent from invalid.
- skills.ts: add `malformed-description` for a description above the
1024-character Agent Skills limit, with boundary coverage.
- inspectorClient-skills.test: justify the `as unknown as` per AGENTS.md —
both fields are `private` with no public setter, the asserted shape is
exactly what the class declares, and the alternative is a live connection.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
- skills.ts: apply the two RFC 3986 §6.2.2 percent-encoding normalizations
the URL parser does not — decode an escape standing for an unreserved
character, and upper-case the hex of every escape that remains. `URL.href`
leaves `%72eference.md` encoded, so a server echoing an RFC-equivalent
form of the URI we asked for was rejected by `onReadSkillFile` as a
different resource, and an encoded name segment produced a false
`name-path-mismatch`. Both are the tool calling a conforming server wrong.
- oauthResume.test: assert the Skills restore path. It was wired into the
setters but never checked, and the comprehensive present-value and
undefined-fallback cases omitted Skills entirely. Adds it to both, plus a
case restoring a REAL saved selection rather than the EMPTY value — a bug
that always wrote the default would have passed the others.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
- skills.ts: validate `frontmatter.name` against the Agent Skills format
(1–64 lowercase alphanumerics and hyphens, no leading, trailing or
consecutive hyphen). Checking only for non-emptiness let `Bad Name` read
as "Conforms" — and the name is not decorative: it must equal the URI
path segment, so a name that cannot appear in a URI is a contradiction
the entry cannot satisfy. Suppressed when the name is absent entirely,
which `missing-name` already reports.
- skills.ts: detect duplicates on the NORMALIZED identity. Containment and
the read that fetches the bytes both treat normalized-equivalents as one
resource, so `skill://demo/SKILL.md` and `skill://demo/x/../SKILL.md`
were passing as two distinct files while naming one. The finding still
reports the raw URI, and two different unparseable URIs stay distinct.
- SkillsScreen: run `checkSkillConformance` on the fetched `skills/get`
entry and treat an error — or an entry answering for a different URI —
as invalid (red), not as "a different snapshot" (yellow). A fresh
point-in-time read excuses a CHANGE; it does not excuse a violation, and
answering with another skill is never a refresh of the one requested.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
- skills.ts: stop sorting `resources` arrays recursively. Only the entry's
OWN manifest is a set; `frontmatter` is verbatim arbitrary JSON, so a
custom `frontmatter.metadata.resources` array was being sorted too and
two genuinely different frontmatters compared equal. The manifest sort is
now done once, at the top level, and every other array keeps its order.
- SkillsScreen: invalidate on the whole selected entry, not just its
manifest. The `skills/get` comparison covers `frontmatter`, so a refresh
that changed only a description left a stale "matches skills/list"
verdict on screen — computed against an entry that no longer existed.
Re-verifying after a metadata-only refresh is the cheap direction to be
wrong in; showing a match against a different entry is not.
- SkillsScreen: give each "Verify all" invocation a token, so a finalizer
can only ever clear its own batch rather than any batch with a matching
manifest key.
- SkillsScreen.stories: supply `onGetSkill`. Every story rendered the
"Fetch with skills/get" button, and clicking it threw.
On the batch token: the A → B → A sequence the review describes is not
actually reachable, because the button is disabled whenever the in-flight
batch's key matches the selection — returning to A finds it disabled, not
clickable. The token is defence in depth for the same property, and the
test asserts what is really guaranteed rather than the unreachable path.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
Two of these reverse round 4, which asked for the opposite. Where the
readings conflict I took the permissive one: a wrong tightening rejects
conforming servers, a wrong loosening only fails to report something.
- skills.ts: stop requiring the `skill:` scheme. SEP-2640 says a server
SHOULD use `skill://` and explicitly allows a domain-native scheme, so
round 4's requirement handed a conforming `github://…` skill a false
`malformed-uri` and skipped its name and root checks. Containment is
scheme-independent — it compares against THIS entry's own root — so the
traversal fix that requirement came packaged with is unaffected.
- skills.ts: the 512-entry and 16 MiB limits are warnings. Both are
SHOULD NOTs for a server and MAYs for a host, so calling them errors
contradicted the "error = a MUST was broken" contract round 4
established and told authors a permitted skill was invalid.
- skills.ts: `skillEntriesMatch` compares canonically — keys sorted,
manifest sorted by URI. `JSON.stringify` treated key order and manifest
order as differences, though neither carries meaning.
- SkillsScreen: a differing `skills/get` is "a different snapshot"
(yellow), not a disagreement (red). The SEP defines it as a fresh
point-in-time read, so a skill that changed since the listing
legitimately differs.
- useServerCommands: route `onGetSkill` through the command-scoped OAuth
recovery like every other server command. An expired authorization
rendered an error and stopped there — no reauthorization, no retry.
- test-servers/skills.ts: stamp the modern base result envelope
(`resultType` / `ttlMs` / `cacheScope`) on every skills result. The SDK
stamps it only for methods in its own codec, and `skills/*` are
consumer-owned, so a 2026-era connection was receiving results without
it. Unconditional rather than per era: the modern leg builds a server
per request, so there is no era to branch on at registration time, and
on legacy they are three members no codec inspects.
- composable-test-server: remove the `skills.directoryRead` option. It was
publicly settable and advertised a method nothing serves, producing the
exact false capability the fixture exists to help catch. `skills` is now
a plain boolean; the flag returns in phase 3 (#2248) with its handler.
Still not changed: whether a modern `skills/list` RESULT SCHEMA must
require that envelope. Stamping it server-side is free; requiring it
client-side would reject servers on a reading nobody has checked against
the normative text. #2248 owns settling it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
- SkillsScreen: actually call `skills/get`. The client method and its
tests existed but no production caller invoked it, so #2234's acceptance
criterion ("skills/get retrieves a single entry") was unmet and a server
author's required handler could not be exercised. The detail pane now
fetches the selected URI on demand and reports whether the fetched entry
AGREES with the one skills/list advertised — both describe the same
skill, so a disagreement is a server bug only a side-by-side fetch shows.
- core/mcp/sha256.ts: a dependency-free SHA-256, used when `crypto.subtle`
is absent. `SubtleCrypto` needs a secure context, and this app is
documented as servable over plain HTTP on a LAN IP — where every skill
verification threw and the UI reported a read failure for files it had
fetched fine. Checked against the FIPS 180-4 vectors and differentially
against WebCrypto.
