Vasp Batch Execution

2. Avoid illegal path layouts

• output_dir must not be inside input_dir.
• Output mapping must not overlap any source calc directory.
• Nested calc folders under an already-valid calc root are rejected.

3. Submit and collect

• The tool stages work under the output tree, injects the bootstrap script, dispatches via DPDispatcher, then collects outputs back into the final output tree.
• Treat the final output tree as a collected snapshot, not an in-place mutation of the input tree.

4. Triage failures minimally

• Read _BATCH_STATE.json first.
• If needed, inspect only the focused scheduler/stdout/stderr evidence for failed jobs.
• Rerun only the failed subset, again into a fresh output root.

5. Hand off to structured analysis

• After collection, prefer analyze_vasp_results over ad hoc manual parsing when the next step needs convergence, energy, or bandgap summaries.
• Treat this skill as execution-only; dispatch success is not the same as usable scientific output.

Method-critical defaults

• Keep the input tree and output tree separate so the collected snapshot stays auditable.
• Do not use launch success as a scientific result; the post-run evidence files are part of the contract.

Output Contract

Return:

• whether the run used single-folder or batch mode
• submitted calc-directory count
• output_root_rel
• batch_state_rel
• representative output path
• whether the batch finished as collected_complete or only collected_partial

References

• Use execute only for focused follow-up reads after the batch state points to a concrete failure target.
• Hand off finished NEB or MD batches to reaction-neb-analysis or md-diffusion-analysis rather than reusing this skill as an analysis layer.