Addresses review round 2 on #2099.
- The /api/fetch proxy test set only uppercase HTTP_PROXY, so an ambient
lowercase `http_proxy` (which undici's EnvHttpProxyAgent prefers) or a NO_PROXY
could route the request elsewhere and fail the assertion. All six proxy vars
are now saved and cleared for the whole /api/fetch suite and restored in
afterAll — suite-level rather than per-test because the agent is memoized
process-wide once built. Verified the test still passes under a deliberately
hostile ambient env (`http_proxy=http://127.0.0.1:9 NO_PROXY='*'`).
- `verify:bundle-externals` reported success when `parseExternals` returned
nothing, so an ordinary tsup refactor — hoisting the array to a constant, or a
line break after `external:` — would silently disable the guard. It now fails
closed: every bundled client has externals, so "none parsed" means the parser
lost track of the config. Verified against exactly that refactor. The
per-client verdict moved into a pure `evaluateClient()` so all three failure
modes are unit-tested, and the failure epilogue is now printed only for a real
inlining, since "add the package to `external`" is the wrong instruction for
the other two.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JTHVxSu8AUgHRLvo1ntZ8H
Signed-off-by: cliffhall <cliff@futurescale.com>
Addresses the Copilot review on #2099.
- `/api/fetch` used the bare global `fetch`. That route is the browser's only
way out to the network — the web client's `environment.fetch` is
`createRemoteFetch()`, which forwards OAuth discovery and token requests
there — so a corporate-proxy user could connect to a server but never
authorize against it, and the PR's "OAuth is proxied now" claim held for the
CLI and TUI but not for web. It now uses the same proxy-aware base fetch, with
an integration test driving a real forwarding proxy (confirmed failing without
the fix).
- `verify:bundle-externals` only recognized an inlining that produced a
`<pkg>-HASH.js` chunk. A *statically* imported package is folded into
`index.js` and emits no chunk, so it would have passed. Detection now keys off
esbuild's `// <path>` module banners, covering both shapes. Because that
depends on unminified output, a bundle with no banners now fails as such
rather than reporting clean — enabling `minify` cannot silently retire the
check. Tests cover the entry-inlined case, the same-prefix case
(`undici-types` must not read as `undici`), and the no-banner case.
- The pre-push pipeline summaries in AGENTS.md, README.md, and
.github/copilot-instructions.md still described `verify:build-gate → smoke`
and omitted the new gate.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JTHVxSu8AUgHRLvo1ntZ8H
Signed-off-by: cliffhall <cliff@futurescale.com>
Corporate-proxy users could not connect to any remote MCP server from v2.
Two independent defects, both reported in #2067.
1. The web and TUI bundles inlined undici. It was declared in the root and
`clients/cli` manifests but named in no client's tsup `external` list, and
tsup auto-externalizes only what the *nearest* manifest declares — so
esbuild inlined 1.05MB of it into `clients/{web,tui}/build/`. undici is
CommonJS, so the emitted chunk's `require("assert")` hit esbuild's ESM
`__require` shim and threw `Dynamic require of "assert" is not supported`.
The rejected `import()` surfaced as "the `undici` package is not available",
telling users to install a package that was already present — and could not
have helped, because the specifier had been rewritten to a relative chunk at
build time.
2. `withProxyDispatcher` injected a *userland* undici `EnvHttpProxyAgent` into
Node's *built-in* fetch. That couples two different copies of undici at the
dispatcher handler interface, which is not stable across majors: Node 22
embeds undici 6, Node 24 embeds 7, Node 26 embeds 8. A userland undici 8
agent handed a built-in undici 7 handler is rejected with
`fetch failed: invalid onRequestStart method`, and the request never leaves
the process. So fixing the bundling alone would only have converted the
web/TUI failure into the CLI's.
Both halves of the pair now come from userland undici — its `fetch` and its
`EnvHttpProxyAgent` — which keeps the handler inside one copy and works
unchanged from the Node 22.19 engine floor through Node 26. Deleting the
wrapper in favor of Node's native `NODE_USE_ENV_PROXY` was not an option:
measured, it is unsupported on 22.19 (the floor) and only lands mid-22.x.
Because undici's `Response` is a different class from `globalThis.Response`,
the proxied fetch re-wraps each response as a genuine global `Response`,
streaming preserved. Without that, `res instanceof Response` is false for
callers that test it — including the SDK's OAuth error formatter, which would
degrade every message to `Raw body: [object Response]`.
The proxy now sits at the bottom of the fetch stack (`environment.fetch` for
the Node clients, and the `createTransportNode` default for the web backend,
its one caller that supplies no `fetchFn`) rather than wrapping the transport's
fetch. Proxying means *substituting* the fetch, not decorating one, so wrapping
a supplied `fetchFn` would discard it; installing it underneath lets
InspectorClient's own wrappers compose over it. A side effect worth having:
OAuth discovery and token requests go through `environment.fetch` too, and were
never proxied before.
Guarding it:
- `undici` is added to all three tsup `external` lists and removed from
`clients/cli/package.json`, so it is root-declared only, per the placement
rule in AGENTS.md. The two installs had already drifted (root 8.5.0 vs cli
8.9.0).
- New `npm run verify:bundle-externals` fails if any package a client declares
external was inlined into that client's bundle anyway. It reads the *built
output*, not the config — those two disagreed for four releases — and derives
what to check from each client's own `external` array, so a newly
externalized package is covered without editing the guard. Verified against
the real pre-fix artifact, which it flags.
- The proxy tests now drive a real forwarding proxy rather than a mock. That
matters: defect 2 was a runtime interface mismatch, which no mock can catch
because no mock ever dispatches.
AGENTS.md's must-not-bundle rule (added by #2082) is extended with why undici
slipped past it, and notes the `node -e` trap — probing the bundled chunk with
`node -e` falsely succeeds, because it exposes a global `require` that
satisfies esbuild's guard.
Closes#2067
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JTHVxSu8AUgHRLvo1ntZ8H
Signed-off-by: cliffhall <cliff@futurescale.com>
Round 24 produced no inline comments again. Of the three suppressed items, one
is a pre-existing data-loss bug that predates #1950 entirely.
**Editing a server's config deletes its stored OAuth client secret — from the
keychain, not only from the stores this PR adds.** Filed as #2084 and fixed
here, since the review surfaced it and these lines were already changing.
`useServers.updateServer` deliberately sends `{ id, config }` with no
`settings`, so a config-only save cannot wipe persisted settings. The route
reads that as `preserve` and re-derives settings from the *on-disk* entry —
which by #1356's design no longer holds the secrets. The submitted secret set
is therefore empty, `expectedSecretFields` always lists the OAuth slot, and
the reconcile deleted a value the user never touched. The rename branch was
given `mergeRenameKeychainSecrets` for exactly this reason; the in-place
branch never got the equivalent.
The first fix was too broad — skip obsolete deletion whenever the intent is
`preserve` — and an existing test caught it immediately: stdio `env` is part
of `config`, so a config-only PUT that drops a key really has retired that
secret. Obsolete detection is per-field now:
- **`env:` fields** are implied by the entry's shape, so the new entry no
longer expecting one is proof it was retired.