- skills.ts: require the `skill:` scheme in `normalizeSkillUri`. Checking
only that a URI was hierarchical let `https://demo/SKILL.md` pass the
name and root checks — a manifest pointing anywhere on the web, reported
as conforming.
- skills.ts: `missing-digest` and `missing-size` are errors. Both are
required fields, and an omitted `size` is what lets a server slip past
the 16 MiB pre-fetch limit while the UI reports zero errors. `warning`
is now reserved for what is legal yet unverifiable — `"dynamic"`.
- SkillsScreen: key the "Verify all" batch guard to the manifest. A global
flag left a newly selected skill's button disabled until the previous
skill's reads settled — forever, if one hung.
Not changed, deliberately: whether a modern-era `skills/list` result MUST
carry the SEP-2549 caching attributes. #2234's analysis records it as open
and the review asserts the opposite; neither reading was checked against
the normative text. Leaving the schema permissive accepts a server that
omits them, while tightening on a wrong reading would reject conforming
responses — the more expensive direction. `skillsSchemas.ts` states this
and #2248 settles it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
- skills.ts: decide root containment on NORMALIZED URIs. A raw prefix
check passed `skill://acme/billing/refunds/../other.md`, which starts
with the advertised root but resolves outside it — a traversal the
conformance report was reporting as clean.
- skills.ts: `normalizeSkillUri` also rejects a relative string, so
`demo/SKILL.md` is `malformed-uri` rather than a skill path, and an
opaque-path URI (`skill:demo/..`), which the parser leaves un-normalized
and on which containment cannot be decided.
- skills.ts: `missing-description` is an error. SEP-2640 requires it, so
an absent one must not read as "0 errors" in the conformance summary.
- managedSkillsState: call `markResponseRejected` for a decode rejection,
as every managed list does. An invalid `skills/list` result was showing
in the Protocol tab as a clean success.
- SkillsScreen: per-row attempt token. The manifest key cannot tell two
verifications of the SAME row apart, so a double click (or a row button
pressed during "Verify all") let an older read finish last and overwrite
the newer verdict.
- SkillsScreen: disable "Verify all" while a batch runs. The concurrency
cap is per invocation, so repeated clicks stacked pools — 4, then 8,
then 12.
- SkillsScreen: include the finding index in each issue alert's key. Three
identical URIs produce two `duplicate-resource` findings with the same
code and URI, and React was free to drop the extras — hiding findings in
exactly the malformed input this view exists to inspect.
- test-servers/skills.ts: register `skills/list` and `skills/get` through
the PUBLIC `setRequestHandler` with explicit param schemas. The private
`_requestHandlers` map is now reached only to wrap `resources/read`,
which has to chain onto the SDK's handler rather than replace it — the
one thing the public API cannot express, and the comment now says so
instead of citing the tasks fixture.
- test-servers/skills.ts: raise `-32602` for an unknown `skills/get` URI.
A plain Error mapped to a generic server failure, making the fixture
non-conforming outside its three documented bad cases.
- clients/web/README.md: the paragraph named four contracts while still
saying the smoke "drives all three".
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
- SkillsScreen: key verdicts and React elements by manifest ROW INDEX,
not URI. The checker deliberately tolerates a duplicated URI so it can
report `duplicate-resource`; a URI key collided those rows into one
verdict, so verifying either updated both and "Verify all" raced two
different digest/size declarations into the same slot.
- SkillsScreen: hold the invalidation generation in React state keyed by
the manifest, not a ref bumped during render. `useValueChange` runs its
callback during render and requires setState-only purity — an abandoned
render left the ref incremented and silently discarded a live
verification.
- SkillsScreen: title a size disagreement "Size mismatch" and render its
reason. `verifySkillResource` catches it before hashing, so the alert
was showing "actual undefined" under "Digest mismatch".
- SkillsScreen: pass `contents` to ContentViewer so a base64 SKILL.md
renders. The text-block form substituted "" and painted a blank preview
for a file verification had just read correctly.
- skills.ts: add `malformed-size` for a size that is not a non-negative
safe integer, and exclude such values from the 16 MiB total. A negative
one could pull the sum back under the limit and hide a violation.
- managedSkillsState: make the overlap guard per-session instead of a
boolean. A disconnect during an in-flight walk left it set, so the
reconnect's own load no-oped and was never retried — permanently, if
the stale request never settled. A stale `finally` can no longer clear
the live session's guard either.
- skillsSchemas: require the `{ skill }` envelope for `skills/get`. The
accepted SEP settles it, and normalizing an inline entry would let a
non-conforming response past the one place that could report it.
- #2248 and the PR description: corrected — they said
`ReadResourceDirectoryResultSchema` was already declared, which round 1
removed. #2248 now owns defining it, and records the frontmatter
cross-check gap too.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
- skills.ts: report a manifest that omits the skill's own SKILL.md
(an empty list included), duplicate URIs, entries outside the skill
root, and a missing size. `Conforms` was reachable for manifests that
break invariants SEP-2640 states.
- skills.ts: cross-check the declared byte length before hashing. A size
that disagrees fails verification on its own — the digest is taken over
the bytes the server served, so agreeing with it says nothing about
whether the manifest describes them.
- skills.ts: copy the view instead of slicing its backing store in
`sha256Digest`. `SharedArrayBuffer.prototype.slice()` returns another
SharedArrayBuffer, which `crypto.subtle.digest` rejects — the cast
claimed to handle the exact input that would have thrown. No cast now.
- skills.ts: state the one obligation NOT checked here — that an entry's
frontmatter matches the fetched SKILL.md's. The digest cannot cover it,
and closing it needs a YAML parser, so it is tracked on #2248.
- managedSkillsState: cap the walk at LIST_MAX_PAGES. The repeated-cursor
guard only catches a server stuck on one cursor; endlessly unique ones
walked forever. Raises rather than truncating, like the salvage walk.
- managedSkillsState: gate every write on a session generation, so a walk
that resolves after a disconnect or destroy cannot repopulate a cleared
store or deliver the previous session's skills into the next.
- inspectorClient: send a cursor when it is `!== undefined`, not when it
is truthy. An opaque cursor may be `""`, and dropping it re-requested
page one — which the store then reported as a repeated-cursor failure.
