From 3d069d4d4b0f97b76160b383a0ff1d5c54009680 Mon Sep 17 00:00:00 2001
From: itelo <itelofilho@gmail.com>
Date: Tue, 11 Aug 2026 15:43:45 -0300
Subject: [PATCH] perf(integrate): choose the embedded agent's model per mode
 (Opus / Haiku), budget $10

The embedded integration agent ran on one model with a $5 per-run cap; a full
Customer Portal integration exhausted the Seam-hosted inference budget mid-run.
Pick the model by the chosen mode instead: the customer-portal path is a small,
high-value surface (embedded portal, reservation deep-links) and runs on Opus
4.8; the broader full-API path runs on cheap Haiku 4.5. Raise maxBudgetUsd to $10.

Pairs with seamapi/seam-connect#17059 (proxy monthly cap -> $10; meter is a
per-model price map so each path is priced by the model that actually ran).
---
 src/lib/steps/integrate.ts | 21 ++++++++++++++++-----
 1 file changed, 16 insertions(+), 5 deletions(-)

diff --git a/src/lib/steps/integrate.ts b/src/lib/steps/integrate.ts
index 5387e49..ed485ab 100644
--- a/src/lib/steps/integrate.ts
+++ b/src/lib/steps/integrate.ts
@@ -61,12 +61,13 @@ export async function runIntegration(args: RunIntegrationArgs): Promise<void> {
       prompt: goal,
       options: {
         cwd: root,
-        // Cost-tuned for a starter integration (Seam pays for this): Sonnet 5 is
-        // near-Opus on coding at ~half the token price, medium effort trims the
-        // thinking spend, and maxBudgetUsd is a hard per-run dollar ceiling.
-        model: 'claude-sonnet-5',
+        // Model depends on the chosen mode (see modelForMode): the customer-portal
+        // path is a small, high-value surface, so it runs on Opus 4.8; the broader
+        // full-API path runs on cheap Haiku 4.5. medium effort trims the thinking
+        // spend, and maxBudgetUsd is a hard per-run dollar ceiling.
+        model: modelForMode(mode),
         effort: 'medium',
-        maxBudgetUsd: 5,
+        maxBudgetUsd: 10,
         env: buildAgentEnv(inference),
         allowedTools: ALLOWED_TOOLS,
         // Auto-apply file edits so the agent runs uninterrupted; the developer
@@ -128,6 +129,16 @@ export async function runIntegration(args: RunIntegrationArgs): Promise<void> {
   }
 }
 
+// The embedded agent's model depends on the mode the developer chose. The
+// customer-portal integration is small but high-value (embedded portal, deep
+// links), so it runs on Opus 4.8; the broader full-API path runs on cheap Haiku
+// 4.5. Defaults to Haiku when the mode is unknown.
+function modelForMode(
+  mode: 'full_api' | 'customer_portal' | undefined,
+): string {
+  return mode === 'customer_portal' ? 'claude-opus-4-8' : 'claude-haiku-4-5'
+}
+
 function buildSystemAppend(
   sdk: Sdk,
   workspaceName: string,