@juspay/neurolink 12.12.16 → 12.13.0

package/dist/providers/anthropic/loopAdapter.js

@@ -24,6 +24,9 @@
 import { resolveDeferredTool } from "../../tools/toolDiscovery.js";
 import { stringifyAnthropicToolOutput } from "./toolOutput.js";
 import { stringifyFinalResultInput } from "./structuredOutput.js";
+import { composeAbortSignalsScoped, createTimeoutController, } from "../../utils/timeout.js";
+import { NeuroLinkError } from "../../utils/errorHandling.js";
+import { ErrorCategory, ErrorSeverity } from "../../constants/enums.js";
 /** Map Anthropic's stop_reason onto the unified finish reason. */
 function mapAnthropicFinishReason(rawStopReason, hadToolCallsAtCap) {
     switch (rawStopReason) {
@@ -37,6 +40,298 @@ function mapAnthropicFinishReason(rawStopReason, hadToolCallsAtCap) {
             return hadToolCallsAtCap ? "tool-calls" : "stop";
     }
 }
+/**
+ * Read one streamed Anthropic step off the wire.
+ *
+ * Module-level rather than nested inside `executeStep` because it IS the step:
+ * folding an SSE event sequence into content blocks, usage and a stop reason.
+ * `executeStep` above it is now only the per-request deadline wiring, which is
+ * the part a reader looking for "what bounds this request" needs to find.
+ */
+async function readAnthropicStep({ params, client, channel, stepSignal, requestTimeoutSignal, noteObservedPromptTokens, finalResultToolName, onTerminalResult, requireTerminalEvent, }) {
+    // Single assertion, not a double: `messages.create` returns a union of
+    // Message and Stream, and Stream<RawMessageStreamEvent> already IS an
+    // AsyncIterable of that event, so the two types overlap and the
+    // compiler still checks the narrowing.
+    const events = (await client.messages.create({ ...params, stream: true }, 
+    // The engine's signal composed with this step's own deadline.
+    // runAgenticLoop already derives the engine signal from the caller's
+    // abortSignal, so preferring config's would ignore engine-initiated
+    // cancellation entirely.
+    { signal: stepSignal }));
+    const textByIndex = new Map();
+    const toolByIndex = new Map();
+    const thinkingByIndex = new Map();
+    const redactedByIndex = new Map();
+    let text = "";
+    let reasoning = "";
+    let rawStopReason;
+    let inputTokens = 0;
+    let outputTokens = 0;
+    let cacheReadTokens = 0;
+    let cacheWriteTokens = 0;
+    // Reported alongside the total, not derivable from it: the two TTL
+    // tiers are priced differently, so a caller that reports them (the
+    // Claude-on-Vertex turn span does) cannot reconstruct the split from
+    // cacheWriteTokens alone.
+    let cacheWrite5mTokens = 0;
+    let cacheWrite1hTokens = 0;
+    let stepOutputTokens = 0;
+    // Anthropic closes every complete message with `message_stop`. Its
+    // absence is the only signal that the response ended early, because the
+    // content blocks that DID arrive are syntactically complete.
+    let sawTerminalEvent = false;
+    for await (const rawEvent of events) {
+        if (stepSignal.aborted) {
+            break;
+        }
+        // Narrowed through the SDK's own discriminated union rather than
+        // re-declared: `RawMessageStreamEvent` already describes every event
+        // shape, so `event.type` checks below are compiler-checked instead of
+        // asserted.
+        const event = rawEvent;
+        if (event.type === "message_start") {
+            const usage = event.message?.usage;
+            inputTokens += usage?.input_tokens ?? 0;
+            const startOutput = usage?.output_tokens ?? 0;
+            outputTokens += startOutput - stepOutputTokens;
+            stepOutputTokens = startOutput;
+            // Anthropic reports cache reads/writes separately from input_tokens
+            // on this same event; without these the stream drops all cache
+            // accounting.