- SkillsScreen: key verdict invalidation on the manifest (URI + digests +
sizes), not the URI alone. A Refresh that changed the manifest left a
green badge attached to a digest nothing had checked.
- SkillsScreen: epoch-guard every read continuation, so a fetch that
resolves after the selection moved on cannot write into the new one.
- SkillsScreen: bound "Verify all" to 4 concurrent reads. A conforming
manifest may hold 512 files.
- skillsSchemas: drop the guessed `resources/directory/read` result
schema. Nothing calls it, so an unverified shape could sit wrong
indefinitely; phase 3 adds it against the normative text.
- skills-http.json: declare the extension bare. It advertised
`directoryRead: true` with no handler, so Connection Info reported
"Supported" for a method that answers -32601.
- types.ts: restore the Tasks doc comment the Skills interface displaced.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
Adds detection, enumeration, and conformance checking for the Skills
extension (`io.modelcontextprotocol/skills`) — phases 1 and 2 of #2234,
which is everything its Acceptance list names.
Skills is a *server*-declared extension, read off the connecting server's
`capabilities.extensions`, so it is deliberately absent from
`ADVERTISABLE_EXTENSIONS` — that registry is what the Inspector advertises
and the user toggles, and an entry there would be a meaningless toggle.
`skills/list` and `skills/get` are consumer-owned extension methods that
neither era codec defines, so the SDK's era gate skips them and they go out
as ordinary `client.request` calls with explicit result schemas. The
raw-wire channel modern `tasks/*` needs is not used, and the Skills tab is
not era-gated: a legacy-era server that declares the extension is serving it.
The checks are the point, not the list view. `core/mcp/skills.ts` reports
each obligation SEP-2640 states — the name/path invariant, the digest
format, the 512-entry and 16 MiB limits, and `resources: "dynamic"`, which
means integrity cannot be verified at all. Digest verification hashes the
fetched bytes with WebCrypto and returns a mismatch with both digests
attached rather than throwing, because showing a mismatch loudly is the
whole value proposition.
Files are fetched on demand: SEP-2640 is explicit that a `resources/read`
of a `SKILL.md` is not a load and confers no standing, so none of the SEP's
host machinery is implemented.
The `skills-http` fixture serves four skills over two pages, three of them
deliberately non-conforming — without those the verification code is
untestable.
Phase 3 (CLI, TUI, `resources/directory/read`, paged mode) is #2248.
Closes#2234
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018kfrH3uRNUM4JAk1reKtyY
Signed-off-by: cliffhall <cliff@futurescale.com>
Every state-subscription hook in core/react/ seeded local React state from
its store prop and re-synced it inside a useEffect. That shape paints one
frame of the previous store's data when the prop changes (switching
servers), and loses any event dispatched between the render that took the
snapshot and the effect that attached the listener.
Convert all 16 to useSyncExternalStore through a new shared
useStoreSnapshot, following useManagedListError (#1954).
useSyncExternalStore needs a referentially stable snapshot, and these
getters hand back a defensive copy, so TypedEventTarget now keeps a
per-event-type dispatch counter (getEventRevision) that the snapshot is
cached against. A contents comparison would not do: MessageLogState folds
a response into its request entry by mutating that entry in place, which a
shallow compare reports as unchanged. The bump lives on the dispatchEvent
override so no dispatch — including a hand-rolled CustomEvent — can bypass
it.
useInspectorClient is included though the issue did not list it; its nine
subscribed slices made the stale frame most visible. Its lastError stays
useState (accumulated from events, no getter to read back) but resets on
client swap during render rather than in an effect.
The settings-draft and async fetch hooks are deliberately untouched: the
former hold genuinely local editable state, the latter are I/O rather than
store subscriptions.
Also answers the issue's open question. react-hooks/set-state-in-effect
never fired here because eslint-plugin-react-hooks has never linted
core/react/ at all — it is a clients/web devDependency and the root config
loads no React plugin. Corrected the config comment that gave "no JSX" as
the reason; the gap itself is #2192, measured at 15 findings, none of them
in the hooks converted here.
Closes#1955
Signed-off-by: cliffhall <cliff@futurescale.com>
Three findings, all valid:
- `beginInteractiveAuthorization` navigated *before* awaiting provider
creation and flow-state recording, so a rejection could arrive after
the browser had been handed over — and the web catch clears the resume
snapshot on one. Fixed at the root: the navigation now resolves last,
after every fallible step, which makes every rejection the method can
produce a pre-navigation rejection. It is also better on its own
terms — a provider that cannot be built means the callback could not
have completed, so navigating first only bought a dead-end round trip.
- `resumePendingReauth` restored the pending slot unconditionally, which
could overwrite a newer challenge deferred while it was awaiting.
Restores only into an empty slot, so latest wins.
- the `authChallengeInteractive` catch re-read the active server rather
than the one the challenge arrived for. `serverId` is now captured
synchronously in the listener and used by both the body and the catch.
Three new tests, one per finding, each driving the interleaving rather
than the happy path.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Quk1hBUtyhkXowwMXzAnAY
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 18. Three of its findings were the same check-then-act:
`planOAuthRevocation` is asynchronous, so the gap between its reads and the
following `storage.clear()` is still a window — shorter than the network one
closed in round 13, but the same shape. An OAuth completion landing at any of
those awaits saved a grant the clear then destroyed.
`OAuthStorage.takeRevocationSnapshot(serverUrl)` reads what revocation needs
and clears the server in ONE synchronous pass over the in-memory state, with
the persist after the mutation. `clearAndPlanRevocation` replaces
`planOAuthRevocation`, and there is no separate `clear` call left in any
caller — the clear cannot be forgotten or reordered.
The snapshot returns values unparsed, deliberately: schema validation is pure
and belongs after the mutation, since running it inside would reintroduce the
`await` this exists to remove.
Two consequences worth naming. A failure of the take-and-clear now REJECTS
rather than returning an outcome, because the state was not cleared and the
caller's contract is that it was; both clients already have a rejection path.
And cross-process atomicity is explicitly not claimed — nothing in this store
takes a lock, every mutation is a read-modify-write over a loaded snapshot, and
`clear` alone always had that property.
Also from round 18:
- The CLI's budget check ran before considering `plan.outcome`, so a key that
needed no network was reported as budget-exhausted — warning that a grant may
be live when that key held none.
