ModelAnalyzerApp

© 2026 Obsvra. This document describes SpiceGrinder and is provided to help you evaluate and use it. It is not a license to reproduce, adapt, or use this material to build a competing product or service. Full terms: the SpiceGrinder EULA.

NAME

ModelAnalyzerApp — standalone Model Analyzer: static complexity plus optional throughput and per-component benchmarking

SYNOPSIS

java -cp spicegrinder.jar com.obsvra.spicegrinder.tools.ModelAnalyzerApp [options] model.xml|model.json ...

On a jpackage install:

analyzer [options] model.xml|model.json ...

DESCRIPTION

For each model file given, ModelAnalyzerApp:

1. Loads the model (XML or JSON) via ModelLoader.
2. Computes static complexity (ModelComplexityAnalyzer).
3. By default, also runs a throughput sample (a warmup pass, then --samples observations timed for real).
4. Prints a report to stdout.
5. By default, embeds the results back onto the model file’s dataset attributes (see EMBEDDED ATTRIBUTES), including analysis.timestamp — --no-write skips this.

Multiple files may be given on one invocation; each is analyzed independently.

Edition comes from the artifact (LicenseBuildEdition) by default, same as Grind. --pro/--free exist to test the other edition’s policy on a build where the classes are still physically present — not the normal way to select an edition.

OPTIONS

ISOLATED BENCHMARK

--isolate runs NodeBenchmarkHarness, timing every component individually and deriving each one’s own cost. The stdout report and the embedded analysis.* attributes stay compact (multiplier/confidence only) regardless of --isolate-raw. The full raw per-node components — subtree cost, weights, per-round samples, intended for collecting data across environments/model complexities to tune the costing methodology itself — are only surfaced via --isolate-raw, and only when the diagnostic stream is actually enabled. The safe default stays safe; the extended data is opt-in.

EMBEDDED ATTRIBUTES

Written onto the model’s dataset attributes unless --no-write is given. Prior analysis.* keys are replaced on each run.

OUTPUT (JSON)

With --json, every file given is analyzed exactly the same way, but the results are collected into one JSON document printed to stdout instead of per-file human-readable text:

{
  "success": true,
  "results": [
    {
      "model": "/abs/path/model.xml",
      "rootName": "root",
      "timestamp": "2026-08-28T12:00:00Z",
      "complexity": {
        "mode": "FULLY_COMPUTED", "nodeCount": 12, "minDepth": 1, "maxDepth": 4,
        "predictedAverageDepth": 2.5, "score": 340,
        "customComponentCount": 0, "customComponentNames": [], "summary": "..."
      },
      "throughput": {"sampleCount": 5000, "sampleDurationMs": 812, "observationsPerSecond": 6157.0},
      "isolatedBenchmark": null,
      "written": true,
      "error": null
    }
  ]
}

results[] has one entry per file given, in order. throughput is null when --no-throughput was given. isolatedBenchmark is null unless --isolate was given, in which case it’s a compact summary — measuredAverageDepth, baselineNanosPerCall, problemNodeCount, and a problemNodes[] array (label, className, core, confidence, relativeMultiplier) — mirroring the same compact section the stdout report shows; the full raw per-node components are still --isolate-raw-only, via the diagnostic stream, exactly as without --json. A file that fails to load or analyze gets a leaner entry instead: {"model": "...", "error": "<message>"}, with top-level success set to false.

--json only changes this payload — the diagnostic-stream logging --isolate-raw triggers still happens the same way, and embedded analysis.* attributes are still written unless --no-write is given. -h/--help/--version output and bad-argument errors (exit code 2) always stay plain text, unaffected by --json.

EXIT STATUS

EXAMPLES

Analyze one model, static + throughput, embed results:

java -cp spicegrinder.jar com.obsvra.spicegrinder.tools.ModelAnalyzerApp model.xml

Static-only, print but don’t modify the file:

analyzer --no-throughput --no-write model.json

Full isolated per-component benchmark with raw diagnostic output:

analyzer --isolate --isolate-raw --diagnostic model.xml

Get a machine-readable report for a script or agent to parse:

analyzer --json --no-write model.xml

SEE ALSO

Grind(1), ModelValidatorApp(1), ComponentLibraryApp(1), McpServerApp(1)