@bitkyc08/opencodex
Version:
Universal provider proxy for OpenAI Codex & Claude Code — use any LLM with Codex CLI/App/SDK and Claude Code
712 lines (665 loc) • 31.7 kB
text/typescript
import type { TranslatorBudget } from "../../lib/translator-budget";
import type { AdapterEvent } from "../../types";
import { repairFreeformToolInput } from "../../responses/apply-patch-envelope";
const OPEN_TAG = "<tool_call>";
const CLOSE_TAG = "</tool_call>";
const FUNCTION_TAG = "<function=";
/**
* Streaming bounds (ingestStreaming only; buffered reconciliation already knows its structured calls).
* MAX_HELD_BYTES is a runaway guard on held text plus queued events, far below the translator budget,
* so a large duplicated write inside the block is still reconciled. MAX_TRAILING_CHARS bounds the
* latency case: a duplicated block is the tail of the content (#5548), so a closed block followed by
* this much prose, with no block open after it, is treated as prose and released.
*/
const MAX_HELD_BYTES = 4 * 1024 * 1024;
const MAX_TRAILING_CHARS = 8 * 1024;
interface TextContext {
fence: string | null;
lineStart: boolean;
inlineTicks?: number;
}
export interface SerializedToolCall {
name: string;
body: string;
start: number;
end: number;
}
/** A freeform tool's declared identity. Echoes are matched against it, never against the wire alias. */
export interface FreeformToolIdentity {
name: string;
namespace?: string;
}
/** The request's freeform tools keyed by the wire name the provider sees for each of them. */
export function freeformToolsByWireName<T extends FreeformToolIdentity & { freeform?: boolean }>(
tools: readonly T[] | undefined,
wireName: (tool: T) => string,
): Map<string, FreeformToolIdentity> {
return new Map(tools?.filter(tool => tool.freeform).map(tool => [wireName(tool), { name: tool.name, namespace: tool.namespace }] as const) ?? []);
}
export interface StructuredToolCallReference {
names: ReadonlySet<string>;
argumentsText: string;
freeformTool?: FreeformToolIdentity;
recoveredParameterWrapper?: boolean;
}
const BLOCK_HEADER = /<tool_call>\s*<function=([^>\r\n]+)>/y;
/** A separate bare block starts a line (see `splitAtPossibleSerializedToolCall`); a header mid-line is body text. */
const NEXT_BLOCK_HEADER = /\n<tool_call>\s*<function=[^>\r\n]+>/g;
const FUNCTION_CLOSE = "</function>";
const PARAMETER_CLOSE = "</parameter>";
function trimmedEnd(text: string, from: number, to: number): number {
while (to > from && /\s/.test(text[to - 1]!)) to--;
return to;
}
function endsWithAt(text: string, from: number, to: number, suffix: string): boolean {
return to - suffix.length >= from && text.startsWith(suffix, to - suffix.length);
}
/**
* The block starting at `offset`, read by delimiter scan so an unterminated block costs linear time.
* MiMo's echo may close a freeform body with a stray `</parameter>` and may omit `</function>`
* (#5724), the grammar the Command Code reader accepts too. The first `</tool_call>` preceded by
* `</function>` closes the block, so a body can still carry a literal `</tool_call>` or header;
* with none before the next line-start block header, the first `</tool_call>` does. That header only
* bounds an unclosed candidate: with no close at all before it, it is body text, and a closed
* `</function></tool_call>` after it still ends the block.