- A disconnect failure after a successful clear was reported as "could not
clear OAuth state" in both clients. It now goes to the disconnect channel,
and the success is still reported.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0177mPHAdECD18nLLCTwR5rh
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 14. Two code fixes and a documentation sweep.
- `buildRevocationRequest` ran outside `revokeToken`'s best-effort try.
`encodeURIComponent` throws on a lone UTF-16 surrogate — valid JSON, so it
can reach there from a persisted client id or secret — and every caller has
already cleared its local state by then, so a rejection instead of a `failed`
outcome would break the documented guarantee. Construction is inside the try
now, and the catch reports `params.endpoint` rather than a `url` that may not
exist on that path.
- The fixture's Basic decode turned a malformed percent escape into an Express
500, reporting a bad credential as a server fault rather than as the
`invalid_client` 401 the endpoint means. Decoding failure is failed
authentication.
- Ten places still said revocation happens *before* the clear, which the
reorder in the previous commit made false: the settings-form description, the
`--no-revoke` help text, both READMEs, `docs/mcp-server-configuration.md`,
three JSDoc blocks, and the PR body. Swept, and the ordering is stated where
it is load-bearing rather than asserted in passing. Also merged the two
adjacent JSDoc blocks on `InspectorClient.clearOAuthTokens`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0177mPHAdECD18nLLCTwR5rh
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 13. Three of its four findings were one root cause, and
it is a design flaw in the premise this PR started from.
Revoking *before* the clear meant the clear waited up to five seconds on the
network — and anything that wrote OAuth state in that window was then deleted
by a clear reasoning about the grant it replaced. A same-server reauthorization
in the TUI (whose Disconnect/Connect accelerators stay live), a stored-only web
clear for a server that becomes active mid-flight, and another CLI/TUI process
writing to the shared file-backed store all hit it.
`revokeStoredOAuthTokens` is therefore split into `planOAuthRevocation` and
`executeOAuthRevocation`, and every caller runs plan → `storage.clear()` →
execute. The snapshot still precedes the clear, because the store is where the
token, the credentials and the endpoint live; but the clear no longer waits on
anything. There is deliberately no single-call wrapper left, so no caller can
reintroduce the racy order by accident.
Fourth finding, same file: `handleConnect` now retires `clearOAuthAttemptRef`
alongside `disconnectAttemptRef`. The server-name check does not cover
disconnect/reconnect to the SAME server — the name still matches on the other
side of it — so a clear still in flight could tear down the session the connect
had just established.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0177mPHAdECD18nLLCTwR5rh
Signed-off-by: cliffhall <cliff@futurescale.com>
Two more review findings, both real.
The shim wrapped `effectiveAuthFetch` only, but the transport is constructed
with the base `this.fetchFn`, and the SDK runs its own discovery from inside
the transport on the 401/refresh leg — so a reconnect with an existing auth
provider could still hit the upstream failure. It now wraps the base fetch,
the same seam `withOAuthEndpointOverrides` uses and the only one that reaches
both paths.
That trades away the previous claim that a captured Network entry shows the
real 404: below the trackers, what they record is the substitution. Rather
than leave that silently misleading, a substituted response now carries
`x-inspector-oauth-metadata-source` naming the URL its body was actually
fetched from — the transport builds its tracker internally, so there is no
seam above it to install into instead. An integration test captures the fetch
the transport is handed and drives it against a real server, since nothing
else proves that wiring.
The wrapper also keyed on URL shape alone, but nothing stops an authorization
or token endpoint from living under `/.well-known/oauth-authorization-server/`;
a POST there returning an ordinary OAuth 4xx would have had its error replaced
by a metadata document. Metadata discovery is a GET, so anything else is now
left alone.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UL9retfvAXRgvi6EWk4SY4
Signed-off-by: cliffhall <cliff@futurescale.com>
The Cancel button POSTed `notifications/cancelled` on every transport. On
a 2026-07-28 Streamable HTTP connection that is the wrong signal — the
spec makes closing the request's SSE response stream the cancellation
signal there and says the notification is neither required nor expected —
so a spec-compliant server acknowledged it 202 and dropped it while the
tool ran to completion.
The SDK already implements the fork, off `transport.hasPerRequestStream`.
Every Inspector connection is wrapped in `MessageTrackingTransport`, which
did not forward that flag, so the SDK read `undefined` and took the stdio
branch on every client — the CLI and TUI included, whose real
`StreamableHTTPClientTransport` advertises it correctly.
The web client needed three more links: its browser transport answers for
the upstream transport that lives on the backend, so it advertises the
flag for `streamable-http` and applies the SDK's `requestSignal` to its
`POST /api/mcp/send`; and `/api/mcp/send`, held open for the whole call,
forwards the browser's disconnect to the upstream `transport.send`.
The test fixture's `toWebRequest` built a Web Request with no `signal`, so
no fixture server could observe a disconnect at all — without that, a test
asserting the cancel reaches the server passes vacuously.
stdio and SSE keep `notifications/cancelled`: they multiplex every request
over one shared channel and have no per-request stream to close. So does
the legacy era, which predates the mechanism.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
Clearing the Inspector's OAuth state deleted the local copy of the tokens
and stopped there. The access token — and the refresh token, which is
long-lived by design — stayed valid at the authorization server until they
expired on their own, so a day of connect/disconnect iteration left it
holding grants for sessions that ended hours ago. RFC 7009 §1 describes
exactly that case, and nothing in the repo constructed a revocation
request, so a server author implementing `revocation_endpoint` had no way
to exercise it with the tool that exists to exercise this.
core/auth/revocation.ts builds and sends the request; the three clear paths
call it, so the behavior is written once — web's "Clear OAuth state and
disconnect" (both the active and stored-only paths), the TUI's Auth tab,
and the CLI's --relogin. It runs BEFORE the local clear, because the token,
the client credentials and the discovered `revocation_endpoint` all live in
the store the clear is about to empty. The request names the refresh token
when there is one: §2.1 asks the AS to also invalidate the access tokens
issued under the same grant, so one request covers both halves.
Best-effort by construction. Every path returns a TokenRevocationOutcome
rather than throwing: no advertised endpoint means nothing is sent and that
authorization server behaves exactly as before, and a network error, a
non-2xx, a 5s timeout or an unreadable store is reported and nothing more.
