Two pieces of hardening on the Claude Code workflow, plus a comment for a
finding that needed no change.
Same-repo gating. `Get PR details` now emits an `is_fork` output, and the
PR-head checkout and the `Run Claude Code` step are both gated on it. A fork
PR is declined outright — no checkout, no agent run, with the reason written
to the step summary — rather than falling through to a "metadata-only" review
of the base tree, which would trade an untrusted-code problem for a
wrong-tree one. A deleted fork (`head.repo` null) counts as a fork.
The head checkout also drops its `repository:` input: only a same-repo head
reaches that step now, and checkout's default `github.repository` is a value
no PR can influence.
SHA pins. actions/checkout, actions/github-script, and claude-code-action
were all on mutable major tags in a job that holds ANTHROPIC_API_KEY and
grants the agent Bash. Each is now pinned to a full commit SHA with the
trailing `# vX.Y.Z` comment Dependabot reads, so upgrade automation is
unaffected. Pinned to the v7 / v9 SHAs so this does not regress the bumps in
the open Dependabot PR #1922.
The failure fallback Copilot flagged turns out not to exist: `Checkout
repository` carries no status-check function, so GitHub applies an implicit
`success()` and skips it after a failed lookup regardless of the `outcome`
comparison. Its condition is now the explicit `== 'skipped'` and says so.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SW1p8E2uiyyLx4RKwrwSrt
From Copilot's fifteenth pass of #1962. `installDirs` filtered on the presence
of a lockfile, so a missing one silently removed that install from the
comparison rather than failing. For the root — the install every shared source
resolves from — that meant the guard could report success from client locks
alone, which is the deny-by-default contract inverted.
Enrolment is now by `package.json` (an install we are meant to compare), with
the root always enrolled, and a missing lockfile is a named, loud failure. A
`clients/` directory with no `package.json` is not an install and is skipped,
so a stray scratch dir isn't asked for a lockfile it should never have.
Three regression cases: a missing client lockfile in a fixture that IS skewed
(so failing open would show as a pass), a missing ROOT lockfile, and a stray
`clients/` dir that must not be enrolled.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
From Copilot's fourteenth pass of #1962. `hasReadableLockShape` and the
diagnostic it drives both promise "lockfileVersion 2+", but the check only
looked for a `packages` table with a root entry. A file declaring v1 while
carrying such a table was accepted, produced an empty version map, and could
make a real skew read as aligned once only one other holder remained — the
same fail-open the shape check was added to close, through the half of the
contract that wasn't enforced.
The declared version is now verified: a finite number, 2 or greater. Absent,
non-numeric, and declared-v1-with-packages cases are all in the table.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
From Copilot's thirteenth pass of #1962. This PR added `verify:dep-lockstep`
to the vouch loop in `verify-format-coverage.mjs` without a test for it. The
existing tests exercise `rootReachesScript` in isolation and the opposite
direction (dep-lockstep vouching for format-coverage), so a typo in a sibling's
name here would have left `test:scripts` green while that guard silently
stopped being enforced — the same "a gate that stops gating" failure the
vouch cycle exists to prevent, one level up.
Three cases, reusing the fixture-repo approach: dropping either sibling from
`validate` exits 1 naming that sibling, and with both wired no sibling is
reported missing. The last asserts on the *reason* rather than the exit
status, since the fixture has no client manifests to harvest globs from and
so fails later for an unrelated reason.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
From Copilot's twelfth pass of #1962. A `/// <reference types="x" />` pulls in
a package's declarations exactly like an import does, but TypeScript reports
it in `preProcessFile`'s `typeReferenceDirectives`, not `importedFiles` — and
only the latter was read. A shared source using one could therefore load the
root's copy of a package alongside a client's without it ever entering the
candidate set, so its skew would pass. Same silent-miss direction as the
earlier syntax gaps, through a channel the parser switch didn't cover because
I only read one of its two outputs.
`typeReferencePackageNames` maps a directive to both candidates, since the
name is the *type* name and not the package: `node` resolves to `@types/node`,
while a package shipping its own declarations resolves to itself. Returning
both over-approximates, the safe direction — whichever isn't installed drops
out downstream. Scoped names mangle `@scope/pkg` → `@types/scope__pkg`,
TypeScript's convention. `path` references name a file, not a package, and are
ignored.
The real candidate set is unchanged at 17: no shared source uses a triple-slash
reference today. As with the other syntax fixes, this closes the channel
against the one that arrives later.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
From Copilot's eleventh pass of #1962, and the most consequential kind of bug
for a gate: it failed *open*.
`topLevelLockVersions` returns an empty map for anything without a v2+
`packages` table — a lockfileVersion 1 file, or a malformed one. `main()` fed
that straight into the comparison, so such an install contributed no holders
at all, and a genuine skew among the remaining installs was reported as
aligned. The guard would have said OK to exactly what it exists to catch.
`hasReadableLockShape` now gates every discovered lockfile before any
comparison, requiring a `packages` object carrying the `""` root entry npm
always writes — "has a packages key" is not enough. An unreadable one is named
and exits 1, telling you to regenerate it. A JSON parse failure likewise
reports the file rather than dying on a bare SyntaxError with no path in it.
The empty-map behavior of the pure helper is kept and its test now says why it
is safe: the shape check rejects those inputs before they can reach
`findSkew`.
Three regression cases: a v1 lockfile in a fixture that IS skewed (so failing
open would show as a pass), a `packages` table with no root entry, and the
unit table for the predicate. Fixtures now write the `""` root entry, matching
real npm output.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
From Copilot's tenth pass of #1962. The suite covered only pure helpers, so a
regression in source enumeration, install discovery, lockfile loading, the
sibling-guard vouch, or the nonzero exit on real skew would have left it
green — and the negative cases had only ever been run by hand, once per review
round.
Five end-to-end cases build a throwaway repo and run the guard as a
subprocess: aligned installs exit 0; a skewed one exits 1 naming the package
and every holder while leaving the agreeing package unreported; a package
imported only by `test-servers/src` still fails (so dropping the second shared
tree can't pass); a shared source matching no tracked file fails; and a
`validate` chain that stopped running `verify:format-coverage` fails the vouch
even when the versions are fine.
