fix: Switch improve.sh to Sonnet, add 1h timeout per cycle

- Use --model sonnet (faster, cheaper, sufficient for gap-filling)
- Add timeout 3600s (1h) on claude commands to prevent hangs
- Reduce RUN_TIMEOUT_MS to 75min (safety net above the 1h timeout)
- Previous run hung for 3+ hours with zero output on Opus

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.claude/skills/setup-trigger-service/improve.sh

@@ -385,7 +385,8 @@ run_team_cycle() {
     prompt=$(build_team_prompt)
     log_info "Launching agent team..."
     echo ""
-    claude -p "${prompt}" --dangerously-skip-permissions
+    timeout --signal=TERM --kill-after=60 3600 \
+        claude -p "${prompt}" --dangerously-skip-permissions --model sonnet
     local rc=$?
 
     # Clean up worktrees after cycle
@@ -405,7 +406,8 @@ run_single_cycle() {
     prompt=$(build_single_prompt)
     log_info "Launching single agent..."
     echo ""
-    claude --print -p "${prompt}"
+    timeout --signal=TERM --kill-after=60 3600 \
+        claude --print -p "${prompt}" --model sonnet
     return $?
 }
 

.claude/skills/setup-trigger-service/trigger-server.ts

@@ -22,7 +22,7 @@ const TRIGGER_SECRET = process.env.TRIGGER_SECRET ?? "";
 const TARGET_SCRIPT = process.env.TARGET_SCRIPT ?? "";
 const MAX_CONCURRENT = parseInt(process.env.MAX_CONCURRENT ?? "1", 10);
 const RUN_TIMEOUT_MS = parseInt(
-  process.env.RUN_TIMEOUT_MS ?? String(4 * 60 * 60 * 1000),
+  process.env.RUN_TIMEOUT_MS ?? String(75 * 60 * 1000),
   10
 );