Forgetting the tokens is what the user asked for. `lost_authorization_state`
recovery deliberately skips it — it clears a half-finished flow to retry it,
and an authorization that never completed has no grant to revoke.
Client authentication resolves the preregistered slot first, mirroring
BaseOAuthClientProvider.clientInformation; reading only the dynamic slot
would send no authentication for a server configured with `oauth.clientId`.
The Basic credential is byte-identical to the SDK's applyBasicAuth so this
request and the token request cannot present the same secret differently.
Opt out per server with `oauth.revokeOnClear` (default on, only `false`
written to disk) or per CLI run with --no-revoke. Read at clear time rather
than connect time, so toggling needs no reconnect. Turning it off is a
testing affordance: a client that walks away still holding live tokens is a
case a server author may want to reproduce.
test-servers now advertises `revocation_endpoint` and serves POST
/oauth/revoke, faithful on §2.2 (an unknown token is a success) and §2.1
(revoking a refresh token invalidates the access tokens under the same
grant), with oauth-revocation-http.json / oauth-no-revocation-http.json as
showcase configs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0177mPHAdECD18nLLCTwR5rh
Signed-off-by: cliffhall <cliff@futurescale.com>
The SDK's `discoverAuthorizationServerMetadata` picks its validation schema
from the well-known filename that resolved rather than from the document that
came back, so anything found at `/.well-known/openid-configuration` is
validated as an OpenID provider document — requiring `jwks_uri`,
`subject_types_supported` and `id_token_signing_alg_values_supported`, three
fields RFC 8414 does not define. RFC 8414 §5 permits that filename for general
OAuth metadata, so a conforming plain OAuth 2.0 authorization server is
rejected; and because the parse throws instead of continuing the candidate
loop, discovery aborts and the connection fails outright.
Filed upstream as modelcontextprotocol/typescript-sdk#2733.
`core/auth/oidcDiscoveryCompat.ts` works around it without fabricating a
field: on a failed RFC 8414 candidate it probes the OIDC candidates the SDK
would try next and, when one returns a document that satisfies
`OAuthMetadataSchema` but fails the OIDC schema, serves that body as the RFC
8414 response so the SDK picks the schema that describes it. A genuine OpenID
provider document is left to the SDK's own OIDC leg. Issuer validation runs
unchanged, since the substituted document is the one the server published.
It wraps `effectiveAuthFetch` — above the fetch tracker, the opposite of
`withOAuthEndpointOverrides` — so the Network tab still records the real 404
and the real probe rather than the substitution.
The candidate derivation mirrors the SDK's `buildDiscoveryUrls`, which a fetch
wrapper cannot call, and a test pins it against that export so the two cannot
drift.
Adds `oauth.asMetadataPath` to the composable test server (mirroring
`oauth.resourceMetadataPath`) and the `oauth-rfc8414-at-oidc-path-http.json`
showcase config that reproduces the failure.
Closes#2172
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UL9retfvAXRgvi6EWk4SY4
Signed-off-by: cliffhall <cliff@futurescale.com>
A modern-era `subscriptions/listen` is a long-lived request: the first
message on its stream must be `notifications/subscriptions/acknowledged`,
and the JSON-RPC `result` for the listen id is reserved as the
graceful-closure marker. A server that answers the listen request with a
bare `result` is saying "acknowledged and closed in the same breath" — the
SDK settles the subscription `graceful` and rejects the pending `listen()`.
The Inspector read that rejection as an ordinary drop: eight re-lists over
roughly a minute, the badge flickering `Reconnecting…`, and a final bare
`Stream ended` that said nothing about why. Accepting the message was
never wrong — it is a well-formed graceful-closure response. The silence
was.
- `core/mcp/subscriptionAck.ts` owns the predicate. The SDK gives this
close no code of its own — it is the same `SdkError(ConnectionClosed)`
any pre-ack close carries — so the predicate reads the message, since
the alternative is refusing to retry a genuinely transient drop.
- The condition is deterministic, so the stream ends on the first
occurrence instead of spending the whole backoff run. Both routes are
covered: the one a subscribe/unsubscribe reaches, and the one a
reconnect reaches when an acknowledged stream drops and the server has
since started refusing.
- A `"never-acknowledged"` status of its own, not `"ended"` — `ended`
means an expected close. It surfaces as the thrown error's message, an
orange "Not acknowledged" badge, and a notice in the panel.
- `subscriptions-never-acknowledged-http.json` reproduces it. It
acknowledges the first listen that subscribes to a resource and refuses
every one after, which is what makes the badge reachable by hand.
Closes#2097
Signed-off-by: cliffhall <cliff@futurescale.com>
parseSSE declared currentEvent/currentData inside the read loop, so an SSE
frame whose data: line arrived in one chunk and whose terminating blank
line arrived in the next was discarded. buffer already carried a partial
LINE across reads; nothing carried a partial FRAME.
Because this is the web client's transport, the dropped frame is a
JSON-RPC response that never settles: the request hangs forever with no
rejection, no console message, and nothing in the Network tab. The
payload does not arrive corrupted, it does not arrive at all, which is
why it presents as a hang rather than an error.
It has not bitten us because Chromium and Firefox happen to deliver whole
frames at the payload sizes this transport currently sees. That is luck,
not a guarantee: chunk boundaries depend on payload size, TCP
segmentation, and any intermediary. The exposure grows with payload size,
so a large resources/read is the most likely first victim.
The fix hoists both out of the loop and adds an end-of-stream flush — a
server that closes without a final blank line has still delivered a
complete frame, and dropping it would lose the last message of every such
stream.
parseSSE is now exported, solely so the test can reach it. That is the
only way to hit this deterministically: driving a real transport cannot
steer where a chunk boundary lands, but a test that supplies the reader
can put it exactly on the terminator.
The test keeps the two PASSING cases (whole frame, split mid-payload)
alongside the three that failed, because they are what make the diagnosis
precise. This is not "SSE parsing is broken" — mid-payload splits are
handled correctly today, and only the frame-terminator boundary loses
data. Verified against the unfixed code: 3 of 5 fail, all 5 pass after.
Signed-off-by: cliffhall <cliff@futurescale.com>
#2114 landed on v2/main, which is the PR this one follows up. Two
conflicts, one of them silent.
The marked conflict was cosmetic: keep v2/main's comment on the
`provider.scope` read, which now notes the getter filters (#2068).
