yyt-2378/automat/stem-reconstruction

STEM Reconstruction

Use this skill for the agent_qwen staged reconstruction harness.

Runtime Surface

• Harness: agent_qwen/harness.py
• Runtime tools: agent_qwen/runtime_skills.py
• Request intake: agent_qwen/request_intake.py
• Qwen tool adapters: agent_qwen/qwen_tools.py
• RPC entrypoint: agent_qwen_vl_rpc.py
• CLI smoke path: python -m agent_qwen.cli

Do not add new root-level agent_*.py scripts for harness steps. Add a runtime tool wrapper in agent_qwen/runtime_skills.py, register it in default_registry(), then call it from AgentQwenHarness.

Before running the harness from RPC or browser demo, normalize the user's natural language with normalize_harness_request(). If ready is false, return need_input_payload() and ask for the missing fields instead of running partial state through the harness.

Required user inputs:

• classic line: raw STEM image path and element list
• direct line: denoised STEM image path and element list
• compare line: raw STEM image path and element list

The public MOE-DIVAESR checkpoint was trained at 0.1 Å/pixel. Real multiscale checkpoints for varying pixel sizes are not included in the public release; real runs with the public checkpoint must use calibrated or resampled 0.1 Å/pixel inputs.

Harness Lines

Classic line:

1. denoise_patch_inference_tool
2. template_match_tool
3. stem2cif_tool
4. Optional confidence_estimation_tool
5. Optional property_prediction_tool

Direct line:

1. direct_reconstruct_tool
2. Optional confidence_estimation_tool
3. Optional property_prediction_tool

Compare line:

1. denoise_patch_inference_tool
2. classic.template_match_tool
3. classic.stem2cif_tool
4. direct.direct_reconstruct_tool
5. Optional compare.confidence_estimation_tool

State handoff:

• denoise returns recon_png
• template matching consumes recon_png, returns label_path and elements
• stem2cif consumes label_path and elements, returns cif_path
• confidence consumes recon_png and elements, returns confidence_report
• property consumes cif_path, returns relaxed_cif and metrics

Validation

Use dry-run first:

python -m agent_qwen.cli \
  --harness-line classic \
  --image examples/sample.png \
  --elements Mo S \
  --dry-run \
  --run-confidence

Use img2struc_agent for real runs:

python -m agent_qwen.cli \
  --harness-line classic \
  --image examples/stem.png \
  --elements Mo S \
  --run-confidence \
  --skip-property