Channel capability matrix
Not every channel supports every feature. Telegram edits messages in place; email cannot. Slack uploads files; WhatsApp does not. The gateway reads each adapter's declared capabilities and degrades gracefully — a streamed reply falls back to a single message where edits are unavailable, and outbound media falls back to text where uploads are unsupported. This page is the authoritative lookup for what each adapter supports today.
Each capability is declared on the adapter itself (the canSendTyping / canEditMessage / canReact / canSendFiles flags and the capabilities manifest), so this table tracks source, not aspiration.
Support matrix
| Capability | Telegram | Slack | Discord | ||
|---|---|---|---|---|---|
| Typing indicator | ✓ | ✓ (probed) | ✓ | ✗ | ✗ |
| Streaming draft edits | ✓ | ✓ | ✓ | ✗ | ✗ |
Message edit (editMessage) | ✓ | ✓ | ✓ | ✗ | ✗ |
| Inbound media | ✓ | ✓ | ✓ | ✓ | ✗ |
| Outbound media | ✓ | ✓ | ✗ | ✗ | ✗ |
| Reactions | ✓ | ✓ | ✓ | ✓ | ✗ |
| Threads / topics | ✓ (forum topics) | ✓ (thread_ts) | ✓ | ✗ | ✗ |
| Reply-to a message | ✓ | ✗ | ✗ | ✓ | ✗ |
| Approval buttons | ✓ | ✓ | ✓ | ✗ | ✗ |
| Slash commands | ✓ (menu) | ✓ | ✓ | ✓ (typed) | ✓ (first line of body) |
| Voice in (transcribed) | ✓ | ✓ | ✓ | ✓ | ✗ |
| Voice out (TTS) | ✓ | ✓ | ✓ | ✓ | ✗ |
| Voice-out rendering | voice bubble | file + inline player | file + inline player | voice bubble (ptt) | — |
| Webhook mode | ✓ | ✗ | ✗ | ✗ | ✗ |
| Max message length | 4096 | 3000 | 2000 | 65536 | 100000 |
✓ (probed) — Slack's typing indicator uses an unofficial API; the adapter probes it once at runtime and reports the real result thereafter, so canSendTyping reflects what actually works on the workspace.
Slash commands. The gateway commands — /stop, /new, /usage, /budget, /personality, /mute and the rest of the gateway surface in SLASH_COMMANDS (packages/surface-kit/src/slash-commands.ts) — are parsed by Gateway.handleMessage with no platform gate, so they work on every channel. WhatsApp has no command menu: type the command as a message. On Email, put the command on the first line of the body and reply in the same thread (the subject picks the conversation); the adapter drops the quoted reply and signature below a command line (commandOrBody in extensions/platform-email/src/index.ts). A message from an unverified email sender is never read as a command: the adapter prefixes its body with a sender warning, so the first word is not a command (pinned by extensions/platform-email/src/__tests__/email-adapter.test.ts, 'EmailAdapter slash commands'). Telegram's / menu is built from the same registry at startup.
Voice in requires the adapter to classify an inbound upload as type: 'audio' — the gateway's transcription gate keys on nothing else. All four chat adapters do: Telegram from voice and audio messages, Slack and Discord from the upload's extension or content type, WhatsApp from audioMessage. A .webm upload on Slack or Discord stays a video and is not transcribed.
Voice out is a separate declaration — voiceCaps plus sendVoiceNote, see Voice output caps below — which is why a channel can speak without being able to listen.
Voice output caps
Voice output is declared per adapter as an AdapterVoiceCaps object, not inferred from method names. The gateway reads it through isVoiceOutboundAdapter and transcodes each reply into the first format the target accepts.
| Platform | Accepted outbound formats (preferred first) | Rendering (kind) | Platform flags | Size cap |
|---|---|---|---|---|
| Telegram | opus, ogg, mp3 | voice_note | — | 50 MB |
| Slack | mp3, m4a, wav | file | — | 25 MB |
| Discord | mp3, ogg, wav | file | — | 25 MB |
opus, ogg | voice_note | ptt: true | 16 MB | |
| — | — | — | — |
voice_noterenders as a playable bubble;filerenders as an upload the platform gives an inline player. Neither Slack nor Discord has a bot-accessible voice-bubble primitive, so both declarefilerather than claiming one.- Producing the accepted container needs
ffmpegon the gateway host unless the TTS provider already emits it. Without ffmpeg, a mismatched reply is skipped (gateway.voice_format_unsupported) rather than delivered unplayable. - A reply over the size cap is skipped with
gateway.voice_too_large; the text reply has already been delivered.
How degradation works
The gateway never assumes a capability. Three behaviors depend directly on this matrix:
- Streaming draft edits (W3.1). When a chat is streaming-enabled and the adapter reports
canEditMessage, the gateway delivers the reply as throttlededitMessageupdates that grow in place. When the adapter cannot edit (WhatsApp, Email), the reply is delivered as a single final message instead. Streaming defaults on for direct messages and off for group chats; setdisplay.streaming_editsin~/.ethos/config.yamltooff,dms, orall. - Outbound media (W3.2). When a tool produces media and the adapter reports
canSendFiles(Telegram, Slack), the gateway maps it to native attachments (sendPhoto/sendDocumenton Telegram,files.uploadV2on Slack). When the adapter cannot send files, the reply degrades to the text summary — no error, no dropped turn. - Spoken replies. When a conversation's voice mode says speak and the adapter declares
voiceCaps, the gateway synthesizes the reply once, transcodes it to the first accepted format, and callssendVoiceNote. When the adapter declares no caps, nothing is synthesized and the turn ends with the text reply already delivered — the skip is recorded as agateway.voice_no_capsevent rather than passing silently. See Send and receive voice notes on a channel.
Source
Each adapter declares its capabilities in source. Consult the adapter for the exact runtime behavior:
The capability contract itself — ChannelCapabilities and the legacy AdapterCapabilities — lives in packages/types/src/platform.ts.