The fixture must be realpath'd. On macOS `tmpdir()` is `/var/…`, a symlink to
`/private/var/…`, and the guard runs `main()` only when `import.meta.url` (the
resolved path) matches `process.argv[1]` — so launching it via the unresolved
path loads the module and does nothing, exiting 0 silently. That is how all
five first "passed" as no-ops, and the comment records it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
From Copilot's ninth pass of #1962. The one-line descriptions in AGENTS.md and
the README claimed "no install-crossing dependency resolves to two different
versions", which is stronger than the gate: candidates cover only packages the
shared sources name directly, and #1965 records a live counterexample
(`@modelcontextprotocol/sdk`, skewed across two install roots while the guard
stays green).
The boundary was already documented at length on the guard and in the AGENTS.md
rule; the summaries just hadn't been narrowed to match, which is how a doc ends
up asserting a guarantee nobody implemented. Both now say "directly imported"
and point at #1965.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
From Copilot's eighth pass of #1962. The no-op guard tested only
`files.length === 0`, which stopped meaning anything once
`SHARED_SOURCE_FILES` was added: `vitest.shared.mts` alone keeps the total
nonzero, so a moved or renamed `core/` would leave the guard deriving
candidates from a near-empty set and passing on skew it exists to catch.
`sourcesWithNoFiles` now requires every configured dir and every named file to
have contributed, and names the ones that didn't. Directory matching is on a
path separator, so a prefix sibling (`core-internal/`) can't vouch for a
renamed `core/`.
Also drops two test comments left stale by the move to `ts.preProcessFile` —
they still described the deleted TRIVIA regex.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
The guard resolves typescript from `clients/web`. The realistic way for that
to fail is a root install run with INSPECTOR_SKIP_CLIENT_INSTALL=1, which
leaves `clients/web/node_modules` empty — and a bare MODULE_NOT_FOUND stack
gives no hint that the fix is `npm install` at the root. Throw with that
guidance and the original error as `cause`.
Deliberately not a silent skip: an unrun guard guards nothing, which is the
failure mode the sibling-vouch ring exists to prevent.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
From Copilot's sixth pass of #1962.
Take the TypeScript-parser suggestion, which was raised across three rounds
and was right. The regex scan got both directions wrong and every widening
traded one failure for the other: it *missed* valid syntax (`import x =
require(…)`, an import-attributes argument, a comment between tokens — each a
silent miss, so the package never entered the candidate set and its skew
passed), and it *invented* names from prose, since `// adapted from "react"`
is indistinguishable from an import to a pattern that cannot tell code from a
comment. The earlier prose test only passed because `cwd omitted` isn't a
valid package name.
`ts.preProcessFile` is the cheap answer I'd missed while arguing against the
full AST: TypeScript's lightweight pre-parse scanner, no parse tree and no
type checking, purpose-built to return module specifiers. It handles every
import form, trivia, strings, and regex literals, and never reads a comment as
code. typescript resolves from `clients/web`, which already carries it, via
the `createRequire` base `smoke-web-browser.mjs` established — a bare import
would resolve relative to `scripts/`, not the cwd.
The derived set drops 18 → 17 real names: every genuine package is retained
and the one prose artifact is gone. Regression tests now use *real* package
names in comments and strings, the case the old test couldn't cover.
Also records the KNOWN BOUNDARY the review's other finding identified.
Candidates cover what the shared sources name directly, so a package reaching
the program only through another package's `.d.ts` is invisible —
`@modelcontextprotocol/sdk` is the live example, skewed 1.29.0 vs 1.30.0 and
present in web's program from both installs, never written in first-party
code. Closing it properly changes what the guard measures, so it is tracked
as #1965, with both candidate derivations measured there: a lockfile closure
is unusable (155 packages, 25 skewed, nearly all irrelevant tooling, and it
misses the SDK anyway), while `tsc --listFilesOnly` is correct and small
(15 for clients/web).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
From Copilot's fifth pass of #1962. `TRIVIA` covered block comments but not
`//`, despite the contract right above it saying a comment is legal wherever
whitespace is — so `import(// lazy\n"pkg")`, `require(// lazy\n"pkg")`, and
`from // reason\n"pkg"` were stepped over, and such a package would never
enter the candidate set. Same silent-miss direction as the block-comment gap.
Both comment syntaxes are now covered. The line-comment branch runs to
end-of-line only; the newline itself is matched by TRIVIA's `\s` branch.
Regression test added alongside the block-comment one.
The derived set is unchanged at 18 real names — widening trivia does not
admit prose, since a specifier must still follow immediately and
`packageNameOf` rejects anything not shaped like one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Both suppressed comments from Copilot's fourth pass of #1962.
TypeScript allows a comment anywhere whitespace is legal, so
`import(/* webpackIgnore: true */ "pkg")` and `from /* why */ "pkg"` are
valid imports the patterns stepped over. That is the dangerous direction: a
missed specifier never enters the candidate set, so the package's skew passes
the guard silently, against the over-approximation the module claims. A
`TRIVIA` fragment now stands in for `\s*` at every inter-token position, with
regression cases for the commented static, dynamic, side-effect, and
`require` forms.
I did not move to the TypeScript parser. This is a root `scripts/` `.mjs`
tool with no bundler and no TS dependency; pulling in a compiler API to read
import specifiers is a lot of machinery for a scan that only needs to
over-approximate, and each missed form is a one-line pattern with a test
beside it. The derived set is unchanged at 18 real names, so nothing here
widens what is checked today.
The README's command row also still described the candidate scan as `core/`
and `test-servers/src` only, omitting the `vitest.shared.mts` the guard now
includes; it also never mentioned that an allowlisted package is tolerated
only within a major. Both stated.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
All four suppressed comments from Copilot's third pass of #1962. Each was
correct.
The candidate boundary omitted `vitest.shared.mts` — root-owned, imported by
every client's vitest config, and already treated as shared non-client source
by `verify:typecheck-coverage`. It imports only Node built-ins today, which
is exactly why the omission would have gone unnoticed until a third-party
import appeared there, resolved from the root, and skewed. Shared *files* are
now a first-class list beside the shared dirs.
The allowlist was a blanket exemption by name. Each entry's rationale
establishes that a patch/minor difference is benign, which is no evidence
that a React 18-vs-19 or Hono 4-vs-5 split would be — that is a different
type surface. Being listed now tolerates skew only *within a major version*;
a cross-major skew fails like anything else. Deriving the major from the
lockfile version keeps this automatic, with no per-package range to maintain
and bump. A version that can't be parsed is treated as "cannot prove same
major" and fails rather than passes.
