cartethyia 2.7.5 → 2.7.7

package/src/providers/integrations/qoder/qoder-text-tools.ts

@@ -1,277 +0,0 @@
-/**
- * Recovery for Qoder's INTL text-mode tool-call echo.
- *
- * Qoder's own wire is structured `tool_calls` deltas, but the DeepSeek model
- * sometimes *also* writes a tool call as prose with the marker:
- *
- *   assistant 请求调用工具
- *   {"name":"write","arguments":"…"},{"name":"write","arguments":"…"}
- *
- * Left alone the JSON reaches the caller as visible content while nothing runs.
- * This module owns the single tolerant parser for that echo family (JSON
- * array/object, bare comma-separated list, unescaped inner quotes). It does
- * NOT parse DSML/XML markup — that family is gone; only the marker+JSON echo
- * the model actually emits is recovered here.
- */
-import { randomUUID } from "node:crypto";
-
-/** One recovered call, in the shape the canonical events want. */
-export interface QoderEchoCall {
-  readonly id: string;
-  readonly name: string;
-  readonly arguments: string;
-}
-
-/** Any spelling of the INTL tool-call leak marker. */
-export const LEAK_MARKER_ANY_RE = /[\[<(]*\s*assistant\s*请求调用工具\s*[\]>)]?/;
-
-/** Canonical marker literals used to seed prefix matching. */
-export const LEAK_MARKER_LITERALS: readonly string[] = [
-  "[assistant 请求调用工具]",
-  "<assistant 请求调用工具>",
-  "assistant 请求调用工具",
-];
-
-/**
- * Escapes literal control characters that occur inside JSON string values.
- * DeepSeek often writes multi-line argument content as real newlines instead
- * of `\n` escapes; only characters inside strings are escaped.
- */
-function escapeControlInsideStrings(text: string): string {
-  let out = "";
-  let inString = false;
-  for (let i = 0; i < text.length; i += 1) {
-    const ch = text[i] ?? "";
-    if (inString) {
-      if (ch === "\\") {
-        out += ch;
-        i += 1;
-        if (i < text.length) out += text[i];
-        continue;
-      }
-      if (ch === '"') {
-        inString = false;
-        out += ch;
-        continue;
-      }
-      if (ch === "\n") {
-        out += "\\n";
-        continue;
-      }
-      if (ch === "\r") {
-        out += "\\r";
-        continue;
-      }
-      if (ch === "\t") {
-        out += "\\t";
-        continue;
-      }
-      if (ch.charCodeAt(0) < 0x20) {
-        out += "\\u" + ch.charCodeAt(0).toString(16).padStart(4, "0");
-        continue;
-      }
-      out += ch;
-      continue;
-    }
-    if (ch === '"') inString = true;
-    out += ch;
-  }
-  return out;
-}
-
-/** Escapes bare quotes that belong inside a JSON string value. */
-function repairJsonQuotes(text: string): string {
-  const ph = "\u0000";
-  return text.replace(/\\"/g, ph).replace(/"/g, '\\"').replaceAll(ph, '\\"');
-}
-
-/**
- * Rebuilds the leaked echo when strict parsing fails: inner quotes unescaped
- * (`[{"name":"read","arguments":"{"path":"a"}"},…]`) or a bare comma-separated
- * object list. Scans top-level objects without relying on JSON validity.
- * Returns null when the text is not that shape at all.
- */
-function salvageEchoArray(rest: string): unknown {
-  const src = rest.trim();
-  if (src.length === 0) return null;
-  const calls: Array<{ name: string; arguments: string }> = [];
-  let i = 0;
-  const bracketed = src[0] === "[";
-  if (bracketed) i = 1;
-  const skipWs = (): void => {
-    while (
-      i < src.length &&
-      (src[i] === " " || src[i] === "\t" || src[i] === "\n" || src[i] === "\r")
-    ) {
-      i += 1;
-    }
-  };
-  const NAME_RE = /"name"\s*:\s*"([^"]{1,200})"/y;
-  const ARG_KEY_RE = /"arguments"\s*:\s*"/y;
-
-  for (;;) {
-    skipWs();
-    if (i >= src.length) break;
-    if (bracketed && src[i] === "]") break;
-    if (src[i] !== "{") return null;
-    i += 1;
-    skipWs();
-    NAME_RE.lastIndex = i;
-    const nameMatch = NAME_RE.exec(src);
-    if (!nameMatch) return null;
-    const name = nameMatch[1] ?? "";
-    i = NAME_RE.lastIndex;
-    skipWs();
-    if (i < src.length && src[i] === ",") {
-      i += 1;
-      skipWs();
-    }
-    ARG_KEY_RE.lastIndex = i;
-    if (!ARG_KEY_RE.exec(src)) return null;
-    const argStart = ARG_KEY_RE.lastIndex;
-    let rawEnd = -1;
-    for (let j = argStart; j + 1 < src.length; j += 1) {
-      if (src[j] === "}" && src[j + 1] === '"') {
-        rawEnd = j;
-        break;
-      }
-    }
-    if (rawEnd < 0) return null;
-    const raw = src.slice(argStart, rawEnd + 1);
-    const parsedArgs = [
-      raw,
-      escapeControlInsideStrings(raw),
-      repairJsonQuotes(raw),
-      escapeControlInsideStrings(repairJsonQuotes(raw)),
-    ].find((candidate) => {
-      try {
-        JSON.parse(candidate);
-        return true;
-      } catch {
-        return false;
-      }
-    });
-    if (parsedArgs === undefined) return null;
-    calls.push({ name, arguments: parsedArgs });
-    i = rawEnd + 2;
-    skipWs();
-    if (i >= src.length || src[i] !== "}") return null;
-    i += 1;
-    skipWs();
-    if (i < src.length && src[i] === ",") {
-      i += 1;
-      continue;
-    }
-    if (i < src.length && bracketed && src[i] === "]") continue;
-    if (i < src.length) return null;
-  }
-  return calls.length > 0 ? calls : null;
-}
-
-/**
- * True when the text is clearly an attempted tool-protocol echo rather than
- * ordinary prose: a leak marker followed by the start of a JSON array/object,
- * or of a tool-style tag (`<tool_call>`, `<execute …/>`, angle-wrapped
- * `<{…}>`). Used by the fail-closed decision: an echo that cannot be recovered
- * fails the turn instead of surfacing protocol text as visible content.