The real one git merged cleanly and would have left dead: #2114's
if (this.oauthConfig.requestRefreshToken === false && requestedScope) {
this.pendingAuthorizationScope = requestedScope;
}
sat directly above this branch's unconditional
`recordAuthorizationRequestScope`, which overwrites the same field a few
lines later. Dropping that gate is what #2117 is for, so it is removed
rather than kept as a no-op.
The two changes compose: the SDK adds `offline_access` only when
`clientMetadata.grant_types` includes `refresh_token`, which the provider
drops while the opt-out is on -- so reading the authorize URL yields the
filtered scope there, exactly what #2114's gated branch persisted.
#2114's third case pinned the gate and inverts here, as its issue
predicted. It now asserts the request is persisted even with the grant
on; that records no new claim, since the request equals what storage
already held.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
`trySilentEmaAuth` and `refreshEmaResourceTokens` saved tokens without
ever saving a scope. Both are reached from real paths -- initial EMA
authentication with a cached IdP session, refreshEnterpriseManagedTokens,
and EmaTransportOAuthProvider.tokens -- and each can be the moment a
newly configured or metadata-derived scope first takes effect, so a
silent token response left the previous stored scope standing.
Rather than a third copy of the rule, every EMA mint now ends in one
place: mintEmaResourceTokens reports the scope it requested, and
saveMintedTokens writes tokens and the RFC 6749 5.1 scope together.
The manager's EMA branch drops its own saveScope, which would now be a
redundant second write of the same value; the step-up branch keeps its
own, since it falls back to the challenge's authorizationScopes union
and runs last.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
Two more paths had the same defect, both found in review.
`handleStandardAuthChallenge` returns an interactive redirect for
`unauthorized` / `token_expired` / `invalid_token` as well, and its
callback lands in the same completeOAuthFlow -- but the recording was
gated on `insufficient_scope`. Those challenges compute their own
`scopeForAuth`, which can differ from what storage holds, so a silent
token response left the stale scope standing. The read is now one
private helper both entry points call, applied to every redirect.
And `scopeForMint` is not always what the EMA mint asks for: with
neither config nor storage supplying a scope, discoverEmaResourceContext
resolves it from the protected resource metadata's `scopes_supported`,
which is what both exchanges send. `completeEmaIdpAuthorizationAndMint`
now resolves that context itself -- passing it into the mint, so no
extra discovery -- and returns it as `requestedScope` for the manager to
use as the RFC 6749 5.1 fallback.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
`provider.scope` is not what the authorization request asked for. The
SDK augments it with `offline_access` when the AS advertises that scope
and the client wants a refresh token, so persisting the provider's copy
after a silent token response understates the RFC 6749 5.1 implied
grant -- the same defect this change exists to fix, one notch smaller.
Read the scope off the captured authorize URL instead, falling back to
`provider.scope` when the URL carries none.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
The Server Settings modal edits a stdio server's `env` and `cwd` as
settings, while the transport reads them off the *config* — so #2091's
tracker, which resolves settings through the last write known to have
reached disk, left those two behind: `server.config` still came straight
off a `servers` entry that stops advancing once a list read fails. A save
that landed during such a window kept spawning the child process with the
pre-save environment, while the modal (re-seeded from the tracker) showed
the new value. Credentials live in `env`, so the symptom is the inspected
server failing to authorize rather than a visibly lost edit.
Adds the settings→config mapping to `core/mcp/serverList.ts` — the one
place both sides of the wire can share — and routes the `/api/servers` PUT
write-through and the web client's connect path through it, so the value
persisted and the value spawned cannot be derived differently.
Closes#2096
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0177mPHAdECD18nLLCTwR5rh
Signed-off-by: cliffhall <cliff@futurescale.com>
RFC 6749 5.1 defines an omitted `scope` in a token response as a grant
identical to what was requested, and `resolvePersistedScopeAfterGrant`
already encodes that rule -- but the ordinary authorization leg never
gave it a `requestedScope` to fall back to, so nothing was persisted and
the previous stored scope kept being reported as what the current token
carries.
`authenticate()` now records `provider.scope` as the pending
authorization scope once the authorization URL has been captured, so it
survives the redirect and reaches `completeOAuthFlow`. The
enterprise-managed leg had the same defect for the same reason and is
fixed one level up: it now falls back to `scopeForMint`, which is what
the mint actually asked for, rather than to the raw
`pendingAuthorizationScope` that is undefined on this leg.
A response that does carry `scope` still wins over the request, on both
legs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 11 on #2114. The docs claimed the persisted scope
"self-heals on the next grant"; on the ordinary authorization leg it
did not.
`authenticate()` never set `pendingAuthorizationScope` — only the
step-up paths did — so `completeOAuthFlow` called
`resolvePersistedScopeAfterGrant(tokens.scope, undefined)`. When the
token response omits `scope` (RFC 6749 §5.1: omitted means the grant
matched the request), the fallback was undefined, nothing was saved,
and the stale `offline_access` stayed in storage where
`resolveEffectiveGrantedScope` keeps reporting it as granted.
`authenticate()` now records the scope it actually requested, read once
from the filtering getter so what is recorded is what went out.
Gated on the filter being active rather than applied unconditionally.
The general version is arguably more RFC-correct, but it would change
what lands in storage for every server that never touched this setting,
inside a PR whose contract is that nothing changes unless the box is
unchecked. With the grant declared, request and storage agree and there
is nothing to heal.
Three tests; the omitted-scope one fails when the recording is removed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 10 on #2114 — a defect in round 8's fix, and the
precise loop that fix's extra argument was added to prevent.
`enrichChallengeWithScopes` narrows `requiredScopes` to the *missing*
subset, so a challenge requiring `offline_access tools:write` against a
stored scope that already carries `offline_access` arrives as just
`["tools:write"]`. Passing that narrowed list made the filter read the
server's explicit requirement as inherited and strip it, so
re-authorization would earn the same challenge forever.
Passes `challenge.requiredScopes` — the raw list — instead. The mixed
case is now covered, and fails against the narrowed list.
Also stops the checkbox promising more than the setting delivers: it
does not register anything for a preconfigured or CIMD client, and the
authorization server decides issuance either way. It now says it
declares the grant so the server *may* issue a refresh token.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 8 on #2114.
