diff --git a/docs/modules/module-21-kaggle-vibe-coding/DAY_5_PRODUCTION_GRADE_PLAN.md b/docs/modules/module-21-kaggle-vibe-coding/DAY_5_PRODUCTION_GRADE_PLAN.md
new file mode 100644
index 0000000..1927b03
--- /dev/null
+++ b/docs/modules/module-21-kaggle-vibe-coding/DAY_5_PRODUCTION_GRADE_PLAN.md
@@ -0,0 +1,125 @@
+# Day 5: Spec-Driven Production Grade Development
+
+## Source Assignment
+
+Kaggle Unit 5 focuses on "Spec-Driven Production Grade Development in the Age of Vibe Coding."
+
+The final assignment includes:
+
+- summary podcast review
+- whitepaper review
+- optional agent deployment to Google Cloud
+- optional frontend web application linked to a cloud-hosted agent
+
+The deployment and frontend codelabs are optional and may require a billing account. In this repository, those steps remain blocked by the Module 21 Stage 3 manual review gate.
+
+## Repo Interpretation
+
+Day 5 is mapped as a production-readiness and specification discipline exercise, not as a live deployment task.
+
+Allowed local/documentation scope:
+
+- map spec-driven development concepts
+- define behavior-first acceptance criteria
+- document Gherkin-style scenarios as source-of-truth examples
+- identify production pipeline controls
+- document zero-trust deployment gates
+- map automated code-review and policy-server concepts to repo guardrails
+
+Blocked pending manual review:
+
+- Cloud Run deployment
+- public endpoint exposure
+- Secret Manager provisioning
+- live AI Studio API integration
+- frontend connected to a live cloud-hosted agent
+- billing-impacting resources
+
+## Spec-Driven Development Mapping
+
+Day 5 reframes vibe-coded output as disposable implementation and treats behavior specifications as the durable contract.
+
+Repository mapping:
+
+| Course concept | Repo-safe interpretation |
+|---|---|
+| Gherkin behavior specs | deterministic acceptance criteria for local agent behavior |
+| code as disposable | generated code must be replaceable if tests/specs stay stable |
+| production-grade pipeline | branch, PR, guardrails, test evidence, and manual review gates |
+| zero-trust development | no unreviewed secrets, endpoints, IAM, or cloud resources |
+| automated review agents | CI guardrails and scoped review checklists |
+| policy servers | documented safety boundaries and manual approval gates |
+
+## Example Acceptance Scenarios
+
+```gherkin
+Feature: Agent safety before production deployment
+
+  Scenario: Unsafe prompt is blocked locally
+    Given a local ADK agent workflow
+    When a prompt contains PHI-like or prompt-injection text
+    Then the workflow routes the request to a security handler
+    And the request does not reach the normal classifier path
+
+  Scenario: Cloud deployment remains gated
+    Given the Module 21 Stage 3 gate is not approved
+    When a task requests Cloud Run deployment
+    Then the deployment is documented as deferred
+    And no cloud resource is provisioned
+
+  Scenario: Frontend integration remains non-live
+    Given there is no approved cloud-hosted agent endpoint
+    When a frontend integration is planned
+    Then the design remains local or documentation-only
+    And no public endpoint is referenced in public docs
+```
+
+## Production Gate Checklist
+
+Before any future live deployment, all items must be explicitly approved:
+
+- cost review completed
+- billing impact accepted
+- Cloud Run exposure reviewed
+- Secret Manager usage approved
+- no `.env` or committed secret files
+- API key scope restricted
+- IAM least-privilege account defined
+- public endpoint policy approved
+- rollback path documented
+- DevTools Network and Console evidence plan defined
+
+## Current Evidence Status
+
+Completed locally:
+
+- Day 1 course mapping and Cloud Run conceptual notes
+- Day 2 MCP/tooling evidence
+- Day 3 ADK customer support agent prototype
+- Day 4 local security guardrails and targeted tests
+- Stage 3 manual review gate
+
+Deferred:
+
+- live Cloud Run deploy
+- cloud-hosted agent URL
+- live frontend-to-agent integration
+- Secret Manager setup
+- AI Studio API runtime integration
+
+## Safety Boundary Confirmation
+
+- No PHI
+- No e-Nabiz export
+- No patient data
+- No secrets
+- No `.env`
+- No cloud resource provisioning
+- No public endpoint
+- No billing-impacting configuration
+
+## Day 5 Completion Decision
+
+Day 5 is complete for this repository as a documentation-first production readiness mapping.
+
+The optional deployment and frontend codelabs are intentionally deferred until the Module 21 Stage 3 manual review gate is approved.
diff --git a/experiments/module-21-kaggle-vibe-coding/customer-support-agent/app/agent.py b/experiments/module-21-kaggle-vibe-coding/customer-support-agent/app/agent.py
index 1c7a71c..7ee35d5 100644
--- a/experiments/module-21-kaggle-vibe-coding/customer-support-agent/app/agent.py
+++ b/experiments/module-21-kaggle-vibe-coding/customer-support-agent/app/agent.py
@@ -88,7 +88,7 @@ def security_check(ctx: Context, node_input: Any):
     yield Event(output=payload, state=payload, route=route)
 
 
-def save_query(node_input: str):
+def save_query(node_input: Any):
     """Saves user query in state for downstream nodes."""
     query = _extract_query(node_input)
     yield Event(output=query, state={"user_query": query})