+            cacheReadTokens += usage?.cache_read_input_tokens ?? 0;
+            cacheWriteTokens += usage?.cache_creation_input_tokens ?? 0;
+            // BEST EFFORT, and the limit is worth stating. The nested TTL
+            // breakdown exists only on `Usage` (this event); `MessageDeltaUsage`
+            // carries the cache TOTALS but not the split, so message_start is
+            // the only place in the raw event stream it can come from. The
+            // pre-migration loop read it off `stream.finalMessage()` — the SDK's
+            // ACCUMULATED message — so if the API leaves `cache_creation` null
+            // here and fills it only on the assembled message, these two stay
+            // zero and the totals above remain correct regardless.
+            // Reported as undefined rather than a false zero when absent.
+            cacheWrite5mTokens +=
+                usage?.cache_creation?.ephemeral_5m_input_tokens ?? 0;
+            cacheWrite1hTokens +=
+                usage?.cache_creation?.ephemeral_1h_input_tokens ?? 0;
+            // The guard calibrates from the FULL prompt size, not input_tokens
+            // alone: on a cache hit the uncached remainder is tiny and using it
+            // would let the guard drift far under the real cost.
+            noteObservedPromptTokens?.((usage?.input_tokens ?? 0) +
+                (usage?.cache_read_input_tokens ?? 0) +
+                (usage?.cache_creation_input_tokens ?? 0));
+            continue;
+        }
+        if (event.type === "content_block_start") {
+            const index = event.index ?? 0;
+            // A redacted_thinking block carries its whole payload here and
+            // produces no deltas, so if it is not captured on this event it is
+            // never seen again. Anthropic validates the thinking chain when
+            // extended thinking continues across a tool-use turn, so a missing
+            // one fails the NEXT request — and only for accounts where safety
+            // redaction actually triggers, which is why it survives testing.
+            if (event.content_block?.type === "redacted_thinking") {
+                const data = event.content_block.data;
+                if (typeof data === "string") {
+                    redactedByIndex.set(index, data);
+                }
+            }
+            if (event.content_block?.type === "tool_use") {
+                toolByIndex.set(index, {
+                    id: event.content_block.id ?? "",
+                    name: event.content_block.name ?? "",
+                    inputJson: "",
+                });
+            }
+            continue;
+        }
+        if (event.type === "content_block_delta") {
+            const index = event.index ?? 0;
+            const delta = event.delta;
+            if (delta?.type === "text_delta" && delta.text) {
+                text += delta.text;
+                textByIndex.set(index, (textByIndex.get(index) ?? "") + delta.text);
+                channel.push({ content: delta.text });
+            }
+            else if (delta?.type === "thinking_delta" && delta.thinking) {
+                const acc = thinkingByIndex.get(index) ?? {
+                    text: "",
+                    signature: "",
+                };
+                acc.text += delta.thinking;
+                thinkingByIndex.set(index, acc);
+                reasoning += delta.thinking;
+                // Reasoning rides its own field; `content` stays a present string
+                // so plain-text consumers are unaffected.
+                channel.push({ content: "", reasoning: delta.thinking });
+            }
+            else if (delta?.type === "signature_delta" && delta.signature) {
+                const acc = thinkingByIndex.get(index) ?? {
+                    text: "",
+                    signature: "",
+                };
+                acc.signature += delta.signature;
+                thinkingByIndex.set(index, acc);
+            }
+            else if (delta?.type === "input_json_delta" && delta.partial_json) {
+                const pending = toolByIndex.get(index);
+                if (pending) {
+                    pending.inputJson += delta.partial_json;
+                }
+            }
+            continue;
+        }
+        if (event.type === "message_delta") {
+            rawStopReason = event.delta?.stop_reason ?? rawStopReason;
+            const cumulative = event.usage?.output_tokens ?? stepOutputTokens;
+            outputTokens += cumulative - stepOutputTokens;
+            stepOutputTokens = cumulative;
+            continue;
+        }
+        if (event.type === "message_stop") {
+            sawTerminalEvent = true;
+        }
+    }
+    // Ordered ahead of the terminal-event check because a deadline that
+    // fired IS why the terminal event is missing, and the timer's identity
+    // is the more useful of the two answers. The SDK's stream iterator exits
+    // without throwing on an aborted read, so nothing else would report it:
+    // the step would return the partial content as though the model had
+    // simply said less.