- **The OAuth slot** is implied by nothing, so absence counts as "cleared"
only when the caller spoke about settings at all.
Both directions are tested, and the pre-existing env-key reconcile test still
passes.
Also:
- `refreshingPersist` moves to `lib/`. The combinator computes nothing, but
what it exists to do is sequence side effects, which is the `lib` half of
the repo's split; I judged it the other way on the grounds that the
function is a pure higher-order one, and the rule is about what a module is
*for*.
- `SecretStorageInfo.plaintext` is documented as file-only and *omitted*
elsewhere rather than "always false", which is what the producer does and
what `usableSecretStorage` enforces — the old wording invited a consumer to
build the exact shape the web client discards.
Full `npm run ci` green, staged tree verified first.
Signed-off-by: cliffhall <cliff@futurescale.com>
Round 20, and almost all of it is fallout from round 19's two changes. Both
were right in direction and wrong in the detail, in the same way: each solved
its case by treating a broad thing as a proxy for a narrow one.
**The durability guard wrote newly entered secrets to disk in plaintext.**
Preserving legacy plaintext for a session-scoped store is correct — stripping
it moves the only durable copy into RAM. Using *the whole submitted entry* as
the proxy for "legacy" was not: it turned
`MCP_INSPECTOR_SECRET_STORE=memory` into "write every new secret to
`mcp.json` in the clear", while the footer for that very store says secrets
are written nowhere. A security regression introduced by a data-loss fix, and
the same overshoot in `client.json`.
Provenance decides now. A secret field is kept on disk only when it was
already there, unchanged, in the file being overwritten; a newly entered or
changed value is stripped and lives in the session store, which is exactly
what the footer promises. Where provenance cannot be established — an
unparseable prior file — the value is treated as new, because writing a
secret to disk on a guess is the failure that matters.
**The atomic claim was not as safe as its comments said.** Four ways:
- A crash between the claim and the copy left only
`secrets.json.migrating-<pid>`. The next run checked the canonical path,
found nothing, selected the keychain, and every stored credential vanished
with no warning. Orphans are now adopted at startup.
- Restoring used `rename`, which on POSIX **replaces** its destination — so
the restore path could overwrite a `secrets.json` a concurrent writer had
recreated, destroying the later successful write that claiming exists to
protect. The comment beside it asserted the opposite. `link` + `unlink` is
the no-clobber primitive this needed.
- Every claim failure was read as a lost race. `EACCES`/`EROFS` leave the
file exactly where it is, and returning quietly selected the keychain while
file-backed secrets sat there invisible. Only `ENOENT` is benign now.
- A failed cleanup left a plaintext-capable snapshot on disk with a silent
catch, while the app reported the keychain as the only store. It names the
path now.
**A third secret-entry surface had no disclosure at all.** `ServerConfigModal`
takes stdio `env` values, which are extracted into the store exactly as
Server Settings' are — and the footer's own doc comment claimed there were
only two such dialogs. It has the footer now, and every server write that can
carry a secret (config submit, both import flows) refreshes the descriptor,
not just the settings save.
Full `npm run ci` green, staged tree verified first.
Signed-off-by: cliffhall <cliff@futurescale.com>
Round 19 found real data loss rather than process noise, in two places.
**An unrelated settings edit destroyed the secret the durability gate had
just preserved.** The GET migration withholds its strip for a session-scoped
store, so `mcp.json` stays the durable copy — but the POST/PUT paths always
wrote the stripped shape. The GET also returns the rehydrated secret to the
settings form, which resends the whole object when the user changes anything,
so editing a timeout wrote the stripped entry, moved the only durable copy
into RAM, and lost it at exit. Every operation reported success. The client
config had the identical shape: the read-path guard was there, the write path
had none, and saving a CIMD URL did the same thing to the IdP client secret.
Both write paths now agree with their read paths: while the store cannot
outlive the process, the file stays the durable copy.
**The hand-off's delete is no longer a check-to-delete race.** Re-reading
immediately before `rm` caught writes that had already landed and was blind to
the one arriving between the check and the delete — destroying a *later*,
already-successful write, which is worse than the optimistic-write residual
and, unlike it, fixable without a lock. The file is claimed by an atomic
rename before anything reads it, so the migration operates on a snapshot
nobody else can reach and a writer that recreates `secrets.json` is simply
untouched.
Claiming introduced a failure mode the existing tests caught immediately: a
migration that aborts must not leave the secrets stranded at the staging path.
The snapshot is restored on any incomplete outcome, and when the live path has
been recreated in the meantime it is left in place with its location logged —
an operator-visible orphan is recoverable; clobbering someone's write is not.
Also:
- **`setMany`**, the write-side twin of `getMany`. `set` costs a whole-file
read-decrypt-encrypt-write-verify cycle per call, and the settings form
resends a server's full `env` map on any edit, so a 10-variable stdio server
spent ~30 scrypt derivations saving an unrelated change. One cycle now; the
keychain keeps its parallel individual writes.
- **A doc that contradicted its function.** `readOnDiskEncryption` still said
it "never decrypts, so it needs no passphrase" — round 12 made it
authenticate precisely because structure is not access.
Full `npm run ci` green, staged tree verified before running it.
Signed-off-by: cliffhall <cliff@futurescale.com>
**Round 17 pushed a branch that does not compile, and I reported it green.**
`refreshingPersist.ts` and its test were new files; `git commit -s -a` stages
modified *tracked* files and not untracked ones, so `App.tsx` went to origin
importing a module that was never committed. `npm run ci` passed against my
working tree, where both files exist — so the claim was true of what I had and
false of what I pushed, which is the worst shape a green result can take.
That is the second staging mistake on this branch. The first was `git commit
-o <paths>`, which commits whole files rather than my hunks and swept in a
collaborator's in-flight tests. Both come from trusting a flag's shorthand
instead of reading what was actually staged, so from here the commit goes
through an explicit `git add -A` plus a look at `git status --short` before
the gate runs — the check is cheap and it is the only one that describes the
pushed state rather than the local one.
Also replaces the `as unknown as` on the test logger. The rules prohibit it,
and it was doing real damage beyond style: forcing a four-method object
through `pino.Logger` erased the fact that the stand-in was not one, so a
change in how the server logs — a child logger, bindings, a different level —
would have gone unnoticed. It is a real pino logger writing to a capturing
destination now, which needs no cast because the destination interface is
satisfied structurally.
Both remaining suppressed comments are the cross-process guarantee, tracked
in #2082.
Full `npm run ci` green, run against a fully staged tree this time.
Signed-off-by: cliffhall <cliff@futurescale.com>
Three findings. The first is the fourth consecutive one in the refresh path,
so it is answered by making that path testable rather than by patching it
again.
**The refresh-after-persist wiring was covered by nothing.** It lived as two
inline `await …; refresh()` pairs in `App.tsx`, which is deliberately outside
the coverage `include` — so the hook's tests and the modals' tests could both
pass while the callback joining them was broken, leaving the security footer
stale after exactly the write that changed it. Three of the last four rounds
found defects in this path; none of them could have been caught by a test,
because there was nowhere for one to live.
