[szymoniwacz]

Summary

• add the installable Ruby 3.3 gem and meetxt executable
• implement meetxt transcribe <audio-file> from local validation through one OpenAI request to atomic Markdown output
• isolate whisper-1 verbose JSON segment handling in the OpenAI adapter, with SDK retries/logging disabled and sanitized error mapping
• add RSpec, RuboCop, GitHub Actions, and the real setup/run/test/lint/build documentation

Closes #6

Scope check

This is the first file-based transcription slice only. It does not add recording, diarization, summaries, integrations, retries, cloud storage, accounts, GUI behavior, or Linux-specific support.

Validation

• bundle exec rake — 37 examples, 0 failures; 22 files linted with no offenses
• gem build meetxt.gemspec — built meetxt-0.1.0.gem
• bash tests/test-adapter.sh — passed
• bash scripts/check-workflow-leak.sh — clean
• bash scripts/ai-workflow-doctor.sh — ready
• git diff --check main...HEAD — clean

Automated validation makes no paid OpenAI request. A manual smoke test exposed the verbose-response SDK shape mismatch addressed in the latest commit; the smoke test must be rerun manually.

Review notes

Independent integrated review identified SDK conversion-error sanitization and missing CLI validation/write-failure coverage. Both were fixed. A follow-up review of the real-response parsing fix was also clean: openai-ruby 0.80.0 preserves verbose-only fields behind its public raw-key accessor, while its multipart encoder correctly sends the bracketed timestamp array.

Diff-risk: Medium
Signals: architecture/API boundary, cross-layer application/CI/docs change, security-sensitive external upload and secret handling, new OpenAI SDK dependency, CI workflow change
Evidence: 33 scoped files establish the first complete CLI slice; provider behavior is isolated in lib/meetxt/providers/openai_adapter.rb; 37 mocked/local examples cover request count, SDK response coercion and multipart encoding, sanitization, validation, rendering, CLI behavior, and atomic no-overwrite output
Reviewer focus: OpenAI request/response mapping and error sanitization; hard-link-based atomic publication; CLI exit/output behavior
Required action: Human review and green CI before merge

Limitations and follow-ups

• Real OpenAI response/upload behavior is not exercised in CI; the documented smoke test is manual and paid and must be rerun after this fix.
• The MVP intentionally pins whisper-1 with verbose_json segment timestamps behind the adapter boundary.
• Manual RubyGems publishing and rollback remain deferred until release preparation.