+    if (requestTimeoutSignal?.aborted) {
+        throw requestTimeoutSignal.reason;
+    }
+    // A response that ended before its terminal event is not a turn. Its
+    // tool_use blocks parse perfectly, so returning them here would dispatch
+    // tools the model never finished asking for, and the turn would go on to
+    // report itself as a normal stop. Not retriable: the request was already
+    // answered, and re-sending it would double any side effect the truncated
+    // half already caused upstream.
+    if (requireTerminalEvent && !sawTerminalEvent && !stepSignal.aborted) {
+        throw new NeuroLinkError({
+            code: "ANTHROPIC_STREAM_TRUNCATED",
+            message: "Anthropic stream ended before its terminal message_stop event; the turn is incomplete and its content blocks were not dispatched.",
+            category: ErrorCategory.NETWORK,
+            severity: ErrorSeverity.HIGH,
+            retriable: false,
+            context: { provider: "anthropic", rawStopReason },
+        });
+    }
+    // JSON.parse happily yields null, an array or a primitive, and an
+    // assertion converts none of them. Both `tool_use.input` and
+    // AgenticLoopToolCall.args require an object, so anything else becomes
+    // {} rather than being passed through as invalid wire content.
+    const parseArgs = (json) => {
+        if (!json) {
+            return {};
+        }
+        try {
+            const parsed = JSON.parse(json);
+            return typeof parsed === "object" &&
+                parsed !== null &&
+                !Array.isArray(parsed)
+                ? parsed
+                : {};
+        }
+        catch {
+            return {};
+        }
+    };
+    // Rebuild the assistant turn in wire order so it can be replayed as the
+    // assistant message on the next step. Text that arrived before a tool
+    // call has to survive here, or the model stops seeing its own reasoning
+    // mid-turn.
+    const blocks = [];
+    const indices = new Set([
+        ...textByIndex.keys(),
+        ...toolByIndex.keys(),
+        ...thinkingByIndex.keys(),
+        ...redactedByIndex.keys(),
+    ]);
+    for (const index of [...indices].sort((a, b) => a - b)) {
+        // Thinking blocks are replayed with their signature. Anthropic
+        // validates that signature when extended thinking continues across
+        // turns, so dropping the block — or keeping the text without the
+        // signature — breaks the next step of a thinking turn.
+        const redacted = redactedByIndex.get(index);
+        if (redacted) {
+            blocks.push({ type: "redacted_thinking", data: redacted });
+        }
+        const thinking = thinkingByIndex.get(index);
+        // Both halves required: Anthropic's thinking block carries a
+        // mandatory signature, and replaying one with an empty string is
+        // rejected outright. A thinking block that never received a
+        // signature_delta is dropped rather than sent unsigned.
+        if (thinking?.text && thinking.signature) {
+            blocks.push({
+                type: "thinking",
+                thinking: thinking.text,
+                signature: thinking.signature,
+            });
+        }
+        const blockText = textByIndex.get(index);
+        if (blockText) {
+            blocks.push({
+                type: "text",
+                text: blockText,
+            });
+        }
+        const pending = toolByIndex.get(index);
+        if (pending) {
+            blocks.push({
+                type: "tool_use",
+                id: pending.id,
+                name: pending.name,
+                input: parseArgs(pending.inputJson),
+            });
+        }
+    }
+    const allCalls = [...toolByIndex.values()].map((pending) => ({
+        id: pending.id,
+        name: pending.name,
+        args: parseArgs(pending.inputJson),
+    }));
+    // A terminal structured-output call ends the turn: its arguments are the
+    // answer. Reporting it as text and leaving it out of `toolCalls` is what
+    // routes it through the engine's ordinary zero-tool-calls exit, so it is
+    // never dispatched, never counted against the breaker, and never shows
+    // up as a tool execution.