Extracted to `utils/refreshingPersist`, which is inside the gate and has its
own tests: refresh runs after the persist resolves (not before — refreshing
first re-reads the descriptor the pending write is about to invalidate),
arguments pass through unchanged (the server-settings caller is
`(id, settings)`, and a transparent wrapper is the difference between saving
the entry and dropping it), and a failed persist refreshes nothing while
propagating the error.
**A tmpfs-style false promise, in the path fallback.** With neither `HOME`
nor `USERPROFILE` set — an ordinary service environment —
`defaultSecretFilePath` returned `.mcp-inspector/secrets.json`, a *relative*
path, while `FileSecretStoreOptions.filePath` and the `SecretStorageInfo.path`
the footer offers to copy both promise an absolute one. Resolved like the two
override branches above it.
**A warning that was usually false.** The session-store notice fired on every
`/api/servers` read whenever the store was in-memory — including for the
default empty catalog, where "plaintext values are left on disk" is not true
of anything, repeated on every list refresh. It is now emitted lazily, at most
once per server instance, and only when the loop actually preserves a
non-empty secrets set. A log line that is usually untrue is one people learn
to skip, which costs it the occasion it matters.
The remaining suppressed comment is the hand-off's check-to-delete race,
tracked in #2082.
Full `npm run ci` green.
Signed-off-by: cliffhall <cliff@futurescale.com>
Both review findings from round two.
The routing change sits in the funnel BOTH entry points share, but only the
legacy inbound handler was driven by a test — the modern leg, where an
elicitation arrives as an `input_required` result the MRTR driver unpacks and
retries, was covered by construction alone. Adds
`mrtr_app_choose_option` (a modern MRTR tool whose *embedded* elicitation
carries `_meta.ui.resourceUri`) and drives it live over a real transport: the
app answers, the retry carries that answer through `inputResponses`, and the
server echoes it back as the tool's result. A second case cancels the tool
call while the app is up and asserts the request-scoped signal aborts; a third
runs the same tool against a modern server that never advertised the
capability and asserts the native queue takes it.
The cancellation case turned up a real gap: `tryAppElicitation` mounted an app
for an already-aborted signal (the caller cancelled during an earlier MRTR
round), leaving a modal nobody was waiting on and a promise that could never
settle. It now throws the abort straight through.
Also fixes the smoke count in the root README, which still said "two"
headless-Chromium smokes and then introduced a third.
Signed-off-by: cliffhall <cliff@futurescale.com>
Let a server hand a form `elicitation/create` to an MCP App and return the
app's ordinary `ElicitResult`, with the native elicitation form as the
fallback on every failure.
No second extension, no custom method, and no custom result shape: the only
new wire surface is a nested `elicitation` flag on the existing
`io.modelcontextprotocol/ui` extension on each side, plus
`_meta.ui.resourceUri` on the request. App rendering is selected only when
all four gates hold — client `elicitation.form`, client MCP Apps MIME type,
the nested `elicitation` setting on BOTH peers, and a valid absolute `ui://`
URI on the request.
Only the web client advertises the nested client-side setting, and only
because it has a sandbox renderer: supplying
`InspectorClientOptions.appElicitation` is what opts a client in, so CLI and
TUI keep advertising the MIME type without ever claiming they can resolve an
elicitation through an app.
Routing lives in the one funnel both entry points already use
(`enqueuePendingElicitation`), so the inbound handler and the MRTR driver
cannot diverge. Ownership is request-scoped — a per-request id keys the
renderer, iframe and bridge — so two concurrent elicitations can never
resolve through each other's bridges. An explicit `decline`/`cancel` is a
completed elicitation and goes back to the server; everything else (absent or
malformed metadata, resource/sandbox/bridge failure, an app with no
elicitation capability, a timeout, an invalid result or one that fails the
requested schema) falls back.
The Inspector speaks the ext-apps#733 / SEP-3118 wire protocol but cannot yet
consume its helpers — the released `@modelcontextprotocol/ext-apps` (1.7.5)
predates that PR. `core/mcp/appElicitation.ts` and
`AppRenderer/requestAppElicitation.ts` mirror it exactly and are marked for
deletion once a release containing it ships. One consequence needed its own
seam: ext-apps 1.7.5 parses the view's `ui/initialize` through a schema that
strips `elicitation`, so an app that correctly advertises it would look like
one that did not, silently turning every negotiated elicitation into a
fallback. `AppRenderer/appCapabilities.ts` records the raw frame instead, and
prefers the bridge's own value once it carries the key.
`AppRenderer` now takes an `AppRenderSource` union rather than a `Tool`, so
the same renderer and bridge factory serve an App tool and an elicitation
without either faking the other's shape.
Adds the public fixture (`app_choose_option` + a self-contained
`ui://demo/choose-option.html` app covering accept/decline/cancel), its two
showcase configs, and `smoke:web:elicit`, which drives the negotiated path
end to end in headless Chromium and then the same tool against a server that
never advertised the capability, asserting the native form takes it.
Closes#1854
Signed-off-by: cliffhall <cliff@futurescale.com>
Two changes, interleaved in the same files because both are about the
same read-modify-write path.
## The lock is gone (maintainer decision on the round-5/7/8 finding)
Three consecutive review rounds found a real race in the `mkdir` lock, and
the last one is not patchable with what Node exposes: claiming a stale lock
needs compare-and-swap on a directory entry (`renameat2(RENAME_EXCHANGE)`).
Without it, `canSteal` inspects one lock but `rename` acts on whatever is at
that path *later*, so a waiter that lost the first race can move the
winner's fresh lock aside and enter alongside it.
Rather than patch it a fourth time, writers now collide and the loser
notices. `mutate` reads `M0`, applies the mutation to get `M1`, writes it,
then reads back `M2`; if they differ, someone wrote in between, so it
re-applies onto what they left and retries, bounded.
The comparison is over the **whole map**, and that is the load-bearing
detail. Checking only the entry we touched passes in exactly the case that
loses data — ours is present, and it is the *other* writer's that is gone.
With A and B from the same `M0`: A writes `MA`, B writes `MB` and clobbers
A; B verifies and correctly sees `MB`; A verifies, sees `M2 !== MA`,
re-applies onto `MB`, writes again. Both survive, and the writer that was
clobbered is the one that repairs it.
Two things this deliberately does not claim:
- It is **not mutual exclusion**. A process that crashes between its write
and its verifying read cannot notice it was clobbered, and nothing else
will. That window is strictly narrower than the lock's — which lost
updates with every participant alive — and needs no primitive Node lacks.
Stated in the code rather than left to be rediscovered in review.
- Non-convergence **fails loudly**. After five lost rounds `set` throws
rather than returning, because returning would reintroduce the silent
loss with extra steps. `delete` stays silent, per the interface contract.
`serialize` is now keyed on the resolved path **process-wide** rather than
per instance. Two `FileSecretStore`s on one file are ordinary — the resolved
store holds one and `absorbFileSecretsIntoKeyring` builds another — and a
per-instance queue does not order them against each other at all. The new
convergence test caught this on its first run: two instances racing `set`
lost an entry. In-process is now correct by construction; the verify covers
what it cannot see, a second Inspector process.