*/
function blockAt(text: string, offset: number): SerializedToolCall | undefined {
BLOCK_HEADER.lastIndex = offset;
const header = BLOCK_HEADER.exec(text);
if (!header) return undefined;
const bodyStart = offset + header[0].length;
NEXT_BLOCK_HEADER.lastIndex = bodyStart;
const next = NEXT_BLOCK_HEADER.exec(text);
const limit = next ? next.index + 1 : text.length;
let unclosed: SerializedToolCall | undefined;
for (let close = text.indexOf(CLOSE_TAG, bodyStart); close >= 0 && (close < limit || !unclosed);
close = text.indexOf(CLOSE_TAG, close + CLOSE_TAG.length)) {
let bodyEnd = trimmedEnd(text, bodyStart, close);
const closed = endsWithAt(text, bodyStart, bodyEnd, FUNCTION_CLOSE);
if (closed) bodyEnd = trimmedEnd(text, bodyStart, bodyEnd - FUNCTION_CLOSE.length);
if (endsWithAt(text, bodyStart, bodyEnd, PARAMETER_CLOSE)) bodyEnd -= PARAMETER_CLOSE.length;
const call = {
name: header[1]!.trim(),
body: text.slice(bodyStart, bodyEnd),
start: offset,
end: close + CLOSE_TAG.length,
};
if (closed) return call;
if (close < limit) unclosed ??= call;
}
return unclosed;
}
/** Finds complete bare blocks outside literal Markdown; ambiguous outer blocks stop the scan. */
function callsIn(text: string, context: TextContext = { fence: null, lineStart: true }): SerializedToolCall[] {
const calls: SerializedToolCall[] = [];
let offset = 0;
while (offset < text.length) {
const split = splitAtPossibleSerializedToolCall(text.slice(offset), context, true);
offset += split.emit.length;
if (!split.hasOpenTag) break;
const match = blockAt(text, offset);
if (!match) break; // An incomplete/ambiguous outer block cannot authorize an inner call.
calls.push(match);
offset = match.end;
context = { fence: null, lineStart: false };
}
return calls;
}
/**
* The first block, and only when the text after it is exactly one repetition of that same block
* (trailing whitespace allowed). The returned range covers the pair and any trailing whitespace.
*/
function repeatedCallIn(text: string, context?: TextContext): SerializedToolCall | undefined {
const first = callsIn(text, context)[0];
if (!first) return undefined;
if (text.slice(first.end).trimEnd() !== text.slice(first.start, first.end).trimEnd()) return undefined;
return { ...first, end: text.length };
}
/** Splits safe visible text from a possible control block while carrying Markdown context across chunks. */
export function splitAtPossibleSerializedToolCall(
text: string,
initialContext: TextContext = { fence: null, lineStart: true },
final = false,
): {
emit: string;
defer: string;
hasOpenTag: boolean;
context: TextContext;
} {
const context = { ...initialContext };
const split = (at: number, hasOpenTag = false) => ({
emit: text.slice(0, at), defer: text.slice(at), hasOpenTag, context,
});
for (let index = 0; index < text.length; index++) {
if (context.lineStart) {
const rest = text.slice(index);
const fence = /^ {0,3}(`{3,}|~{3,})([^\n]*)/.exec(rest);
if (!context.inlineTicks && fence && (!context.fence || (fence[1]![0] === context.fence[0]
&& fence[1]!.length >= context.fence.length && /^[ \t\r]*$/.test(fence[2]!)))) {
if (!final && !rest.includes("\n")) return split(index);
context.fence = context.fence ? null : fence[1]!;
index += fence[0].length - 1;
context.lineStart = false;
continue;
}
if (!final && /^ {0,3}(`*|~*)$/.test(rest)) return split(index);
// Only bare control markup qualifies. Prose, quotes, indented examples and
// fenced code stay user-visible even when their body matches a real call.
if (!context.fence && !context.inlineTicks) {
if (rest.startsWith(OPEN_TAG)) {
const header = rest.slice(OPEN_TAG.length).trimStart();
if (/^<function=[^>\r\n]+>/.test(header)) return split(index, true);
if (!final && (FUNCTION_TAG.startsWith(header)
|| (header.startsWith(FUNCTION_TAG) && !/[>\r\n]/.test(header.slice(FUNCTION_TAG.length))))) {
return split(index);
}
} else if (!final && OPEN_TAG.startsWith(rest)) return split(index);
}
}
if (!context.fence && text[index] === "`") {
let end = index + 1;
while (text[end] === "`") end++;
if (!final && end === text.length) return split(index);
const ticks = end - index;
if (!context.inlineTicks) context.inlineTicks = ticks;
else if (context.inlineTicks === ticks) context.inlineTicks = undefined;
index = end - 1;
}
context.lineStart = text[index] === "\n";
}
return split(text.length);
}
/**
* The line, fence and inline-code state after `text`. Serialized blocks are neutralised first so
* the scan runs through the whole text instead of stopping at the first opening tag; a block ends
* mid-line, which is exactly what the neutral spelling reports too.
