Orchestrate multi-simulation campaigns — generate parameter sweep configurations (grid, linspace, or Latin Hypercube sampling), initialize and track batch job campaigns, monitor job completion status, and aggregate results with summary statistics across all runs. Use when running a parameter study across dt, kappa, or other simulation inputs, managing dozens or hundreds of simulation configurations, combining outputs from completed batch runs to find the best result, or automating the generate-run-collect workflow for systematic studies, even if the user only says "I need to try many parameter combinations" or "how do I organize a sweep."
Orchestrate multi-simulation campaigns — generate parameter sweep configurations (grid, linspace, or Latin Hypercube sampling), initialize and track batch job campaigns, monitor job completion status, and aggregate results with summary statistics across all runs. Use when running a parameter study across dt, kappa, or other simulation inputs, managing dozens or hundreds of simulation configurations, combining outputs from completed batch runs to find the best result, or automating the generate-run-collect workflow for systematic studies, even if the user only says "I need to try many parameter combinations" or "how do I organize a sweep."
Provide tools to manage multi-simulation campaigns: generate parameter sweeps, track job execution status, and aggregate results from completed runs.
Requirements
Python 3.10+
No external dependencies (uses Python standard library only)
Works on Linux, macOS, and Windows
Inputs to Gather
Before running orchestration scripts, collect from the user:
Input
Description
Example
Base config
Template simulation configuration
base_config.json
Parameter ranges
Parameters to sweep with bounds
dt:[1e-4,1e-2],kappa:[0.1,1.0]
Sweep method
How to sample parameter space
grid, lhs, linspace
Output directory
Where to store campaign files
./campaign_001
Simulation command
Command to run each simulation
python sim.py --config {config}
Decision Guidance
Choosing a Sweep Method
Need every combination (full factorial)?
├── YES → Use grid (warning: exponential growth with parameters)
└── NO → Is space-filling coverage needed?
├── YES → Use lhs (Latin Hypercube Sampling)
└── NO → Use linspace for uniform sampling per parameter
Note on swept parameter names: sweep_generator.py writes each swept value into the base config by key path. A bare name (e.g. kappa) overwrites a top-level key; a dot-notation name (e.g. parameters.kappa) targets a nested key. The swept key path must match where the solver reads the value — sweeping kappa against a config that nests parameters.kappa would add an unused top-level key and silently leave the base value in place. See references/sweep_strategies.md.
Workflow
Step 1: Generate Parameter Sweep
Create configurations for all parameter combinations:
result_aggregator.pyminimizes by default: best_run is the run with the
lowest metric value (and summary.minimize is true). If higher is better
(e.g. yield, accuracy, throughput), pass --maximize so best_run becomes the
highest value:
# Higher is better -> select the maximum
python3 scripts/result_aggregator.py \
--campaign-dir ./campaign_001 \
--metric yield \
--maximize \
--json
Decision guidance: If higher is better (yield, accuracy, throughput), pass
--maximize; otherwise the reported best_run is the minimum.
CLI Examples
# Generate 5x3=15 runs varying dt (5 values) and kappa (3 values)
python3 scripts/sweep_generator.py \
--base-config sim.json \
--params "dt:1e-4:1e-2:5,kappa:0.1:1.0:3" \
--method linspace \
--output-dir ./sweep_001 \
--json
# Generate LHS samples for 4 parameters with budget of 20 runs
python3 scripts/sweep_generator.py \
--base-config sim.json \
--params "dt:1e-4:1e-2,kappa:0.1:1.0,M:1e-6:1e-4,W:0.5:2.0" \
--method lhs \
--samples 20 \
--output-dir ./lhs_001 \
--json
# Check campaign status
python3 scripts/campaign_manager.py \
--action status \
--config-dir ./sweep_001 \
--json
# List jobs (read-only), optionally filtered by status
python3 scripts/campaign_manager.py \
--action list \
--config-dir ./sweep_001 \
--status-filter failed \
--json
# Get summary statistics from completed runs (minimize: best = lowest)
python3 scripts/result_aggregator.py \
--campaign-dir ./sweep_001 \
--metric final_energy \
--json
# Maximization metric: best = highest value (yield, accuracy, throughput)
python3 scripts/result_aggregator.py \
--campaign-dir ./sweep_001 \
--metric yield \
--maximize \
--json
Conversational Workflow Example
User: I want to run a parameter sweep on dt and kappa for my phase-field simulation. I want to try 5 values of dt between 1e-4 and 1e-2, and 4 values of kappa between 0.1 and 1.0.
Use parameter-optimization/doe_generator.py to get sample points
Use simulation-orchestrator/sweep_generator.py to create configs
Run simulations (user's responsibility)
Use simulation-orchestrator/result_aggregator.py to collect results
Use parameter-optimization/sensitivity_summary.py to analyze
Verification checklist
Before trusting a campaign's best_run or summary statistics, record concrete evidence for each item:
Confirmed the swept key path actually changed the value the solver reads: opened at least one generated config_NNNN.json and verified the swept parameter (e.g. parameters.kappa) holds the expected value at the expected nesting level, not a duplicate unused top-level key (sweep_generator.py writes by key path).
Reconciled job accounting from result_aggregator.py --json: recorded summary.total_jobs, summary.completed, and summary.failed, and confirmed completed + failed == total_jobs. Any shortfall means runs were silently skipped (missing result file or extract_metric returned None) and must be investigated, not ignored.