`verify-format-coverage.mjs`'s comment overstated the vouching graph: the
three guards do not each check the others. This one checks both siblings;
each sibling checks only this one. Described as it actually works.
The AGENTS.md rule said to bump a shared dependency in "root + all four
clients", but launcher declares no zod and `findSkew` ignores installs where
a package is absent — so the rule as written asked for spurious dependencies.
Reworded to every install that *declares* it, and mirrored in
`.github/copilot-instructions.md`.
The failure output follows: it no longer advises adding a package that is
already allowlisted, saying "MAJOR skew" instead, and it names all three
shared sources rather than the two dirs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
- The v1 recipe published on 127.0.0.1 and the v2 one silently dropped
that. Docker's -p binds host interfaces independently of the
container's HOST, and a no-Origin request skips the allow-list
(core/mcp/remote/node/server.ts:241), so the token is the only guard on
a process-spawning backend. Restore the loopback prefix on both
recipes and say why.
- MCP_PROXY_AUTH_TOKEN is not removed: web-server-config.ts:329 still
falls back to it. Call it a deprecated fallback rather than a rename,
and distinguish it from the query param, which has no fallback
(App.tsx getAuthToken reads only the new name).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFXDQBEjvjCEBhmYkmw79a
Copilot review follow-ups:
- `applyProtocolVersion` took the value straight from an unvalidated JSON
body and went to `RegExp.test`, which coerces — `123` and `true` both
matched the token pattern and reached `Transport.setProtocolVersion` with
the wrong type. It now takes `unknown` and checks `typeof === "string"`
first, with cases for the coercible values.
- Replace `as unknown as Transport` in the new session fixtures with a
`makeTransport()` helper returning a structurally valid `Transport`, so an
interface change stays visible to the tests.
- Drop `as unknown as typeof fetch` in the new transport test —
`vi.fn<typeof fetch>()` is already the right type.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0119LaPQ3v4NyBK3ZTdM9j84
- The "nothing combines with an ad-hoc target" rule has a web exception:
run-web.ts:130 exempts `--transport stdio`, so web ignores it rather
than rejecting. Only cli/tui reject every `--transport`.
- mcp-server-configuration.md still said a leading-dash argument cannot
reach a stdio server under --cli, the opposite of what the migration
guide now says. The parser puts every pre-`--` token in targetArgs, so
the guide is right; corrected the older section to match.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFXDQBEjvjCEBhmYkmw79a
- `--` under --cli: everything before the separator IS forwarded to the
target, flags included (cli.ts:536-541), so "no way to pass a
leading-dash argument" was wrong. Show the correct placement instead
of sending users to a catalog file.
- Stdout: v2's emitResult appends "\n" where v1's awaitableLog wrote the
JSON with no terminator, so the output is not byte-identical.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFXDQBEjvjCEBhmYkmw79a
From the suppressed comment on Copilot's second pass of #1962. Under-
approximating is the dangerous direction here: a package the scan misses
never enters the candidate set, so its skew passes the guard silently.
Three forms went unmatched. `import x = require("pkg")` and a bare
`require("pkg")` are the ordinary import syntax in `.cts` — which the
previous commit just admitted to the scan, so leaving them out would have
been an inconsistency introduced by that widening. A dynamic import carrying
import attributes (`import("pkg", { with: … })`) was missed because the
pattern required the closing paren. A static template literal
(`import(\`pkg\`)`) was missed for want of a backtick in the quote class.
Backticks are accepted only in the *call* forms. A `from` clause requires a
string literal, so allowing them there would buy nothing while reopening the
prose hazard the anchored specifier pattern closes: inline code in comments
is written with backticks throughout this codebase, and admitting it inflated
the derived set from 18 names to 41 with prose fragments (`tools`,
`connect()`, `messages`). Inert today, since a name absent from every
lockfile contributes nothing — but a prose word colliding with a real package
name would silently widen the set. Pinned with a regression test.
The real set is unchanged at 18, so these forms add no package today; they
close the syntax against the one that arrives later.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
The web client's SDK Client owns the `initialize` handshake, so the
`setProtocolVersion()` call it makes after initialize lands on
RemoteClientTransport — a pass-through to the Node backend, not the real
HTTP transport. The backend's StreamableHTTPClientTransport therefore never
learned the negotiated version and stamped no `Mcp-Protocol-Version` on any
request after initialize, starting with `notifications/initialized`. A
stateful server that requires the header rejects that notification with
HTTP 400.
RemoteClientTransport now records the version and rides it on every
`/api/mcp/send`; the backend applies it to its transport before the send, so
that request and everything the transport issues afterwards (the standalone
SSE GET, the session DELETE) carry the header. The version is validated
against a token pattern on the backend so a client can't push arbitrary
bytes into an upstream header.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0119LaPQ3v4NyBK3ZTdM9j84
- Sandbox port: the Apps tab needs a second, dynamic listener, so
"no second port" was wrong for Docker/remote deployments. Qualify the
architecture section, add a pinned-port Docker recipe, and add a
troubleshooting entry.
- Node floor: npm only warns EBADENGINE unless engine-strict is set, so
npx does not refuse to run on an older Node.
- v1 default list: the web UI persisted its list in localStorage; only
the on-disk default catalog is new.
- The v1→v2 --server example needs --cli (and a --method) to select an
entry; note there is no one-entry web equivalent.
- New behavior change #4: an ambiguous URL path (neither /mcp nor /sse)
now errors instead of falling back to SSE.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFXDQBEjvjCEBhmYkmw79a
Both from Copilot's review of #1962.
The candidate-set enumeration matched only `.ts`/`.tsx`, while
`verify:format-coverage` and `verify:typecheck-coverage` both gate all four
TypeScript extensions. No `.mts`/`.cts` exists under `core/` or
`test-servers/src` today, which is exactly why the omission would have gone
unnoticed until a shared dependency arrived through one and skewed — the
failure this guard exists to prevent. Extract the rule as the exported
`isSharedSourceFile` predicate, cover all four extensions, and pin it with
regression tests (including path-boundary anchoring, so a `core-internal/`
sibling isn't swept in).
The README's `validate` row also omitted `test:scripts` and claimed typecheck
was cli/tui only; launcher runs one too, and web typechecks via `tsc -b`
inside its `build`. Restate the row to match the actual script chain.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
The ToolControls regressions build their `Tool[]` by hand, so they would still
pass if the new server option broke — config loading, resolveConfig, the
ordering, the title marker, or its composition with pagination. Copilot flagged
exactly that gap.
