fix(azure): match the generation, not one release, for max_completion_tokens (#13007)

Boarded with 13 sibling PRs into one worktree off release/v3.8.51 and validated as a set: 132 focused tests pass across all 15 test files in the batch, typecheck:core is clean, check-changelog-integrity reports no lost base bullets, and check-file-size is green. Your PR merged without conflict against its siblings.

Thank you — the write-up made this reviewable: measuring the behaviour on the release tip and showing the before/after table meant the defect could be confirmed rather than taken on faith.

changelog.d/fixes/12981-azure-generation-range.md

@@ -0,0 +1 @@
+- **fix(azure):** Deployments from GPT-6 onward now send `max_completion_tokens` instead of `max_tokens`, which Azure rejects with HTTP 400. The rule matched a literal `gpt-5`, so each new generation arrived broken; it now matches the generation range, while `gpt-35-turbo` still keeps `max_tokens`.

open-sse/executors/azureParamRules.ts

@@ -20,15 +20,24 @@
 /**
  * Deployments that require `max_completion_tokens` instead of `max_tokens`.
  *
- * Matches the GPT-5 family and the o1/o3/o4 reasoning series at a token
+ * Matches GPT-5 and later, and the o1/o3/o4 reasoning series, at a token
  * boundary, so a deployment named `my-gpt-5-prod` matches while an unrelated
  * `piston-o4-legacy`-style name does not match by accident. `gpt-chat-latest`
  * is listed explicitly: it is a moving alias that currently resolves to a
  * GPT-5-era model and rejects `max_tokens`, but carries no version number for
  * the boundary pattern to key on.
+ *
+ * The generation is a range rather than a literal `gpt-5`, because the rule is
+ * a property of the generation and not of one release: `gpt-6-astra` rejects
+ * `max_tokens` for exactly the reason `gpt-5` does, and pinning the literal
+ * meant every new family arrived broken (#12981).
+ *
+ * It is a range and not `\d+` on purpose. Azure's own name for GPT-3.5 is
+ * `gpt-35-turbo`, which takes `max_tokens` and would be caught by a digit-run.
+ * `1\d` keeps a future `gpt-10` working without letting `gpt-35` in.
  */
 export const AZURE_COMPLETION_TOKEN_DEPLOYMENT =
-  /(?:^|[/_-])(?:gpt-5|o(?:1|3|4))(?:[._-]|$)|^gpt-chat-latest$/i;
+  /(?:^|[/_-])(?:gpt-(?:[5-9]|1\d)|o(?:1|3|4))(?:[._-]|$)|^gpt-chat-latest$/i;
 
 /**
  * Apply the Azure param rules to an already-translated Chat Completions body.

tests/unit/azure-param-rules.test.ts

@@ -46,6 +46,37 @@ test("gpt-5 family converts max_tokens too", () => {
   }
 });
 
+test("generations after GPT-5 convert max_tokens too (#12981)", () => {
+  // The rule belongs to the generation, not to one release. gpt-6-astra is the
+  // deployment from the report; the rest are the next names Azure will use.
+  for (const model of ["gpt-6-astra", "gpt-6", "azure/gpt-7-mini", "gpt-9.1", "gpt-10-turbo"]) {
+    const out = applyAzureParamRules(model, { max_tokens: 100 }, { max_tokens: 100 }) as Record<
+      string,
+      unknown
+    >;
+    assert.equal(out.max_tokens, undefined, `${model} should drop max_tokens`);
+    assert.equal(out.max_completion_tokens, 100, `${model} should set max_completion_tokens`);
+  }
+});
+
+test("gpt-35-turbo is not a GPT-3.5 deployment caught by the generation range", () => {
+  // Azure's own name for GPT-3.5 has no dot, so a digit-run like `gpt-\d+`
+  // would match it and strip the max_tokens it actually requires. This is why
+  // the pattern is a range and stops at 19.
+  for (const model of ["gpt-35-turbo", "gpt-35-turbo-16k", "azure/gpt-35"]) {
+    assert.equal(
+      AZURE_COMPLETION_TOKEN_DEPLOYMENT.test(model),
+      false,
+      `${model} must keep max_tokens`
+    );
+    const out = applyAzureParamRules(model, { max_tokens: 100 }, { max_tokens: 100 }) as Record<
+      string,
+      unknown
+    >;
+    assert.equal(out.max_tokens, 100, `${model} should pass through untouched`);
+  }
+});
+
 test("reasoning_effort is dropped when tools are present", () => {
   const out = applyAzureParamRules(
     "gpt-5.1",