Switch chat agents to GPT-5.6 Luna (max reasoning) (#499)

src/server/features/onboarding/OnboardingChatAgent.ts

@@ -171,11 +171,12 @@ export class OnboardingChatAgent extends AIChatAgent {
       // Cancel the (billable) LLM call if the user aborts/navigates away.
       abortSignal: options?.abortSignal,
       // Budget shared by reasoning + visible output. Reasoning tokens (enabled
-      // on the model) eat into this, so it's well above what the ~350-word
+      // on the model) eat into this, and at max effort they dwarf the ~350-word
-      // strategy needs — otherwise the answer truncates mid-table once the
+      // strategy — the old 4000 cap left almost no answer headroom and
-      // model has spent the budget thinking. It's a ceiling, not a target: the
+      // truncated mid-table once the model had spent the budget thinking.
-      // model only generates (and we only bill) what it actually uses.
+      // Deliberately roomy: it's a per-step ceiling, not a target — the model
-      maxOutputTokens: 4000,
+      // only generates (and we only bill) what it actually uses.
+      maxOutputTokens: 32_000,
       stopWhen: stepCountIs(5),
       // Meter LLM spend against the same credit pool as DataForSEO: sum the real
       // per-step cost OpenRouter reports and deduct it. Best-effort, hosted-only.

src/server/features/sam/SamChatAgent.ts

@@ -302,9 +302,13 @@ export class SamChatAgent extends Think {
         // SAM is meant to run complex multi-step work in one turn (site-read
         // intake plus a full research chain, multi-competitor sweeps), so give
         // it generous headroom — cost is bounded by per-step metering and the
-        // model stopping on its own, not by this cap.
+        // model stopping on its own, not by this cap. The per-step budget is
+        // shared by max-effort reasoning + visible output; a tight cap risks
+        // reasoning eating the reply (the issue #161 failure mode), so it's
+        // deliberately roomy — ~10x measured reasoning use — while keeping the
+        // worst-case turn (48 steps at the full cap) under ~$2.
         maxSteps: 48,
-        maxOutputTokens: 6000,
+        maxOutputTokens: 32_000,
       };
     });
   }

src/server/lib/openrouter.ts

@@ -9,30 +9,12 @@ import {
 
 // OpenRouter model slug used for the in-app chat agents (onboarding + SAM).
 // Override with OPENROUTER_MODEL to swap models without a code change.
-const DEFAULT_CHAT_AGENT_MODEL = "minimax/minimax-m3";
+const DEFAULT_CHAT_AGENT_MODEL = "openai/gpt-5.6-luna";
+
+// Previous default; kept reachable via OPENROUTER_MODEL for rollback. Its
+// routing needs the ZDR/provider tuning below.
+const MINIMAX_M3 = "minimax/minimax-m3";
 
-/**
- * Returns the AI SDK LanguageModel for the chat agents. `usage: { include: true }`
- * turns on OpenRouter usage accounting so each response carries its real USD
- * cost (providerMetadata.openrouter.usage.cost) — which we meter against the
- * shared usage-credit pool. `provider.order` prefers Together, then Atlas
- * Cloud (fp8); `zdr: true` restricts routing to Zero-Data-Retention endpoints
- * (prompts are never retained), which is the actual constraint — it excludes
- * MiniMax first-party without a hand-maintained allowlist. The account also
- * enforces this ("Non-frontier requires ZDR" data policy); the request-level
- * flag is belt-and-braces so the constraint survives a dashboard change.
- * Fallbacks stay on within the ZDR set because pinning providers caused a
- * prod outage (Jul 2026: Together upstream-rate-limited m3 and every chat
- * turn 429'd); as of Jul 2026 the ZDR set for m3 is Together/AtlasCloud/
- * Novita/Parasail at the same price plus Morph at 2x output as a last resort.
- *
- * `reasoning` turns on OpenRouter's reasoning-token channel so the model's
- * chain-of-thought comes back as a separate reasoning stream instead of
- * leaking into the visible answer text (MiniMax M3 otherwise dumps its
- * `<think>` trace inline). `effort: "medium"` is OpenRouter's default —
- * stated explicitly only because the SDK type requires one once the channel
- * is configured.
- */
 export async function getChatAgentModel(): Promise<LanguageModelV3> {
   const apiKey = await getRequiredEnvValue("OPENROUTER_API_KEY");
   const modelId = await getOptionalEnvValue("OPENROUTER_MODEL");
@@ -40,21 +22,49 @@ export async function getChatAgentModel(): Promise<LanguageModelV3> {
 }
 
 /**
- * Synchronous variant for callers that already hold the env values. Think's
+ * Returns the AI SDK LanguageModel for the chat agents. `usage: { include: true }`
- * `getModel()` hook is sync and runs on every turn, so the SAM agent reads the
+ * turns on OpenRouter usage accounting so each response carries its real USD
- * key/model from its DO env and builds the model here.
+ * cost (providerMetadata.openrouter.usage.cost) — which we meter against the
+ * shared usage-credit pool.
+ *
+ * Default model: GPT-5.6 Luna at `reasoning.effort: "max"` — "max" is valid at
+ * the OpenRouter API for GPT-5.x but missing from the SDK's effort union, so
+ * the reasoning config rides in `extraBody`. Reasoning tokens stream on the
+ * separate reasoning channel and are billed as output tokens, which the usage
+ * accounting above captures.
+ *
+ * Sync on purpose: Think's `getModel()` hook is sync and runs on every turn,
+ * so the SAM agent reads the key/model from its DO env and builds here.
  */
 export function buildChatAgentModel(
   apiKey: string,
   modelId?: string,
 ): LanguageModelV3 {
-  return createOpenRouter({ apiKey })(modelId ?? DEFAULT_CHAT_AGENT_MODEL, {
+  const model = modelId ?? DEFAULT_CHAT_AGENT_MODEL;
+  const openrouter = createOpenRouter({ apiKey });
+
+  // MiniMax M3 (env-override path only): `provider.order` prefers Together,
+  // then Atlas Cloud (fp8); `zdr: true` restricts routing to Zero-Data-
+  // Retention endpoints, which excludes MiniMax first-party — the account's
+  // "Non-frontier requires ZDR" data policy enforces the same, this flag is
+  // belt-and-braces. Fallbacks stay on within the ZDR set because pinning
+  // providers caused a prod outage (Jul 2026: Together upstream-rate-limited
+  // m3 and every chat turn 429'd). The explicit reasoning channel keeps m3's
+  // `<think>` trace out of the visible answer text.
+  if (model === MINIMAX_M3) {
+    return openrouter(model, {
+      usage: { include: true },
+      reasoning: { effort: "medium" },
+      provider: {
+        order: ["together", "atlas-cloud/fp8"],
+        zdr: true,
+        allow_fallbacks: true,
+      },
+    });
+  }
+
+  return openrouter(model, {
     usage: { include: true },
-    reasoning: { effort: "medium" },
+    extraBody: { reasoning: { effort: "max" } },
-    provider: {
-      order: ["together", "atlas-cloud/fp8"],
-      zdr: true,
-      allow_fallbacks: true,
-    },
   });
 }