- */
-export function isQoderEchoAttempt(text: string): boolean {
-  const fence = text.match(/^```[a-z]*\s*\n?/i);
-  const src = fence && fence[0] ? text.slice(fence[0].length) : text;
-  const marker = LEAK_MARKER_ANY_RE.exec(src);
-  if (!marker) return false;
-  const rest = src.slice(marker.index + marker[0].length).trim();
-  return (
-    rest.startsWith("[") ||
-    rest.startsWith("{") ||
-    /^<[a-zA-Z0-9_.:-]+\b/.test(rest) ||
-    (rest.startsWith("<[") || rest.startsWith("<{"))
-  );
-}
-
-/**
- * Recovers tool calls the model echoed back as text. Only a well-formed echo
- * whose every entry names a tool and carries parseable arguments is accepted;
- * anything else returns no calls and the original text.
- */
-export function parseQoderTextToolCalls(
-  text: string,
-  allowedNames?: ReadonlySet<string>,
-): { calls: QoderEchoCall[]; clean: string } {
-  if (text.length === 0) {
-    return { calls: [], clean: text };
-  }
-  const fence = text.match(/^```[a-z]*\s*\n?/i);
-  const src = fence && fence[0] ? text.slice(fence[0].length) : text;
-  const marker = LEAK_MARKER_ANY_RE.exec(src);
-  if (!marker) {
-    return { calls: [], clean: text };
-  }
-  const head = src.slice(0, marker.index).trim();
-  let rest = src.slice(marker.index + marker[0].length).trim();
-  rest = rest.replace(/\s*```\s*$/, "").trim();
-
-  let parsed: unknown;
-  try {
-    parsed = JSON.parse(rest);
-  } catch {
-    try {
-      parsed = JSON.parse(escapeControlInsideStrings(rest));
-    } catch {
-      // A bare comma-separated object list (`{...},{...}`) is not valid JSON
-      // on its own, but is exactly the echo the model emits. Wrap it in an
-      // array before falling back to salvage.
-      let wrapped: unknown;
-      try {
-        wrapped = JSON.parse(`[${rest}]`);
-      } catch {
-        wrapped = undefined;
-      }
-      const salvaged = salvageEchoArray(rest);
-      if (salvaged === null && wrapped === undefined) {
-        return { calls: [], clean: text };
-      }
-      parsed = salvaged ?? wrapped;
-    }
-  }
-  if (Array.isArray(parsed) && parsed.length === 0) return { calls: [], clean: text };
-
-  const items: unknown[] = Array.isArray(parsed) ? parsed : [parsed];
-  const calls: QoderEchoCall[] = [];
-  for (const item of items) {
-    if (typeof item !== "object" || item === null) return { calls: [], clean: text };
-    const rec = item as Record<string, unknown>;
-    const name = typeof rec["name"] === "string" ? rec["name"].trim() : "";
-    if (name.length === 0) return { calls: [], clean: text };
-    if (allowedNames && !allowedNames.has(name)) return { calls: [], clean: text };
-    const rawArgs = rec["arguments"];
-    let args: string;
-    if (typeof rawArgs === "object" && rawArgs !== null) args = JSON.stringify(rawArgs);
-    else if (typeof rawArgs === "string") args = rawArgs;
-    else return { calls: [], clean: text };
-    const argsText = args.trim();
-    if (argsText.length > 0) {
-      let accepted = argsText;
-      let ok = false;
-      try {
-        JSON.parse(accepted);
-        ok = true;
-      } catch {
-        // fall through to normalization
-      }
-      if (!ok) {
-        accepted = escapeControlInsideStrings(argsText);
-        try {
-          JSON.parse(accepted);
-          ok = true;
-        } catch {
-          // still invalid
-        }
-      }
-      if (!ok) return { calls: [], clean: text };
-      args = accepted;
-    }
-    calls.push({ id: `call_${randomUUID().replace(/-/g, "").slice(0, 16)}`, name, arguments: args });
-  }
-  return { calls, clean: head };
-}