docs: #149 — draft ROS Discourse post for community feedback

Adds docs/community/ros-discourse-post.md with the complete, copy-paste-ready
post for https://discourse.ros.org/. Covers the current MuJoCo/Gymnasium
status, embeds the grasp demo GIF, links the live CI report, and asks four
concrete open questions about the Gazebo integration path.

https://claude.ai/code/session_01MMF7mhLKHeAfaP8HcvGpbv

Author: claude

docs/community/ros-discourse-post.md

@@ -0,0 +1,115 @@
+# ROS Discourse Post: Visual Testing Harness for AI Coding Agents
+
+_Context: Issue [miaodx/roboharness#149](https://github.com/MiaoDX/roboharness/issues/149) — post on ROS Discourse for community feedback._
+
+**Post in:** https://discourse.ros.org/ → **General** category
+
+**Title:** `Visual testing harness for AI coding agents in robot simulation — looking for feedback`
+
+---
+
+## Full Post Content
+
+Hi ROS community,
+
+**A question for those using Claude Code, Codex, or other AI coding agents for robotics work:** how do you debug simulation behavior when the agent is writing the control code?
+
+My specific problem: when I used Claude Code to write MuJoCo control scripts, it could read error logs and joint angles, but it couldn't _see_ what the robot was actually doing. The agent would iterate on code that looked plausible but produced obviously wrong behavior — wrong grasp orientations, unstable footing, arm trajectories that clipped through geometry. These failures are trivial for a human to spot in a viewer, but invisible to a text-only agent.
+
+---
+
+### What I built: roboharness
+
+[roboharness](https://github.com/MiaoDX/roboharness) is a visual testing harness that pauses simulation at named checkpoints and captures multi-view screenshots alongside structured JSON state. The agent reads these files directly — no separate VLM inference step needed.
+
+**MuJoCo grasp demo** (front view — Plan → Pregrasp → Approach → Close → Lift → Holding):
+
+![MuJoCo grasp front view](https://raw.githubusercontent.com/MiaoDX/roboharness/main/assets/X32_Y28_Z13_front_view.gif)
+
+**[→ Live Report (auto-generated from CI on every push)](https://miaodx.com/roboharness/grasp/)**
+
+The report shows 6 checkpoints from pre-grasp to object-in-hand. CI regenerates it via GitHub Actions on every push using `MUJOCO_GL=osmesa` for headless rendering.
+
+The core pattern is two lines to wrap any Gymnasium environment:
+
+```python
+from roboharness.wrappers import RobotHarnessWrapper
+
+env = RobotHarnessWrapper(env,
+    checkpoints=[{"name": "pre_grasp", "step": 50}, {"name": "lift", "step": 120}],
+    output_dir="./harness_output",
+)
+```
+
+At each checkpoint, roboharness saves:
+- PNG screenshots from all configured cameras (front, side, wrist, top-down)
+- `state.json` — joint positions, velocities, ctrl commands
+- `metadata.json` — sim_time, step index, camera list
+
+The AI agent reads these files and reasons about what to change next.
+
+Or with the lower-level API:
+
+```python
+from roboharness import Harness
+from roboharness.backends.mujoco_meshcat import MuJoCoMeshcatBackend
+
+backend = MuJoCoMeshcatBackend(model_path="robot.xml", cameras=["front", "side"])
+harness = Harness(backend, output_dir="./output")
+
+harness.add_checkpoint("pre_grasp")
+harness.add_checkpoint("lift")
+result = harness.run_to_next_checkpoint(actions)
+# result.views → multi-view screenshots, result.state → joint angles + velocities
+```
+
+---
+
+### Current status
+
+| Simulator | Status |
+|---|---|
+| MuJoCo (native backend) | ✅ Implemented — headless via `MUJOCO_GL=osmesa` or `egl` |
+| Gymnasium wrapper | ✅ Works with Isaac Lab, ManiSkill, LeRobot, etc. |
+| LeRobot (`make_env()` factory) | ✅ Implemented |
+| **Gazebo / ROS2** | 📋 Planned — not yet implemented |
+
+---
+
+### Questions for this community
+
+The Gazebo integration is the obvious next step for ROS users, and I'd genuinely value input before committing to an approach:
+
+1. **Capture method:** For screenshot capture from Gazebo, which do you prefer in practice — subscribing to `/camera/image_raw` via a ROS2 node, or using Gazebo's native snapshot API (where available)? Or is there a third approach that's more CI-friendly?
+
+2. **State source:** Is TF2 the right source for robot state (end-effector poses, joint frames) in ROS2 context? Would you also expect `/joint_states` to be exposed, or is TF sufficient?
+
+3. **CI renderer:** If you run headless Gazebo in CI today, which renderer are you using — `--headless-rendering` with WebGL, OSMesa, or something else? Osmesa works reliably for MuJoCo; curious whether the same holds for Gazebo.
+
+4. **Use case fit:** Does AI-agent-driven robot simulation development cause debugging pain in your workflow? Or is this solving a problem you don't actually have? Honest answer appreciated — I'd rather know the use case is wrong than optimize for the wrong thing.
+
+---
+
+GitHub: https://github.com/MiaoDX/roboharness  
+MIT License, Python 3.10+, numpy-only core (MuJoCo/Meshcat optional)
+
+Happy to answer technical questions. Looking for feedback, not promoting — I'm genuinely trying to understand whether this is useful to ROS developers and what the right Gazebo integration path looks like.
+
+---
+
+## Posting Notes
+
+- Post in **General** category on https://discourse.ros.org/
+- ROS Discourse account required (login with GitHub)
+- Add tags: `simulation`, `testing`, `ai-agents` (if tag system allows)
+- The GIF is embedded via raw GitHub URL — it will render inline on Discourse
+- Tone check: the four questions at the end must feel like real questions, not softening before a pitch. Edit them if they read as rhetorical
+- After posting, note the Discourse thread URL in a comment on issue #149 so the response thread is tracked
+
+---
+
+## Related issues
+
+- **#152** — Gazebo/ROS2 showcase (the implementation that this post is seeking feedback for)
+- **#150** — ros2_mcp collaboration discussion (complementary tool, different angle)
+- **#146** — awesome-ros2 listing (deferred — requires forking external repo)