feat(platforms): add Carbon Voice as a platform plugin

[cristianmgm7]

Summary

Adds Carbon Voice as a bundled platform plugin under plugins/platforms/carbonvoice/, following the same pattern as the other 21 platform plugins. Carbon Voice is a voice-first platform: it transcribes inbound voice to text (STT) before delivery and can re-synthesize the agent's replies as voice memos (TTS), so from Hermes' side it's a text-in / text-out platform. The adapter connects over Socket.IO (primary) with a REST polling fallback — no public webhook or tunnel required — and persists a disk cursor so messages received while offline are processed on next startup.

Rework per review: the previous revision of this PR integrated Carbon Voice as a native platform (enum branch, factory branch, auth maps, cron map, send-tool routing, toolset, wizard entries — 9 core files). As requested, it is now a platform plugin registered via ctx.register_platform(): env enablement, cron delivery, standalone sending, auth metadata, and prompt hints all come from the existing PlatformEntry hooks. The PR now contains zero core changes — the diff is the plugin subpackage, its tests, and docs.

Review feedback addressed

1. voice_out_carries_text was read by the adapter but ignored by core (audio memo + duplicate text). The general delivery contract lives in fix(gateway): thread auto-TTS reply + support voice_out_carries_text adapters #32655 (fix(gateway): thread auto-TTS reply + support voice_out_carries_text adapters), which is further along in review and hardened per that review's feedback: a completeness guard (a truncated/stripped TTS rendition never suppresses the full text), reply_to threading for the voice memo, and behavioral tests. This PR only declares the flag on the adapter; until fix(gateway): thread auto-TTS reply + support voice_out_carries_text adapters #32655 lands the attribute is ignored and voice-out delivers audio plus a duplicate text bubble — degraded UX, never lost content. Keeping the mechanism in one PR avoids two competing implementations of the same base.py block.
2. Voice-out configuration not wired. The plugin's env_enablement_fn seeds extra["voice_out"] from CARBONVOICE_VOICE_OUT, and the chain env → seed → PlatformConfig.extra → adapter flag is now covered by tests (tests/gateway/test_carbonvoice_plugin.py::TestVoiceOutWiring).
3. No credential-scoped lock. connect() now acquires _acquire_platform_lock("carbonvoice-pat", …) before opening the API client (same pattern as the telegram/discord/slack/whatsapp adapters) and disconnect() releases it. Tested (deny → bail out before api.open(); release on disconnect).

What's included

• plugins/platforms/carbonvoice/ — plugin subpackage (adapter, api, transport, state, dedupe, reactions, channels, audit, permits, parse, gate, conversations, constants) + plugin.yaml with requires_env/optional_env for the setup wizard.
• register(ctx) provides: env_enablement_fn (auto-enable from CARBONVOICE_PAT), standalone_sender_fn (out-of-process cron delivery), cron_deliver_env_var="CARBONVOICE_HOME_CHANNEL", auth env vars, platform_hint, setup_fn.
• Docs: website/docs/user-guide/messaging/carbonvoice.md + messaging index entries.
• Tests: tests/gateway/test_carbonvoice_plugin.py (15 — plugin registration hooks, voice-out env wiring, credential lock).

Dependencies

None added. httpx is already core; python-socketio is optional — without it the adapter runs in polling-only mode (install_hint covers it). No pyproject.toml / uv.lock changes.

Access control

Deny-by-default: the bot answers only the owner (whoami.created_by, auto-detected at connect) plus CARBONVOICE_ALLOWED_USERS and runtime-approved users. Unknown senders are dropped silently; the owner gets a rate-limited prompt and can approve with a one-tap reaction (💯 allow / 👎 block) or /cv-allow-user. Approvals reuse the existing PairingStore, so they authorize in core too.

Testing

• tests/gateway/test_carbonvoice_plugin.py (15) — all passing.
• Regression: tests/plugins/ + startup no-eager-install (1855 tests) and the voice/telegram/drop-recovery suites (212 tests) pass; ruff clean.

[teknium1]

Thank you for the substantial Carbon Voice implementation. I found blocking integration issues before this can be safely salvaged.

