Retail Virtual Try-On Agent

Virtual Try-On Agent

Creates virtual try-on (VTO) agents on Google Cloud supporting image and video (catwalk animation) try-on modes.

STOP -- READ THIS BEFORE RESPONDING

Check your operating mode first — there are two distinct modes:

Mode A: Deployed Try-On Agent

If the system context tells you that setup is complete (e.g. it says "You are a DEPLOYED try-on agent", or the conversation history shows setup has already been completed) — skip Q-MODE entirely. Respond directly to the user's try-on query. Do NOT output the Q-MODE block. Do NOT mention setup.

Mode B: First-time Setup (default)

If there is no such context (fresh invocation, no prior setup) — your VERY FIRST response MUST be the Q-MODE block below. Nothing else.

Do NOT ask about products, industry, GCP project, or anything else first. Do NOT propose a plan. Do NOT explain what you will do.

Your first message to the user must be EXACTLY this (copy-paste, no changes):

[skill: retail-virtual-tryon] active.
Q-MODE: Pick a setup mode? [default: 1]
  1. Quick start -- Local testing sandbox, interactive 6-question config, ~90s. Best for demos.
  2. Export Web App & GCS Catalog Sync -- Generate standalone containerized codebase, GCS catalog sync, and Cloud Run config, ~3 min.

Then STOP and wait for the user's answer.

Accept: 1, quick, empty/Enter (= Quick Start), 2, export, sync or webapp (= Export Web App & GCS Catalog Sync).

Execution Rules

1. Q-MODE first, always. No exceptions. No preamble.
   • CRITICAL WARNING: Do NOT automatically run setup or deployment scripts (e.g. setup_tryon.py, export_app.py, deploy_cloudrun.sh) upon receiving a general request like "I want to create/deploy a VTO app on GCP". You MUST first present the Q-MODE setup menu choice and wait for the user to select Mode 1, 2, or 3.
2. One question at a time. Show the default. Accept empty input. Format: Q: <question text>? [default: <value>] Pressing Enter = use the default. NEVER ask multiple questions in one turn.
3. Execute steps in order. Do NOT jump ahead or skip steps.
4. Verify each step succeeded before moving to the next.
5. Save all answers to ./design-spec.md (in the workspace) as you collect them. After the interview, run .venv/bin/python "$SKILL_DIR/scripts/setup.py" --config ./design-spec.md (see the Workspace Setup section to resolve $SKILL_DIR).
6. Confirm completion of each step before proceeding.

Before Quick Start has launched setup, the user can say "export", "webapp", or "deploy" to switch to Mode 2 (Export Web App & GCS Catalog Sync). Carry over answers already given for project, mode, and region; ask only the remaining Mode-2 questions (GCS catalog bucket, export directory). After setup.py has already started buckets/APIs, the workflow is committed -- to run Mode 2 instead, start a fresh workspace.

Workspace Setup

The skill has two locations:

• Install dir -- where SKILL.md and scripts live (varies by host)
• Workspace -- the agent's cwd; design-spec.md, .venv, and per-run state live here

By the end of this section the workspace must have .venv/ (with the skill installed editable), design-spec.md, and SKILL_DIR exported in the shell.

Run this as ONE shell command -- splitting it across tool calls loses state:

SKILL_DIR=$(for d in ~/.claude/skills ~/.agents/skills ~/.gemini/skills ~/.cursor/skills; do
  [ -f "$d/retail-virtual-tryon/SKILL.md" ] && echo "$d/retail-virtual-tryon" && break
done)
bash "$SKILL_DIR/scripts/bootstrap.sh"

bootstrap.sh finds a Python 3.11+ interpreter (with absolute-path fallback for sandboxed shells), creates .venv, installs the skill editable, and copies design-spec.md into the workspace.

All scripts run from the install dir against the workspace config. Use .venv/bin/python, not bare python -- bare python may resolve to a Python without the skill's editable install on sys.path.

.venv/bin/python "$SKILL_DIR/scripts/setup.py" --config ./design-spec.md

Edit ./design-spec.md and set gcp_project_id (the agent will do this based on the user's answers in Q-MODE).

Mode 1: Quick start (4-5 questions)

Question Formats & Accepted Choices:

• Q-B: Try-on mode? Format to print:

  Q: Try-on mode? [default: 2]
    1. image_only (Faster, static images only)
    2. image_and_video (Catwalk video animations via Veo)

  Accept: 1 (= image_only), 2 (= image_and_video), image_only, image_and_video.

• Q-D: Catalog Path? Accept: demo (uses bundled catalog), local directory path, or GCS URI starting with gs://.

• Q-E: Upload local catalog to GCS? Format to print:

  Q: Upload local catalog to GCS? [default: 1]
    1. Yes (Sync and host catalog in GCS)
    2. No (Run locally using local folder assets)

  Accept: 1 (= Yes), 2 (= No), Yes, No.

After collecting these answers, do this automatically:

1. Write the answers to ./design-spec.md (in the workspace), filling in:
   • gcp_project_id
   • tryon_mode
   • tryon_model (always use flash)
   • tryon_output_bucket (default: {project_id}-tryon-output)
   • tryon_upload_bucket (default: {project_id}-tryon-uploads)
   • gcp_region
   • tryon_catalog_path
   • tryon_catalog_upload (set to true by default, set to false only if Q-E is answered as No)
2. Tell the user: "Setting up local sandbox environment resources..."
3. Run: .venv/bin/python "$SKILL_DIR/scripts/setup.py" --config ./design-spec.md
4. Stream the output. On success, start the local sandbox server in the background: .venv/bin/python "$SKILL_DIR/scripts/start_sandbox.py" --config ./design-spec.md
5. Provide the user with the clickable localhost link to test: "VTO fitting room sandbox is running! Open http://localhost:8080 in your browser to test it."

Mode 2: Export Web App & GCS Catalog Sync (4 questions)