sorokin.oleg.edu/learnharness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 6 Wiki Activity

fix(bridge): gate tool-input-available on clean execution — invalid tool inputs never reach the dock

Browse Source 
Live crash (modal-realism session): GLM double-encoded the quiz options
array as a JSON STRING inside the tool-call arguments. The bridge emitted
tool-input-available at toolcall_end — BEFORE validation runs — so the
client briefly saw the malformed quiz as a pending question and the dock
crashed in normalizeOptions ((string).map). The later tool-input-error
chunk could not revive the tree.

The old SDK gated this with InvalidToolInputError: validation ran before
the input-available chunk. The bridge now holds the part in
input-streaming until tool_execution_end and only then emits
tool-input-available (clean) or tool-input-error (invalid, structural
JSON Schema failure or the zod re-run inside execute).

Server-side behavior was already correct (the record shows the error
part + a successful model retry) — this is purely the wire-ordering hole.

Pinned in chat-flow.test.ts: the refinement-layer retry (correctAnswer
mismatch) and the exact live structural shape (string options) both
assert the invalid call never becomes input-available on the wire.

Gates: tsc, lint, build, 306/306.
This commit is contained in:
amosblomqvist
2026-09-12 03:59:43 +03:00
parent 2d717fd932
commit 6af7a38684
2 changed files with 89 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/server/pi/ui-stream.ts
+24 -18
View File
@@ -17,9 +17,12 @@ import { QUESTION_TOOLS } from "../../core/pending.ts";
* NEVER tool-output-available (the client throws on an output for a call it
* owns). The synthetic "question posed" result lives in memory only; the
* record's part stays output-less until the learner answers.
* - Errored calls (authoring validation) surface as tool-input-error — the
* exact part shape the SDK produced for InvalidToolInputError (the
* output-error state the chat flow asserts).
* - tool-input-available is emitted ONLY after the call cleared validation
* at the boundary (clean tool_execution_end) — the SDK's
* InvalidToolInputError gating. An invalid call surfaces as
* tool-input-error instead and NEVER becomes input-available, so a
* malformed question can never mount the dock (the live modal-realism
* crash: GLM double-encoded `options` as a JSON string).
* - The persisted message is assembled from the SAME parts the client
* assembles (the processUIMessageStream shapes), so every downstream
* consumer (phases, plan blocks, pending, extraction) is unchanged.
@@ -239,17 +242,12 @@ export function createPiTurnResponse(spec: PiTurnSpec): Response {
case "toolcall_end": {
const toolCall = streamEvent.toolCall;
toolInputs.set(toolCall.id, { toolName: toolCall.name, input: toolCall.arguments });
// Input stays input-streaming here: it only becomes available
// once the call clears validation at the execution boundary
// (tool_execution_end) — an invalid input must never reach
// input-available, where the dock would mount on it.
const part = newToolParts.get(toolCall.id);
if (part) {
part.state = "input-available";
part.input = toolCall.arguments;
}
send({
type: "tool-input-available",
toolCallId: toolCall.id,
toolName: toolCall.name,
input: toolCall.arguments,
});
if (part) part.input = toolCall.arguments;
break;
}
}
@@ -274,12 +272,20 @@ export function createPiTurnResponse(spec: PiTurnSpec): Response {
input,
errorText,
});
} else if ((QUESTION_TOOLS as readonly string[]).includes(toolName)) {
// Client-owned: the answer arrives via addToolOutput;
// the synthetic placeholder result never reaches wire
// or record.
break;
} else {
// Validation cleared: only NOW does the input become available
// to the client (the InvalidToolInputError gating the SDK owned).
if (part) {
part.state = "input-available";
part.input = input;
}
send({ type: "tool-input-available", toolCallId: event.toolCallId, toolName, input });
if ((QUESTION_TOOLS as readonly string[]).includes(toolName)) {
// Client-owned: the answer arrives via addToolOutput;
// the synthetic placeholder result never reaches wire
// or record.
break;
}
const output = (event.result as { details?: unknown } | undefined)?.details;
if (part) {
part.state = "output-available";