Ma Sandbox Test Plan

A pipeline that gates on failure at every stage. The output of each stage IS the evidence — don't summarize it away. See .claude/skills/ma-sandbox-references/evidence-quality.md for what makes evidence good vs. hollow.

Prerequisites

Stage 4's JavaScript scenarios need a Playwright MCP server connected in this session. This repo doesn't ship one in .mcp.json — it's a per-user setup, not shared repo infra. If you don't have one connected yet, add it once with:

claude mcp add playwright -- npx -y @playwright/mcp@latest

MCP servers only load at session startup, so this doesn't take effect in the session that's currently running the skill — end this session and start a new one, then re-invoke /ma-sandbox-test-plan. Check /mcp to confirm the connection before running this skill.

Stage 0: Detect stack

CHANGED_FILES=$(git diff --name-only origin/main...HEAD)

Categorize by path prefix: go/ → Go, javascript/ → JS, python/ → Python. Mixed changes run every applicable stage below, once per stack, each gated independently — a Go build failure doesn't block the JS pipeline from also reporting its own result.

Stage 1: Build + lint (gate)

On any failure: stop that stack's pipeline, report the command and full error output, and do not proceed to Stage 2 for that stack. Other stacks (in a mixed change) still run independently.

Stage 2: Unit tests (gate)

Scope to the packages/dirs actually touched — derive this from Stage 0's $CHANGED_FILES directly, there's no fixed mapping table to maintain:

On failure: stop, report the failing test output, do not proceed to Stage 3 for that stack.

Stage 3: Integration setup

Only reached for stacks whose Stage 1–2 passed.

Go backend:

bash $(git rev-parse --show-toplevel)/.claude/skills/ma-sandbox-scripts/check-backend.sh

JavaScript UI:

VITE_INFO=$(bash $(git rev-parse --show-toplevel)/.claude/skills/ma-sandbox-scripts/find-vite-port.sh)
VITE_PORT=$(echo "$VITE_INFO" | awk '{print $1}')
VITE_PID=$(echo "$VITE_INFO" | awk '{print $2}')
VITE_MODE=$(echo "$VITE_INFO" | awk '{print $3}')  # "existing" or "started"

• Exit 0 → use $VITE_PORT in place of 5173 in all URLs below. Record $VITE_MODE/$VITE_PID for cleanup.
• Non-zero exit → stop immediately and report the failure. Do not fall back to guessing a port via curl/lsof — a successful response doesn't confirm which branch is being served.
• If the Go backend is also deployed (Stage 3 said deployed <service>), use the full sandbox at http://localhost:8090 instead of Vite, and skip straight to Stage 4.

Python:

poetry run ma sandbox health

Confirm the services relevant to the change are running.

Stage 4: Integration scenarios

Before running scenarios, list the code paths identified from the diff and note any paths skipped — this list becomes the "Known gaps" input for Stage 5.

Go

For each controller/reconcile path touched by the diff:

1. Apply a named test CR (test-<feature>-<scenario> — see .claude/skills/ma-sandbox-references/sandbox-data.md for demo-data conventions; prefer creating a new entity over disturbing shared demo state, per evidence principle 7).
2. Capture logs scoped to that resource:

   bash $(git rev-parse --show-toplevel)/.claude/skills/ma-sandbox-scripts/capture_service_logs.sh <service> --resource <test-name>

3. Verify state independently via kubectl get/kubectl describe — don't rely on logs alone.
4. Show the full lifecycle: creation, intermediate phase transitions, terminal state, and (if the diff touches them) immutability enforcement or garbage collection. See .claude/skills/ma-sandbox-references/evidence-quality.md #8–9.