fix(translator): sanitize Read tool args from non-Anthropic models

CHANGELOG.md

@@ -15,6 +15,7 @@ _In development — bullets added per PR; finalized at release._
 ### 🐛 Fixed
 
 - **fix(embeddings):** forward output dimensions to Gemini for consistent embedding dims. (thanks @nguyenha935)
+- **fix(translator):** sanitize Read tool args from non-Anthropic models to prevent retry loops. (thanks @GodrezJr2)
 - **fix(combo): round-robin members fail over faster under concurrency saturation via a configurable queue depth** — when a round-robin combo member was saturated, requests sat in the per-model semaphore's **unbounded** queue and only failed over to the next member after the full `queueTimeoutMs` (default 30s) elapsed — so a burst of agentic requests deep-queued one hot member instead of spilling to healthy ones. The per-model semaphore now accepts a bounded queue depth and emits `SEMAPHORE_QUEUE_FULL` once it is full (the round-robin loop already cascades on that code), so a configured low depth fails over immediately. A new `queueDepth` combo-config knob (global default / provider override / per-combo, default **20** for backward compatibility; **0** = never queue → fail over now) is exposed in Settings → Combo Defaults. ([#3872](https://github.com/diegosouzapw/OmniRoute/issues/3872) — thanks @KooshaPari)
 
 ---

open-sse/translator/helpers/toolCallShim.ts

@@ -27,16 +27,51 @@ function coerceToArray(v: unknown): unknown[] {
   return [];
 }
 
+// Claude Code's Read tool caps `limit` at 2000 lines per call. Non-Anthropic models
+// (GPT-5.5, DeepSeek …) occasionally emit absurd values (e.g. `limit: 25999999999999999`)
+// that Claude Code rejects, causing a retry loop that wastes tokens. Clamp here.
+const READ_MAX_LIMIT = 2000;
+
+// `pages` is only meaningful for PDFs and only as `"N"` or `"N-M"` (1-based).
+// Reference: claude-code-tools docs + upstream decolua/9router#1144.
+function isValidPdfPagesArg(filePath: unknown, pages: unknown): boolean {
+  return (
+    typeof filePath === "string" &&
+    filePath.toLowerCase().endsWith(".pdf") &&
+    typeof pages === "string" &&
+    /^\d+(?:-\d+)?$/.test(pages)
+  );
+}
+
+function sanitizeReadArgs(args: Record<string, unknown>): void {
+  // Coerce numeric-string limit/offset (some non-Anthropic models stringify everything).
+  if (typeof args.limit === "string" && /^\d+$/.test(args.limit)) {
+    args.limit = Number(args.limit);
+  }
+  if (typeof args.offset === "string" && /^-?\d+$/.test(args.offset)) {
+    args.offset = Number(args.offset);
+  }
+
+  if (typeof args.limit === "number") {
+    if (args.limit > READ_MAX_LIMIT) args.limit = READ_MAX_LIMIT;
+    if (args.limit < 1) delete args.limit;
+  }
+  if (typeof args.offset === "number" && args.offset < 0) args.offset = 0;
+
+  if ("pages" in args && !isValidPdfPagesArg(args.file_path, args.pages)) {
+    delete args.pages;
+  }
+}
+
 const TOOL_SHIMS: Record<string, ShimFn> = {
-  // Claude Code Read accepts `pages` only for PDFs and rejects an empty string.
-  // Some non-Anthropic models emit optional `pages: ""` for ordinary files.
-  // Buffer and emit one cleaned JSON delta so the client never sees the bad field.
+  // Claude Code Read rejects bad params and retries — wasting tokens with non-Anthropic
+  // models that emit oversized limits, negative offsets, stringified numbers, or stray
+  // `pages` on non-PDF files. Buffer and emit one cleaned JSON delta so the client never
+  // sees the bad fields. See `sanitizeReadArgs` for the per-field rules.
   Read: (input) => {
     if (typeof input !== "object" || input === null || Array.isArray(input)) return input;
     const patched = { ...(input as Record<string, unknown>) };
-    if (patched.pages === "" || (Array.isArray(patched.pages) && patched.pages.length === 0)) {
-      delete patched.pages;
-    }
+    sanitizeReadArgs(patched);
     return patched;
   },
   submit_pr_review: (input) => {

tests/unit/translator-tool-call-shim.test.ts

@@ -75,6 +75,96 @@ test("applyToolCallShimToBuffer: Read removes empty pages but preserves valid ra
   assert.deepEqual(withValidPages, { file_path: "/tmp/a.pdf", pages: "1-5" });
 });
 