+    const terminal = finalResultToolName
+        ? [...toolByIndex.values()].find((pending) => pending.name === finalResultToolName)
+        : undefined;
+    const toolCalls = terminal ? [] : allCalls;
+    // The RAW accumulated input_json, never the parsed-then-restringified
+    // args. `parseArgs` yields {} for a payload the token cap cut off
+    // mid-string, so re-stringifying would turn a truncated answer into
+    // "{}" and lose it outright. `stringifyFinalResultInput` canonicalizes
+    // when the JSON parses and returns it verbatim when it does not, which
+    // is what lets the caller's coercion layer repair a partial payload
+    // into a partial object instead of nothing.
+    const finalText = terminal
+        ? stringifyFinalResultInput(terminal.inputJson)
+        : text;
+    if (terminal) {
+        onTerminalResult?.(finalText);
+    }
+    return {
+        text: finalText,
+        ...(reasoning ? { reasoning } : {}),
+        toolCalls,
+        usage: {
+            inputTokens,
+            outputTokens,
+            cacheReadTokens,
+            cacheWriteTokens,
+            // Omitted entirely when the stream never reported a split, so a
+            // consumer can tell "no TTL breakdown available" from "zero tokens
+            // in that tier".
+            ...(cacheWrite5mTokens ? { cacheWrite5mTokens } : {}),
+            ...(cacheWrite1hTokens ? { cacheWrite1hTokens } : {}),
+        },
+        rawStopReason,
+        raw: blocks,
+    };
+}
 export function createAnthropicLoopAdapter(config) {
     return {
         providerLabel: "anthropic",
@@ -79,254 +374,67 @@ export function createAnthropicLoopAdapter(config) {
         },
         async executeStep(request, channel, signal) {
             const params = request.raw;
-            // Single assertion, not a double: `messages.create` returns a union of
-            // Message and Stream, and Stream<RawMessageStreamEvent> already IS an
-            // AsyncIterable of that event, so the two types overlap and the
-            // compiler still checks the narrowing.
-            const events = (await config.client.messages.create({ ...params, stream: true }, 
-            // The engine's signal, not config's: runAgenticLoop already derives
-            // it from the caller's abortSignal, so preferring config's would
-            // ignore engine-initiated cancellation entirely.
-            { signal }));
-            const textByIndex = new Map();
-            const toolByIndex = new Map();
-            const thinkingByIndex = new Map();
-            const redactedByIndex = new Map();
-            let text = "";
-            let reasoning = "";
-            let rawStopReason;
-            let inputTokens = 0;
-            let outputTokens = 0;
-            let cacheReadTokens = 0;
-            let cacheWriteTokens = 0;
-            // Reported alongside the total, not derivable from it: the two TTL
-            // tiers are priced differently, so a caller that reports them (the
-            // Claude-on-Vertex turn span does) cannot reconstruct the split from
-            // cacheWriteTokens alone.
-            let cacheWrite5mTokens = 0;
-            let cacheWrite1hTokens = 0;
-            let stepOutputTokens = 0;
-            for await (const rawEvent of events) {
-                if (signal.aborted) {
-                    break;
-                }
-                // Narrowed through the SDK's own discriminated union rather than
-                // re-declared: `RawMessageStreamEvent` already describes every event
-                // shape, so `event.type` checks below are compiler-checked instead of
-                // asserted.
-                const event = rawEvent;
-                if (event.type === "message_start") {
-                    const usage = event.message?.usage;
-                    inputTokens += usage?.input_tokens ?? 0;
-                    const startOutput = usage?.output_tokens ?? 0;
-                    outputTokens += startOutput - stepOutputTokens;
-                    stepOutputTokens = startOutput;
-                    // Anthropic reports cache reads/writes separately from input_tokens
-                    // on this same event; without these the stream drops all cache
-                    // accounting.
-                    cacheReadTokens += usage?.cache_read_input_tokens ?? 0;
-                    cacheWriteTokens += usage?.cache_creation_input_tokens ?? 0;
-                    // BEST EFFORT, and the limit is worth stating. The nested TTL
-                    // breakdown exists only on `Usage` (this event); `MessageDeltaUsage`
-                    // carries the cache TOTALS but not the split, so message_start is
-                    // the only place in the raw event stream it can come from. The
-                    // pre-migration loop read it off `stream.finalMessage()` — the SDK's
-                    // ACCUMULATED message — so if the API leaves `cache_creation` null
-                    // here and fills it only on the assembled message, these two stay
-                    // zero and the totals above remain correct regardless.