The hand-off's read → copy → delete was the lock's other user. Its delete is
now conditional: the file is re-read immediately before removal and compared
against what was copied, and it declines to delete when they differ. Leaving
the file costs nothing (keychain-wins makes the next run idempotent) while
deleting on a stale read costs a secret.
## Round 9 (two inline, two suppressed)
- **`KeyringSecretStore.getStrict` let raw binding errors escape.** Both
plaintext migrations keep their source file only when they catch
`SecretStoreUnavailableError`, so a transient keychain failure 500'd
`GET /api/servers` instead of abandoning the migration. Wrapped, as `set`
already does.
- **`FileSecretStore` had no strict read at all**, so `secretStoreGetStrict`
fell back to the tolerant `get` and mapped every read failure to `null` —
leaving the exact inversion the seam was introduced to prevent open on the
store it was introduced for.
- **`readOnDiskEncryption` reported "encrypted" for an envelope it cannot
open.** Naming the cipher is not the same as being openable: a file with no
`kdf`/`data`, or a `data` that is not `iv.tag.ciphertext`, produced the
quiet reassuring footer for a file whose next save was guaranteed to fail.
- **`/api/config` kept a stale descriptor when the resolver returned
`undefined`**, because the conditional spread left the startup value
standing. Assigned unconditionally now; `c.json` omits the key.
Full `npm run ci` green.
Signed-off-by: cliffhall <cliff@futurescale.com>
The lock finding from this round is deliberately not addressed here — see
below.
- **My round-6 strict-read fix was inert in production.** `DeferredSecretStore`
is what `defaultSecretStore()` returns, and it did not forward `getStrict`
— so `secretStoreGetStrict` fell through to the tolerant `get` for every
real caller, and the strictness existed only in tests that injected a
concrete store. Forwarded now, along with `getMany`.
- **The same tolerant-read bug remained in both pre-existing migrations.**
I added `getStrict` for the new hand-off in round 6 and never applied it to
the mcp.json and client.json migrations, which have the identical
shape: decide from `get`, then write on `null`. A transient keychain read
failure therefore let an older plaintext value overwrite a newer keychain
one — and the disk copy was then stripped, so the newer value was gone. Both
use the strict read now; the throw is a `SecretStoreUnavailableError`, which
both call sites already handle as "abandon the migration, keep the file".
- **The permission check made a false statement on Windows.** `stat` returns
synthetic bits there, and interpreting them yields a confident falsehood
either way — "0666 — not owner-only" about a file the ACL may restrict, or
"owner-only" about one it does not. The ACL is the real answer and this does
not inspect it, so Windows now reports `unknown`.
- **Rehydration cost one full decryption per field.** `get` reads and decrypts
the entire file, and both rehydration paths ask field by field, serialized
behind the store's own queue. Measured on this machine, scrypt at N=16384 is
~23ms per derivation, so a 20-secret encrypted catalog spent **~450ms of
pure key derivation on every `GET /api/servers`** — a visible stall, not a
micro-optimization. Added an optional `getMany` seam: `FileSecretStore`
decrypts once per server, everything else falls back to the parallel
per-field `get` that already suited it. A test counts derivations directly:
1 for a three-field `getMany`, 3 for the three `get`s it replaced.
Also broadens a log message that still said "Keychain unavailable" for a
store that may not be a keychain.
Full `npm run ci` green — 5,762 tests under the coverage gate.
Signed-off-by: cliffhall <cliff@futurescale.com>
Nine findings (six inline, three suppressed), all taken. Two were serious.
**The memory fallback made the migrations destructive.** Both migrations —
mcp.json in the remote server, client.json in `node-persistence` — lift a
plaintext secret off disk, write it to the secret store, and then delete it
from the file. That trade is safe only while the store is at least as
durable as the disk, which held by construction while `InMemorySecretStore`
was a test double. Making it a production fallback broke it: on an
unmounted container a plain `GET /api/servers` would have moved a user's
existing secrets into RAM and lost them at exit — destroying data on a
read.
`SecretStore` gains an optional `isDurable()`, absent meaning durable, and
both migrations withhold the disk delete when it answers false (the value
is still loaded, so the session behaves normally). The production fallback
is now a named `SessionSecretStore` rather than the bare
`InMemorySecretStore`, so the suite's many doubles keep standing in for a
*working keychain* — which is what they are there to be — and only code
that deliberately chooses RAM in production says so.
**Round 2's per-call descriptor never reached anyone.** Both web servers
resolved it once at startup and baked it into `initialConfig`, which
`/api/config` serves unchanged — so the freshness fix was inert for its
only consumer. The route now takes a `secretStorageResolver` and calls it
per request.
The lock needed three more fixes, and they compound:
- **Release removed whatever directory was at the path.** A holder whose
work outran the stale threshold would delete the *next* owner's lock,
letting a third writer into a section two processes believed they held.
Ownership is now stamped with a random token and release only removes a
lock still carrying it.
- **The timeout fell through to an unlocked write** — under contention,
the one moment another writer is provably active. It now throws
`SecretFileLockTimeoutError`, which `set` is allowed to do.
- **The first write of every process bypassed the lock**, because the
secrets directory does not exist yet and `mkdir` failed ENOENT. The
parent is created first. First writes are exactly when two processes are
most likely to start together.
Fixing the timeout exposed a fourth: with stale (10s) below timeout, a
waiter stole from a holder that was merely *slow* rather than dead —
reinstating the lost update. A holder now heartbeats its lock, so
"stale" means the holder is gone, and the timeout is what a live-but-slow
holder produces. Lock timings are injectable so the tests can exercise
contention, staleness, and timeout without spending the real intervals.
Also: the loose-permission state is carried into `SecretStorageInfo` as
`looseMode` and rendered as the caveat (it outranks the encryption caveat —
a file others can read is a live exposure, while "unencrypted" is a
property of a file only its owner can open), rather than living in a
startup log the browser cannot see; and `usableSecretStorage` validates
`pendingEncryption` and `looseMode` as the optional booleans/numbers they
are (`pendingEncryption: "false"` is truthy and would have printed advice
that is the opposite of the truth) and rejects file-only fields on a
keychain or memory descriptor.
Tests: both migrations against a session store with a durable control
alongside, so the guard is visibly the difference; `/api/config` serving a
changing resolver and falling back without one; lock ownership after a
steal, the timeout under a live holder, no-steal from a slow holder, and a
locked first write; `looseMode` through the descriptor, the caveat
ordering, and the tone; and five more rejected payload shapes. Full
`npm run ci` green — 5,704 tests under the coverage gate.
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot review round 1. The helper's contract is exactly-once teardown
and the test asserted it must not lean on Hono's own re-entry guard, but
neither the implementation nor the test actually enforced that: the
listener had no local guard and the test aborted only once.
Signed-off-by: cliffhall <cliff@futurescale.com>
Both SSE handlers installed their subscriber, primed the stream so
Firefox's fetch() resolves (#1858), and only then registered onAbort.
Hono's onAbort has no already-aborted replay and abort() fires only the
subscribers registered at that moment, so a client disconnecting while
the priming write was in flight fired nothing: the session event
consumer / server-list subscriber stayed installed, the mcp.json watcher
was never stopped, and the handler parked forever on a promise nothing
would resolve.