The provider's `scope` getter is not the only request path. Mid-session
`insufficient_scope` recovery builds its union from *raw* storage in
`enrichChallengeWithScopes` and hands it to `mcpAuth` directly, so it
never reads that getter. An `offline_access` inherited from an earlier
grant therefore reappeared during step-up and the SDK re-added
`prompt=consent` with the setting off — the failure this option exists
to prevent, on the one path that bypassed it. The comment claiming the
getter was "the one seam both readers go through" was wrong, and is
corrected to name both.
`handleAuthChallenge` now applies the same filter, gated on the setting
so the default path is untouched. The forced-retry reuses that value,
so filtering once at the computation site covers it.
`scopeForDeclinedRefreshGrant` gains the scopes a challenge requires,
honored exactly like a configured one: stripping a scope the server is
asking for would loop — re-authorize, earn the same challenge, strip it
again.
Three regressions; the inherited-scope one fails when the filter is
reverted.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 4 on #2114 — a real gap, not a wording one.
Declining the grant stops the SDK *adding* `offline_access`, but the
Inspector's own storage puts it back. A successful default-on
authorization persists the scope the AS granted
(`resolvePersistedScopeAfterGrant`), `offline_access` included;
`prepareForAuth` reloads it and `authenticate` hands it to the SDK as
`scope`, which sees the token in scope and appends `prompt=consent`
again. So any server that had authorized once could uncheck the box and
still hit AADSTS90094 — the users most likely to reach for the setting
were the ones it did not help.
`scopeForDeclinedRefreshGrant` filters it out of the *request* only,
applied in the provider's `scope` getter — the one seam both
`clientMetadata.scope` and `authenticate`'s `scope:` argument read
through. An `offline_access` the user configured is honored; storage is
never rewritten, so re-checking the box restores the old behavior
instead of finding the value destroyed, and scope-satisfaction checks
still see the real grant. Verified by reverting the filter: the new e2e
regression fails.
Also from that round, both previously suppressed:
The checkbox looked effective under enterprise-managed authorization,
which never builds this provider at all — it authorizes against the
enterprise IdP with a fixed `openid offline_access`
(`ema/constants.ts:16`). It now annotates itself the way the endpoint
overrides do. The existing count assertion covered two such notices and
was named for the overrides alone; it is renamed and expects three, with
a dedicated case pinning the notice to this control so the count cannot
be met by three copies elsewhere.
And "obtains no new refresh token" was another overstatement — the AS
decides issuance, and a preconfigured client is reused with whatever its
registration already permits. Replaced with what the setting actually
controls.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
- The SSE branch selects a caller-supplied `eventSourceInit.fetch` as-is, so
it bypassed the challenge observer and a first-time legacy SSE
authorization lost the advertised URL the same way. Observing is passive,
so it is applied to that fetch too while the deliberate no-proxy-wrap rule
stays. Covered behaviorally against a 401-serving server.
- Dropped an `as unknown as typeof fetch` from the new integration test
(banned by AGENTS.md, and it erased the signature); the spy is typed as
`typeof fetch` directly. The `as MCPServerConfig` beside it was
unnecessary too.
- The integration test reused a per-pid OAuth storage file. SDK `auth()`
persists the advertised URL in its discovery state, so a rerun could
satisfy discovery from that cache and pass with the plumbing removed —
which would have quietly falsified the regression check. Fresh mkdtemp
per run, removed in teardown.
Adds observer unit tests (challenge reported, body untouched, 403 included,
silent on success).
Signed-off-by: cliffhall <cliff@futurescale.com>
The challenge-driven plumbing did not reach the default CLI/TUI flow. With
no stored token in the legacy era, InspectorClient deliberately builds the
transport with no authProvider (so the SDK cannot open a browser before the
callback server is listening), which also leaves challenge interception off;
the 401 then surfaces as the SDK's headerless UnauthorizedError and the
runner calls authenticate() with no challenge in hand. So a server serving
metadata only at the advertised path still failed there — including for the
legacy showcase the README instructs.
Add a passive counterpart to the challenge interceptor: an observer fetch
that reports every 401/403 without touching control flow or the body. The
OAuthManager records the advertised URL from it and uses it in
authenticate() and its CIMD pre-registration. Latest-wins, including
clearing when a later challenge advertises none.
Also: the fixture helper dropped an explicitly supplied empty
resourceMetadataPath by truthiness, bypassing validation and silently
reverting to the well-known route.
Covered by a direct-client integration test that drives connect ->
authenticate() against the real server and asserts the advertised path was
requested and the well-known one never was; verified failing without the
fix.
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 3 on #2114.
The unit tests asserted our own `grant_types` array and stopped there.
What actually unblocks an Entra tenant is what the SDK puts in the
authorize URL, and nothing covered that — so an SDK change that stopped
keying its augmentation off `grant_types` would leave AADSTS90094
intact with every test green.
Adds the complementary case to the existing e2e suite, against a real
AS advertising `offline_access`: with the grant declined, the authorize
URL carries neither `offline_access` nor `prompt=consent`, while the
configured scope still goes out. Verified non-vacuous by reverting the
provider conditional — it fails. The positive test alongside it now
also asserts `prompt=consent` is present, which was unasserted despite
being the literal failure in the issue.
Drops an unjustified `as unknown as ServerEntry` I added in the App
test — AGENTS.md forbids it, and it hid a real mismatch (a partial
settings object against a complete one). Typed `ServerEntry` /
`InspectorServerSettings` fixtures instead, so a field added to that
interface later is a compile error rather than a missing value at
runtime. The `constructedOAuth` helper reads through the constructor's
own parameter type and needs no cast at all now.
Scopes the two doc comments that still claimed the flag removes every
`offline_access` — `StoredMCPServer.oauth` and `OAuthProviderConfig`,
both contradicting the caveat already stated elsewhere.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 2 on #2114. The setting was described as
guaranteeing "no offline_access, no forced prompt", which is not what
the SDK does.
`startAuthorization` appends `prompt=consent` for any `offline_access`
in the effective scope and never consults `grant_types`, and
`determineScope` skips its augmentation when the scope already carries
it. So the prompt still goes out when the user configured
`offline_access` in Scopes, or left Scopes blank and the resource
advertises it in `scopes_supported`. The flag removes the SDK's
*automatic* augmentation and nothing more; the docs, the help text and
the type say that now.
The form also flags the case a user can act on: opted out while
`offline_access` is still listed in Scopes. Whole-token match, so
`offline_access_extra` does not trip it.