*/
function contextAfter(text: string, context: TextContext): TextContext {
if (text.length === 0) return context;
return splitAtPossibleSerializedToolCall(text.replaceAll(OPEN_TAG, "<tool-call>"), context, true).context;
}
/**
* A deferred header/fence can itself be long. While its unresolved portion only grows, a
* chunk-local check suffices; the first resolving character goes back through the full splitter.
* Each unbounded phase is scanned once on entry and once on exit, preserving its Markdown rules.
*/
function deferredContinuationBreak(text: string, context: TextContext): RegExp | undefined {
if (text.startsWith(OPEN_TAG)) {
if (/^<tool_call>\s*$/.test(text)) return /\S/;
if (/^<tool_call>\s*<function=[^>\r\n]*$/.test(text)) return /[>\r\n]/;
}
if (!context.inlineTicks && context.lineStart) {
const fence = /^ {0,3}(`{3,}|~{3,})([^\n]*)$/.exec(text);
if (fence) {
if (!context.fence) return /\n/;
// Closing-fence ticks may grow only before its optional whitespace suffix.
if (!fence[2]) return fence[1]![0] === "`" ? /[^`]/ : /[^~]/;
return /[^ \t\r]/;
}
}
if (/^`+$/.test(text)) return /[^`]/;
return undefined;
}
/**
* Holds possible duplicate text within the shared translator budget until the dispatch outcome is
* known. While a block candidate is open, any other event (reasoning) is queued at its position in
* the held text rather than overtaking it or forcing the block out early, and `drain` restores the
* original order.
*/
export class SerializedToolCallContentBuffer {
releasedAnswerText = false;
private text = "";
private bytes = 0;
private hasOpenTag = false;
private continuationBreak: RegExp | undefined;
private delimiterSuffix = "";
private sawCloser = false;
private openAfterCloser = false;
private trailingProseChars = 0;
private context: TextContext = { fence: null, lineStart: true };
private queued: { offset: number; event: AdapterEvent }[] = [];
constructor(private readonly budget: TranslatorBudget) {}
/** Remembers visible text released before a later structured call can be reconciled. */
private observeText(events: AdapterEvent[]): AdapterEvent[] {
if (events.some(event => event.type === "text_delta" && event.text.trim() !== "")) this.releasedAnswerText = true;
return events;
}
/** Measures an append exactly even when the runtime prices isolated UTF-16 surrogates differently. */
private appendedByteLength(delta: string): number {
let bytes = Buffer.byteLength(delta);
if (this.text.length === 0 || delta.length === 0) return bytes;
const tail = this.text[this.text.length - 1]!;
const head = delta[0]!;
const tailCode = tail.charCodeAt(0);
const headCode = head.charCodeAt(0);
if (tailCode >= 0xd800 && tailCode <= 0xdbff && headCode >= 0xdc00 && headCode <= 0xdfff) {
bytes += Buffer.byteLength(tail + head) - Buffer.byteLength(tail) - Buffer.byteLength(head);
}
return bytes;
}
/** Reserves the replacement before releasing the old text, preserving it if the budget rejects growth. */
private replace(next: string, hasOpenTag: boolean): void {
const nextBytes = Buffer.byteLength(next);
const reservation = this.budget.reserveTransient(nextBytes, { kind: "live_transient" });
try {
reservation.commitRetained();
this.budget.releaseRetained(this.bytes, { kind: "live_transient" });
this.text = next;
this.bytes = nextBytes;
this.hasOpenTag = hasOpenTag;
this.continuationBreak = undefined;
this.delimiterSuffix = "";
this.sawCloser = false;
this.openAfterCloser = false;
this.trailingProseChars = 0;
if (hasOpenTag) this.observeDelimiters(next);
} catch (error) {
reservation.release();
throw error;
}
}
/** Charges only the appended bytes, so holding an open block never needs twice its retained size. */
private append(delta: string): void {
const deltaBytes = this.appendedByteLength(delta);
this.budget.reserveTransient(deltaBytes, { kind: "live_transient" }).commitRetained();
if (this.hasOpenTag) this.observeDelimiters(delta);
this.text += delta;
this.bytes += deltaBytes;
}
/** Scan new text plus a fixed overlap, never the retained body (which may be a large rope). */
private observeDelimiters(delta: string): void {
const scan = this.delimiterSuffix + delta;
const closer = scan.lastIndexOf(CLOSE_TAG);
const afterCloser = closer + CLOSE_TAG.length;
if (closer >= 0) {
this.sawCloser = true;
this.openAfterCloser = false;