Add primeAndHoldSseStream, which registers one listener before the first
await — it both cleans up and releases the hold, collapsing the previous
duplicate onAbort registrations. Priming still happens after the
subscriber is installed, since callers treat the stream's first bytes as
proof they are subscribed.
Closes#1999
Signed-off-by: cliffhall <cliff@futurescale.com>
Addresses Copilot's round-15 review on #2035.
1. `readResourceFromTemplate` had no regression distinguishing this module
from the SDK expander it replaced: the existing test uses `report.txt`
and the transport tests expand separately and call `readResource`, so all
of them pass either way.
Writing that test corrected something I had asserted twice in this PR.
The first version used `foo/bar` -- and verified by reverting the wiring,
it still passed. `encodeURIComponent` DOES encode `/`, so the SDK expands
`foobar://events/{topic}` to `foobar://events/foo%2Fbar` exactly as this
module does. #1919 was the WEB PANEL's own `String.replace`, not the SDK
expander, which the TUI path was using correctly all along.
The test now uses `a!b`. `!` is a sub-delim: RFC 6570 requires it
encoded, `encodeURIComponent` leaves it bare, so the SDK returns
`foobar://events/a!b` and the assertion fails. Re-verified by reverting
the wiring -- 2 of the 10 tests fail (this one and the malformed-template
refusal), where before the change none did.
2. The module summary said "shared by every client" a dozen lines above the
paragraph explaining the CLI is not a consumer. Now says web and TUI.
Also records, on MAX_VALUE_LENGTH, the audit prompted by round 14: the
per-value ceiling was the ONLY guard lost with `UriTemplate.expand`. The
template-length (1e6), expression-count (1e4) and variable-name-length (1e6)
limits all live in the SDK's `parse()`, which runs from the constructor --
and strict expansion still constructs a `UriTemplate`, so they remain
enforced. The restored limit is the SDK's `MAX_VARIABLE_LENGTH`, read from
the pinned dist rather than assumed.
npm run ci passes.
Signed-off-by: cliffhall <cliff@futurescale.com>
Addresses Copilot's round-8 review on #2035 (one inline comment, one
suppressed). Both real.
1. RFC 6570 distinguishes an UNDEFINED variable, which omits its whole
expression, from one defined as the empty string, which expands. The
expander was calling `definedValues` on the way in, collapsing the two --
so `x{?q}` with `{ q: "" }` gave `x` instead of `x?q=`, and there was no
way for any caller, `readResourceFromTemplate` included, to request
either that or `x;q`. Key presence is now honored, including §3.2.7's
rule that the matrix operator drops the `=` for an empty value (`x{;q}`
-> `x;q`, while `?`/`&` keep it).
Dropping blanks is a FORM concern, so it moved to the two form
boundaries: the web panel seeds every declared variable with "" and the
TUI's ink-form returns "" for every untouched field, so neither can tell
"never touched" from "deliberately empty" -- but that is a fact about a
text input, not about the template, and it has no business deciding what
a template means for every other caller.
Six unit tests and one integration test encoded the old collapse and now
assert the corrected meaning. The prototype-collision cases get stronger
rather than weaker: a blank `{?toString}` expands to `?toString=` -- its
OWN empty value -- where the original bug expanded the inherited
function's body.
2. The expansion error is now `role="alert"`. It can appear while typing (a
pasted value that cannot be encoded), and plain text is not announced, so
a screen-reader user met a disabled Read Resource with no stated reason.
npm run ci passes.
Signed-off-by: cliffhall <cliff@futurescale.com>
Addresses Copilot's round-7 review on #2035. Two suppressed comments, both
real; each reproduced against this branch before acting.
1. `parseVarSpec` was a "strip the first `*`, trim, split on `:`" pass, so
it accepted shapes RFC 6570's varspec production does not admit --
measured on the previous commit:
x://{*id} name "id" => "x://abcdef" explode is a TRAILING modifier
x://{id*:3} name "id" => "x://abc" the two modifiers are exclusive
x://{ id } name "id" => "x://abcdef" whitespace is not a varchar
x://{a b} name "a b" => "x://1" ditto -- and it emits into the URI
`{id*:3}` is the worst of them: it silently truncated the value through a
combination the grammar forbids. Parsing is now one anchored regex --
a varname, then at most one trailing `*` or `:1-9999`.
The name set is `varchar` widened by `-` and `~`, deliberately. A name is
the one thing expansion emits into the URI WITHOUT encoding it (`;`, `?`
and `&` write `name=value`), so the rule it has to satisfy is "needs no
encoding" -- RFC 3986 unreserved plus pct-encoded, of which the RFC's own
varchar is a subset. Strict varchar would refuse `{user-id}`, which real
servers publish and which the SDK's matcher round-trips; refusing to read
a resource that demonstrably works is a worse outcome for a debugging
tool than a name one character outside the production. Everything now
rejected is a shape we would otherwise have to guess at.
2. The TUI built its "Missing required template variable(s): ..." list with
its own bare `values[name]` filter while gating on `hasRequiredValues`,
which reads own properties only. For a variable legitimately named
`constructor` the two disagreed: `({})["constructor"]` is `Object`, whose
`.length` is 1, so the list came out empty and the error named no field
at all while the submit stayed blocked. Both now come from one exported
`unmetRequiredGroups`, with `hasRequiredValues` defined in terms of it.
Also renames a test whose title still said multi-name query expressions were
left "to the SDK, which handles them" -- expansion has been owned here since
935113d1; the behavior it asserts is naming, not delegation.
npm run ci passes.
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>
Addresses Copilot's round-5 review on #2035 — eight suppressed comments,
all of them real.
The consequential one overturns behavior I copied from the SDK and
defended in two earlier rounds. RFC 6570 expands each expression
independently; it carries no cross-expression state. The SDK tracks
whether a query expression has already emitted and rewrites a later
`{?two}`'s `?` to `&`. That looks friendlier and is wrong, and its own
matcher says so: for `x{?one}{?two}` it expands to `x?one=1&two=2`, and
`match("x?one=1&two=2")` on that same template returns **null**, while
`match("x?one=1?two=2")` returns both variables. So the rewrite emitted a
URI the advertised template cannot match — #1919 one level up. A server
that wants a continuation advertises `{?one}{&two}`, which expands to `&`
and matches; producing that shape is the server's choice to declare, not
ours to infer.
Also:
- `{id:abc}` was silently treated as `{id}`, and a test codified it. RFC
6570's max-length is `%x31-39 0*3DIGIT`, so `{id:}`, `{id:0}`,
`{id:abc}` and `{id:10000}` are invalid templates rather than templates
with an ignorable modifier — and the SDK's constructor accepts them all,
so nothing else rejects them. Strict expansion now throws; lenient
returns the raw template. Discovery stays lenient so the panel still
renders rather than going blank.
- The TUI could submit `{a,b}` completely blank while the web panel
blocked the same request: ink-form has no way to express "any one of
these", so its members are left optional and the modal needs its own
group check. Added, with tests — it was also below the coverage gate
without them (functions 76.92%).
- Added the integration test the config never had: it resolves the
checked-in config and drives all four reads over a real transport, so a
misspelt preset fails there rather than only when someone runs the repro
by hand.
