fix(providers): stop zed-hosted claude-haiku-4-5 thinking from inflating max_tokens (#13364) (#13780)

Merged in the 2026-09-16 sweep of the maintainer's own open PRs, at the owner's explicit instruction. No push was made to the PR branch: the merge took the head as the owning session left it (verified OPEN, non-draft and MERGEABLE against the release tip immediately before merging).

changelog.d/fixes/13364-zed-hosted-haiku-thinking-inflation.md

@@ -0,0 +1 @@
+- **fix(providers):** stop zed-hosted `claude-haiku-4-5` extended-thinking requests from inflating `max_tokens` past the model's real 64000 output cap (#13364) — thanks @ThiagoMafra-Integrare

open-sse/services/model.ts

@@ -73,6 +73,14 @@ const PROVIDER_MODEL_ALIASES: ProviderModelAliasMap = {
     "claude-sonnet-4-5": "claude-sonnet-4.5",
     "claude-haiku-4-5": "claude-haiku-4.5",
   },
+  // #13364: zed-hosted's passthrough catalog exposes short hyphenated Claude ids
+  // that don't match modelSpecs' dotted canonical alias, so capMaxOutputTokens()
+  // resolves no cap and thinking+tools requests inflate max_tokens unbounded.
+  // Scoped to claude-haiku-4-5 (the reported/reproduced model) — add Sonnet/Opus
+  // entries only once confirmed against the live Zed catalog.
+  "zed-hosted": {
+    "claude-haiku-4-5": "claude-haiku-4.5",
+  },
 };
 
 const CROSS_PROXY_MODEL_ALIASES: Record<string, string> = {

tests/unit/issue-13364-zed-hosted-haiku-thinking-inflation.test.ts

@@ -0,0 +1,39 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { openaiToClaudeRequest } from "../../open-sse/translator/request/openai-to-claude.ts";
+
+// #13364 — zed-hosted/claude-haiku-4-5: extended thinking + tools inflates
+// max_tokens to 163072 (> Anthropic's real 64000 cap for this model), causing
+// Zed's proxy to reject the request in-stream with:
+//   "max_tokens: 163072 > 64000, which is the maximum allowed number of
+//    output tokens for claude-haiku-4-5-20251001"
+//
+// Root cause: zed-hosted's passthrough catalog exposes the short hyphenated id
+// "claude-haiku-4-5", but modelSpecs' registered alias for that model is the
+// dotted "claude-haiku-4.5" — a spelling mismatch (not a missing spec) means
+// capMaxOutputTokens() resolves no cap for zed-hosted, so fitThinkingToMaxTokens
+// adds the requested budget instead of fitting it under the model's real ceiling.
+test("zed-hosted/claude-haiku-4-5 thinking+tools must not inflate max_tokens past the real 64000 output cap", () => {
+  const body = {
+    model: "claude-haiku-4-5",
+    max_tokens: 16000,
+    reasoning_effort: "high", // what claude-to-openai produces for the client's budget_tokens:15999
+    tools: [
+      {
+        type: "function",
+        function: { name: "read_file", parameters: { type: "object", properties: {} } },
+      },
+    ],
+    messages: [{ role: "user", content: "hello" }],
+  };
+  const credentials = { _provider: "zed-hosted" };
+
+  const result = openaiToClaudeRequest("claude-haiku-4-5", body, false, credentials);
+  const claudeHaiku45OutputCap = 64000; // src/shared/constants/modelSpecs.ts "claude-haiku-4-5-20251001"
+
+  assert.ok(
+    (result.max_tokens as number) <= claudeHaiku45OutputCap,
+    `max_tokens (${result.max_tokens}) must not exceed the model's real output cap ` +
+      `(${claudeHaiku45OutputCap}) — Anthropic/Zed rejects the request otherwise`
+  );
+});