this.trailingProseChars = 0;
}
if (scan.lastIndexOf(OPEN_TAG) >= (closer >= 0 ? afterCloser : 0)) {
this.openAfterCloser = true;
}
const tail = closer >= 0 ? scan.slice(afterCloser) : delta;
if (this.sawCloser) this.trailingProseChars += tail.replace(/\s+/g, "").length;
this.delimiterSuffix = scan.slice(-(Math.max(OPEN_TAG.length, CLOSE_TAG.length) - 1));
}
/** Returns immediately safe text and retains only the suffix that still needs reconciliation. */
ingest(delta: string): string {
if (this.hasOpenTag || (this.continuationBreak && !this.continuationBreak.test(delta))) {
this.append(delta);
return "";
}
const split = splitAtPossibleSerializedToolCall(this.text + delta, this.context);
this.replace(split.defer, split.hasOpenTag);
this.context = split.context;
if (!split.hasOpenTag) this.continuationBreak = deferredContinuationBreak(split.defer, split.context);
return split.emit;
}
/**
* Streaming ingest with bounded retention. Past either bound the stream prefers delivering text
* over suppressing a possible duplicate: everything held is released in order, nothing is
* suppressed (the behaviour before #5548 for that block), and scanning resumes from the carried
* context. The size bound is checked before the delta is retained.
*/
ingestStreaming(delta: string): AdapterEvent[] {
const deltaBytes = this.appendedByteLength(delta);
if (this.hasOpenTag && this.bytes + deltaBytes > MAX_HELD_BYTES) {
const released = this.drain([]);
// A delta that alone passes the bound is delivered as text rather than retained.
if (Buffer.byteLength(delta) > MAX_HELD_BYTES) {
this.context = contextAfter(delta, this.context);
return this.observeText([...released, ...textEvents(delta)]);
}
return this.observeText([...released, ...textEvents(this.ingest(delta))]);
}
const text = this.ingest(delta);
// Checked after ingest too: one delta can open a block and already carry more than a bound.
if (this.hasOpenTag && (this.bytes > MAX_HELD_BYTES || (this.sawCloser && !this.openAfterCloser && this.trailingProseChars > MAX_TRAILING_CHARS))) {
return this.observeText([...textEvents(text), ...this.drain([])]);
}
return this.observeText(textEvents(text));
}
/** Exposes held text as evidence for narrowly repairing duplicated argument prefixes. */
current(): string {
return this.text;
}
/**
* Passes a non-text event through, in order. With an open block candidate held, the event is
* queued behind the held text and a heartbeat stands in for it; with only a partial prefix held (no complete opening tag yet),
* that prefix cannot be a whole duplicate and is released ahead of the event.
*/
hold(event: AdapterEvent): AdapterEvent[] {
if (!this.hasOpenTag) return [...this.drain([]), event];
const eventBytes = Buffer.byteLength(JSON.stringify(event));
// Queued events count toward the same runaway bound as held text.
if (this.bytes + eventBytes > MAX_HELD_BYTES) return [...this.drain([]), event];
this.budget.reserveTransient(eventBytes, { kind: "live_transient" }).commitRetained();
this.bytes += eventBytes;
this.queued.push({ offset: this.text.length, event });
// The consumer still sees activity, so a stall watchdog never mistakes a held turn for a dead one.
return [{ type: "heartbeat" }];
}
/**
* Drains held text and queued events in their original order, suppressing only blocks that
* duplicate a dispatched call; pass an empty list on failure to preserve everything.
*/
drain(structuredCalls: readonly StructuredToolCallReference[]): AdapterEvent[] {
const removed = duplicatedSerializedToolCallRanges(this.text, structuredCalls, this.context);
const kept = (from: number, to: number): string => {
let piece = "";
let cursor = from;
for (const range of removed) {
if (range.end <= cursor || range.start >= to) continue;
piece += this.text.slice(cursor, Math.max(cursor, range.start));
cursor = Math.min(to, range.end);
}
return piece + this.text.slice(cursor, to);
};
const out: AdapterEvent[] = [];
let cursor = 0;
for (const boundary of [...this.queued, { offset: this.text.length, event: undefined }]) {
const text = kept(cursor, boundary.offset);
if (text.length > 0) out.push({ type: "text_delta", text });
if (boundary.event) out.push(boundary.event);
cursor = boundary.offset;
}
// Later text continues after what was drained, so its line and fence state carry forward.