Problems

• gateway/platforms/carbonvoice/adapter.py:131 relies on voice_out_carries_text, but current gateway/platforms/base.py:5000-5027 suppresses follow-up text only for a successful Telegram caption. Carbon Voice voice-out therefore delivers both the audio memo and text.
• adapter.py:164 reads only config.extra['voice_out']; the PR's config bridge does not seed that field, while website/docs/user-guide/messaging/carbonvoice.md tells users to set CARBONVOICE_VOICE_OUT=true.
• adapter.py:289-360 has no credential-scoped lock. Current adapters use the base helpers at gateway/platforms/base.py:2752-2779 to prevent two profiles from claiming one credential.

Suggested changes

• Please rework this as a platform plugin. gateway/platforms/ADDING_A_PLATFORM.md:5-15 recommends that route for community/third-party platforms, and gateway/platform_registry.py:113-159 already provides env enablement, cron delivery, standalone sending, auth metadata, and prompt hints without new core branches.
• Wire and test voice-out configuration, then add a tested general delivery contract for Carbon Voice audio/text suppression.

Automated hermes-sweeper review.

[teknium1]

voice_out_carries_text is not consumed by current main. BasePlatformAdapter suppresses the follow-up text only for a successful Telegram caption (gateway/platforms/base.py:5000-5027), so voice-out will produce both an audio memo and a text reply. Please add the general base-layer contract plus coverage, or remove this duplicate-suppression path.

[teknium1]

This reads only config.extra['voice_out'], but the documented CARBONVOICE_VOICE_OUT=true is never bridged into extra by this PR's gateway/config.py changes. Wire the supported configuration path and add an env/config propagation test.

[teknium1]

Acquire a credential-scoped lock before opening the API/Socket.IO connection and release it in disconnect(). Current BasePlatformAdapter._acquire_platform_lock() / _release_platform_lock() provide the profile-safe mechanism used by token-backed adapters.

[cristianmgm7]

Thanks for the review — reworked as requested. The branch has been rebuilt from scratch on current main as a platform plugin (force-pushed; PR title/body updated to match).

Per-point:

1. voice_out_carries_text ignored by core (audio + duplicate text) → added the general delivery contract: a documented voice_out_carries_text: bool = False capability flag on BasePlatformAdapter, honored in the auto-TTS flow next to the existing Telegram caption check. Default False (no behavior change for other platforms); a failed play_tts() still falls back to the text send. Covered by tests/gateway/test_voice_out_carries_text.py (suppression, default-unchanged, TTS-failure fallback). This is now the only core change in the PR (gateway/platforms/base.py, +21/−2).
2. CARBONVOICE_VOICE_OUT not seeded → the plugin's env_enablement_fn seeds extra["voice_out"]; the env → seed → adapter chain is tested in tests/gateway/test_carbonvoice_plugin.py::TestVoiceOutWiring.
3. No credential-scoped lock → connect() acquires _acquire_platform_lock("carbonvoice-pat", …) before opening the API client and disconnect() releases it, same as the telegram/discord/slack/whatsapp adapters. Tested (lock denied → bail before api.open()).

Plugin rework: everything lives in plugins/platforms/carbonvoice/ and registers via ctx.register_platform() with env_enablement_fn, standalone_sender_fn, cron_deliver_env_var, and the auth env vars — no enum/factory/authz/cron/send-tool/wizard branches in core anymore. No dependency changes either: python-socketio is optional (polling-only fallback without it).

Regression: tests/plugins/ + startup no-eager-install (1855) and the voice/telegram/drop-recovery suites (212) pass.

[cristianmgm7]

Follow-up: dropped this PR's own voice_out_carries_text implementation in base.py — the same delivery contract already exists in #32655, where it's further along in review and hardened per that feedback (completeness guard + reply_to threading + behavioral tests). Keeping the mechanism in one place avoids two competing implementations of the same block. This PR is now strictly zero core changes: the plugin subpackage, its tests, and docs. The adapter declares the flag; #32655 makes core honor it.