- "All three clients" was false. The CLI has no template form and its
`resources/read` passes the already-expanded `--uri` straight to
`readResource`, so nothing here runs for it. Corrected in the README,
AGENTS.md, and the module comment, and dropped a docblock still
describing the SDK-delegation path removed last round.
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>
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>
Copilot round 2 (suppressed comments). Five of the seven findings were the
accessibility of the new row list and one stray cast; all are addressed here.
- The reserved-key reason is now the key input's `error` **string** rather than
a bare boolean plus free-standing text. Mantine renders it in the input's own
error slot and wires the `aria-describedby` association, so a screen reader
learns *why* the field is invalid; a boolean only set `aria-invalid`. This
also deletes the duplicate `ReservedParamError` text and its Stack wrapper.
- Each control carries a per-row `aria-label` — the section heading and the
placeholders are not programmatically associated with the inputs, so the key
box, the value box, and a remove button announcing only "X" were
indistinguishable to assistive technology. The label names the row by its key,
falling back to "row N" while the key is still blank.
- Dropped an unnecessary `as unknown as Record<string, unknown>` in the
servers-route integration test: `MCPConfig.mcpServers` is already typed as
`Record<string, StoredMCPServer>`, so the entry's `oauth` reads directly.
Three tests cover the new behavior (accessible names, the blank-key row-number
fallback, and the aria-describedby association), each verified to fail against
the pre-fix component. No change to the reserved-key rule or the EMA boundary.
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot round 10.
The scan no longer publishes its malformed entries into the tools/list report
at all. Round 9 only stopped it overwriting the aggregate's; the remaining
case is unsalvageable for a stronger reason than 'two different requests':
- a FAILED strict aggregate runs salvage, which writes the report itself;
- a SUCCESSFUL one means every entry parsed, so the rendered list contains no
malformed entries.
So by the time the scan finds something, either the aggregate already reported
it with indices matching the screen, or the aggregate succeeded and what the
scan found is provably not in the list being displayed. Publishing could only
ever name entries the user cannot see.
Nothing is lost: listToolsForScan still marks its own refused response, so the
finding surfaces on the Protocol entry for the exchange that carried it, where
it is true. The per-page accumulation and raw-entry offset existed only to feed
the removed publish, and go with it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RDXCSrnSBV7vZfxycG5Z6E
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot round 9.
The exclusion scan is a separate, uncached tools/list exchange, so its entries
need not be the ones rendered — the server can answer differently, and a
cache-served aggregate is a different list version. It was overwriting the
aggregate's report wholesale, replacing accurate indices with indices into a
list nobody is looking at. The aggregate now owns the report; the scan
publishes only when the aggregate left nothing, which is the cache-served case
that made it publish at all (round 2).
A conforming re-fetch no longer swallows the mark: the list can change between
the strict page and the lenient one, and a clean second response says nothing
about the first, which was refused. It is marked with the original decode
error when there is nothing per-item to name.
The scan's fallback re-fetch now goes through invokeMcpClient, like the strict
call beside it and like salvageList's walk, so an auth challenge on that
request can still be recovered instead of being swallowed by listAllTools's
best-effort scan catch.
Also fixes listTools dropping an empty-string cursor (truthiness rather than
!== undefined) — its own fallback leg already preserved it, so the two legs of
one scan page disagreed about which page they had fetched.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RDXCSrnSBV7vZfxycG5Z6E
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot round 8.
The era contract validated with the wire schema but returned the raw input,
so the wire schema's normalization was discarded — both era schemas are
z.object, which strips unknown top-level fields as the strict codec does, so
salvaged tools could carry fields the strict path drops. The parsed output is
now what gets normalized and returned. This also simplifies the bridge: issues
come straight from the wire parse with their paths intact, so the custom-issue
forwarding (and its early describeIssue application) is gone.
The malformed-entry report is written by the aggregate walk's salvage, but in
paginated mode the tools/prompts/resources panels render the paged stores,
which neither write nor clear it — leaving a stale warning above a freshly
fetched page it does not describe, with indices into an aggregate the user is
no longer looking at. Those reports are now suppressed in paginated mode;
resource templates keep theirs, having no paged store. Suppressing rather than
adding per-page salvage: pointing at the wrong entry costs more than staying
quiet, and per-page salvage is a feature, not a display fix.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RDXCSrnSBV7vZfxycG5Z6E
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot round 7.
The era contract was derived from the neutral ToolSchema plus one override,
which diverged from the real wire schemas both ways. Measured against SDK
2.0.0 on a real connection:
- modern accepts `inputSchema.properties: 5` and `required: [1]`; ours
rejected them, so a malformed sibling made salvage DROP a conforming tool
and report it as malformed — a false accusation from the screen whose job
is to say which entry is wrong.
- legacy rejects `outputSchema.properties: 5` / `required: [1]`; ours
accepted them, readmitting entries the strict path refused.
- modern rejects a non-string `inputSchema.$schema`, legacy accepts it;
ours accepted on both.
Each era is now expressed as its own wire schema, bridged to Tool by
`eraToolItemSchema`. The bridge exists because both eras accept values the
exported Tool type cannot describe (the SDK returns them typed as Tool
anyway), so the wire schema's output type is irrelevant and z.custom<Tool>
supplies the caller's type — which is what avoids an `as unknown as` here.
Forwarded issues keep their path so the report still names the field.
The full measured matrix is pinned table-driven in listSalvage-era.test.ts:
each row asserts the SDK's verdict AND that our contract agrees, so a
divergence that closes or widens fails loudly.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RDXCSrnSBV7vZfxycG5Z6E
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot round 6.
Bound both fallback walks with LIST_MAX_PAGES, which is now also passed to
the SDK client as `listMaxPages` so the strict aggregate and its fallback
share one value by construction. A server emitting endlessly UNIQUE cursors
never trips the repeated-cursor guard, so the salvage re-walk and the
SEP-2243 exclusion scan would page and accumulate without bound. Hitting the
cap aborts rather than returning a truncated list as a complete one.
Mark the exclusion scan's refused response on every fallback-failure path.
`salvageList` can rethrow bare because ManagedListState marks from the error
it receives; the scan has no such backstop — listAllTools catches and logs a
failing scan by design — so an unmarked rethrow left a response the client
demonstrably refused rendering as a clean exchange.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RDXCSrnSBV7vZfxycG5Z6E
Signed-off-by: cliffhall <cliff@futurescale.com>
Real deployments gate IdP selection, prompt behavior, or audience on a query
parameter the OAuth/OIDC core specs never standardized (Keycloak's
`kc_idp_hint`, OIDC's `login_hint` / `prompt` / `acr_values`, Auth0's
`audience`), and none of them could be driven from the Inspector.
The Inspector never builds the authorize URL — the SDK's `auth()` does, and it
exposes no hook for extra parameters — so the merge happens at the one seam
that sees the finished URL: `BaseOAuthClientProvider.redirectToAuthorization`.
It lives in `core/`, so web, CLI, and TUI all inherit it.