this.context = contextAfter(this.text, this.context);
this.queued = [];
this.replace("", false);
return this.observeText(out);
}
/** Text-only drain for callers that never queued an event. */
flush(structuredCalls: readonly StructuredToolCallReference[]): string {
return this.drain(structuredCalls)
.map(event => (event.type === "text_delta" ? event.text : ""))
.join("");
}
/** Releases retained bytes when the stream ends or its consumer cancels iteration. */
dispose(): void {
this.budget.releaseRetained(this.bytes, { kind: "live_transient" });
this.text = "";
this.bytes = 0;
this.hasOpenTag = false;
this.continuationBreak = undefined;
this.queued = [];
}
}
/** Reads a wrapped input or raw arguments from a declared freeform tool; neither path rewrites them. */
function inputFromArguments(
argumentsText: string,
freeformTool?: FreeformToolIdentity,
): string | undefined {
try {
const parsed = JSON.parse(argumentsText) as unknown;
const input = parsed !== null && typeof parsed === "object" && !Array.isArray(parsed)
? (parsed as Record<string, unknown>).input
: undefined;
if (typeof input === "string") return input;
} catch {
// Chat gateways can send custom-tool input as raw text.
}
if (!freeformTool) return undefined;
// Match raw input only when the Responses bridge dispatches these same bytes. It can unwrap
// alternate fields or repair freeform input using the declared tool's name and namespace.
return repairFreeformToolInput(argumentsText, freeformTool.name, freeformTool.namespace) === argumentsText
? argumentsText : undefined;
}
/** One wrapping newline after the function header is template layout, not input (vLLM `_trim_wrapping_newlines`). */
function freeformBody(value: string): string {
return value.replace(/^\r?\n/, "").trimEnd();
}
/**
* Whether a structured call's freeform input already equals the body of `repeated`. Such a call
* explains the repeated pair on its own, which is what competes with a doubled call in the same
* batch: both readings account for the two blocks, and the response never says which one it meant.
*/
function agreesWithRepeatedBlock(
structured: StructuredToolCallReference,
repeated: SerializedToolCall,
): boolean {
const input = structured.names.has(repeated.name) ? inputFromArguments(structured.argumentsText, structured.freeformTool) : undefined;
return input !== undefined && freeformBody(input) === freeformBody(repeated.body);
}
/** A second call can explain a repeated pair even when its arguments cannot safely be rewritten. */
function hasDoubledInput(structured: StructuredToolCallReference, repeated: SerializedToolCall): boolean {
if (!structured.names.has(repeated.name)) return false;
const input = inputFromArguments(structured.argumentsText, structured.freeformTool);
if (input === undefined) return false;
const body = freeformBody(repeated.body);
const normalized = freeformBody(input);
return normalized === body + body || normalized === body + "\n" + body;
}
/** The `[start, end)` ranges of blocks whose function identity and freeform input match a dispatched call. */
function duplicatedSerializedToolCallRanges(
text: string,
structuredCalls: readonly StructuredToolCallReference[],
context?: TextContext,
): { start: number; end: number }[] {
if (structuredCalls.length === 0) return [];
const repeated = repeatedCallIn(text, context);
if (repeated) {
// A doubled call beside an agreeing call leaves the pair ambiguous. If reduction was
// refused, keep the markup too: otherwise the visible text and executable call disagree.
const matching = structuredCalls.filter(structured => agreesWithRepeatedBlock(structured, repeated));
const doubled = structuredCalls.some(structured => structured !== matching[0] && hasDoubledInput(structured, repeated));
return matching.length === 1 && !doubled ? [{ start: repeated.start, end: repeated.end }] : [];
}
return callsIn(text, context).filter(call => {
const body = freeformBody(call.body);
return structuredCalls.some(structured => {
const input = structured.names.has(call.name) ? inputFromArguments(structured.argumentsText, structured.freeformTool) : undefined;
if (input === undefined) return false;
const normalized = freeformBody(input);
return normalized === body || (structured.recoveredParameterWrapper === true
&& body.startsWith("<parameter=") && normalized === freeformBody(body.slice("<parameter=".length)));
});
});
}
/** Removes eligible blocks only when both the function identity and freeform input match a dispatched call. */
export function stripDuplicatedSerializedToolCalls(
text: string,
structuredCalls: readonly StructuredToolCallReference[],
context?: TextContext,
): string {
let result = "";
let cursor = 0;
for (const range of duplicatedSerializedToolCallRanges(text, structuredCalls, context)) {
result += text.slice(cursor, range.start);
cursor = range.end;
}
return result + text.slice(cursor);
}
/**
* The reduced arguments when the freeform body of the repeated block is written twice in the
* single string "input" field, or undefined for any other shape. Only the batch reconciler may
* apply it: the reduction rewrites executable arguments, so it needs a uniqueness proof.
