fix: text_to_speech saves to Desktop and returns inline audio

[maharshi-smallest]

Summary

Fixes TTS tool so generated audio is actually accessible to the user.

Changes

• Save to ~/Desktop by default instead of temp dir
• Return audio as MCP AudioContent (base64) inline — Claude Desktop can render or offer it directly
• Both file path AND inline audio in response

Problem

Claude Desktop runs the MCP server in a sandbox — temp files on the server side are inaccessible. Users couldn't get the generated audio.

🤖 Generated with Claude Code

[entelligence-ai-pr-reviews]

EntelligenceAI PR Summary

Improves the transcribe_audio tool in src/tools/transcribe-audio.ts with UX, validation, and reliability enhancements.

• Updated tool and parameter descriptions to explicitly instruct the LLM to ask for language and a real machine file path before invocation
• Removed default value for language, making it a required parameter
• Added tilde (~) expansion for file paths using process.env.HOME
• Wrapped file read operations in try/catch with descriptive ENOENT-aware error messages
• Converted file buffer to Uint8Array before sending as the API request body

---

Confidence Score: 5/5 - Safe to Merge

Safe to merge — this PR makes targeted, well-scoped improvements to src/tools/transcribe-audio.ts including tilde expansion via process.env.HOME, ENOENT-aware try/catch error handling, and stricter parameter validation by removing the default language value. The changes are defensive in nature and improve reliability without introducing new logic paths that could regress existing functionality. No review comments were generated and heuristic analysis flagged zero issues across the changed file.

Key Findings:

• Tilde expansion using process.env.HOME is a common and safe pattern for resolving user home-directory paths, and the implementation avoids shell injection since it does not invoke a subprocess.
• Wrapping file reads in try/catch with ENOENT-aware messaging improves UX by surfacing actionable errors rather than letting unhandled promise rejections propagate to the caller.
• Making language a required parameter (removing its default) is a deliberate correctness improvement that forces the LLM to prompt users explicitly, reducing silent mis-transcriptions.
• Coverage was complete (1/1 changed files reviewed) and the heuristic ceiling returned 5/5 with zero critical, significant, or medium issues identified.

Files requiring special attention

• src/tools/transcribe-audio.ts