Corrects the Clear stored OAuth state claim too. It clears the local
tokens and client information — it is not a revocation and cannot
reach the AS. A dynamically registered client re-registers on the next
connect (the new registration declaring only `authorization_code`); a
preconfigured `oauth.clientId` is reused unchanged.
Adds the two missing seam tests, both confirmed to fail without the
code they cover: App's connect wiring (deleting the spread fails two of
the three) and an /api/servers POST followed by a read-back, since a
200 alone never proved the opt-out persisted.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review on #2114. The setting was described as taking effect
"immediately", which is wrong twice over.
`applyLiveServerSettings` treats OAuth as connection-time-only, so a
connected client keeps the metadata it was built with — deliberately,
and identically to every other field in that section (client id,
secret, scopes, authorizationParams, endpoint overrides). Not wiring
this one field live; saying "next connect" instead, in the docs, the
checkbox description, and the type.
And two pieces of state outlive the setting: the registration already
held at the authorization server still lists the grant, and a refresh
token issued earlier stays usable, because the SDK's
`refreshAuthorization()` posts the stored token without consulting
`grant_types`. So the stated cost is not guaranteed until stored OAuth
state is cleared. Documented, rather than revoking credentials as a
side effect of a checkbox — Clear stored OAuth state already does that
explicitly.
Also adds the manager->provider seam test the other bridges have:
without it, deleting `requestRefreshToken:` from `createOAuthProvider`
would leave the feature dead with every other test green.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
The Inspector always declared `refresh_token` in its OAuth client
metadata. The SDK's `determineScope()` then appends `offline_access`
whenever the authorization server advertises it and the client declares
that grant, and `startAuthorization()` appends `prompt=consent` whenever
`offline_access` is in scope. Against Microsoft Entra ID that forced
consent prompt routes a non-admin user into the admin-consent workflow
and fails with AADSTS90094 — even after a tenant admin has consented.
Adds a per-server "Request refresh token" setting (default on) that
drops the grant from the client metadata, breaking that chain at its
source. Persisted as `oauth.requestRefreshToken` and written only when
off, so entries that never touch it keep a byte-stable round-trip. Web
edits it in Server Settings → Authorization; the CLI and TUI pick it up
from the same per-server settings via `buildRunnerClientAuthOptions`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>
An RFC 9728 `WWW-Authenticate: Bearer resource_metadata="…"` challenge was
parsed and then dropped by the normalized `AuthChallenge`, so every
challenge-driven `mcpAuth()` call ran discovery without it and probed
locations derived from the MCP server URL instead. OAuth failed outright
against any server whose metadata document lives off the well-known path.
The SDK handles this correctly inside its own transports, but the Inspector
never reaches that path — it intercepts the 401 and drives `auth()` itself
to capture the authorization URL and run the step-up scope union.
Carry the advertised URL on `AuthChallenge` as a string (it crosses the web
client's remote-backend boundary as JSON), convert it at the OAuth boundary,
and pass it to the standard challenge leg, the forced step-up leg, and CIMD
pre-registration — which runs before `auth()` and did its own discovery. A
malformed value is ignored rather than surfaced, matching the SDK's parser.
The callback leg needs nothing: SDK `auth()` persists the URL in its
discovery state.
Adds an `oauth.resourceMetadataPath` test-server mode that moves the
document off the well-known path entirely, so the integration test proves
discovery genuinely fails without the fix.
Closes#2071
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 2 on #2100, plus a collision the review did not find.
The default app-origin port was 6276 — which is already
RUNNER_OAUTH_CALLBACK_DEFAULT_PORT, the CLI/TUI loopback OAuth callback.
That port is fixed precisely so an OAuth app can pre-register
http://127.0.0.1:6276/oauth/callback, which makes it the one listener here
that cannot move when something takes it first. A running --web would hold
it and break a later --cli/--tui OAuth flow at its registered redirect URI,
and docs/mcp-app-review.md has those two running side by side. This
listener has a documented dynamic fallback, so this listener yields: the
default is now 6278 (6277 skipped — v1's retired proxy port, which the
migration guide tells people to stop forwarding). A test asserts the two
defaults stay distinct.
Review comments:
- POST /api/app-document now applies hono's bodyLimit BEFORE parsing. The
8MiB check ran after c.req.json() had buffered and parsed the whole
request, and bounded only `html` — so bulk in `csp`, or in keys the route
never reads, was fully materialized before the 413.
- `csp` is validated as an HTTP header value and length-bounded. The
controller hands it to res.writeHead() verbatim, where Node throws
ERR_INVALID_CHAR synchronously — inside the handler for a later,
unrelated request, outside this route's error handling.
- The app-origin port now propagates everywhere the other two do: the CLI
--print-handoff port-forward command (and its test), the Dockerfile's ENV
and EXPOSE, and the remote-review SSH recipe in docs/mcp-app-review.md.
Without it an app declaring _meta.ui.domain is unreachable through every
documented remote workflow.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0177mPHAdECD18nLLCTwR5rh
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 1 on #2100.
- POST /api/app-document: `null` and `[1,2]` are valid JSON, so the
destructure that followed threw OUTSIDE the parse try/catch and turned a
malformed body into a 500 instead of the documented 400.
- app-origin controller: the error handler cleared `server` but not
`origin`, and `publish` gates on `origin` — so an error after a successful
listen kept minting URLs on a dead port and the browser never took the
srcdoc fallback. Clear both, and drop the now-unfetchable documents.
- sandbox proxy: drop a stale `srcdoc` before navigating to a dedicated
origin (srcdoc wins over src), and restore the opaque-origin posture —
strip `allow-same-origin`, reset `innerOrigin` — before assigning `srcdoc`
on the fallback path. Not reachable today (each render remounts the
iframe), but the cost of being wrong about that later is an isolation
boundary.
- publishAppDocument tests: replace two `as unknown as Response` casts with
real `Response`s.
- docs: `frame-ancestors` admits the proxy AND the Inspector page (every
ancestor is checked); the shared helper falls back to loopback, not
`'none'`; the showcase origin is `127.0.0.1`, the default bind address.
- tests: null/array body, post-listen error, and both factories receiving a
working `publishAppDocument`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0177mPHAdECD18nLLCTwR5rh
Signed-off-by: cliffhall <cliff@futurescale.com>
