fix(compaction): use medium reasoning for codex summarizer instead of minimal

GPT-5.6 generation models dropped the "minimal" reasoning effort
(valid levels: low/medium/high/xhigh/max), and the pi-ai catalog does
not declare minimal as null in thinkingLevelMap, so clampThinkingLevel
passes it through and the OpenAI API rejects the compaction request
with a 400, surfacing as "context compaction failed".

Hardcode medium instead: supported by both old and new OpenAI models,
still keeps long thinking from eating the summary budget.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: su-fen

crates/agent-gui/src/lib/chat/compaction/summarizer.ts

@@ -71,8 +71,9 @@ function isTransientError(error: unknown) {
 }
 
 function buildSummarizerRuntime(providerId: ProviderId, runtime: ProviderRuntimeConfig) {
-  // Codex 用最低推理档做摘要，避免长思考挤占摘要预算。
-  return providerId === "codex" ? { ...runtime, reasoning: "minimal" as const } : runtime;
+  // Codex 用 medium 档做摘要，避免长思考挤占摘要预算；不能用 minimal——
+  // GPT-5.6 世代已砍掉该档且 pi-ai 目录未标 null，clamp 不会兜底，API 会直接 400。
+  return providerId === "codex" ? { ...runtime, reasoning: "medium" as const } : runtime;
 }
 
 type SummarizerRequest = {