Skip to main content

Chat

The Chat tab is a persistent OpenCode conversation for one agent. It maps to a dedicated OpenCode session and is independent of scheduled sub-agent runs. Workspaces register a local stdio HyperVibes MCP server that uses workspace- scoped agent credentials and never receives a Hyperliquid private key.

Start a conversation​

Choose a provider, model, and optional thinking mode at creation. Later turns retain that selection while OpenCode records per-message attribution. The transcript includes messages, tool activity, errors, token/context usage, cost, and compaction telemetry through live updates.

Compaction and deletion are OpenCode session operations available only while the session is idle. Deleting an agent removes its idle mapped sessions and agent-owned conversation records.

You can keep typing while the agent responds. Messages sent during a running turn are queued, shown as queued in the transcript, and submitted in order once the running turn finishes. Stop abandons messages still queued behind the running turn. Telegram messages also queue, with an acknowledgement in the chat. Up to 20 messages can wait in a conversation's queue. The queue lives on the server, so it survives closing or reloading your browser, but queued messages are lost if the server restarts.

The agent-conversations profile has no native shell, filesystem, web, or task tools and never reads the workspace .env. Its default read tools are:

  • hypervibes_get_account
  • hypervibes_list_strategy_prompts
  • hypervibes_get_strategy_prompt
  • hypervibes_get_trading_context
  • hypervibes_get_memory_detail
  • hypervibes_list_memories
  • hypervibes_list_orders
  • hypervibes_list_account_transactions
  • hypervibes_get_order
  • hypervibes_list_analysis_instruments
  • hypervibes_list_indicators
  • hypervibes_get_indicator
  • hypervibes_get_indicator_results

The header shows each conversation permission's current status. Use Manage permissions to change Orders, Memory, Notifications, Trade notes, Indicators, or Strategy prompts, then Save to apply them. Cancel discards your edits. Permissions can be changed while the conversation is idle.

Each permission supports Deny, Confirm (one-time approval for each action), or Allow. Configure an agent's starting permissions in Settings → Chat permission defaults, select Manage permissions, adjust the permissions in the modal, then select Save. Cancel discards your edits. New Web and Telegram conversations copy these defaults when they are created. Changing the defaults does not change existing conversations; use Manage permissions in a conversation to change its own permissions.

New agents initially default to Confirm for Orders, Memory, Trade notes, Indicators, and Strategy prompts, and Deny for Notifications. Existing agents keep their saved defaults. Read-only tools remain allowed. The Strategy prompts permission controls hypervibes_update_strategy_prompt.

Chat can also create or update an agent-owned indicator using validated PineScript-subset source. Each mutation follows the Indicators permission and records the conversation as its provenance. Chat cannot grant an indicator access to instruments outside the agent's analysis set. It first lists selected analysis instruments and uses those IDs unchanged. If none are selected, Chat directs the operator to Settings rather than guessing instrument IDs or attempting the mutation.

Notifications initially default to Deny. To let Chat send notifications, set Notifications to Allow for the conversation, or in the agent's defaults before creating a new conversation.

Discuss a prompt​

The Discuss prompt action on a role page opens Chat with the draft:

  1. The user selects a sub-agent prompt target and edits its draft.
  2. HyperVibes creates a conversation using the latest Chat model.
  3. The draft is included in the opening message when it fits the context limit.
  4. Chat can load the saved target prompt when the draft is too long.
  5. Any requested MCP update follows the conversation's Strategy prompts permission.

Chat prompt tools only access the current agent's targets. Scheduled sub-agents receive their prompt revision as input but cannot modify it.