-                    // Reported as undefined rather than a false zero when absent.
-                    cacheWrite5mTokens +=
-                        usage?.cache_creation?.ephemeral_5m_input_tokens ?? 0;
-                    cacheWrite1hTokens +=
-                        usage?.cache_creation?.ephemeral_1h_input_tokens ?? 0;
-                    // The guard calibrates from the FULL prompt size, not input_tokens
-                    // alone: on a cache hit the uncached remainder is tiny and using it
-                    // would let the guard drift far under the real cost.
-                    config.noteObservedPromptTokens?.((usage?.input_tokens ?? 0) +
-                        (usage?.cache_read_input_tokens ?? 0) +
-                        (usage?.cache_creation_input_tokens ?? 0));
-                    continue;
-                }
-                if (event.type === "content_block_start") {
-                    const index = event.index ?? 0;
-                    // A redacted_thinking block carries its whole payload here and
-                    // produces no deltas, so if it is not captured on this event it is
-                    // never seen again. Anthropic validates the thinking chain when
-                    // extended thinking continues across a tool-use turn, so a missing
-                    // one fails the NEXT request — and only for accounts where safety
-                    // redaction actually triggers, which is why it survives testing.
-                    if (event.content_block?.type === "redacted_thinking") {
-                        const data = event.content_block.data;
-                        if (typeof data === "string") {
-                            redactedByIndex.set(index, data);
-                        }
-                    }
-                    if (event.content_block?.type === "tool_use") {
-                        toolByIndex.set(index, {
-                            id: event.content_block.id ?? "",
-                            name: event.content_block.name ?? "",
-                            inputJson: "",
-                        });
-                    }
-                    continue;
-                }
-                if (event.type === "content_block_delta") {
-                    const index = event.index ?? 0;
-                    const delta = event.delta;
-                    if (delta?.type === "text_delta" && delta.text) {
-                        text += delta.text;
-                        textByIndex.set(index, (textByIndex.get(index) ?? "") + delta.text);
-                        channel.push({ content: delta.text });
-                    }
-                    else if (delta?.type === "thinking_delta" && delta.thinking) {
-                        const acc = thinkingByIndex.get(index) ?? {
-                            text: "",
-                            signature: "",
-                        };
-                        acc.text += delta.thinking;
-                        thinkingByIndex.set(index, acc);
-                        reasoning += delta.thinking;
-                        // Reasoning rides its own field; `content` stays a present string
-                        // so plain-text consumers are unaffected.
-                        channel.push({ content: "", reasoning: delta.thinking });
-                    }
-                    else if (delta?.type === "signature_delta" && delta.signature) {
-                        const acc = thinkingByIndex.get(index) ?? {
-                            text: "",
-                            signature: "",
-                        };
-                        acc.signature += delta.signature;
-                        thinkingByIndex.set(index, acc);
-                    }
-                    else if (delta?.type === "input_json_delta" && delta.partial_json) {
-                        const pending = toolByIndex.get(index);
-                        if (pending) {
-                            pending.inputJson += delta.partial_json;
-                        }
-                    }
-                    continue;
-                }
-                if (event.type === "message_delta") {
-                    rawStopReason = event.delta?.stop_reason ?? rawStopReason;
-                    const cumulative = event.usage?.output_tokens ?? stepOutputTokens;
-                    outputTokens += cumulative - stepOutputTokens;
-                    stepOutputTokens = cumulative;
-                }
+            // Per-REQUEST deadline, armed fresh for this step and disposed the moment
+            // it settles. Being armed per step is what makes it unresettable from a
+            // step boundary: the boundary runs between two steps, when no request
+            // deadline is running at all.
+            //
+            // Scoped composition rather than AbortSignal.any: the engine signal lives
+            // for the whole turn, and `any` keeps its registration on that signal
+            // until the derived one is collected, so composing per step accumulates
+            // listeners for the turn's whole length.