Confirmed completed > 0 and that the recorded summary.metric matches the field the solver actually writes. A typo'd or absent metric makes extract_metric return None, yielding zero completed runs with no error.
Recorded summary.minimize and confirmed it matches the intended direction (default minimize; --maximize for yield/accuracy/throughput) before quoting best_run.
Did NOT treat job_tracker.py "completed" as physical success: it flags a job completed purely from a result-file's existence and stamps exit_code 0 — independently checked the run's real exit status / solver logs for non-zero codes or NaN/Inf output.
Applied an outlier/sanity check to the metric values (e.g. Tukey 1.5x IQR from references/aggregation_methods.md) and confirmed best_run.value is physically plausible, not a crashed run that emitted a spurious extremum.
For LHS sweeps, recorded the --seed used and saved manifest.json (parameter bounds, total_runs, parameter_space) so the sample set is reproducible.
Common pitfalls & rationalizations
Tempting shortcut
Why it's wrong / what to do
"The job tracker says completed, so the run succeeded."
job_tracker.py marks "completed" whenever a result file exists and hard-codes exit_code 0 — it never reads the actual exit code. A crashed run that wrote a partial result file looks identical to a clean one. Check the solver's real exit status and output validity.
"completed is high, so I have all my results."
Jobs with a missing result file or a metric that extract_metric can't read are silently skipped — neither counted as completed nor failed. Reconcile completed + failed against total_jobs; a gap means lost runs.
"Aggregation returned a best_run, so that's the optimum."
By default the aggregator minimizes. If higher is better you must pass --maximize, or best_run is the worst point. Always record summary.minimize and confirm the direction.
"I swept kappa, so the runs vary."
sweep_generator.py writes by key path. If the base config nests the value under parameters.kappa but you sweep the bare name kappa, every config keeps the original nested value and gains an unused top-level key — the sweep is scientifically meaningless. Sweep the exact dotted path the solver reads.
"The metric name is close enough."
A misspelled or absent metric makes extract_metric return None for every run, so completed is 0 and statistics are empty — with no error raised. Verify the metric matches the solver's output field exactly.
"Grid covers everything, so use it for all my parameters."
Grid is n^d — it explodes exponentially (4 params x 10 = 10,000 runs). For 4+ dimensions use lhs with a deliberate budget; reserve grid for 1-3 parameters.
"LHS is random, so I don't need to record anything."
LHS is reproducible only with a fixed --seed. Without recording the seed (and manifest.json), the sample set cannot be regenerated or defended.
Security
Input Validation
Metric names (result_aggregator.py --metric) are validated against [a-zA-Z_][a-zA-Z0-9_.]* to prevent traversal or injection via crafted keys
Swept parameter names (sweep_generator.py --params) are validated against [a-zA-Z_][a-zA-Z0-9_]*(.[a-zA-Z_][a-zA-Z0-9_]*)* (dot notation for nested keys); invalid names are rejected
--params format strings are parsed and validated (name:min:max:count with finite numeric bounds — NaN/Inf rejected — min < max, and positive integer counts capped at 100,000); at most 32 parameters per sweep
--method is validated against a fixed allowlist (grid, linspace, lhs)
--samples is validated as a positive integer with an upper bound (max 1,000,000)
--action is validated against a fixed allowlist (init, status, list); for the read-only list action, --status-filter is validated against pending, running, completed, failed
File Access
sweep_generator.py reads a single base config file (JSON) specified by --base-config and writes generated configs to --output-dir
result_aggregator.py enforces a 10 MB file-size limit per result file, maximum JSON nesting depth, and strict numeric type checking (rejects bool, NaN, Inf)
All string values from result files are sanitized (truncated, control characters stripped) before surfacing them
Config paths interpolated into shell commands are validated against a safe-character allowlist and escaped with shlex.quote()
Tool Restrictions
Read: Used to inspect script source, references, base configs, and campaign status files
Write: Used to save generated sweep configs, campaign manifests, and aggregated results; writes are scoped to the user's working directory
Grep/Glob: Used to locate campaign files, result files, and search references
The skill's allowed-tools excludes Bash to prevent the agent from executing arbitrary commands when processing untrusted simulation outputs
Safety Measures
No eval(), exec(), or dynamic code generation
All subprocess calls use explicit argument lists (no shell=True)
Reduced tool surface (no Bash) limits the agent to read/write operations only
Command templates are validated but never executed by the skill itself; execution is the user's responsibility
Limitations
Not a job scheduler: Does not submit jobs to SLURM/PBS; generates configs and tracks status
No parallel execution: User must run simulations externally (can use GNU parallel, SLURM, etc.)
File-based tracking: Status tracked via files; no database or real-time monitoring
Local filesystem: Assumes all files accessible from local machine
References
references/campaign_patterns.md - Common campaign structures
references/aggregation_methods.md - Result aggregation techniques
Version History
See CHANGELOG.md for the authoritative, dated history. Summary:
v1.1.3 (2026-06-24): Added a Verification checklist and a Common pitfalls & rationalizations section grounded in the scripts' real behavior (result-file-only "completed" detection, silent skip of unreadable metrics, minimize-by-default direction, key-path merge semantics)
v1.1.1 (2026-06-23): Dot-notation nested overrides in sweep_generator.py, input-validation hardening (--params name/finite/count caps, --samples bounds), documented --maximize and the list action, corrected Script Outputs table and worked-example numbers
