feat: Voice Dev Terminal — multi-turn voice-controlled terminal via Local Link

[hassan1731996]

What This Does

A voice-native developer terminal session for OpenHome. Say "terminal", then keep talking — run git commands, check disk space, find files, tail logs — all hands-free, staying in session across multiple commands.

Why It's Unique

Every other ability in this repo is either one-shot (execute one thing, exit) or uses external APIs. This is the only ability that uses exec_local_command() via the Local Link bridge — voice → LLM-generated shell command → real terminal execution on the user's machine → spoken LLM summary of the output.

Key Differentiators vs the openhome-local-link Template

	openhome-local-link template	Voice Dev Terminal
Session	One command, exits	Stays open for multiple commands
LLM context	None — single turn	Full conversation history
Output	Raw string spoken verbatim	LLM-summarized into natural speech
Dangerous commands	No guard	Voice confirmation required
Inline trigger	No	"terminal, git status" runs immediately

UX Flow

"terminal"
→ "Terminal ready. What do you need?"

"what's my git status?"
→ "Running that now."
→ "You're on branch main, 2 files modified and not staged."

"show recent commits"
→ "Running that now."
→ "3 commits in the last day. Latest: fix auth bug, 2 hours ago."

"remove the build folder"
→ "That runs: rm -rf ./build. Confirm?"
→ [user: "yes"] → "Running that now." → "Done."

"done"
→ "Got it."

Technical Notes

• Uses exec_local_command(cmd, timeout=30.0) — requires openhome local start running on the user's machine
• LLM (sync text_to_text_response) used twice per turn: command generation + output summarization
• Conversation history passed to both LLM calls so relative references ("do it again with verbose") resolve correctly
• Danger detection: token-set intersection on {rm, rmdir, sudo, shutdown, reboot, dd, mkfs, kill, pkill}
• No persistent storage — terminal sessions are ephemeral by design
• resume_normal_flow() called in finally covering all exit paths

Validator

✅ All checks passed!

[github-actions]

🔀 Branch Merge Check

PR direction: feature/voice-dev-terminal → dev

✅ Passed — feature/voice-dev-terminal → dev is a valid merge direction

[github-actions]

✅ Community PR Path Check — Passed

All changed files are inside the community/ folder. Looks good!

[github-actions]

✅ Ability Validation Passed

📋 Validating: community/voice-dev-terminal
  ✅ All checks passed!

[github-actions]

🔍 Lint Results

✅ __init__.py — Empty as expected

Files linted: community/voice-dev-terminal/main.py

✅ Flake8 — Passed

✅ All checks passed!

[uzair401]

Hello @hassan1731996, thanks for submitting this PR.

We already have templates available for this functionality: Local Agents Controller, Hermes Controller, and OpenClaw Controller. They handle these tasks through a single WebSocket endpoint, which has already been implemented in Local Link.
Because of this, I think there may be some overlap with existing functionality. Please review these templates and clarify what unique capabilities this PR adds.

[hassan1731996]

Thanks for the review, Uzair. I've looked at all three templates you mentioned and updated the README to document the distinction clearly.

OpenClaw and Hermes are WebSocket bridges to persistent agent processes running on the user's machine — they're designed for agent-based delegation workflows and require the OpenClaw or Hermes software to be running. Voice Dev Terminal doesn't use that path at all.

openhome-local-link is the closest comparator, and the README now explicitly addresses it. The core differences:

• Multi-turn session — the local-link template is one-shot (runs one command and exits); Voice Dev Terminal keeps the session open so the user can run multiple commands without re-triggering
• Conversation history — commands are resolved in context ("do it again but verbose", "now in the src folder"), which requires carrying history across turns
• Danger confirmation — tokens like rm, sudo, kill, dd trigger a voice confirmation step before the command runs; the template has no safety guard

I've pushed the updated README to the branch. Happy to adjust anything.