Local AI: an inspectable JavaScript function

Try the original output

This isolated preview cannot access your Donatello account. Use fictional sample details only; this is not a real business or contact service.

Open full-window demo ↗

Exact request

Write only JavaScript source, without Markdown fences. Define function slugify(value). It must reject non-string input with TypeError, trim whitespace, lowercase, use Unicode NFKD normalization and remove combining accent marks, replace each run of characters other than ASCII a-z and digits 0-9 with a single hyphen, and trim hyphens from both ends. Empty or punctuation-only strings return an empty string. Do not use dependencies or external services. Do not export the function.

Follow-up requests

The previous response was incorrect: require("unicode-normalization") is an external dependency and JavaScript trim("-") does not remove hyphens. Return the complete corrected function only, no Markdown. Use the built-in value.normalize("NFKD"). Remove leading and trailing hyphens using a regular expression. Preserve the TypeError for non-strings. Do not use require or import.

Checks and limits

• Operator regression checks: 12/12 cases passed on the unchanged corrected function.
• Actual local completion times: 4463 ms first attempt; 4169 ms correction. Model-server startup is recorded separately.
• Windows 10.0.26200; AMD Ryzen 7 4800H with Radeon Graphics; 16 logical CPUs; 15 GiB system RAM rounded. Not a minimum RAM requirement.
• Model file SHA-256 matched the Agent manifest; runtime bound to loopback with --offline; this harness made no cloud-model calls.

Timing separates model startup and completion in each inference record. It excludes operator review and is not comparable to full project completion time.

• This is a narrow local-inference example, not an end-to-end desktop chat or automatic-fallback test.
• The operating-system network was not disabled. Offline runtime and loopback requests do not prove that every Agent tool works offline.
• Single-machine, single-task timing samples are not a general performance benchmark.
• ASCII URL slugs only; scripts without ASCII transliteration may produce empty output.
• The first generation was incorrect and required a targeted follow-up.

Human and operator interventions

• Codex built an isolated harness around the production Agent LocalModelManager, supplying only the Electron application path. No installed chat/profile was changed.
• The operator reviewed the first answer, found an external dependency and incorrect hyphen trimming, and supplied an explicit correction prompt. Both raw answers are preserved.
• The operator tested the corrected output without editing it.
• The interactive inspector and test JSON were written by the operator; only the displayed slugify function was generated by the local model.

Download and verify the actual files

Generated source is preserved unchanged. Operator-created test records and presentation wrappers are identified above. Compare each file's SHA-256 with this record.

Download provenance manifest (JSON)