*/
function doubledInputReduction(
argumentsText: string,
functionNames: ReadonlySet<string>,
repeated: SerializedToolCall,
): string | undefined {
if (!functionNames.has(repeated.name)) return undefined;
const body = freeformBody(repeated.body);
try {
const parsed = JSON.parse(argumentsText) as unknown;
if (parsed !== null && typeof parsed === "object" && !Array.isArray(parsed)
&& Object.keys(parsed).length === 1
&& ((parsed as Record<string, unknown>).input === body + body
|| (parsed as Record<string, unknown>).input === body + "\n" + body)) {
return JSON.stringify({ input: body });
}
} catch {
// A malformed concatenation is handled by the prefix repair instead.
}
return undefined;
}
/** Removes a malformed argument prefix only when a bare block and the JSON suffix prove identical input. */
export function repairArgumentsDuplicatedBesideSerializedCall(
argumentsText: string,
functionNames: ReadonlySet<string>,
serializedText: string,
): string {
try {
JSON.parse(argumentsText);
return argumentsText;
} catch {
// Continue only for the exact duplication shape emitted by some Chat gateways.
}
const bodies = callsIn(serializedText)
.filter(call => functionNames.has(call.name))
.map(call => call.body.trimEnd());
if (bodies.length === 0) return argumentsText;
for (const body of bodies) {
if (!argumentsText.startsWith(body)) continue;
let start = body.length;
while (start < argumentsText.length && /\s/.test(argumentsText[start]!)) start += 1;
const candidate = argumentsText.slice(start);
let parsed: unknown;
try {
parsed = JSON.parse(candidate);
} catch {
continue;
}
if (parsed === null || typeof parsed !== "object" || Array.isArray(parsed)) continue;
const input = (parsed as Record<string, unknown>).input;
if (typeof input !== "string") continue;
if (body !== input.trimEnd()) continue;
return candidate;
}
return argumentsText;
}
/** One structured call as the reconciler sees it, before its arguments meet the visible text. */
export interface StructuredToolCallInput {
wireName: string;
restoredName: string;
argumentsText: string;
freeformTool?: FreeformToolIdentity;
}
/**
* Repairs the arguments of every structured call in one response against the visible text that
* response carried, and returns them in input order. The per-call prefix repair stands alone,
* because the markup it proves is matched against that call's own repaired input. The
* doubled-input reduction is applied only when exactly one call in the batch explains the repeated
* pair, including doubled shapes that cannot themselves be rewritten. It rewrites executable
* arguments, so a second explanatory call leaves the pair ambiguous.
*/
export function reconcileStructuredToolCalls(
calls: readonly StructuredToolCallInput[],
serializedText: string,
recoverEmptyMiMoFreeformInput = false,
): StructuredToolCallReference[] {
const references: StructuredToolCallReference[] = calls.map(call => {
const names = new Set([call.wireName, call.restoredName]);
return { names, freeformTool: call.freeformTool, argumentsText: repairArgumentsDuplicatedBesideSerializedCall(call.argumentsText, names, serializedText) };
});
if (recoverEmptyMiMoFreeformInput && calls.length === 1 && calls[0]!.freeformTool
&& emptyObjectArguments(references[0]!.argumentsText)) {
// A standalone MiMo block and one empty call to the exact wire tool provide the only
// unambiguous recovery shape. Prose before/after it or another call keeps text inert.
const blocks = callsIn(serializedText);
if (blocks.length === 1 && serializedText.slice(0, blocks[0]!.start).trim() === ""
&& serializedText.slice(blocks[0]!.end).trim() === "") {
const block = blocks[0]!;
const wrappedBody = freeformBody(block.body);
// MiMo can put the freeform input after a malformed `<parameter=` opener.
