sima-model-compiler-issue-triage

name	sima-model-compiler-issue-triage
description	Use when triaging model-compiler GitHub issues, including install failures, offline package reports, CI/package publishing failures, dependency conflicts, smoke test failures, model surgery reports, and quantize/compile problems.

Model Compiler Issue Triage

Purpose

Use this skill to classify model-compiler issues, identify the minimum missing evidence, and route reports to the right investigation path without over-claiming a root cause.

Use When

• A GitHub issue or support report concerns model-compiler installation or activation.
• Offline packages fail to build, download, extract, or install without network access.
• GitHub Actions packaging, smoke tests, or publication fail.
• pip dependency resolution fails for the model compiler package.
• Quantization, model surgery, or compile reports need first-pass triage.

Workflow

1. Identify the area: install, offline-package, build-ci, dependencies, smoke-test, model-surgery, quantize-compile, documentation, or unknown.
2. Extract hard evidence from the report: commands, package target, architecture, Python version, logs, workflow links, and whether the run is online or offline.
3. Determine whether the report is actionable:
   • Actionable: includes exact command and failing error block.
   • Needs repro details: missing command, environment, package target, or logs.
   • Needs human review: includes private model details, credentials, proprietary artifacts, or infrastructure-specific access.
4. Ask only for missing information that materially unblocks investigation.
5. Do not state a root cause unless the logs directly support it.

Required Evidence by Area

Install

Ask for:

• Host OS and version.
• uname -m.
• Whether the install ran on the host or inside the SDK container.
• Exact sima-cli neat install target.
• Full failing error block.

Offline Package

Ask for:

• Offline package target and branch/tag.
• Whether the offline zip was downloaded and extracted.
• Exact install_modelsdk_wheels.sh command.
• Whether --offline-package was used.
• pip output if network access happened during an offline install.

Build or Publish CI

Ask for:

• GitHub Actions run URL.
• Failed job name.
• Branch/tag.
• Architecture.
• Step where the failure occurred: online build, offline closure, metadata, package test, or publish.

Dependency Conflict

Ask for:

• Conflicting package/version lines from pip.
• Python version.
• Platform and architecture.
• Online vs offline install path.

Model Surgery or Compile

Ask for:

• Model format and a shareable repro model or minimal repro.
• Target device: MLSoC (mlsoc) or Modalix (modalix).
• Input/output names and shapes.
• Quantization/calibration settings.
• Exact command and compiler or AFE logs.

Related Skills

• Use skills/model_surgery/SKILL.md when the issue involves unsupported ONNX operators, graph rewrites, or model compatibility.
• Use skills/quantize_compile/SKILL.md when the issue involves standard quantize/compile commands or generated artifacts.