Adds an integration test that connects a real InspectorClient to a real
streamable-HTTP test server and asserts the wire shape: repeats appended (not
adjacent — the ordering the defect depends on), only the appended copy marked
"(duplicate)", an unregistered name ignored, an empty list a no-op, and
duplication running before the page slice so a pair straddles a boundary. The
last case loads `duplicate-tool-names-http.json` through loadConfig +
resolveConfig, so the JSON plumbing the manual repro uses is covered too.
Verified non-vacuous: mutating withDuplicates to place repeats adjacent fails
four of the five.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SW1p8E2uiyyLx4RKwrwSrt
Closes#1896
`clients/web`'s `tsc -b` exhausted the 4GB default heap at zod 4.4.x. The
cause is not a zod regression nor a pathological inference site of ours: it
is two copies of zod at *mismatched versions* inside one tsc program.
`clients/web/tsconfig.test.json` compiles `test-servers/src`, whose files
resolve zod (and the SDK's zod) from the **root** install, alongside web's
own sources, which resolve from `clients/web/node_modules`. At 4.3.6 vs
4.4.3 TypeScript must relate two structurally-distinct declarations of every
`@modelcontextprotocol/*` schema type, which is exponential over a recursive
generic surface — `TS2589 Type instantiation is excessively deep`, then OOM.
Bumping root zod to 4.4.3 and changing nothing else returns the build to
baseline: 11.4s and 2,567,882 instantiations, against 11.2s / 2,567,503 at
4.3.6. Two copies at the same version are harmless; skew is what explodes.
So drop the `~4.3.6` hold on `clients/web` and move all four manifests and
lockfiles to zod 4.4.3 together, and add `verify:dep-lockstep` to `validate`
to keep them there. The guard derives its candidate set from what `core/`
and `test-servers/src` import, compares the committed lockfiles' top-level
entries, and fails deny-by-default on any skew not in the annotated
`TOLERATED_SKEW` allowlist. It joins the existing mutual-vouch ring so
dropping it from `validate` is caught by `verify:format-coverage`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Nothing in the repo could serve the shape this fix is about: every preset
registers a unique name, and the SDK's `registerTool` rejects a repeat. So the
only reproduction was a unit test — there was no way to smoke the real app
against a server that returns duplicates.
Adds `ServerConfig.duplicateToolNames` (a list of registered tool names) to the
composable test server, plumbed through the JSON config, and a showcase config
`duplicate-tool-names-http.json`. The named tools are emitted a second time in
`tools/list` with the same `name` and a "(duplicate)" title.
The repeats are appended after the whole list rather than placed beside their
twin. That is both the realistic shape (two tool sources concatenated) and what
makes the defect observable: React matches a leading run of same-key children
first, so a head-adjacent duplicate happens to line up and the bug hides. Only
a separated pair collides in the keyed map and orphans a row — verified both
ways against the prod bundle.
The `tools/list` override is shared with the existing pagination handler (which
already hand-builds the list); with no `maxPageSize` it returns a single
unbounded page. Duplication happens before pagination, so a duplicated pair can
straddle a page boundary the way a real server's would.
Also adds a ToolControls test for this appended ordering, distinct from the
reporter's interleaved fixture, and documents the config in the README showcase
table.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SW1p8E2uiyyLx4RKwrwSrt
`ToolControls` keyed both the main list and the SEP-2243 excluded list by
`tool.name`. A server may legitimately return the same tool name more than
once, and colliding React keys let a filtered-out row survive reconciliation:
searching narrows the array, React matches the surviving rows against the
stale keys, and an unrelated row stays mounted (plus a "two children with the
same key" console warning).
Stamp each row's position in the unfiltered list before filtering and use
`<index>:<name>` as the key, on both lists. The position is captured before
the filter, so the key stays stable as the search narrows. No deduplication,
no change to what `tools/list` returns, and the name/title search behavior is
unchanged.
Adds regression tests for both lists using the reporter's fixture.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SW1p8E2uiyyLx4RKwrwSrt
Closes#1822
Adds docs/v1-to-v2-migration.md — the map for users arriving from v1:
- CLI flag mapping (every v1 flag survives; lists what v2 adds)
- `--config` vs `--catalog` semantics with before/after examples
- Node engine bump (>=22.7.5 → >=22.19.0)
- what no longer ships (the three sub-packages; v2 is one tarball)
- the three behavior changes most likely to bite: exit code 5 on a
failing tool call, target-must-come-first under --cli, and the
reversed `--` separator
- env-var mapping (MCP_PROXY_AUTH_TOKEN → MCP_INSPECTOR_API_TOKEN,
MCP_PROXY_FULL_ADDRESS removed, SERVER_PORT repurposed), the v1 UI
config settings, web query params, and Docker
Linked from the root README (banner + docs list), the CLI README, and
docs/mcp-server-configuration.md; AGENTS.md's docs/ tree entry updated.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SW1p8E2uiyyLx4RKwrwSrt
Review feedback (round 3). The predicate absorbed *any* CONNECTION_CLOSED,
while the doc comment claimed it absorbed only the teardown's — so a
transport dropping on its own after the progress notifications but before
`disconnect()` would have been swallowed. These tests assert only on the
notifications, so that regression would have passed green.
Gate the absorption on a teardown flag owned by the helper, which now
returns an `InFlightCall` exposing `disconnectAndSettle()`. Keeping the flag
inside means it can't be raised too early (which would silently reopen the
hole) and the settle can't be forgotten; call sites also collapse from two
statements to one.
Verified: a pre-teardown ConnectionClosed now fails the test and stays
observed (no run-level error), while a teardown-time one is still absorbed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Review feedback (round 2). Narrowing the handler in the previous commit
fixed the silent-pass but reintroduced #1947's own failure class on the
unexpected path: `then` returns a *derived* promise, and the re-throw
rejects that one rather than `call`. The caller doesn't await it until
after `disconnect()`, so a rejection arriving while the test is still
waiting on progress notifications sits unobserved for seconds and Node
reports it as an unhandled rejection.
Attach an observer to the derived promise the moment it exists. Nothing is
swallowed — it stays rejected, so the caller's `await` still fails the test.
Verified both directions with a probe rejecting with a non-ConnectionClosed
error and awaiting after a 1.5s gap: without the observer the run reports
"Unhandled Rejection" + "Errors 1 error"; with it, the test fails and the
run stays clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Review feedback: the blanket rejection handler also converted a genuine
`callTool` failure into a pass. That matters here because these tests
assert on the progress notifications, which arrive before the result — so
a call that emitted all three progress events and then failed would have
satisfied every assertion and been swallowed.