+// Port of decolua/9router#1144: non-Anthropic models (GPT-5.5, DeepSeek …) sometimes
+// emit absurd Read-tool args (e.g. limit: 99999999999) that Claude Code rejects and
+// retries, wasting tokens. The shim clamps/normalizes those args before re-emitting.
+test("applyToolCallShimToBuffer: Read clamps limit to 2000 (non-Anthropic models)", () => {
+  const out = JSON.parse(
+    applyToolCallShimToBuffer(
+      "Read",
+      JSON.stringify({ file_path: "/etc/hosts", limit: 99999999999 })
+    )
+  );
+  assert.equal(out.limit, 2000);
+});
+
+test("applyToolCallShimToBuffer: Read drops non-positive limit", () => {
+  const zero = JSON.parse(
+    applyToolCallShimToBuffer("Read", JSON.stringify({ file_path: "/etc/hosts", limit: 0 }))
+  );
+  assert.equal("limit" in zero, false);
+
+  const negative = JSON.parse(
+    applyToolCallShimToBuffer("Read", JSON.stringify({ file_path: "/etc/hosts", limit: -50 }))
+  );
+  assert.equal("limit" in negative, false);
+});
+
+test("applyToolCallShimToBuffer: Read clamps negative offset to 0", () => {
+  const out = JSON.parse(
+    applyToolCallShimToBuffer("Read", JSON.stringify({ file_path: "/etc/hosts", offset: -5 }))
+  );
+  assert.equal(out.offset, 0);
+});
+
+test("applyToolCallShimToBuffer: Read coerces numeric-string limit/offset", () => {
+  const out = JSON.parse(
+    applyToolCallShimToBuffer(
+      "Read",
+      JSON.stringify({ file_path: "/etc/hosts", limit: "100", offset: "5" })
+    )
+  );
+  assert.equal(out.limit, 100);
+  assert.equal(out.offset, 5);
+});
+
+test("applyToolCallShimToBuffer: Read strips pages for non-PDF files", () => {
+  const out = JSON.parse(
+    applyToolCallShimToBuffer(
+      "Read",
+      JSON.stringify({ file_path: "/etc/hosts", pages: "1-3" })
+    )
+  );
+  assert.equal("pages" in out, false);
+});
+
+test("applyToolCallShimToBuffer: Read strips malformed pages even on PDFs", () => {
+  const out = JSON.parse(
+    applyToolCallShimToBuffer(
+      "Read",
+      JSON.stringify({ file_path: "/tmp/doc.pdf", pages: "abc" })
+    )
+  );
+  assert.equal("pages" in out, false);
+});
+
+test("applyToolCallShimToBuffer: Read accepts a single page on PDFs", () => {
+  const out = JSON.parse(
+    applyToolCallShimToBuffer(
+      "Read",
+      JSON.stringify({ file_path: "/tmp/doc.PDF", pages: "7" })
+    )
+  );
+  assert.equal(out.pages, "7");
+});
+
+test("applyToolCallShimToBuffer: Read combined absurd args from non-Anthropic model", () => {
+  // Simulates the upstream issue exactly: GPT-5.5-style giant limit, negative offset,
+  // and a stray empty-string pages on a non-PDF.
+  const out = JSON.parse(
+    applyToolCallShimToBuffer(
+      "Read",
+      JSON.stringify({
+        file_path: "F:/repo/file.js",
+        offset: -5,
+        limit: 25999999999999999,
+        pages: "",
+      })
+    )
+  );
+  assert.deepEqual(out, { file_path: "F:/repo/file.js", offset: 0, limit: 2000 });
+});
+
 test("applyToolCallShimToBuffer: submit_pr_review with valid arrays preserved", () => {
   const raw = JSON.stringify({
     summary: "ok",