space

Appearance

Agent runner system ​

Summary ​

Runtime agents that use LLM providers, local deterministic functions, and other agents to perform tasks. Sidebar chat sessions are now workflow-backed: each chat is one long-lived workflow run, and the runner API projects current session state from workflow events while preserving existing panel behavior.

Motivation ​

LLM-based agents needed a runtime layer beyond the raw OpenAI conversation client. This system provides:

  • Agent runner with workflow-backed session lifecycle, tool execution, and approval flow
  • Preset system exposing context-aware capabilities (MCP tools) with risk badges
  • Streaming replies from OpenCode and other providers
  • Agent scope within reloadable units

Design ​

  • Workflow-backed runner facade: Preserves the sidebar runner API (create-session, get-session, list-sessions, run-turn, cancel-turn, delete-session, flush, drop) while storing current chats as workflow runs.
  • WorkflowStore: Owns agent chat workflow definitions, long-lived runs, run steps, and event history.
  • Event projection: Projects transcript, status, and session data from workflow events such as :agent-session-created, :agent-status-changed, and agent item events. Old JSON session files are not the current chat source of truth.
  • PresetRegistry: Defines available capabilities as presets with three-state control (Auto/On/Off) and risk badges
  • MCP transport: Remote OpenCode MCP transport for tool exposure across processes
  • Conversation-first supervisor: Streaming with tee for parallel consumers
  • Approval: Toggle-driven — preset toggle is user action; tool execution remains model action with approval checks

Workflow-backed session flow ​

  1. The sidebar calls create-session, which ensures the editable default agent workflow exists and starts a long-lived workflow run.
  2. The workflow waits for :agent-user-input.
  3. Each run-turn appends the user message as a workflow event, resumes the waiting step, and lets the provider-specific turn controller stream items/status back as more workflow events.
  4. The sidebar and graph read the same projected workflow run/session state.
  5. Legacy agent-sessions/*.json files are migrated only by the explicit tools.agent-session-migrate command, not during app startup.

Tasks ​

  • [x] Agent runner with session lifecycle
  • [x] Preset system with MCP tool exposure
  • [x] Risk-based approval model
  • [x] OpenCode streaming integration
  • [x] Remote MCP transport
  • [x] Conversation-first supervisor
  • [x] Workflow-backed sidebar sessions and explicit legacy migration