fix/Full voice functionality in Discord when deployed via Modal/using modal backend

[twilwa]

Summary

This fixes the remaining Discord voice-mode regressions on the hosted Modal gateway and hardens the hosted runtime around the follow-up review feedback.

Changes in this PR:

• preserve Discord voice-session continuity by reusing the linked text-channel source metadata for voice input
• treat active voice-linked text channels as free-response contexts, so Hermes does not require a fresh @mention after /voice channel
• restrict that voice-linked exemption to the exact bound text channel, so unrelated threads under the same parent channel do not bypass mention-gating
• add the hosted Modal gateway runtime and install ffmpeg + libopus0 in the Modal container so Discord voice playback and transcription work in production
• make the Modal deploy wrapper import-safe when optional modal / fastapi dependencies are unavailable
• harden hosted bootstrap env handling for malformed base64 payloads and GitHub token alias normalization

Root cause

There were two primary runtime issues behind the broken hosted behavior:

1. The Modal container was missing voice runtime dependencies.
   Before the fix, hosted logs showed:

   • Opus codec not found — voice channel playback disabled
   • discord.errors.ClientException: ffmpeg was not found

   That prevented Hermes from speaking in voice and blocked the ffmpeg-based audio conversion path used for Discord voice input.

2. Voice-linked text channels were not being treated as the active conversation context.
   After /voice channel, Hermes could join voice, but the linked Discord text channel still fell back to mention-gating and failed auto-thread creation with:

   • 400 Bad Request (error code: 50024): Cannot execute action on this channel type

   That split voice turns from text turns and made continuity feel broken.

The follow-up review comments also caught two hardening gaps:

• importing the Modal deploy wrapper pulled in optional hosted dependencies too early for tests and non-Modal workflows
• hosted env bootstrap and token normalization paths were a little too trusting of malformed inputs and alias ordering

What changed

• Discord voice input now reuses the bound source metadata from the /voice channel session, which keeps voice and linked text replies in the same session.
• Active voice-linked text channels are now handled as free-response channels instead of trying to create unsupported auto-threads.
• The voice-linked free-response exemption now only applies to the exact linked text channel, not sibling threads under the same parent.
• Added a hosted Modal gateway wrapper/runtime.
• The Modal image now installs:
  • ffmpeg
  • libopus0
• Moved the deploy-time secret-name helper out of the Modal wrapper so unit tests do not depend on importing the hosted deploy module.
• Made scripts/modal_gateway.py import cleanly even when modal or fastapi are not installed.
• _decode_base64_env() now safely ignores malformed base64 / non-UTF-8 payloads instead of crashing bootstrap.
• normalize_github_token_env() now prefers already-present GITHUB_TOKEN / GH_TOKEN values before falling back to alternate secret names.

Testing

Automated:

• .venv/bin/python -m pytest tests/gateway/test_discord_free_response.py tests/gateway/test_voice_command.py tests/gateway/test_modal_runtime.py -q

Result:

• 200 passed

Manual:

• redeployed the Modal gateway
• verified /voice channel join
• verified two-way voice behavior in Discord after redeploy
• verified the linked text channel no longer needed a fresh @hermes to continue the active conversation
• verified the Modal deploy wrapper can still be imported cleanly without optional hosted dependencies installed

Notes

The hosted status dashboard retains historical log tail content from the persisted Modal volume, so older pre-fix ffmpeg / Opus errors may still appear in dashboard history. The new runtime itself starts cleanly and the voice path was manually verified after redeploy.

What does this PR do?

see above

Related Issue

#1559 is related, this addresses gaps in full-duplex voice mode when running via the Modal backend runtime.

Fixes #

Type of Change

• [ X] 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• [X ] 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

see above

How to Test

see above

Checklist

Code

• [X ] I've read the Contributing Guide
• [X ] My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• [X ] I searched for existing PRs to make sure this isn't a duplicate
• [ X] My PR contains only changes related to this fix/feature (no unrelated commits)
• [X ] I've run pytest tests/ -q and all tests pass
• [ X] I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• [X ] I've tested on my platform:

Documentation & Housekeeping

• [ X] I've updated relevant documentation (README, docs/, docstrings) — or N/A
• [ X] I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• [ X] I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• [ X] I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• [X ] I've updated tool descriptions/schemas if I changed tool behavior — or N/A

Screenshots / Logs

logs would be best included here but mine are a bit messy, if needed i'll unapply/test>save-logs>reapply/test>save-logs and include.

[Copilot]

Pull request overview

This PR fixes Discord voice-mode regressions when running Hermes via the hosted Modal gateway by preserving voice/text session continuity, relaxing mention-gating only for the active voice-linked text channel, and adding a Modal runtime wrapper (including required voice dependencies) plus bootstrap hardening.

Changes:

• Preserve Discord voice-session continuity by storing/reusing linked text-channel SessionSource metadata for voice input events.
• Treat the active voice-linked text channel as a free-response context (skip mention requirement + skip auto-thread creation), without exempting sibling threads.
• Add a Modal hosted runtime + deploy wrapper, including config/env sanitization, base64 bootstrap handling, and installing ffmpeg + libopus0.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
gateway/run.py	Reuses stored voice session source metadata; uses provider-aware STT defaults; avoids signal handlers outside main thread.
gateway/platforms/discord.py	Tracks voice-linked source metadata; exempts the exact voice-linked text channel from mention-gating and auto-threading.
gateway/modal_runtime.py	Adds Modal bootstrap/sanitization helpers and a hosted gateway service + dashboard HTML renderer.
scripts/modal_gateway.py	Adds an import-safe Modal deploy wrapper that builds a single-container ASGI dashboard and installs voice deps.
tests/gateway/test_voice_command.py	Adds coverage for source metadata reuse and provider-aware STT default behavior.
tests/gateway/test_discord_free_response.py	Adds coverage for voice-linked mention-gating behavior (and ensures threads still require mention).
tests/gateway/test_stt_config.py	Adds coverage ensuring transcription uses provider-aware defaults (no forced model kwarg).
tests/gateway/test_modal_runtime.py	Adds coverage for Modal runtime sanitization, bootstrap, dashboard rendering, and env normalization.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

yaml.safe_load(config_text) can raise yaml.YAMLError for malformed (but still base64/UTF-8 valid) payloads, which would crash Modal bootstrap and prevent the gateway from starting. Consider wrapping the YAML parse in a try/except (or using a helper that returns {} on any YAML error) and optionally recording a diagnostic in modal-bootstrap.json or logs.

[Copilot]

os.chdir(self.project_root) happens outside the try block, so a bad project_root (or missing directory) would terminate the gateway thread without setting _last_error or committing state. Move the chdir into the try and set _last_error on failure so the dashboard/health endpoint reflects the real startup error.

Suggested change

	os.chdir(self.project_root)
	try:
	try:
	os.chdir(self.project_root)

[twilwa]

copilot nits have been addressed, ready for review when yall are

[teknium1]

Merged via PR #8984. Your voice session continuity fix and signal handler guard were cherry-picked onto current main with your authorship preserved. The Modal hosted runtime portions were excluded per maintainer decision. Thanks for the contribution!