Fix: bound DeepSeek release-test requests

Explicitly disable thinking for the bounded JSON edit harness and give steer the same timeout/retry limits as manual edits. Keep real provider calls and all browser assertions intact.

Prepared with AI assistance under maintainer direction.

tests/live-e2e-llm-agent.test.mjs

@@ -7,6 +7,7 @@ import {
   MANUAL_EDIT_SYSTEM_INSTRUCTIONS,
   VARIANT_SYSTEM_INSTRUCTIONS,
   createLlmAgent,
+  llmRequestSettings,
   parseManualEditResponse,
   parseVariantResponse,
   progressiveVariantGuidance,
@@ -19,6 +20,17 @@ import {
   validateVariantVisibleCopy,
 } from './live-e2e/agents/llm-agent.mjs';
 
+describe('live-e2e LLM request settings', () => {
+  it('explicitly disables DeepSeek thinking for bounded JSON edit requests', () => {
+    assert.deepEqual(llmRequestSettings('deepseek'), { thinking: { type: 'disabled' } });
+  });
+
+  it('leaves other providers unchanged', () => {
+    assert.deepEqual(llmRequestSettings('anthropic'), {});
+    assert.deepEqual(llmRequestSettings('openai'), {});
+  });
+});
+
 describe('live-e2e LLM agent provider config', () => {
   it('defaults to OpenAI gpt-5.6-terra at medium reasoning effort', () => {
     const config = resolveLlmAgentConfig({}, {});

tests/live-e2e/agents/llm-agent.mjs

@@ -248,6 +248,14 @@ function resolveProvider(opts, env) {
   return 'openai';
 }
 
+export function llmRequestSettings(provider) {
+  // DeepSeek defaults to high-effort thinking, which can consume the entire
+  // bounded response before emitting the JSON these edit tests exercise.
+  // Keep real model-generated edits, but make this harness mode explicit.
+  // https://api-docs.deepseek.com/guides/thinking_mode/
+  return provider === 'deepseek' ? { thinking: { type: 'disabled' } } : {};
+}
+
 /**
  * Anthropic-SDK-shaped shim over the `ai` SDK for OpenAI models, so the
  * three text-only call sites in this file stay provider-agnostic. system
@@ -331,6 +339,7 @@ export async function createLlmAgent(opts = {}) {
         try {
           response = await client.messages.create(
             {
+              ...llmRequestSettings(provider),
               model,
               temperature: 0,
               max_tokens: 16000,
@@ -476,6 +485,7 @@ export async function createLlmAgent(opts = {}) {
         try {
           response = await client.messages.create(
             {
+              ...llmRequestSettings(provider),
               model,
               temperature: 0,
               max_tokens: 16000,
@@ -605,10 +615,14 @@ export async function createLlmAgent(opts = {}) {
       ].join('\n');
 
       const response = await client.messages.create({
+        ...llmRequestSettings(provider),
         model,
         max_tokens: 4096,
         system: systemBlocks(STEER_SYSTEM_INSTRUCTIONS),
         messages: [{ role: 'user', content: userMessage }],
+      }, {
+        maxRetries: LLM_REQUEST_MAX_RETRIES,
+        timeout: MANUAL_EDIT_REQUEST_TIMEOUT_MS,
       });
 
       const cacheRead = response.usage?.cache_read_input_tokens ?? 0;