Channel Integrations

Tandem channels let users chat with the same engine sessions from Telegram, Discord, or Slack.

For the full runtime model behind this behavior, see How Tandem Works Under the Hood.

What channels do

Receive inbound channel messages.
Map user/channel identity to a Tandem session.
Send prompt parts (text, optionally file) to /session/{id}/prompt_async.
Stream run events and post replies back to the channel.

Supported channels

Telegram
Discord
Slack

Slash Commands

Channels expose a lightweight command surface so users can manage sessions and draft automations directly from chat.

Core commands

/help
/new [name]
/sessions
/resume <id or name>
/rename <name>
/status
/run
/cancel
/todos
/requests
/answer <question_id> <text>
/approve <tool_call_id>
/deny <tool_call_id>

Model commands

/providers
/models [provider]
/model <model_id>

Workflow-planning commands

/schedule help
/schedule plan <prompt>
/schedule show <plan_id>
/schedule edit <plan_id> <message>
/schedule reset <plan_id>
/schedule apply <plan_id>

Operator commands

/automations ...
/runs ...
/memory ...
/workspace ...
/mcp ...
/packs ...
/config ...

/help schedule and /schedule help both show the workflow-planning guide, and /schedule with no extra text defaults to the same help output. /help automations, /help memory, /help workspace, /help mcp, /help packs, and /help config show the namespaced guides for the new operator commands.

Configure and inspect status with:

GET /channels/status
PUT /channels/{name}
DELETE /channels/{name}

KB-first channel grounding

When a channel enables an MCP server marked as a knowledgebase, Tandem treats that connector as the grounding source for factual or project-specific channel answers.

Knowledgebase MCPs are identified by either:

purpose: "knowledgebase"
grounding_required: true
the hosted default MCP server name kb

For those sessions, the runtime injects a compact grounding policy and reshapes the prompt request to tool_mode: "required" with a KB-only MCP allowlist. The model must inspect KB evidence before it can produce a final answer. If the KB has no matching evidence, the channel reply should say that the enabled knowledgebase did not contain the answer instead of falling back to model memory.

Strict KB grounding

For hosted knowledge-bot demos or any channel that must answer only from retrieved KB material, enable the channel-level setting:

{
  "strict_kb_grounding": true
}

You can set it through PUT /channels/{name} alongside the rest of the channel config.

When strict_kb_grounding is enabled and the channel request explicitly allows a KB MCP such as ["mcp.kb.*"], Tandem:

still forces the KB tool turn first
rewrites the final channel answer from retrieved KB excerpts only
refuses with I do not see that in the connected knowledgebase. when retrieval has no supported answer
avoids adding generic platform instructions, inferred owners, or best-practice guidance that are not present in the KB
adds a compact Source: footer when the KB tool output exposes source paths

This mode is channel-scoped. It does not change ordinary chat behavior for channels that do not enable a KB MCP or leave strict_kb_grounding off.

Enable the hosted KB MCP for a channel scope with the normal channel MCP command:

/mcp enable kb

Or configure the channel tool preferences so the prompt request receives an explicit allowlist such as:

["mcp.kb.*"]

This grounding behavior is channel/session scoped. It does not automatically force KB usage for ordinary web chat or automations unless their request/tool policy explicitly enables the KB MCP.

Media and file uploads

When adapters are configured for media ingestion:

Files are stored under engine storage root in channel_uploads/....
Stored file references are attached to prompts as file parts:
- type: "file"
- mime
- filename (optional)
- url (local path, file://..., or remote URL)
Prompt also includes user text part (type: "text").

List uploaded files:

curl -s "http://127.0.0.1:39731/global/storage/files?path=channel_uploads&limit=200" \
  -H "X-Agent-Token: tk_your_token"

Storage layout

Typical pattern:

<state_root>/channel_uploads/<channel>/<chat_or_user>/<timestamp>_<filename>

Example:

/srv/tandem/channel_uploads/telegram/667596788/1772310564423_photo_305646779.jpg

Model and media compatibility

Image-capable providers/models can analyze image file parts.
Non-vision or unsupported models should still complete the run with a fallback response.
Channel adapters should avoid hanging on unsupported media and return clear user guidance.

Formatting notes (Telegram)

Telegram outbound formatting should use MarkdownV2-safe rendering:

Prefer parse_mode: "MarkdownV2"
Escape Telegram-reserved characters outside valid entities
Keep fallback retry as plain text on Telegram parse errors

Channel Memory

Channel sessions now keep memory in two layers:

Raw transcript history stays in normal Tandem session storage.
Retrieval memory stores exact user-visible completed user+assistant exchanges in global memory.

This lets future channel sessions recall prior work without loading full transcript history into every run.

For the storage-level breakdown of these layers, see Memory Internals.

What gets archived

After a successful channel reply, Tandem writes one chat_exchange memory entry containing:

the latest user message
the latest user-visible assistant reply
provenance such as session ID, project ID, workspace root, and message IDs

The engine dedupes retries of the same exchange automatically.

Why this stays small and safe

Tandem archives one completed exchange, not every partial stream event.
Raw session history remains the source of truth even if retrieval memory is compacted or re-ranked later.
Prompt context still stays bounded by normal memory search limits rather than unbounded chat log replay.

For the engine-side explanation of this layering, see How Tandem Works Under the Hood.

Workflow planning from channels

The /schedule command family is a thin wrapper over Tandem’s workflow-plan endpoints.

/schedule plan <prompt> creates a workflow draft from a natural-language request.
/schedule edit <plan_id> <message> revises the draft conversationally.
/schedule apply <plan_id> saves the draft as an automation.

When the current channel session is already bound to a workspace, Tandem automatically forwards that workspace root into the planner so the resulting workflow targets the right repo or project by default.

Expanded operator commands

Channels now support a broader operator surface on top of existing engine APIs:

/automations to list, inspect, run, pause, resume, and delete saved automations
/runs to inspect recent automation runs and their artifacts
/memory to search, review, save, and delete memory entries
/workspace to inspect the current workspace binding, search files, and check the git branch
/mcp, /packs, and /config for integration/runtime inspection and lightweight control

Destructive commands require --yes, while read/list/show/search commands execute immediately.