fix(sse): do not 502 Claude Code /model probes (content:[] + max_tokens) (#11568)

Merged via /merge-batch (lote 2026-08-26, v3.8.51). Boarded no worktree combinado junto com outras ~30 PRs; validação única: typecheck/complexity/cognitive-complexity/changelog-integrity verdes, file-size rebaseado onde necessário (crescimento legítimo), lint com os mesmos 228 achados pré-existentes confirmados via sonda contra o tip puro (não introduzidos por este lote), e ~370 testes focados (unit + vitest) passando. Obrigado pela contribuição.

changelog.d/fixes/claude-code-model-probe-max-tokens.md

@@ -0,0 +1 @@
+- fix(sse): accept Claude `content:[]` + `stop_reason: max_tokens`/`tool_use` in `detectMalformedNonStream` (match `isEmptyContentResponse`) so Claude Code `/model` probes with `max_tokens: 1` no longer become a false 502

open-sse/utils/diagnostics.ts

@@ -256,18 +256,20 @@ export function detectMalformedNonStream(resp: unknown): MalformedReason | null
     //     sentinel, or only null entries) — the model genuinely produced no
     //     usable output. That is a MALFORMED-200 empty_choices regardless of
     //     stop_reason (parity with the OpenAI content:"" path).
-    //  2) `content: []` — no block at all. Only a genuinely *terminal* response
-    //     (a final stop_reason with no output) is empty_choices. #9971: a
-    //     truncated / non-terminal body — the Claude Code OAuth upstream cutting
-    //     a long generation mid-turn, or a content-less thinking-only stream
-    //     that never emitted a terminal event — carries content:[] with no
-    //     reachable end, so flagging it would turn an upstream truncation into a
-    //     false 502. Require a terminal stop_reason before calling a block-less
-    //     response genuinely empty.
+    //  2) `content: []` — no block at all. #9971: a truncated / non-terminal
+    //     body (no stop_reason) must not become empty_choices. A terminal
+    //     stop_reason with no output usually is empty_choices — except the
+    //     same legitimate empty stops that `isEmptyContentResponse` already
+    //     accepts (`max_tokens`, `tool_use`). Claude Code's `/model` probe
+    //     sends `max_tokens: 1`; Opus can burn that budget on thinking and
+    //     return content:[] + stop_reason max_tokens. Treating that as
+    //     empty_choices turns a valid 200 into MALFORMED-200 → 502 even
+    //     though errorClassifier would have let it through.
     if (content.length === 0) {
       const stopReason = typeof body.stop_reason === "string" ? body.stop_reason : "";
-      const isTerminal = stopReason.length > 0;
-      return isTerminal ? "empty_choices" : null;
+      if (stopReason.length === 0) return null;
+      if (stopReason === "max_tokens" || stopReason === "tool_use") return null;
+      return "empty_choices";
     }
     return "empty_choices";
   }

tests/unit/diagnostics-claude-thinking-5108.test.ts

@@ -51,6 +51,14 @@ test("#5108 genuinely empty Claude content:[] is still flagged malformed", () =>
   assert.equal(detectMalformedNonStream(claudeMsg([])), "empty_choices");
 });
 
+test("Claude Code /model probe: content:[] + max_tokens is not empty_choices", () => {
+  const body = {
+    ...claudeMsg([]),
+    stop_reason: "max_tokens",
+  };
+  assert.equal(detectMalformedNonStream(body), null);
+});
+
 test("#5108/#9971 Claude thinking block with neither text nor signature is valid output (was empty_choices)", () => {
   const body = claudeMsg([{ type: "thinking", thinking: "", signature: "" }]);
   // #9971: an empty thinking block is valid structural output — the upstream can

tests/unit/diagnostics.test.ts

@@ -293,6 +293,30 @@ test("detectMalformedNonStream returns 'empty_choices' for Claude message with e
   assert.equal(detectMalformedNonStream(body), "empty_choices");
 });
 
+test("detectMalformedNonStream returns null for Claude content:[] with stop_reason max_tokens (Claude Code /model probe)", () => {
+  // Claude Code probes model switches with Hi + max_tokens:1. Opus can burn the
+  // token on thinking and return an empty content array with max_tokens — valid
+  // upstream 200, must not become empty_choices / 502.
+  const body = {
+    type: "message",
+    role: "assistant",
+    content: [],
+    stop_reason: "max_tokens",
+    usage: { input_tokens: 33, output_tokens: 1 },
+  };
+  assert.equal(detectMalformedNonStream(body), null);
+});
+
+test("detectMalformedNonStream returns null for Claude content:[] with no stop_reason (#9971 non-terminal)", () => {
+  const body = { type: "message", role: "assistant", content: [] };
+  assert.equal(detectMalformedNonStream(body), null);
+});
+
+test("detectMalformedNonStream returns null for Claude content:[] with stop_reason tool_use (parity with errorClassifier)", () => {
+  const body = { type: "message", role: "assistant", content: [], stop_reason: "tool_use" };
+  assert.equal(detectMalformedNonStream(body), null);
+});
+
 test("detectMalformedNonStream returns 'empty_choices' for Claude message with empty-text block", () => {
   const body = {
     type: "message",