Run GLM-5.2 for the bulk, escalate the hard turns to Opus 4.8

The Stack

Steps

1. 1

   Write the cost-routing config and validate it

   Set GLM-5.2 as the default on an OpenAI-compatible host (OpenRouter id z-ai/glm-5.2, or Z.ai direct) and escalate to Opus 4.8 only on hard turns. CI checks the config shape; pointing it at real traffic needs a key and is fenced.

   cat > routing.json <<'JSON'
   {
     "base_url": "https://openrouter.ai/api/v1",
     "default_model": "z-ai/glm-5.2",
     "escalate_model": "anthropic/claude-opus-4.8",
     "escalate_when": "task_difficulty == hard || category == general_coding"
   }
   JSON
   node -e '
   const fs = require("fs");
   const c = JSON.parse(fs.readFileSync("routing.json", "utf8"));
   function bad(m) { console.error("BAD: " + m); process.exit(1); }
   const host = String(c.base_url || "");
   if (!host.includes("openrouter.ai") && !host.includes("z.ai")) bad("base_url is not an OpenAI-compatible GLM host (openrouter.ai or z.ai)");
   const def = c.default_model || "";
   if (!def.includes("glm-5.2")) bad("default_model is not GLM-5.2");
   const esc = c.escalate_model || "";
   if (!esc.includes("opus-4.8")) bad("escalate_model is not Opus 4.8");
   const when = String(c.escalate_when || "");
   if (when.length < 3) bad("escalate_when must gate escalation, not match everything");
   const where = host.includes("openrouter.ai") ? "openrouter.ai" : "z.ai";
   console.log("config OK: bulk -> " + def + " via " + where + ", escalate to " + esc + " on hard turns");
   '

2. 2

   Point it at your work and measure cost per task (the model step, not checked by CI)

   Add your key and run real prompts. Watch cost per finished task, not per token: if GLM-5.2 lands your bulk work in the same number of turns, the savings are real; if it needs extra loops, the gap shrinks. Keep Opus for the short, hard, general-coding turns. The runs are fenced.