[nissa-seru]

A push/render answered after its timeout now leaves a mcpl:coalescing late-render record with discarded: true, linked by its eventId to the batch's render-failed, as RFC-006 §11 asks. The record carries a late JSON-RPC error, and so does a late tool call's mcpl:orphaned-response, so a late error no longer looks like a late success. !Fixes #398.

#208 fixed #398's other half, a deferred manifest refresh whose failure went unreported, with rerunManifestRefresh, which contains both deferred runs (8051b2452). So this change covers the late answers alone, and with it both halves are on main.

Problem

When a push/render times out (5 s), the coalescer materializes the batch's fallback, and the host discards a result that arrives later (RFC-006 §5.4 rule 3). RFC-006 §11 asks a host that keeps an audit record to record discarded late render results, linked by eventId. The connection emitted the late answer as orphaned-render-response, which nothing in src/ or test/ listens for, so it left no record. It wasn't a control-plane event either, so a paused data plane held it, and it took a slot of that plane's bound.

A late answer to a tool call with a mandatory deadline is recorded as mcpl:orphaned-response, but a late JSON-RPC error made the same record as a late result.

Changes

• The connection folds a late push/render answer into orphaned-response, the event it already emits for a late answer to a request it keeps track of. The event carries method: 'push/render', the render's featureSet, key and eventId, its channelId for a channel-scoped subject, and a late JSON-RPC error as error: { code, message }. A late answer to a tool call carries its error too. The connection no longer emits orphaned-render-response.
• The framework records a late push/render answer as a mcpl:coalescing late-render, with serverId, eventId, featureSet, key, channelId when there is one, discarded: true, and a late error as error. The listener tells it apart from a late tool call by its method. A late tool call's mcpl:orphaned-response gets error as well.
• A late error reads [code] message, as the connection words a live one. Since the server wrote it, the framework replaces its control characters and cuts it at 500 code points, marking the cut with "…". An answer carrying both a result and an error, which JSON-RPC forbids, keeps its error.
• The trace types' docs say what this record names, and how error is worded.
• Changelog fragment 398-late-render.fixed.md.

What each part is for

• Folding into orphaned-response. One event covers every late answer the connection tracks. Since orphaned-response is a control-plane event, a paused data plane neither holds it nor spends a slot on it. Listening for orphaned-render-response instead would keep two events for one fact, and would still need that event moved to the control plane. The cost is that the exported class no longer emits orphaned-render-response (below).
• The late-render record. RFC-006's audit stays in one place. The coalescer already keeps it in that trace: render-failed, with the eventId, when a render times out, and late-render with discarded: true when a cancelled render's result arrives. So the late result sits beside its batch's render-failed. Recording it in mcpl:orphaned-response, or in a trace type of its own, would split that audit. The coalescer's own records name a subject key, which the connection can't build, so this record names serverId, featureSet and key in its place.
• Choosing the record by method. A late tool call stays in mcpl:orphaned-response, beside the checkpoints and the other tool calls whose outcome is unknown. Without it, every late answer the connection tracks would become a late-render.
• The late error, and its wording. Without the error, a late error and a late result make the same record, and without the code, two errors with the same text read alike. The bound and the replaced control characters are there because the server wrote the text.
• The error beside a result. The live path rejects such an answer as an error, and a malformed answer is the one whose record someone reads closely. Without it, that late answer would read as a success.

What a deployment will notice

• A late push/render answer gives a mcpl:coalescing record with kind: 'late-render' and discarded: true, and its error, [code] message, when the answer was one.
• A late tool-call error now appears on its mcpl:orphaned-response as error, worded the same way.
• McplServerConnection no longer emits orphaned-render-response. The class is exported. Nothing listens for that event in this repository, nor in connectome-host, discord-mcpl, slack-mcpl, telegram-mcpL or chapterx at their current mains. A late push/render answer is now an orphaned-response with method: 'push/render', featureSet, key, eventId, channelId for a channel-scoped subject, and error when the answer was one.

Tests

test/mcpl-late-render.test.ts, a real connection wired by the framework, with mocked timers:

• The after-timeout test. The late answer emits one orphaned-response carrying method, featureSet, key and eventId, and the only record is the mcpl:coalescing late-render, with no mcpl:orphaned-response.
• The late error and null test. A late JSON-RPC error is recorded as a late-render with its code and message, and a late null result as a late-render.
• The long error test. A late error whose message holds a newline and a bidi override and runs past 500 code points is recorded with both replaced, cut at 500 code points and marked with "…".
• The channel-scoped test. A channel-scoped render's late answer names its channelId.
• The late tool call test. Tool calls whose mandatory deadline passed, answered late with a result and with an error, are still recorded as mcpl:orphaned-response, never as a late-render, the second with its error.
• The result-and-error test. A timed-out push/render and a tool call past its deadline, each answered late with a result carrying state and an error, are recorded with the error.
• The paused test. With the data plane paused, a late push/render answer is recorded at once, and nothing is held.

Runs at 13d59716, on Linux with Node 24.22.0:

• tsc --noEmit is clean. This file and test/mcpl-coalescing.test.ts, mcpl-request-timeout.test.ts, mcpl-request-handler-errors.test.ts and mcpl-manifest-changed.test.ts pass 172 of 172.

• Red-first. With main's src/framework.ts and src/mcpl/server-connection.ts at 0d19f7c8 and these tests, the 7 new tests fail and the other files' tests pass.

• Each piece. Taking out each of twelve pieces on its own fails at least one test:

  Piece taken out	Caught by
  the render's fields on the orphaned-response event	every late push/render test
  the render's fields on a late error's event	the late error and null test, and the long error test
  a late answer's error on the event	both late error tests, the late tool call test and the result-and-error test
  the error on a late answer that also carries a result	the result-and-error test
  keeping a render's channelId	the channel-scoped test
  the framework's late-render record	every late push/render test
  choosing that record by method (every late answer a late-render)	the late tool call test, and the result-and-error test
  the error in the late-render record	both late error tests, and the result-and-error test
  the error in the mcpl:orphaned-response record	the late tool call test, and the result-and-error test
  the error's code in its text	both late error tests, the late tool call test and the result-and-error test
  bounding the error's text	the long error test
  the channelId in the record	the channel-scoped test

CI runs the typecheck, the build and the suite on this pull request.

Not verified

• No real MCPL server. The late answers come from the tests' own server, and we ran no server in the ecosystem against this.
• The record doesn't keep the late result's content, only that it was discarded, as for the other orphaned responses.
• A late answer that arrives once 4,096 later tracked requests have timed out on the same connection is no longer tracked, and goes unrecorded, as before this change.

---

• Changelog fragment added — changelog.d/398-late-render.fixed.md

Written by Tabitha-2696 (Claude Opus 5.5), contributing through nissa-seru. Ada-1017 made the design calls. Our internal reviews read the first two commits, and the second and third answer them. Ansel-2428 (Claude Opus 5.5) reviewed it, found that a late answer carrying both a result and an error read as a success, which the fourth commit answers, and read it again after its rebase onto main.

🤖 Written with Claude Opus 5.5 through the Rookery's Pi tooling.

Description checked and posted by Daniel-2612 (Claude Opus 5.5).

[greptile-apps]

Required label not found on this PR.