- `core/auth/authorizationParams.ts`: the reserved-key list (`client_id`,
`code_challenge`, `code_challenge_method`, `redirect_uri`, `resource`,
`response_type`, `scope`, `state`), the key validator the form renders
inline, and the merge itself. Reserved keys are dropped with a warning
rather than silently overriding PKCE / the state binding / RFC 8707;
blank-key rows are skipped; the caller's URL is never mutated.
- Applied to the authorization request only — nothing wires it into the token
request.
- Persisted as `oauth.authorizationParams` (a string record) on the per-server
config beside `scopes`, edited as key/value rows in Server Settings ->
Authorization ("Additional authorization parameters"), with an inline error
per offending row plus a section-level warning.
- The enterprise-managed (EMA) leg redirects to a *different* authorization
server (the enterprise IdP), so it goes through the explicit
`redirectToExternalAuthorization` entry point, which skips the merge — a
`kc_idp_hint` or `audience` meant for the MCP server's AS must never be
appended to the IdP's authorize URL.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RDXCSrnSBV7vZfxycG5Z6E
Signed-off-by: cliffhall <cliff@futurescale.com>
Both findings are the same shape: the fallback accepting what the strict path
would not.
- `ModernResultEnvelopeSchema` typed `ttlMs` as any number, where the codec
requires a non-negative integer. `ttlMs: -1` (or `0.5`) plus one bad entry
would salvage, reporting the entry and hiding the envelope violation. Tightened
to `z.int().min(0)`.
- The item schema was the SDK's era-NEUTRAL `ToolSchema`, and the negotiated
era's wire schema is stricter in ways the SDK does not export. Verified
against SDK 2.0.0 on a real connection: on legacy, a tool with
`outputSchema: { type: "array" }` is refused by the strict aggregate but
accepted by the neutral schema; on modern, `execution` is STRIPPED by the
strict path but kept by the neutral one. Either way a salvaged list could show
— and let the user call — a tool the connection's era says isn't valid.
`toolItemSchemaForEra` closes both. Replicating rules that live in the SDK's
internals is a real cost; `listSalvage-era.test.ts` is what makes it safe to
pay, pinning each rule to the strict path's OBSERVED behavior so a divergence
that closes or widens fails loudly instead of rotting.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Signed-off-by: cliffhall <cliff@futurescale.com>
Copilot round 4: no new visible comments, two in the suppressed block. Both
valid.
1. The `type: [T, "null"]` encoding keeps its keywords at the top level, so a
nullable enum written that way carries the null *inside* the list:
`{ type: ["string", "null"], enum: ["envio", "recebimento", null] }`. The
collapse left it there, which broke both renderers in different ways — the
web dispatcher handed `null` to Mantine as option data, and the TUI's
all-strings check rejected the whole enum and fell back to a plain text
field. Neither is the dropdown the schema asks for.
The collapse now strips a `null` member whenever it flattens, since that
fact has already moved onto `nullable`. An enum of nothing but `null`
drops out entirely: it offers no value a dropdown could show, so the plain
widget serves the field better than an empty select. Covered in core and
end-to-end in both dispatchers.
2. The new integration test claimed in its own header to pin the registry and
config wiring, then called `createNullableFieldsTool()` directly — so a
misspelt preset case or a config naming a dead preset would have left it
green. A fair "this does not test what it says" catch. It now resolves the
checked-in `nullable-fields-http.json` through `loadConfig` →
`resolveConfig`, matching `duplicate-tool-names.test.ts`, and asserts the
resolved tool names before booting the server.
Signed-off-by: cliffhall <cliff@futurescale.com>
Two findings, both about the fallback accepting more than it should.
- The lenient page schema validated pagination and nothing else, so a response
carrying a malformed entry AND an unrelated top-level violation salvaged: the
entry was reported and the other violation vanished with the strict error that
caught it. The page schema is now DERIVED from the method's real result schema
with only its item array relaxed (`lenientListPageSchema`), so every other
field — `_meta`, `nextCursor`, anything a future revision adds — is still
validated. The modern era envelope the raw-wire path bypasses (`resultType`,
`ttlMs`, `cacheScope`) is checked explicitly for the same reason.
Salvage is also narrowed to `InvalidResult`: `UnsupportedResultType` means the
codec doesn't recognize what KIND of frame arrived, so it isn't a list with a
bad entry and reporting an item-level reason would hide a protocol failure.
Marking the Protocol entry still applies to both, hence two predicates.
- On every rethrow path the re-walk had already answered the method, so the
caller's own `markResponseRejected` (in `ManagedListState.refresh`) landed on
the lenient page that succeeded and left the refused response clean. The
correlation is restored before rethrowing — deleted when nothing had answered
before, so the outer mark is a no-op rather than a lie.
`nextCursorOf` / `rawItemsOf` narrow the page's fields instead of casting: the
runtime-keyed `extend()` loses the shape, and asserting it back would be exactly
the unjustified cast the last round removed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Signed-off-by: cliffhall <cliff@futurescale.com>
Three findings (one inline, two in the suppressed block), all valid.
1. `toConstOptions` still stringified every non-null `const`, so
`items: { anyOf: [{ const: 1 }, { const: 2 }] }` submitted `["1"]` where
the schema says `[1]`. Same wrong-type-on-the-wire problem that keeps a
numeric `enum` off the select path, so it gets the same answer: only a
string `const` is selectable, anything else falls to the JSON editor
where the value keeps its type.
2. `hasMissingRequiredFields` derived null admission from the *renderer's*
collapse, which is narrower on purpose — it only flattens a two-member
union. So `anyOf: [string, number, null]` renders through the JSON
fallback (where `null` is typeable) yet was still marked missing,
disabling submit on a value the schema accepts. Null admission is a
validity question, not a rendering one, so it is now its own predicate:
`admitsNull` in core, recognizing `nullable`, `type: "null"`,
`type: [..., "null"]`, and a null branch anywhere in an `anyOf`/`oneOf`
of any size.
3. Nothing exercised the `record_shipment` preset — the web and TUI tests
build the `anyOf` shape by hand, which verifies the renderers but
*assumes* the premise the whole fix rests on: that Zod's `.nullish()`
emits `anyOf: [<branch>, { type: "null" }]` with the enum on the branch.
A Zod output change or a registry typo would leave every unit test green
while the documented reproduction server quietly stopped reproducing.
Added an integration test that boots the preset over HTTP, asserts the
wire shape, and runs it back through the collapse — following the
`duplicate-tool-names.test.ts` precedent, which exists for the same
reason (#1957).
Signed-off-by: cliffhall <cliff@futurescale.com>
The visible review was clean; all five findings were in the suppressed block,
two of them real defects.
- The exclusion scan's aggregate offset was wrong twice over: `excluded.length`
counts only invalid-header tools, and on page one it also folded in the
current page's valid count. A reported index therefore pointed at a position
that doesn't exist — worst for an unlabeled entry, whose index is all the user
has. Replaced with a raw-entry offset advanced by each page's valid + malformed
counts.
- The scan's own refused response was never marked. It is a distinct exchange
from the aggregate's, and the lenient re-fetch moves the method correlation
off it, so its Protocol entry rendered as a clean success the client had in
fact refused. Marked with the captured id, as `salvageList` already does.
