Skip to content

Workflow Architecture

Every message runs through a four-step Agno Workflow:

on_message (ChatCog)
    └── Workflow.arun()
            ├── Step 1: FetchHistory
            ├── Step 2: LLMChat
            ├── Step 3: ExtractRenderTables
            └── Step 4: SendResponse

The workflow is defined in dango/workflow.py and created once at startup via create_discord_workflow().

Step 1 — FetchHistory

File: dango/steps/fetch_history.py

Pulls recent messages from the channel using the Discord API and converts them to Agno Message objects. Key behaviours:

  • Stops at a [new chat] marker (dropped by /newchat)
  • Respects history_limit from runtime.yml
  • Resolves dango_deep_*.json attachments back to proper user turns (used by /deep)
  • Builds a mention_map (<@123> → "Alice") for mention resolution in the next step

Step 2 — LLMChat

File: dango/steps/call_agent.py

Runs the Agno Agent against the formatted history plus the current message. Key behaviours:

  • Selects fast or deep agent based on AUTO_ROUTE, _force_deep flag, or URL-upgrade logic
  • Resolves @mention tokens in message content using the mention_map
  • Downloads image attachments and passes them to the model
  • Trims history to CONTEXT_TOKEN_BUDGET before sending
  • Handles retries (Gemini: model-level; others: 2× at agent level) and bidirectional fallback

The Agno Agent uses a dynamic instructions callback that reads session_state on every call, injecting the user's display name and current time into the system prompt when ENABLE_CONTEXTUAL_SYSTEM_PROMPT=on.

Step 3 — ExtractRenderTables

File: dango/steps/table_steps.py

Scans the LLM response for Markdown tables and renders each one as a PNG using Pillow. Noto Sans CJK fonts are used so Chinese, Japanese, and Korean text renders correctly.

If no tables are found, this step passes through transparently.

Step 4 — SendResponse

File: dango/steps/send_response.py

Sends the final response back to Discord:

  • Splits messages longer than Discord's 2000-character limit across multiple messages
  • Attaches rendered table PNG files (and cleans up the temp files afterward)
  • Replies to the original message when it can still be fetched, otherwise falls back to channel.send()
  • Appends any [dango-sysinfo] notes (e.g. fallback notifications)
  • Handles error messages forwarded from earlier steps

Entry points

The workflow is invoked from two discord.py Cogs:

Cog File Triggers
ChatCog dango/commands/chat_commands.py on_message, /newchat, /deep
AdminCog dango/commands/admin_commands.py All admin slash commands

The workflow is driven by ChatCog.on_message: it fires for every Discord message, checks whether the bot should respond (mention, allowed channel, or allowed DM user), builds the message_data structure, and calls workflow.arun(). The /deep command runs the same pipeline with the deep model forced on.

Custom slash commands and agent tools added via dango.extensions plug in separately — see Custom Commands & Tools.