Improve session spawning UX and dev environment

backend/src/mainloop/services/chat_handler.py

@@ -48,11 +48,10 @@ def build_chat_system_prompt(recent_repos: list[str] | None = None) -> str:
 - Planning or brainstorming that needs multiple steps
 - Any work that can run in the background
 
-Usage:
-1. For code work: suggest a recent repo or ask for the GitHub repository URL
-2. Always get explicit confirmation before spawning a session
-3. Call spawn_session with just a title (and repo_url for code work)
-   - The session automatically receives the user's original request from the conversation
+IMPORTANT - When spawning a session:
+1. IMMEDIATELY acknowledge and spawn - don't ask for confirmation
+2. Respond with a brief acknowledgment like "On it - spawning a session to [task summary]"
+3. Then call spawn_session with the title (and repo_url for code work)
 
 Do NOT use spawn_session for:
 - Answering simple questions
@@ -66,14 +65,12 @@ def build_chat_system_prompt(recent_repos: list[str] | None = None) -> str:
 The user has recently worked with these repositories:
 {repos_list}
 
-If the request involves code work, suggest using one of these repos. For example:
-"I can spawn a session to work on this. Should I use {recent_repos[0]}?"
+If the request involves code work and includes a repo URL, use it. Otherwise use the most relevant recent repo.
 """
     else:
         base_prompt += """
 
-If the request involves code work, ask for the GitHub repo URL like:
-"I can spawn a session to work on this. Would you like me to proceed? Please provide the GitHub repo URL."
+If the request involves code work but no repo is provided, ask for the GitHub repo URL.
 """
 
     return base_prompt

k8s/apps/mainloop/overlays/dev/backend-patch.yaml

@@ -42,8 +42,10 @@ spec:
           env:
             - name: IS_TEST_ENV
               value: 'true'
-            # Use cheapest/fastest model for dev
+            # Use opus for main thread, haiku for background agents
             - name: CLAUDE_MODEL
+              value: opus
+            - name: CLAUDE_WORKER_MODEL
               value: haiku
             # Mock GitHub API
             - name: USE_MOCK_GITHUB

k8s/apps/mainloop/overlays/test/backend-patch.yaml

@@ -20,6 +20,8 @@ spec:
             # Use cheapest/fastest model for tests
             - name: CLAUDE_MODEL
               value: haiku
+            - name: CLAUDE_WORKER_MODEL
+              value: haiku
             # Mock GitHub API (no real GitHub calls in tests)
             - name: USE_MOCK_GITHUB
               value: 'true'

scripts/kind/reset-data.sh

@@ -8,12 +8,18 @@ CONTEXT="kind-${CLUSTER_NAME}"
 echo "=== Using context: ${CONTEXT} ==="
 
 echo "=== Cleaning up k8s task namespaces ==="
-# Delete all task-* namespaces (created by worker workflows)
+# Delete all task-* namespaces (legacy worker workflows)
 for ns in $(kubectl --context "${CONTEXT}" get ns -o name 2>/dev/null | grep "^namespace/task-" | cut -d/ -f2); do
   echo "Deleting namespace: ${ns}"
   kubectl --context "${CONTEXT}" delete ns "${ns}" --wait=false 2>/dev/null || true
 done
 
+# Delete all mainloop-session-* namespaces (session workers)
+for ns in $(kubectl --context "${CONTEXT}" get ns -o name 2>/dev/null | grep "^namespace/mainloop-session-" | cut -d/ -f2); do
+  echo "Deleting namespace: ${ns}"
+  kubectl --context "${CONTEXT}" delete ns "${ns}" --wait=false 2>/dev/null || true
+done
+
 echo "=== Resetting database ==="
 # Drop both public and dbos schemas to fully reset state
 kubectl --context "${CONTEXT}" exec -n mainloop postgres-0 -- psql -U mainloop -d mainloop -c "