Narrow the handler to absorb only `SdkError` / `CONNECTION_CLOSED` and
re-throw anything else. The handler is still attached at call time, so an
unexpected rejection stays *handled* and surfaces as an ordinary test
failure at the `await inFlight` after `disconnect()` rather than as another
unhandled rejection. Fulfillment is still accepted: whether the call beats
the teardown is a race, so asserting either outcome would reintroduce the
original bug as a flake.
Verified the narrowed predicate is actually exercised rather than dead
code — both call sites reject with SdkError/CONNECTION_CLOSED on every run.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Two progress tests fire `client.callTool()` without holding the promise so
they can assert on the notifications the call streams while it is still in
flight. `disconnect()` then closes the SDK client, which rejects every
pending request with "Connection closed" — with nothing holding the
promise that lands as an unhandled rejection, which vitest counts as a run
error and fails `npm run ci` at `coverage` even though all tests pass.
Add a `settleInFlight()` helper that attaches the handler at call time (not
after the assertions, which would leave a window the rejection can escape
through) and returns a promise the test awaits after `disconnect()` so
teardown stays ordered. Neither outcome is asserted: the call may
legitimately complete before the teardown or reject with it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
* fix: load @napi-rs/keyring lazily so an unsupported platform degrades
`@napi-rs/keyring` ships one prebuilt binary per platform triple and
throws on import where it has none. On Android / Termux there is no
`@napi-rs/keyring-android-arm64`, so the static top-level import in
`core/auth/node/secret-store.ts` threw during module evaluation and the
Inspector exited at startup with "Cannot find native binding" — before
any of the keychain-unavailable handling in that same file could run.
Replaces the static import with a cached dynamic one. The outcome is
cached, not just the module, so a box without a binary doesn't
re-attempt (and re-throw) resolution on every secret operation —
`expectedSecretFields` means that would otherwise be once per server per
`GET /api/servers` — and so `set` can name the underlying cause in its
`KeychainUnavailableError`.
An unloadable package now folds into the same degradation contract as an
unreachable keychain: `get` returns null, `delete` and
`deleteAllForServer` no-op, and `set` throws `KeychainUnavailableError`
(not double-wrapped — the load failure is already typed by the time the
catch sees it). The class doc comment now enumerates all three ways
"unavailable" can arrive, since each escaped the contract at some point.
The shared `vi.mock` stub can't express this — it models a keyring that
loads — so the new tests build on `vi.resetModules()` + `vi.doMock` with
a throwing factory and re-import the module, covering the load failure,
each method under it, and the load-once caching.
Verified the dynamic import survives bundling in all three clients
(cli, tui, and the web runner) rather than being inlined back to a
static one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XAR2Nr9kXbrywFNUVoTe9F
* fix: steer a missing native binding to a reinstall hint
`KeychainUnavailableError` gave the Linux libsecret / gnome-keyring
advice for every cause, including the one where the keychain is fine and
the @napi-rs/keyring platform package is what's missing — an unsupported
triple (#1905) or npm's optional-deps bug dropping it on a supported one
(npm/cli#4828, the Windows report on #1852). Detect the loader's "Cannot
find native binding" phrasing and point those users at a reinstall /
npx-cache clear instead.
Ported from the superseded #1943, which fixed this alongside the same
lazy-load change.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
* fix: name a bad keyring namespace instead of "is not a constructor"
The lazy import accepted whatever the resolution produced. The named
exports arrive through CJS interop, so a resolution that stopped
yielding them — a default-only export upstream, a bundler changing
interop, a platform where named-export detection fails — would hand us
`undefined`, and `new mod.AsyncEntry(...)` would throw a TypeError
inside the very try that implements graceful degradation.
Shape-check the resolved module and treat a bad shape as unavailable.
Scope, because it is narrower than it first appears: this does NOT
prevent a bad shape from emptying the secret list. `get` returns null
either way — that is its read-tolerance contract, and the TypeError
lands in the same catch a dead keychain does. What changes is the
diagnosis. Without the check the only signal is `set` reporting
"keyring.mod.AsyncEntry is not a constructor", which reads like an
Inspector bug; with it, `set` names the real problem once, at the load
boundary, in the same actionable 503 as every other unavailability.
Catching the silent-empty-list case itself would need a round-trip
against the unmocked package, which nothing in the suite does today.
The member access sits inside the try because reading a missing export
is not always a harmless undefined — a Proxy-backed namespace can throw,
as vitest's module mocks do — and letting that escape would reject the
cached promise, which is otherwise guaranteed to always resolve.
Of the four new tests only the cause-message one fails without the
guard; the rest pin the surrounding contract.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
* docs: correct the test header's count of unavailability entry points
The shape check added a fourth way "unavailable" can arrive, and the
enumeration in secret-store.ts was updated to match, but the test file's
header still said three — so it read as an exhaustive list that silently
omitted the case a whole describe below it covers.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
* fix: give the wrong-shape cause its own remediation hint
The shape check added a cause that fell through to the libsecret
advice — so a packaging mismatch told the user to install a keyring
daemon, which cannot help, and on Windows or macOS points somewhere that
does not exist. That is the same wrong-advice bug the native-binding
branch was added to fix, reintroduced by the guard itself.
Hint selection moves into `hintFor`, where libsecret is the explicit
fallback rather than the default, and the shape failure becomes a
distinct `KeyringModuleShapeError` matched by type — it is our own error,
so there is no reason to re-parse text we wrote (the native-binding
branch matches a string only because that message comes from napi-rs).
Also retitles the wrong-shape test: it claimed the guard avoids
"returning null secrets", but `get` returns null either way by the
read-tolerance contract. The title now says what actually differs —
`set` hard-fails — matching the correction already made in the comments.
Both raised by Copilot review.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
* fix: give both shape-check failure paths the packaging hint
The shape check can fail two ways — the members are absent, or reading
them throws — and only the first carried KeyringModuleShapeError. The
catch path returned the raw error, so a Proxy-backed namespace throwing
on access fell through to the libsecret advice: the same wrong-remedy
bug as the previous two commits, one branch further along.
Both paths now carry the type. KeyringModuleShapeError takes a detail
string plus a standard `cause`, so the throwing path reads "its exports
could not be read: <original>" and keeps the underlying failure legible
while still earning the packaging hint.
