One card, end to end — MIKE
architecture node:
User input:
Representation at this boundary
Source modules
Loop map
One run — a card dragged into In Progress starts the executor → the provider slot and the memory gate admit it → its worktree and gVisor sandbox are created → the conversation is rebuilt from SQLite rows → the loop streams one completion, persists the turn, runs its tool calls over the stdio bridge, and calls again → the run settles in Review.
The store is the state — assistant rows and tool calls are persisted before anything runs on them, so a restart rebuilds the same conversation from rows. Nothing about a run lives only in a process: the board, the chat and the diff are reads of the same records.
Three ways a run stops —
request_review lands the card in Review for the person; ask_question stops the executor until the person answers (an auto-started subtask may instead be answered from its parent card); and the consecutive-failure limit (default 3) fails the run into Review. There is no iteration cap.