const hasParameterOpener = wrappedBody.startsWith("<parameter=");
const body = hasParameterOpener
? freeformBody(wrappedBody.slice("<parameter=".length)) : wrappedBody;
if (block.name === calls[0]!.wireName && body.trim() !== ""
&& !/(?:^|\n)<tool_call>\s*<function=[^>\r\n]+>/.test(block.body)
&& !(hasParameterOpener && /^<parameter=[A-Za-z_$][\w$.-]*>/.test(wrappedBody))
&& !body.includes("<parameter=") && !body.includes("</parameter>")) {
references[0] = { ...references[0]!, argumentsText: JSON.stringify({ input: body }),
recoveredParameterWrapper: hasParameterOpener };
}
}
}
return reduceUnambiguousDoubledInput(references, serializedText);
}
/** An empty JSON object means the structured freeform call supplied no input. */
function emptyObjectArguments(text: string): boolean {
try {
const parsed: unknown = JSON.parse(text);
return parsed !== null && typeof parsed === "object" && !Array.isArray(parsed)
&& Object.keys(parsed).length === 0;
} catch {
return false;
}
}
/**
* Applies the doubled-input reduction across the batch. The doubled shape is valid JSON, so the
* prefix repair returns it untouched, and the reduction only ever rewrites a call the repair left
* alone. A call whose input already equals the repeated body is a competing explanation, not a
* bystander: both readings account for the pair and the response never picks one, so a batch with
* two qualifying calls keeps every argument exactly as sent. The reduction then rewrites nothing,
* and the range matcher keeps the markup when a doubled call competes with an agreeing call.
* An already-agreeing call leaves nothing to reduce.
*/
function reduceUnambiguousDoubledInput(
references: readonly StructuredToolCallReference[],
serializedText: string,
): StructuredToolCallReference[] {
const repeated = repeatedCallIn(serializedText);
if (!repeated) return [...references];
const candidates = references.map(reference => ({
reduction: doubledInputReduction(reference.argumentsText, reference.names, repeated),
explains: agreesWithRepeatedBlock(reference, repeated),
doubled: hasDoubledInput(reference, repeated),
}));
if (candidates.filter(candidate => candidate.explains || candidate.doubled).length !== 1) {
return [...references];
}
return references.map((reference, index) => {
const reduction = candidates[index]!.reduction;
return reduction === undefined ? reference : { ...reference, argumentsText: reduction };
});
}
/**
* One structured call as the reconciler sees it: both the wire name and its restored client name
* identify it, and its arguments are repaired against the visible text the same response carried.
* A single call is its own batch, so the doubled-input reduction still applies here; a caller
* holding several calls of one response must pass them together to
* `reconcileStructuredToolCalls` so the reduction sees all of them.
*/
export function reconcileStructuredToolCall(
wireName: string,
restoredName: string,
argumentsText: string,
serializedText: string,
): StructuredToolCallReference {
return reconcileStructuredToolCalls([{ wireName, restoredName, argumentsText }], serializedText)[0]!;
}
/**
* Buffered-response counterpart of the streaming path, applied in place to the content events in
* `events[start, end)`. It replays them through the same buffer the stream uses, so both paths
* share one rule set: text carries its line and fence context across events (the inline-think
* splitter may cut one answer into several), any other event keeps its place relative to held
* text, and only text still held at the end is matched against the structured calls.
*/
export function reconcileSerializedToolCallEvents(
events: AdapterEvent[],
start: number,
end: number,
structuredCalls: readonly StructuredToolCallReference[],
budget: TranslatorBudget,
): void {
if (structuredCalls.length === 0) return;
const buffer = new SerializedToolCallContentBuffer(budget);
const reconciled: AdapterEvent[] = [];
try {
for (const event of events.slice(start, end)) {
if (event.type !== "text_delta") { reconciled.push(...buffer.hold(event).filter(held => held.type !== "heartbeat")); continue; }
const text = buffer.ingest(event.text);
if (text.length > 0) reconciled.push({ type: "text_delta", text });
}
reconciled.push(...buffer.drain(structuredCalls));
} finally {
buffer.dispose();
}
events.splice(start, end - start, ...reconciled);
}
function textEvents(text: string): AdapterEvent[] {
return text.length > 0 ? [{ type: "text_delta", text }] : [];
}