[nissa-seru]

Three small fixes to the MCPL connection, one commit each with its tests and changelog fragment, so each can be taken or left on its own. #449 describes the defects, with the runs on main that show them. The new tests fail on main at 3adbdb8 and pass at this head. !Fixes #449.

Changes

1. mcpl/manifestChanged without params (18ff80e). handleManifestChanged read the announcement's domains for its receipt only after it had applied the fetched manifest. A notification sent without params reaches it as undefined, so it threw once the grant had changed, and emitted no mcpl:manifest-change-receipt and refreshed no tools list. It now reads the announcement first, through announcedDomainsOf, which gives the strings of its params' domains array, or none when the params are missing or malformed. The receipt's announcedDomains is string[] in its trace type, and now always is one.
2. A WebSocket server's close reason in errors (8605165). Two errors put a WebSocket peer's close reason into their text as the server wrote it: the connect error for a transport that closes during the handshake, and the error each pending request is rejected with when the connection drops. mcpl:server-closed already leaves that text out (reasonFromPeer). Both errors now give the transport's own words for such a close, websocket closed. They keep the close code, which says what the server meant whether or not it sent a reason (4001 marks a refused credential). A request's error already had code=. The handshake's error now ends (websocket closed, code 4001) for any WebSocket close, since, by our reading, a close before the handshake completes reaches no 'close' event, and on main it gave no code. The connect error reaches the host's log, the mcpl:server-connect-failed trace and the mcpl-down alert, and, by our reading, a request's error reaches an MCPL tool call's result. So a reason with a line break could pass for a host log line, or mislead the model. A stdio child's exit reads as before.
3. The buffered-event cap's drop count (c15fd9d). While a plane is paused, a server's responder-less events are buffered up to MAX_BUFFERED_EVENTS, and past it the oldest is dropped. droppedBufferedEvents counted those drops over the connection's whole life. So a later pause's flush line reported earlier pauses' drops as its own, and the drop log, which speaks at the first drop and at every hundredth after it, counted across pauses too. ready(), which ends a pause, now reads the count for its line and resets it.

What each part is for

Each fix was weighed against a simpler or a wider one.

• announcedDomainsOf. Reading the announcement before anything is applied ends the throw after the grant has changed, and taking only the strings of an array keeps the receipt's announcedDomains the string[] its type says. Refusing an announcement without params instead would leave a server's real manifest change unapplied for want of a hint the handler doesn't need, since the handler reads the change itself from the fetched manifest.
• closeReasonText. One helper serves both errors, so the rule mcpl:server-closed already follows, that a peer's close reason stays out, holds wherever a close reason becomes an error. The close code stays in each, since it's a number the server can't use to write a line, and it tells a refused credential from another close. The handshake tells a WebSocket close from a stdio exit by the transport's existing kind, so the close info needs no new field. Escaping the peer's text instead, as host log lines do with logDetail, would keep server-written text in errors that reach the model and the alert.
• The reset in ready(). ready() ends a pause and reads its count for the flush line, so the read and the reset are one step. Resetting in pauseDataPlane() instead would zero a pause's count whenever the data plane is paused again before it opens, and four call sites pause it: two reconnect paths (reconnectAfterFailure, reconnectOnce), a quiesce (runQuiesce) and the data-plane gate (installMcplDataPlaneGate).

Overlaps

#208 rewrites the manifest-changed listener as onRequest('manifest-changed', ...), so the two conflict once, at that listener in src/framework.ts. Merged together, the listener keeps #208's onRequest form with params: unknown. src/mcpl/server-connection.ts merges without a conflict.

Tests

• mcpl-manifest-changed. An announcement without params gets one receipt naming no domains, with the fetched manifest applied and the tools list refreshed. Params that are null or a string, or domains that isn't an array, name none, and an array reports only its strings.
• mcpl-ws-transport. A server that closes with code 4001 during the handshake fails the connect with (websocket closed, code 4001), both with a two-line reason and with none. One that closes with that reason while a tools/call waits rejects it with reason=websocket closed. Neither error holds any of the server's text. Two pauses past the cap report their own drops, 150 and 75, the second through a control-only flush and a second pauseDataPlane(), and the drop log speaks at each pause's first drop.
• Runs. On main at 3adbdb8 with these tests added, the six new tests fail on Linux with Node 20.20.2 and 24.22.0, each on its defect. At this head typecheck and build are clean, and our runner runs the suite eight files at a time with --test-force-exit. On Node 24.22.0 it reports 2433 of 2434 tests passing in 152 files. The one failure is tool-result-guard's "fix(guard): a withheld tool result says where its original is (#277) #281: batches behind one stalled write cost one deadline: no write starts until it settles", in a file this PR doesn't change, and history-module-overview stopped early at 22 of 28. On Node 20.20.2 it reports 2373 of 2373 in the 151 files that finished, and mcpl-coalescing didn't finish within 1200 seconds. Run one at a time without that flag, history-module-overview passes 28 of 28 and mcpl-coalescing 67 of 67 on both runtimes, and tool-result-guard passes 124 of 124 on Node 24 and 122 of 124 on Node 20, where two of its fix(guard): a withheld tool result says where its original is (#277) #281 tests fail. CI's unforced step passes 105 of 105 on both.

---

Implemented by Kestrel-2409 and independently reviewed by Esther-2584, both using Claude Opus 5.5. Ada-1017 filed the three defects from an internal review of #208 and made the calls on their scope.

[greptile-apps]

Required label not found on this PR.