Three instances of one mistake in a row (a new cause added without
deciding which remediation it inherits) says the shape was wrong, not
just the branch: making libsecret an explicit fallback in `hintFor` did
not stop me adding a second exit that skipped it. The tests now pin the
invariant directly — one asserts both construction paths produce the
same hint — so the next cause that forgets is a failing test rather than
a reviewer's catch.
Raised by Copilot review.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix: construct AsyncEntry inside the try so keychain degradation engages
`KeyringSecretStore` built its `AsyncEntry` outside the `try` in `get`,
`set`, and `delete`. `AsyncEntry::new` performs the platform-store setup
(on Linux, the Secret Service connect with a keyutils fallback) and
throws when no backend is reachable, so the documented degradation
contract never engaged for a construction-time failure: the raw keyring
error escaped, 500ing every `GET /api/servers` on a box without a Secret
Service (the published container, which has no D-Bus session).
`expectedSecretFields` always includes the OAuth slot, so `rehydrateConfig`
constructs an entry for every server — the default seeded catalog was
enough to 500 the first list load, before any server was added. And
because the escaping error wasn't a `KeychainUnavailableError`, it
bypassed both the routes' 503 translation and the `migratePlaintextSecrets`
skip branch, yielding a generic 500 instead of the actionable message.
Moving construction inside the existing `try` restores the contract for
that failure mode: `get` returns null, `delete` no-ops, and `set` is the
only operation that throws — as `KeychainUnavailableError`, which is what
the 503 translation and the migration skip both match on.
The test stub's constructor could not fail, which is why the coverage
gate never saw the gap. Adds a `constructorThrows` hook alongside the
existing method-level `failures` flags and covers all three methods plus
`deleteAllForServer` under it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XAR2Nr9kXbrywFNUVoTe9F
* test: pin why the OAuth secret field is unconditional
`expectedSecretFields` always includes the OAuth slot, and #1848's report
frames that as why the 500 fired so broadly — which invites a follow-up
that skips the keychain read for servers with "no OAuth config". That
change would lose data.
`extractSecretsFromStored` deletes the `oauth` block outright when
`clientSecret` was its only property, so such a server carries NO marker
on disk that a secret exists — the keychain is the sole record, and the
unconditional slot is what finds it again. Gating the read on a
disk-visible `oauth` block would silently stop rehydrating exactly that
shape.
The existing tests do fail if the slot is made conditional (verified by
applying the change: 4 red), but only under names that explain nothing
about the consequence — `always lists the OAuth slot first` reads like a
tautology worth updating rather than a trap. Add a round-trip case that
states the consequence, so the next reader sees the field is load-bearing
rather than defensive.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAt8rqxysNbhYWLhoRm3fU
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two fixes from Copilot's review of #1953.
useManagedListError re-synced React state from the `state` prop inside a
useEffect, which renders one frame carrying the PREVIOUS store's error
after `state` changes (switching servers) before the effect corrects it —
the pattern AGENTS.md forbids. useSyncExternalStore reads the snapshot
during render, so a store swap lands in the same frame, and it also
closes the window where an error recorded between render and subscribe
was missed. The snapshot returns the stored Error instance (or null), so
it is referentially stable as the API requires.
The `markResponseRejected` correlation maps were never cleared. Ids of
requests that never got a response — a timeout, a dropped connection —
are the only entries trackResponse can't remove, so they accumulated
across reconnects. Cleared on the start-clean connect path, per the
convention documented on resetSessionState (one route out, `onerror`
with no `onclose`, tears down nothing).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SW1p8E2uiyyLx4RKwrwSrt
markResponseRejected was called for ANY refresh failure. The id it
recovers is "the last response received for this method", which is the
failing exchange only when a response actually arrived and was refused
while decoding. A transport drop, a timeout, or an abort produces no
response frame at all — the last-answered id then still points at an
EARLIER successful call, so marking it stamped "Rejected by the
Inspector" onto an exchange that worked. That is the same class of lie
this issue exists to remove.
A server-sent JSON-RPC error is excluded for a different reason: the id
would be right, but the failure is the server's and its entry already
renders as an error from the error frame.
Gated on SdkErrorCode.InvalidResult / UnsupportedResultType — decisions
the client made about a frame in hand. The failure is still recorded as
list state either way; only the Protocol attribution is gated.
Caught in Copilot's review of #1953.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SW1p8E2uiyyLx4RKwrwSrt
extractStatus returned none for any non-request entry, so a clientError
annotation that landed on a standalone response frame — the fallback
messageLogState uses when no request entry was there to fold into, e.g. a
trimmed log or a reconnect boundary — rendered with no status badge at
all. The clientError check now runs before the request-only lifecycle.
Caught in Copilot's review of #1953.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SW1p8E2uiyyLx4RKwrwSrt
A list load that failed — a transport error, or a result the SDK's era
codec rejects as invalid — reported nothing. The connect-time refresh is
fired with no caller to await it (`void this.refresh()`), so its rejection
became an unhandled rejection: the header said "Connected", the panel
rendered an empty list indistinguishable from a server that has none, and
the failed exchange showed in the Protocol panel as a clean success.
Two surfaces now tell the truth:
- ManagedListState records the failure as observable state (`getError` +
an `errorChange` event) and still re-throws, so App's auth-recovery
wrapper keeps working. The four `useManaged*` hooks expose it via a
shared `useManagedListError`, and the Tools/Prompts/Resources sidebars
render a `ListLoadError` alert with the reason and a Retry.
- A response the client refused is attributed back to its Protocol entry
(`markResponseRejected` → `responseRejected` → `MessageEntry.clientError`),
which now renders an Error status and a "Rejected by the Inspector"
alert rather than the green resultType badge alone. The SDK gives no
request id with a decode failure, so the id is recovered by correlation;
see the comment on `markResponseRejected` for why that is exact.
Also stops `listAllTools` from dropping a failing excluded-tools walk on
the floor — it stays non-fatal but is logged.
Closes#1953
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SW1p8E2uiyyLx4RKwrwSrt
Firefox does not hand a streaming `fetch()` response to JS until the
first *body* byte arrives; Chromium resolves the promise as soon as the
headers do. Both SSE endpoints flushed headers immediately and then
stayed silent until there was something to report, which deadlocked
`/api/mcp/events`: `RemoteClientTransport.openEventStream()` awaits that
fetch *before* the MCP client sends `initialize`, so no `initialize` ->
no event to report -> no body byte -> the fetch never resolves. The web
UI hung on "Connecting..." forever with nothing in the UI, the console,
or the node logs.