+            const requestTimeout = config.requestTimeoutMs
+                ? createTimeoutController(config.requestTimeoutMs, "anthropic", "stream")
+                : null;
+            const composed = composeAbortSignalsScoped(signal, requestTimeout?.controller.signal);
+            try {
+                return await readAnthropicStep({
+                    params,
+                    client: config.client,
+                    channel,
+                    // The engine's signal, not config's: runAgenticLoop already derives
+                    // it from the caller's abortSignal, so preferring config's would
+                    // ignore engine-initiated cancellation entirely.
+                    stepSignal: composed.signal ?? signal,
+                    requestTimeoutSignal: requestTimeout?.controller.signal,
+                    ...(config.noteObservedPromptTokens
+                        ? { noteObservedPromptTokens: config.noteObservedPromptTokens }
+                        : {}),
+                    ...(config.finalResultToolName
+                        ? { finalResultToolName: config.finalResultToolName }
+                        : {}),
+                    ...(config.onTerminalResult
+                        ? { onTerminalResult: config.onTerminalResult }
+                        : {}),
+                    ...(config.requireTerminalEvent
+                        ? { requireTerminalEvent: config.requireTerminalEvent }
+                        : {}),
+                });
             }
-            // JSON.parse happily yields null, an array or a primitive, and an
-            // assertion converts none of them. Both `tool_use.input` and
-            // AgenticLoopToolCall.args require an object, so anything else becomes
-            // {} rather than being passed through as invalid wire content.
-            const parseArgs = (json) => {
-                if (!json) {
-                    return {};
-                }
-                try {
-                    const parsed = JSON.parse(json);
-                    return typeof parsed === "object" &&
-                        parsed !== null &&
-                        !Array.isArray(parsed)
-                        ? parsed
-                        : {};
-                }
-                catch {
-                    return {};
-                }
-            };
-            // Rebuild the assistant turn in wire order so it can be replayed as the
-            // assistant message on the next step. Text that arrived before a tool
-            // call has to survive here, or the model stops seeing its own reasoning
-            // mid-turn.
-            const blocks = [];
-            const indices = new Set([
-                ...textByIndex.keys(),
-                ...toolByIndex.keys(),
-                ...thinkingByIndex.keys(),
-                ...redactedByIndex.keys(),
-            ]);
-            for (const index of [...indices].sort((a, b) => a - b)) {
-                // Thinking blocks are replayed with their signature. Anthropic
-                // validates that signature when extended thinking continues across
-                // turns, so dropping the block — or keeping the text without the
-                // signature — breaks the next step of a thinking turn.
-                const redacted = redactedByIndex.get(index);
-                if (redacted) {
-                    blocks.push({ type: "redacted_thinking", data: redacted });
-                }
-                const thinking = thinkingByIndex.get(index);
-                // Both halves required: Anthropic's thinking block carries a
-                // mandatory signature, and replaying one with an empty string is
-                // rejected outright. A thinking block that never received a
-                // signature_delta is dropped rather than sent unsigned.
-                if (thinking?.text && thinking.signature) {
-                    blocks.push({
-                        type: "thinking",
-                        thinking: thinking.text,
-                        signature: thinking.signature,
-                    });
-                }
-                const blockText = textByIndex.get(index);
-                if (blockText) {
-                    blocks.push({
-                        type: "text",
-                        text: blockText,
-                    });
-                }
-                const pending = toolByIndex.get(index);
-                if (pending) {
-                    blocks.push({
-                        type: "tool_use",
-                        id: pending.id,
-                        name: pending.name,
-                        input: parseArgs(pending.inputJson),
-                    });
-                }
+            finally {
+                composed.dispose();
+                requestTimeout?.cleanup();
             }
-            const allCalls = [...toolByIndex.values()].map((pending) => ({
-                id: pending.id,
-                name: pending.name,
-                args: parseArgs(pending.inputJson),
-            }));
-            // A terminal structured-output call ends the turn: its arguments are the
-            // answer. Reporting it as text and leaving it out of `toolCalls` is what
-            // routes it through the engine's ordinary zero-tool-calls exit, so it is
-            // never dispatched, never counted against the breaker, and never shows
-            // up as a tool execution.
