fix(claude): reconcile max_tokens vs thinking budget and lift per-model ceiling (#2381)

On the translated OpenAI->Claude path, adjustMaxTokens capped max_tokens
before applyThinking set thinking.budget_tokens, so max-effort budget
(128000) could exceed a 64k-clamped max_tokens -> Anthropic 400.
prepareClaudeRequest now reconciles after the budget is known: prefer
raising max_tokens, only shrink budget when it meets/exceeds the ceiling.

Also lift the global 64000 cap: the ceiling is now the model's real
maxOutput, so high-output models (fable/mythos, opus-4.8/sonnet-4.6) get
their full budget. adjustMaxTokens gains an optional ceiling arg (default
unchanged, callers untouched); openai-to-claude passes the model maxOutput.

Native Claude Code passthrough is unaffected.

Co-Authored-By: Claude <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

tests/translator/bugs-toClaude-context.test.js

@@ -67,6 +67,101 @@ describe("OpenAI → Claude context mapping", () => {
     expect(JSON.stringify(out), "remote image dropped").toContain("pic.png");
   });
 
+  // prepareClaudeRequest reconciles max_tokens vs thinking.budget_tokens.
+  // applyThinking runs after adjustMaxTokens caps max_tokens, so a claude-budget
+  // model at "max" effort (budget 128000) can exceed the clamped max_tokens and
+  // trip Anthropic's "max_tokens > budget_tokens" rule (400). See claude.js.
+  describe("max_tokens vs thinking.budget_tokens reconciliation", () => {
+    // 64k-ceiling model (maxOutput 64000) + max-effort budget 128000: budget alone
+    // exceeds the ceiling → cap max_tokens at 64000 and shrink budget below it.
+    it("max effort budget on a 64k model → budget < max_tokens ≤ 64000", () => {
+      const out = prepareClaudeRequest({
+        model: "claude-opus-4-20250514",
+        max_tokens: 64000,
+        thinking: { type: "enabled", budget_tokens: 128000 },
+        messages: [{ role: "user", content: "q" }],
+      }, "anthropic");
+      expect(out.max_tokens).toBe(64000);
+      expect(out.thinking.budget_tokens).toBeLessThan(out.max_tokens);
+      expect(out.thinking.budget_tokens).toBeGreaterThan(0);
+    });
+
+    // Budget fits under the ceiling but exceeds a small client max_tokens →
+    // raise max_tokens to fit, preserving the requested thinking depth.
+    it("xhigh budget with a low client max_tokens → raise max_tokens, preserve budget", () => {
+      const out = prepareClaudeRequest({
+        model: "claude-opus-4-20250514",
+        max_tokens: 16000,
+        thinking: { type: "enabled", budget_tokens: 32768 },
+        messages: [{ role: "user", content: "q" }],
+      }, "anthropic");
+      expect(out.thinking.budget_tokens).toBe(32768);
+      expect(out.max_tokens).toBe(33792); // 32768 + 1024, under the 64000 ceiling
+    });
+
+    // Budget already below max_tokens → nothing to reconcile.
+    it("high budget under max_tokens → both unchanged", () => {
+      const out = prepareClaudeRequest({
+        model: "claude-opus-4-20250514",
+        max_tokens: 64000,
+        thinking: { type: "enabled", budget_tokens: 24576 },
+        messages: [{ role: "user", content: "q" }],
+      }, "anthropic");
+      expect(out.max_tokens).toBe(64000);
+      expect(out.thinking.budget_tokens).toBe(24576);
+    });
+
+    // Non-budget thinking shapes (adaptive / disabled) carry no budget_tokens →
+    // the reconciliation must never touch them.
+    it("adaptive thinking (no budget_tokens) is left untouched", () => {
+      const out = prepareClaudeRequest({
+        model: "claude-opus-4-20250514",
+        max_tokens: 64000,
+        thinking: { type: "adaptive" },
+        messages: [{ role: "user", content: "q" }],
+      }, "anthropic");
+      expect(out.max_tokens).toBe(64000);
+      expect(out.thinking).toEqual({ type: "adaptive" });
+    });
+
+    // Lifted ceiling: a claude-budget model whose caps declare maxOutput 128000
+    // (e.g. fable) may use the full budget at max effort instead of being pinned
+    // to the conservative 64000 default.
+    it("max effort budget on a 128k model → max_tokens up to 128000, budget preserved just under", () => {
+      const out = prepareClaudeRequest({
+        model: "claude-fable-5",
+        max_tokens: 64000,
+        thinking: { type: "enabled", budget_tokens: 128000 },
+        messages: [{ role: "user", content: "q" }],
+      }, "anthropic");
+      expect(out.max_tokens).toBe(128000);
+      expect(out.thinking.budget_tokens).toBe(126976); // 128000 - 1024
+      expect(out.thinking.budget_tokens).toBeLessThan(out.max_tokens);
+    });
+
+    // Regression: a default 64k-ceiling model still clamps an over-large client
+    // max_tokens down to 64000 (the lift is per-model, not global).
+    it("over-large client max_tokens on a 64k model is still clamped to 64000", () => {
+      const out = prepareClaudeRequest({
+        model: "claude-opus-4-20250514",
+        max_tokens: 120000,
+        messages: [{ role: "user", content: "q" }],
+      }, "anthropic");
+      expect(out.max_tokens).toBe(64000);
+    });
+
+    // Lifted ceiling for a 128k model: a large client max_tokens is now allowed
+    // through instead of being clamped to 64000.
+    it("large client max_tokens on a 128k model is allowed up to maxOutput", () => {
+      const out = prepareClaudeRequest({
+        model: "claude-fable-5",
+        max_tokens: 100000,
+        messages: [{ role: "user", content: "q" }],
+      }, "anthropic");
+      expect(out.max_tokens).toBe(100000);
+    });
+  });
+
   it("DeepSeek Claude transport adds a thinking placeholder before tool_use in thinking mode", () => {
     const out = prepareClaudeRequest({
       model: "deepseek-v4-pro",