Write an inert `:` SSE comment the instant each stream opens. Conforming
parsers ignore a comment line, so this unblocks the read without
inventing a wire event and needs no client-side protocol change.
`X-Content-Type-Options: nosniff` was also proposed on the issue but does
not fix it -- verified against Firefox 153, the fetch still never
resolves with the header set and no body byte. The delay is not MIME
sniffing.
The priming write is deliberately ordered *after* each handler registers
its consumer (the session event consumer; the file-watch subscriber plus
`ensureWatcher()`). Callers treat the arrival of the stream's first bytes
as proof they are subscribed, so priming first would hand out that proof
across an `await` and drop an edit made in the gap.
`/api/servers/events` gets the same treatment: its fetch was equally
stuck on Firefox until the first real change event, which for an unedited
`mcp.json` never comes. Its client-side reader counted any `\n\n` as a
change, so the inert frame would have fired a spurious background
re-fetch on every connection -- it now skips frames carrying no
`event:`/`data:` field. The transport's own `parseSSE` already drops
comment frames correctly and needed no change.
Verified end to end in real Firefox against a live streamable-HTTP test
server: before, stuck on "Connecting..." past 25s; after, initialize ->
notifications/initialized -> tools/list, connected.
Closes#1858
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAt8rqxysNbhYWLhoRm3fU
Copilot's third pass (suppressed comment) caught the remaining hole in the
child-process handling: `spawn()` reports a failure to start via an `error`
event, not `exit`. With no `error` listener Node throws it uncaught, replacing
the smoke's diagnostic with a raw stack — and because `exit` never fires in that
case, the readiness poll would otherwise spin for the full 30s before reporting
a timeout that misattributes the cause. `close` is now listened to alongside
`exit` for the same reason: it fires in cases `exit` does not.
prod-web-server.mjs already documents this exact hazard for the launcher child;
this brings the test-server child in line.
Both new branches verified by hand:
- unspawnable executable → "could not spawn the MCP test server (…): spawn
/nonexistent/node-binary ENOENT", immediately, instead of an uncaught throw
- missing module → "MCP test server exited early" with the loader stack, fast,
with no 30s hang
Full `npm run ci` green.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XAR2Nr9kXbrywFNUVoTe9F
Copilot's second pass (suppressed comments) caught a real gap: the smoke
listened only for `pageerror`, which is the *synchronous* half of the
uncaught-crash class. Its async twin — an unhandled rejection, or a failed
dynamic import (this app lazy-loads chunks) — is not a `pageerror`; Chromium
reports it on the console channel. So an async crash during the app-open path
could leave the smoke green.
smoke-web-browser.mjs already solved this; this now mirrors it exactly — same
FATAL_CONSOLE pattern, same split between hard failures and diagnostics, so
benign noise (a font-CDN miss, a React warning) still can't flake CI. Fatal
console errors are now included both in the catch-block diagnostics and in the
final assertion, which is where they were missing.
Confirmed working on a passing run: a benign 409 subresource error is reported
as a non-fatal diagnostic rather than failing the smoke.
Full `npm run ci` green.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XAR2Nr9kXbrywFNUVoTe9F
Copilot review on #1934, both comments actionable:
- smoke:web:app defaulted to 6299 — the same port smoke:web uses — while its
own comment claimed the port was "distinct … so back-to-back runs can't
collide". The three web smokes run sequentially in `npm run smoke`, so it
passed, but a slow teardown, a TIME_WAIT socket, or a parallel run would
EADDRINUSE it. Moved to 6297 (smoke:web 6299, smoke:web:browser 6298) and
rewrote the comment to name the actual values rather than assert distinctness.
- The showcase table tells readers to connect with Protocol Era = Modern unless
noted, and mcp-app-http.json is legacy-only. That was stated in the prose
below but not in the row, so a reader scanning the table would try Modern
first. Marked the row.
Verified: smoke:web:app passes standalone on 6297, and the full `npm run smoke`
chain passes with all three web smokes on distinct ports.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XAR2Nr9kXbrywFNUVoTe9F
`smoke:web:browser` stops at first paint and never connects to a server, so
everything downstream of the connect — the Apps tab, the sandbox controller,
the UI-protocol bridge — was unexercised by any smoke. That is the code #1859
broke, and nothing would have caught a regression in it.
smoke:web:app boots the same prod --web server, spawns the mcp-app-http.json
composable server, and drives connect → open app → widget ready through one
deep-link navigate. The assertion is the documented data-app-status="ready"
contract, which the renderer only reports once the widget has loaded inside the
sandbox iframe AND completed its bridge handshake — so a single attribute
covers the proxy being served, the UI resource loading, and the handshake.
Two mechanics found by running it, both silent-failure shaped:
- server-composable.ts announces readiness on stderr, not stdout, so watching
stdout alone times out with an empty diagnostic. Both streams are scanned.
- the bound port is not the configured one — createTestServerHttp resolves via
findAvailablePort(), which walks upward when the port is taken. The smoke
parses the announced URL instead of assuming 3130.
Scope: this runs against the repo build tree like every other smoke, so it would
NOT have caught #1859 itself (a packaging failure; the file is always present
in-repo). pack:verify owns that dimension — the two are complements, and the
header comment says so. A cheap structural pre-check does assert the proxy page
sits where sandbox-controller.ts resolves it, so a move/rename fails fast.
Both failure modes verified by hand: with clients/web/static removed, the
pre-check fires; with the pre-check bypassed, the data-app-status assertion
catches it independently.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XAR2Nr9kXbrywFNUVoTe9F
The `mcp_app_demo` tool and `mcp_app_demo_widget` UI resource presets already
existed in the preset registry, but no config wired them into a server — so
there was no runnable way to reach a rendered MCP App, and therefore no way to
exercise the sandbox proxy path that #1859 broke.
`mcp-app-http.json` composes the two over plain streamable-HTTP. With it, the
Apps tab renders a real widget and the packaging bug is reproducible on demand:
without clients/web/static in the tarball the widget area shows "Sandbox not
loaded: ENOENT … /clients/web/static/sandbox_proxy.html" instead.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XAR2Nr9kXbrywFNUVoTe9F
`clients/web/static/sandbox_proxy.html` is a committed source file, not a
build artifact, and it was never listed in the root package.json "files"
allowlist. `sandbox-controller.ts` reads it at runtime as
`<runner dir>/../static/sandbox_proxy.html`, so it resolved fine in-repo but
was absent from every published tarball — the read threw and the controller
served its "Sandbox not loaded" fallback page, breaking the Apps tab for
anyone running `npx @modelcontextprotocol/inspector`.
v1 shipped the equivalent as `server/static` in both the root and the server
workspace "files" lists; that entry has no counterpart in v2.
