simulation-orchestrator

---
name: "simulation-orchestrator"
description: "Provide tools to manage multi-simulation campaigns: generate parameter sweeps, track job execution status, and aggregate results from completed runs."
license: "Apache-2.0"
---

# Simulation Orchestrator

## Goal

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
```

| Method | Best For | Sample Count |
|--------|----------|--------------|
| `grid` | Low dimensions (1-3), need exact corners | n^d (exponential) |
| `linspace` | 1D sweeps, uniform spacing | n per parameter |
| `lhs` | High dimensions, space-filling | user-specified budget |

### Campaign Size Guidelines

| Parameters | Grid Points Each | Total Runs | Recommendation |
|------------|------------------|------------|----------------|
| 1 | 10 | 10 | Grid is fine |
| 2 | 10 | 100 | Grid acceptable |
| 3 | 10 | 1,000 | Consider LHS |
| 4+ | 10 | 10,000+ | Use LHS or DOE |

## Script Outputs (JSON Fields)

| Script | Output Fields |
|--------|---------------|
| `scripts/sweep_generator.py` | `configs`, `parameter_space`, `sweep_method`, `total_runs` |
| `scripts/campaign_manager.py --action init` | `campaign_id`, `total_jobs`, `config_dir`, `command_template` |
| `scripts/campaign_manager.py --action status` | `campaign_id`, `status`, `jobs`, `progress`, `total_jobs`, `created_at` |
| `scripts/campaign_manager.py --action list` | `jobs` (array of job records) |
| `scripts/job_tracker.py` | `job_id`, `status`, `start_time`, `end_time`, `exit_code` |
| `scripts/result_aggregator.py` | `summary` (incl. `minimize`), `statistics`, `best_run`, `failed_runs` |

> **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:

```bash
python3 scripts/sweep_generator.py \
    --base-config base_config.json \
    --params "dt:1e-4:1e-2:5,kappa:0.1:1.0:3" \
    --method linspace \
    --output-dir ./campaign_001 \
    --json
```

### Step 2: Initialize Campaign

Create campaign tracking structure:

```bash
python3 scripts/campaign_manager.py \
    --action init \
    --config-dir ./campaign_001 \
    --command "python sim.py --config {config}" \
    --json
```

### Step 3: Track Job Status

Monitor running jobs:

```bash
python3 scripts/job_tracker.py \
    --campaign-dir ./campaign_001 \
    --update \
    --json
```

### Step 4: Aggregate Results

Combine results from completed runs:

```bash
python3 scripts/result_aggregator.py \
    --campaign-dir ./campaign_001 \
    --metric final_energy \
    --json
```

`result_aggregator.py` **minimizes 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:

```bash
# 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

```bash
# 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.

**Agent workflow**:
1. Calculate total runs: 5 x 4 = 20 runs
2. Generate sweep configurations:
   ```bash
   python3 scripts/sweep_generator.py \
       --base-config simulation.json \
       --params "dt:1e-4:1e-2:5,kappa:0.1:1.0:4" \
       --method linspace \
       --output-dir ./dt_kappa_sweep \
       --json
   ```
3. Initialize campaign:
   ```bash
   python3 scripts/campaign_manager.py \
       --action init \
       --config-dir ./dt_kappa_sweep \
       --command "python phase_field.py --config {config}" \
       --json
   ```
4. After user runs simulations, aggregate results:
   ```bash
   python3 scripts/result_aggregator.py \
       --campaign-dir ./dt_kappa_sweep \
       --metric interface_width \
       --json
   ```

## Error Handling

| Error | Cause | Resolution |
|-------|-------|------------|
| `Base config not found` | Invalid file path | Verify base config file exists |
| `Invalid parameter format` | Malformed param string | Use format `name:min:max:count` or `name:min:max` |
| `Output directory exists` | Would overwrite | Use `--force` or choose new directory |
| `No completed jobs` | No results to aggregate | Wait for jobs to complete or check for failures |
| `Metric not found` | Result files missing field | Verify metric name in result JSON |

## Integration with Other Skills

The simulation-orchestrator works with other simulation-workflow skills:

```
parameter-optimization          simulation-orchestrator
        │                              │
        │ DOE samples ────────────────>│ Generate configs
        │                              │
        │                              │ Run simulations
        │                              │
        │<──────────────────────────── │ Aggregate results
        │                              │
        │ Sensitivity analysis         │
        │ Optimizer selection          │
```

### Typical Combined Workflow

1. Use `parameter-optimization/doe_generator.py` to get sample points
2. Use `simulation-orchestrator/sweep_generator.py` to create configs
3. Run simulations (user's responsibility)
4. Use `simulation-orchestrator/result_aggregator.py` to collect results
5. 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
- `campaign_manager.py` validates command templates to reject shell chaining operators (`;`, `|`, `&`, backticks, `$`)
- `--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/sweep_strategies.md` - Parameter sweep design guidance
- `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
- **v1.1.0** (2026-03-26): Standardized metadata, evaluation suite, security review, CHANGELOG
- **v1.0.0** (2026-02-25): Initial release with sweep, campaign, tracking, and aggregation