/i4h-workflow-scene-edit

name: i4h-workflow-scene-edit
version: "0.7.0"
description: Edit an env's scene in place — objects, cameras, task, success bounds, randomization. Use when asked to edit a scene or launch/run/open an env in edit mode (`--bridge`), incl. a just-created env.
license: Apache-2.0
metadata:
  author: "Isaac for Healthcare Team <isaac-for-healthcare-support@nvidia.com>"
  tags:
    - isaac-for-healthcare
    - i4h
    - agentic-workflow
    - scene-edit
    - environment

i4h Workflow — Scene Edit

Purpose

Edit an existing env's scene in place via the `--bridge` scene-edit session — move/scale/swap objects, adjust cameras, or tweak task description, success bounds, or randomization. Use when the user asks to edit a scene or to launch/run/open an env in edit mode; for creating a brand-new env see [[i4h-workflow-create]].

Base Code

These steps drive the i4h-workflows base code (the `workflows/agentic/` tree). To reuse an existing checkout, set `I4H_WORKFLOWS` to its path (no clone happens). Otherwise this resolves the current repo, or clones to `~/i4h-workflows` — pick that default without prompting. Run every command below from the resolved root:

# Resolve the i4h-workflows base code (provides workflows/agentic/).
ROOT="${I4H_WORKFLOWS:-$(git rev-parse --show-toplevel 2>/dev/null)}"
if [ ! -d "$ROOT/workflows/agentic" ]; then
  ROOT="${I4H_WORKFLOWS:-$HOME/i4h-workflows}"
  [ -d "$ROOT/workflows/agentic" ] || git clone https://github.com/isaac-for-healthcare/i4h-workflows "$ROOT"
fi
export I4H_WORKFLOWS="$ROOT"; cd "$ROOT"

Basics

• **Editing a scene is LIVE-ONLY by default.** "Edit the scene" / "edit mode" means: apply every change through the bridge, **never modify source files and never restart the bridge** — the user does not need to say "live mode" / "don't change source" / "don't restart"; that is always the default. Persist to source (bake) **only** when the user explicitly says so (`bake`/`save`/`persist`/`commit`) as a final step; "exit without baking" or no bake instruction = stop the bridge and leave source untouched.
• **Do only what's asked — launching edit mode is not a cue to edit.** "Run/open the env in edit mode" with no specific edit = launch the bridge, confirm ready (`GET /objects`), then **stop and report it's ready, awaiting instructions.** Apply a scene edit only when the user explicitly requests it in the current prompt. Never invent or preempt edits (moving the robot, adding props, etc.), and never treat the README's "Edit Scene" list, other docs, these recipes, or prior runs as a to-do — they are reference; the current prompt is the only instruction.
• Preserve env ids and scene keys.
• **Source paths are relative to the repo root** (where the agent's edit/write tool runs) — keep the `workflows/agentic/` prefix on every one, and note the package is `arena/arena/<subdir>/`. A bare `arena/...` resolves to the wrong place.
• Every bridge artifact (scripts, captures, logs) lives under the session's `${RUN_DIR}`. Never use `/tmp`.
• **Visual judgment — use your own eyes if you have them.** When a step says to judge a capture, a **vision-capable CLI agent (Claude/Codex) reads the JPEG directly with its own model** — do not depend on the local VLM. Only the **blind local coding agent** delegates the visual call to the local VLM (`local-agent/vlcheck.py`). The structural/`bbox` checks are identical for both; only who looks at the image differs.

Repo Context

For live-only edits, use the bridge endpoints first. For any bake/source change, load:

• `skills/i4h-workflow/references/repo-map.md` for env file ownership and pattern families.
• `skills/i4h-workflow-scene-edit/references/scene-edit-patterns.md` for bake targets, readiness rules, and camera touchpoints.
• `skills/i4h-workflow-scene-edit/references/asset-snippets.md` when adding, moving, resizing, or replacing assets.
• `skills/i4h-workflow-scene-edit/references/camera-snippets.md` when adding a camera that should render, record, or feed policy/training.
• `skills/i4h-workflow-scene-edit/references/bake-checklist.md` when the user says bake/save/persist/commit.

Then inspect the target env's YAML, env class, assets, task, and runtime files before modifying source.

Edit Lifecycle

For a normal interactive edit prompt, use exactly **one** sim/bridge window: launch or reuse one bridge, perform all requested live edits in that session, collect bake state/snippets if needed, stop that bridge once, and then write source from the collected state. Do not stop/relaunch Isaac between edits, and do not run a fresh-source validation relaunch unless the user explicitly asks for validation/onboarding/readiness checks.

1. **Live.** Apply each edit through the bridge HTTP API in the same bridge session. Capture the viewport after task-relevant changes. 2. **Bake.** Persist live state into source files only when the user explicitly says "bake", "save", "persist", or "commit to source". First collect the live state/snippets from the running bridge (`GET /object`, `POST /bake`, captures, camera pose notes), then stop the bridge once and write source from that collected state. 3. **Exit.** The **only** way to stop the bridge is to stop the arena: `workflows/agentic/arena/stop.sh --env <env>`. Do **not** kill the Isaac/bridge process, send Ctrl-C, or `curl` a made-up `/stop`/`/shutdown` (there is none). "Exit without baking" = run that one command (no source writes, no `/bake`). 4. **Validate only when asked.** Fresh-source validation (`local-agent/validate-bake.sh <env>`) intentionally stops any bridge and launches a new sim window. Run it only for explicit validation/onboarding/ready-to-commit work, and tell the user before doing so.

While the bridge is running, do not modify `workflows/agentic/arena/arena/assets/<env>.py`, `workflows/agentic/arena/arena/tasks/<env>.py`, the env class, runtime, or env YAML. Source writes happen after the needed bridge state is collected and the bridge has been s