Add `clients/web/static` to the allowlist, and assert it in pack:verify twice
over — once in the tarball packlist and once on disk after install, since what
the runtime needs is the path *relative to* clients/web/build rather than mere
presence in the tarball.
No .npmignore change is needed: clients/web/.gitignore does not list `static`,
so the nested-gitignore packlist hazard that hid `build/` does not apply here.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XAR2Nr9kXbrywFNUVoTe9F
The version label says which line work belongs to; nothing said what KIND of
work it is. In practice that meant reaching for bug or enhancement and
pressing everything else into one of them — a migration guide and a
tsup->tsdown swap both landing as "enhancement", which degrades the label to
"not a bug" and makes filtering by it meaningless.
Every issue created or triaged now carries exactly one of bug, enhancement,
documentation, chore, or question, independent of its v1/v2 label. Adds a
table of what each is and is not for, and states the anti-pattern explicitly
so the two familiar labels don't absorb the other three. A PR needs no type
label — it is classified through the issue it closes, the same way it is
tracked through that issue's board card.
Wires the requirement into the three places that would otherwise drift: the
"not created until" callout, triage pass 1 (an outside reporter can set
neither label, so both are applied in the same step), and the board audit,
which gains an "open, no type label" check and a matching row in the
invariant table.
Note `chore` did not exist as a repo label and was created for this; the
existing `dependencies` label is Dependabot's and scoped to PRs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAt8rqxysNbhYWLhoRm3fU
A board card in Done asserts that the work shipped, and Done is read as the
record of what a milestone actually delivered. Nothing said so, so an issue
closed as a duplicate or not-planned was being moved to Done like any other
close — which makes that record wrong in a way no later reader can detect:
counting Done cards can no longer tell a shipped fix from a report closed as
a duplicate of one.
New rule: exactly two things earn a card a place in Done — its PR merged, or
it is a parent whose last sub-issue closed. Every other close (duplicate,
won't fix, not planned, obsolete, superseded) means nothing shipped, so the
card is deleted instead. Deleting a card touches only the board; the issue
keeps its labels and comments and stays searchable, so nothing is lost.
Also notes that the close reason is the machine-readable form of the same
distinction, and that `gh issue close --reason` cannot express `duplicate` —
it accepts only completed/not planned, so duplicate must be set through the
API or the web UI's "Mark as duplicate" (which also records a duplicate-of
link).
Adds the invariant to the board audit as a tenth check. That required the
audit to read closed issues too, so its fetch moves to --state all with
stateReason, `open(n)` becomes a state test rather than "present in the map",
and --limit must now clear the repo's total issue count, not just the open
ones. The two wrong-board label checks are scoped to open issues in the same
pass — over all history they flag four long-closed cards whose labels nobody
intends to rewrite.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAt8rqxysNbhYWLhoRm3fU
Follow-up on the same issue: the triage rules were clear on the happy path
but three gaps meant a bare "triage new issues" would diverge from what a
maintainer actually does.
Define triage by STATE, not authorship. An issue needs triage when it
arrives with no board card and no milestone — whoever filed it. Framing it
as "externally-filed" was wrong: a maintainer opening an issue by hand in
the GitHub UI lands in exactly the same state, because write access makes
the board reachable, not automatic. What puts an issue in Todo is somebody
performing the approval, not who owns the account. Renamed the section and
reworded every dependent passage accordingly.
Fix the discovery query, which diffed open issues against board #28 only.
Verified against the live boards: it reports #1929 as unboarded when that
issue is correctly carded on #11, so following it literally double-boards
the issue — recreating a defect a past sweep introduced. Now diffs against
the union of both boards, filtered by .content.repository (an org project
can hold other repos' issues; #11 carries one from modelcontextprotocol/
servers), and prints each hit's destination, since an unboarded issue that
already has a milestone is approved and belongs in Todo, not Incoming.
Add "The board audit" — the other drift classes a single-issue rule can't
catch (double-boarded, non-Issue items, statusless cards, the Incoming ⇔
milestone invariant in both directions, wrong board for label, missing
version label, missing Priority), as a table of invariant + fix plus one
runnable jq check that should print 0 across the board. Both snippets were
executed verbatim against the live boards.
Add "Recording the score". The rubric claimed its reasoning "survives in a
form someone can argue with later", but the board stores only the result
and is private, so nothing kept that promise. Triage now posts the axes,
the bonuses claimed, and the total as a comment.
Also scope pass 2 explicitly to a human — deciding what ships in which
release is not inferable from a rubric — and correct the stale note about
three statusless cards on #11.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAt8rqxysNbhYWLhoRm3fU
AGENTS.md had no rule linking a milestone to a board Status, and its two
existing rules contradicted each other: every new issue was to be milestoned
at create time, yet new items were to start in Incoming, with Todo reserved
for work a maintainer had approved. If every issue is milestoned at birth,
"has a milestone" carries no approval information and nothing records the
moment of sign-off.
Make assigning a milestone the approval act, giving one invariant:
Incoming <=> no milestone; everything past it <=> milestoned. External
reports and maintainer-filed issues then differ only in whether approval has
already happened -- an outside reporter cannot set a label, milestone, or
board field, so their issue arriving with none of them is normal rather than
a defect to fix on arrival.
- scope the "not created until five things" callout to maintainer-filed issues
- split the Status bullets into the two cases
- scope the milestone mandate to issues you create, stating the external
exception explicitly
- add a "Triaging externally-filed issues" section: the two passes, plus a
gh/jq snippet that diffs open issues against the board to find unboarded ones
- reword both prose descriptions of the Incoming/Todo line
- reword the rubric's "Linked to a milestone" bonus as "already approved" and
note it is the re-scoring case only; under the old rules it was a constant
offset on every issue, shifting every band boundary down by one
- update both gh recipes (#28 and #11) to default to Todo, with an Incoming
variant for triage
No mirror to .github/copilot-instructions.md: board and milestone mechanics
are on that file's deliberately-absent list.
Closes#1930
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAt8rqxysNbhYWLhoRm3fU
