fix(sse): give extended-thinking targets the reasoning readiness budget

open-sse/utils/streamReadinessPolicy.ts

@@ -95,6 +95,27 @@ function isHighReasoningEffort(
   return effort.toLowerCase() === "high";
 }
 
+/**
+ * Extended-thinking aliases (`claude-sonnet-5-thinking`, `claude-opus-4-6-thinking-1m`,
+ * `gpt-5.6-luna-free-thinking`, ...) run the same cold reasoning warm-up that earns
+ * codex high-effort targets their unconditional bump — the client asked for extended
+ * thinking, so the first non-ping event is gated behind it regardless of prompt size.
+ *
+ * This is keyed on the model alias rather than the provider's registry `format`
+ * because the providers that serve these models translate them themselves:
+ * kiro (`format: "kiro"`, Anthropic models over CodeWhisperer), devin, antigravity
+ * and chatgpt-web all publish `-thinking` ids while sitting outside the
+ * claude-format bump. #11922 was exactly that gap — `kiro/claude-sonnet-5-thinking`
+ * 504'd at the 125s readiness window (80s base + 45s large-history) because nothing
+ * in this policy recognised the request as a reasoning target.
+ */
+function isExtendedThinkingModel(model?: string | null): boolean {
+  if (!model) return false;
+  // Matches `-thinking` at the end and before a further qualifier (`-thinking-1m`),
+  // without matching unrelated ids that merely contain the word.
+  return /-thinking(?:-|$)/.test(model.toLowerCase());
+}
+
 export function resolveStreamReadinessTimeout(
   input: StreamReadinessPolicyInput
 ): StreamReadinessPolicyResult {
@@ -114,6 +135,7 @@ export function resolveStreamReadinessTimeout(
   const estimatedChars = estimateBodyChars(input.body);
   const codexGpt5x = isCodexGpt5x(input.provider, input.model);
   const codexHighReasoning = codexGpt5x && isHighReasoningEffort(input.model, input.body);
+  const extendedThinking = isExtendedThinkingModel(input.model);
 
   if (itemCount > VERY_LARGE_ITEM_THRESHOLD) {
     timeoutMs += 45_000;
@@ -157,7 +179,17 @@ export function resolveStreamReadinessTimeout(
   // first SSE event — enough that the default 80s readiness window 504s before
   // the upstream speaks. Mirror the codex_gpt_5_5_high_reasoning bump so this
   // class of provider cannot be misidentified as a stalled connection.
-  if (isClaudeFormatReasoningProvider(input.provider) && !codexHighReasoning) {
+  // #11922: an explicitly-requested extended-thinking target warms up before it
+  // emits anything, on whichever provider serves it. Mirror the codex high-effort
+  // bump so the readiness watchdog cannot mistake that warm-up for a stalled
+  // stream. Kept exclusive with the other reasoning bumps below — these all model
+  // the same one-off warm-up, so they must not stack into a multi-minute window.
+  if (extendedThinking && !codexHighReasoning) {
+    timeoutMs += 30_000;
+    reasons.push("extended_thinking");
+  }
+
+  if (isClaudeFormatReasoningProvider(input.provider) && !codexHighReasoning && !extendedThinking) {
     timeoutMs += 30_000;
     reasons.push("claude_format_heavy_reasoning");
   }

tests/unit/stream-readiness-policy.test.ts

@@ -163,7 +163,7 @@ test("bumps small requests to third-party Claude-format replicas (agentrouter, Z
   );
 });
 
-test("does NOT bump Minimax (M3) — #3110 moved it from claude to openai format so images work, and the readiness bump is keyed off the registry's `format: \"claude\"` field", () => {
+test('does NOT bump Minimax (M3) — #3110 moved it from claude to openai format so images work, and the readiness bump is keyed off the registry\'s `format: "claude"` field', () => {
   // Minimax's replica quirk (long reasoning warm-up) hasn't changed, but this
   // policy intentionally keys off the translator format, not the provider
   // name — the registry is the single source of truth (see isClaudeFormatReasoningProvider
@@ -266,3 +266,74 @@ test("treats unknown provider names as non-Claude-format (no false positives)",
   assert.equal(result.timeoutMs, 80_000);
   assert.ok(!result.reasons.includes("claude_format_heavy_reasoning"));
 });
+
+test("gives extended-thinking model aliases the reasoning readiness bump (#11922)", () => {
+  // #11922: kiro/claude-sonnet-5-thinking 504'd with
+  // "Stream produced no non-ping SSE event within 125000ms" — the 80s base plus
+  // the 45s very-large-history bump, capped there because nothing recognised the
+  // request as a reasoning target. Kiro serves Anthropic thinking models through
+  // its own CodeWhisperer translator, so `format: "kiro"` (not "claude") kept it
+  // out of the claude_format_heavy_reasoning bump, and the `-thinking` alias was
+  // never a reasoning signal the way `-high` is.
+  const result = resolveStreamReadinessTimeout({
+    baseTimeoutMs: 80_000,
+    provider: "kiro",
+    model: "claude-sonnet-5-thinking",
+    body: { messages: items(401) },
+  });
+
+  assert.equal(result.timeoutMs, 155_000);
+  assert.ok(result.reasons.includes("extended_thinking"));
+});
+
+test("extended-thinking bump is provider-agnostic and fires for small requests", () => {
+  const result = resolveStreamReadinessTimeout({
+    baseTimeoutMs: 80_000,
+    provider: "devin",
+    model: "claude-opus-4-6-thinking",
+    body: { messages: items(2) },
+  });
+
+  assert.equal(result.timeoutMs, 110_000);
+  assert.ok(result.reasons.includes("extended_thinking"));
+});
+
+test("does NOT stack extended-thinking with the Claude-format replica bump", () => {
+  // Both bumps model the same one-off reasoning warm-up. A claude-format replica
+  // serving a `-thinking` alias must get 30s once, not 60s twice.
+  const result = resolveStreamReadinessTimeout({
+    baseTimeoutMs: 80_000,
+    provider: "agentrouter",
+    model: "claude-sonnet-4-6-thinking",
+    body: { messages: items(2) },
+  });
+
+  assert.equal(result.timeoutMs, 110_000);
+  assert.ok(result.reasons.includes("extended_thinking"));
+  assert.ok(!result.reasons.includes("claude_format_heavy_reasoning"));
+});
+
+test("does NOT stack extended-thinking with the codex high-reasoning bump", () => {
+  const result = resolveStreamReadinessTimeout({
+    baseTimeoutMs: 80_000,
+    provider: "codex",
+    model: "gpt-5.5-thinking",
+    body: { messages: items(2), reasoning_effort: "high" },
+  });
+
+  assert.equal(result.timeoutMs, 110_000);
+  assert.ok(result.reasons.includes("codex_gpt_5_5_high_reasoning"));
+  assert.ok(!result.reasons.includes("extended_thinking"));
+});
+
+test("does not treat an unrelated id containing 'thinking' as an alias suffix", () => {
+  const result = resolveStreamReadinessTimeout({
+    baseTimeoutMs: 80_000,
+    provider: "openai",
+    model: "thinking-machines-lab-model",
+    body: { messages: items(2) },
+  });
+
+  assert.equal(result.timeoutMs, 80_000);
+  assert.ok(!result.reasons.includes("extended_thinking"));
+});