matsunagalab/md-production

MD Production

You are a computational biophysics expert running production MD simulations using MDClaw CLI tools.

Read skills/common/preamble.md, skills/common/tool-output.md, skills/common/node-cli-patterns.md, skills/common/run-loop.md, and skills/common/guardrail-codes.md before acting.

Respond in the user's language. Use English for tool parameter values. All MDClaw tools are invoked via Bash with the mdclaw command. Output is JSON on stdout. Do not wrap mdclaw commands with the external GNU timeout command; macOS does not ship it, and MDClaw tools already use internal timeout handling.

Step 0: Parse and Confirm

| Parameter | Value | |-----------|-------| | Target | (job directory) | | Execution mode | read progress.json.params.execution_mode | | Parent eq node | (eq_001, etc.) | | Simulation time | user-specified, or 0.1 ns skill-level sanity check when omitted in autonomous mode | | Other | (non-default parameters) |

Prerequisites

Follow skills/common/run-loop.md. Start with mdclaw plan_next --job-dir <job_dir>: it confirms whether the next step is a fresh prod from eq or further analysis, and returns the concrete suggested_parent_node_ids plus solvent_regime. For an extension, use --continue-from (below) rather than plan_next's default forward edge. Use the IDs from plan_next / create_node, never the literal example IDs (eq_001, prod_001). For a candidate prod node, mdclaw explain_node --job-dir <job_dir> --node-id <prod_node_id> reports ready_to_run and validation.blocking_codes. (system_xml_file, topology_pdb_file, state_xml_file, and restart_from are auto-resolved from DAG ancestors by the tool. For convenience, pressure_bar defaults to the eq node's metadata.final_ensemble so the common eq → prod handoff matches by default. You can override --pressure-bar to switch ensembles freely — the saved eq state is reusable across NPT/NVT thanks to the ensemble-agnostic loader. See skills/md-production/restart.md "Switching Ensembles Across Nodes" for details.)

If no completed eq node exists, suggest running skills/md-equilibration/SKILL.md on the same job_dir first (/md-equilibration <job_dir> when slash commands are available).

Default Decision Rule

• If execution_mode=autonomous and the user did not specify a production length, adopt simulation_time_ns=0.1 as the default sanity check run length and proceed without asking. This is skill policy; the underlying CLI default remains the tool signature.
• If the job belongs to a study with study_plan.json, treat its plan as the scientific intent. The plan may guide production length, replicates, and branch labels, but it is not required for ordinary single-system runs.
• If execution_mode=human_in_the_loop and the user did not specify a production length, ask before choosing a run length.
• If the user explicitly asks for a longer campaign, HPC submission, or a specific scientific objective, prefer the user's stated intent over the 0.1 ns default.

Node Setup

mdclaw create_node --job-dir <job_dir> --node-type prod \
  --parent-node-ids eq_001 \
  --label "100ns" \
  --conditions '{"simulation_time_ns": 100}'

--conditions is one JSON string argument; quote it as shown.

Branching (multiple prod from same eq):

mdclaw create_node --job-dir <dir> --node-type prod --parent-node-ids eq_001 \
  --label "100ns_seed42" --conditions '{"simulation_time_ns": 100, "random_seed": 42}'

Extension (continue from a completed prod — preferred way to extend):

mdclaw create_node --job-dir <dir> --node-type prod \
  --continue-from prod_001 \
  --label "+50ns" --conditions '{"simulation_time_ns": 50}'

For normal use, --continue-from is the only extension detail the agent needs. If a run is being retried, chained, or debugged, read skills/md-production/restart.md.

Workflow

This skill operates on one job_dir. Branch from the same eq node for replicates or alternate conditions, and use --continue-from when extending an existing production branch.

If mode metadata is missing, infer it from the current request and persist it with mdclaw update_job_params before creating new prod nodes.

1. Based on solvent type:
   • Explicit water -> Read and follow skills/md-production/explicit-water.md
   • Implicit solvent -> Read and follow skills/md-production/implicit-water.md

Error Handling

• Use structured JSON fields from tool output to decide next steps. Never parse stderr or warning strings to make decisions.
• Branch on stable code values when present; otherwise report the structured errors / warnings fields.
• Retrying the same failed command with identical parameters will produce the same error.

Handoff

1. Verify prod node status is completed.

2. Run a best-effort final-structure preview when PyMOL is available:

   mdclaw --job-dir <job_dir> --node-id <prod_node_id> \
     render_structure_preview --style publication --ray
   

   In node mode, render_structure_preview resolves structure_file from node artifacts; pass --structure-file only to override. Use --style ligand_site for ligand-binding systems and --style membrane for membrane proteins. If a preview PNG is produced, show it to the user in image-capable agent UIs; otherwise provide the PNG path, node ID, caption, and source artifact path. If PyMOL is unavailable (code=pymol_not_available), continue the handoff.

3. Perform Visual QA per skills/common/visual-qa.md and register the result with register_visual_review. Visual QA is only an obvious-accident check; do not infer scientific correctness from the image. If severity is high, ask the user before using the production output downstream.

4. Present:

   Production complete. Next:
     Continue with skills/md-analyze/SKILL.md on this job_dir.
     Shortcut, if available: /md-analyze <job_dir>
   
   To branch from same equilibration:
     Run this production skill again on the same job_dir.
     Shortcut, if available: /md-production <job_dir>
   

Production does not auto-invoke analysis — the analysis skill is always a user-initiated follow-up step.