-            const terminal = config.finalResultToolName
-                ? [...toolByIndex.values()].find((pending) => pending.name === config.finalResultToolName)
-                : undefined;
-            const toolCalls = terminal ? [] : allCalls;
-            // The RAW accumulated input_json, never the parsed-then-restringified
-            // args. `parseArgs` yields {} for a payload the token cap cut off
-            // mid-string, so re-stringifying would turn a truncated answer into
-            // "{}" and lose it outright. `stringifyFinalResultInput` canonicalizes
-            // when the JSON parses and returns it verbatim when it does not, which
-            // is what lets the caller's coercion layer repair a partial payload
-            // into a partial object instead of nothing.
-            const finalText = terminal
-                ? stringifyFinalResultInput(terminal.inputJson)
-                : text;
-            if (terminal) {
-                config.onTerminalResult?.(finalText);
+        },
+        appendPlanningNudge(conversation, text) {
+            // Merged into the trailing user turn when there is one. The step boundary
+            // runs immediately after `buildToolResultMessages`, so history ends on a
+            // user tool_result message; opening a second consecutive user message
+            // there is a shape the wire format does not require anyone to accept.
+            const nudgeBlock = { type: "text", text };
+            const last = conversation[conversation.length - 1];
+            if (last?.role === "user" && Array.isArray(last.content)) {
+                const merged = {
+                    ...last,
+                    content: [...last.content, nudgeBlock],
+                };
+                return [...conversation.slice(0, -1), merged];
             }
-            return {
-                text: finalText,
-                ...(reasoning ? { reasoning } : {}),
-                toolCalls,
-                usage: {
-                    inputTokens,
-                    outputTokens,
-                    cacheReadTokens,
-                    cacheWriteTokens,
-                    // Omitted entirely when the stream never reported a split, so a
-                    // consumer can tell "no TTL breakdown available" from "zero tokens
-                    // in that tier".
-                    ...(cacheWrite5mTokens ? { cacheWrite5mTokens } : {}),
-                    ...(cacheWrite1hTokens ? { cacheWrite1hTokens } : {}),
-                },
-                rawStopReason,
-                raw: blocks,
+            const message = {
+                role: "user",
+                content: [nudgeBlock],
             };
+            return [...conversation, message];
         },
         buildToolResultMessages(conversation, stepResult, toolResults) {
             const assistantMessage = {

package/dist/providers/googleAiStudio/client.js

@@ -851,6 +851,13 @@ export class GoogleAIStudioProvider extends BaseProvider {
                         const { stream: engineStream, resultPromise } = runAgenticLoop(adapter, currentContents, {
                             tools: engineTools,
                             ...(composedSignal ? { abortSignal: composedSignal } : {}),
+                            // This loop guards no executor of its own, so the engine's
+                            // per-tool bound is the only thing standing between a wedged
+                            // tool and a turn that never ends. Honour the caller's value
+                            // when there is one; the engine defaults otherwise.
+                            ...(options.toolTimeoutMs !== undefined
+                                ? { toolTimeoutMs: options.toolTimeoutMs }
+                                : {}),
                         });
                         const pump = (async () => {
                             for await (const chunk of engineStream) {
@@ -1161,6 +1168,11 @@ export class GoogleAIStudioProvider extends BaseProvider {
                 const { stream: engineStream, resultPromise } = runAgenticLoop(adapter, currentContents, {
                     tools: engineTools,
                     ...(composedSignal ? { abortSignal: composedSignal } : {}),
+                    // This loop guards no executor of its own — see the streaming
+                    // path above for why the engine's bound has to be reachable.
+                    ...(options.toolTimeoutMs !== undefined
+                        ? { toolTimeoutMs: options.toolTimeoutMs }
+                        : {}),
                 });
                 // Drained, not consumed: nothing streams out of generate(), but an
                 // undrained channel would stall the engine mid-turn.