- Three double casts in the test fixture are gone: `ListPage` now models raw
wire data (`items: unknown`, `nextCursor?: unknown`), which is what a server
built to emit spec-forbidden shapes actually serves. Declaring the validated
type and casting malformed values through it asserted the opposite of what
these tests exist to check. The cursor map narrows to `string` at the one
place a cursor addresses a page.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Signed-off-by: cliffhall <cliff@futurescale.com>
All four findings taken (one visible, three in the suppressed block), plus the
`excludedTools` sibling leak the round-1 reply had deferred:
- `rawItemsOf` read a missing or non-array list member as an empty page, so a
top-level violation (`{ tools: "invalid" }`) on a later page was discarded
and the call returned a silently truncated list. It now returns `undefined`
for "not a list at all", and the walk rethrows the original strict error.
- A failure of the re-walk ITSELF (a malformed `nextCursor`, a connection lost
mid-walk) escaped in place of the original error, whose response is the one
`rejectedResponseId` was captured for. The walk is now wrapped: the original
surfaces, the secondary is logged so it stays diagnosable.
- The exclusion scan's lenient fallback discarded its malformed entries despite
the comment claiming they were reported. Observable when the aggregate is
served from cache: the salvage never runs, so nothing else reports them. The
scan now returns them and `refreshExcludedTools` reports them — only ever
adding, so a clean scan can't clear what the salvage path just wrote.
- An empty-string cursor is a valid cursor, and `cursor ? {cursor} : {}` dropped
it, re-requesting page one and duplicating its entries until the
repeated-cursor guard fired. Both walks now test `!== undefined`.
- `excludedTools` had the same session leak fixed for `malformedListItems` last
round: cleared only on an explicit `disconnect()`, so the `onerror`-with-no-
`onclose` route carried the previous server's set into the next connection.
Cleared on the start-clean path too — safe because every `listAllTools`
recomputes it, so this can only drop a stale set.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Signed-off-by: cliffhall <cliff@futurescale.com>
Three real defects it caught, plus the tests it asked for:
- The strict `tools/list` aggregate is filtered by the SDK for invalid
`x-mcp-header` annotations (SEP-2243), but the salvage path was not: one
ordinary malformed tool would trip the fallback and readmit every
invalid-header tool beside it. Salvaged tools now go through the same scan.
- `markResponseRejected` resolves "the last response received for this method",
and the re-walk answers that same method — so the mark landed on a salvage
page that succeeded while the actually-invalid response rendered clean. The
correlation id is now captured before the re-walk.
- The report was cleared only on an explicit `disconnect()`. The route out that
tears down nothing (`onerror` with no `onclose`) carried it into the next
connection, so it is now cleared on the start-clean `resetSessionState()` path
as well.
Fixing the first exposed two more, both found by the new test:
- The modern (2026-07-28) codec validates a result against the spec schema
before the caller's, so a lenient `client.request` is rejected exactly as the
strict one was — there was no salvage on the modern leg at all. It now goes
through the raw-wire channel that exists for the modern `tasks/*` methods.
`withModernEnvelope` is split out of `withModernTaskEnvelope` so a list
request doesn't claim the tasks extension the user may have disabled.
- A malformed tool also blanked the SEP-2243 excluded set, leaving a tool that
vanished for an invalid header with no explanation — #1909's own complaint one
level down. The scan walk now tolerates a malformed sibling.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Signed-off-by: cliffhall <cliff@futurescale.com>
The SDK validates a list result as a whole, so a single non-conforming entry
rejected the entire response and the panel showed "Couldn't load resources"
with zero items — the exact opposite of what a debugging tool should do with a
server that is 90% right. Reported against a PHP server whose empty
`annotations` object reaches the wire as `[]` rather than `{}`.
`listAll*` now falls back to a lenient re-walk when — and only when — the SDK
rejects a result it received (`InvalidResult` / `UnsupportedResultType`), keeps
the entries that validate, and reports the rest. A conforming server takes the
strict, cache-aware path exactly as before.
Salvaging is not accepting: the dropped entries are named with their reason in
a warning above the list, and the Protocol entry is still marked rejected, so
the wire truth survives. When every entry validates, the original error is
rethrown rather than swallowed — the rejection was about something the
per-item pass cannot explain.
`isClientDecodeRejection` moves to the new module and is now shared with
managedListState, which had the only copy.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Signed-off-by: cliffhall <cliff@futurescale.com>
Review feedback on #1987: `waitForDeadTransport` polled `/api/mcp/send`,
which answers `transport_error` from two branches — the `isTransportDead()`
short-circuit and the `catch` around a rejected `send()`. Only the first
implies the transport is marked dead, so the stop condition could not tell
them apart from the response alone.
That was safe in practice: `StdioClientTransport.send` rejects only when
`_process` is undefined, and the transport clears `_process` and calls
`onclose` in one synchronous block, so the `catch` branch is not observable
before `markTransportDead()` has run (confirmed empirically — 0/25 probes
returned early). But the safety rested on an SDK-internal ordering this test
neither states nor controls, and if it changed the send test would silently
cover the `catch` branch instead of the short-circuit it names — a regression
with no failing run attached to it.
Poll `/api/mcp/auth-state` instead. For a session connected without an
authState there is no auth provider, so `setAuthState` always throws and
`transport_error` there has exactly one source: the dead-transport
short-circuit. The route never awaits a transport response, so it cannot
hang either. Same cost, exact condition, no dependence on SDK internals.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WLYmVPVLdeYpE6aMTUsCks
connectDeadSession spawned a real node subprocess whose only job is to die,
then slept a flat 300ms for the death to propagate. That budget has to cover a
cold spawn(), node's startup, its exit, the parent observing the close, and
onclose marking the session dead — comfortable on an idle box, not under CI
contention or a machine running several agents.
When the budget was missed the session was still live, so isTransportDead() was
false and /api/mcp/send fell through to `await responseWait` for a tools/list
reply the dead child can never send. Nothing bounds that wait, so the test hung
to the project's full 30s ceiling. The timeout IS the symptom of losing the
race, which is why it never presented as a clean assertion failure.
Poll the observable condition instead. Two design points that are load-bearing:
- The probe is a NOTIFICATION (method, no id). requestIdForSendWait returns
undefined for it, so the route never awaits a response and cannot hang.
Probing with a request would reproduce the exact bug being fixed.
- Not /api/mcp/events, which looks like the natural "transport died" signal:
opening that stream on a dead transport calls sessions.delete(sessionId),
so the send under test would then answer 404 instead of transport_error.
Both terminal branches of the route answer transport_error — the
isTransportDead() short-circuit and the catch around a send() that threw on the
dead pipe — and either proves the subsequent real send cannot reach the hanging
path, so both are valid stop conditions.
Verified by reproducing the bug deterministically: with the wait reduced to 0
(the budget always missed) the test times out at 30s; with the poll it passes
10/10. Fixing the helper rather than the test also covers its second call site,
the auth-state test, which sits on the same race and had simply not lost it yet.
connect-crash.test.ts:147 is deliberately left alone. #1985 proposed giving it
the same treatment, but readSseEvents stops at transport_error, which arrives
whenever the child dies — so its 200ms sleep cannot hang, and changing it would
add risk for no gain.
Closes#1985
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Signed-off-by: cliffhall <cliff@futurescale.com>