agent-workflow
A test run that cannot reach production
Local end-to-end runs get their own throwaway database and a shared lock, and the retrospective that followed turned four agent mistakes into rules the repository enforces.
An agent that cannot play a word
Phase 1 of player agent tools exposes read-only game state and dry-run previews over WebMCP, while leaving typing and submission strictly to the human.
The missing step before MP4
Restoring a missing image package lets the gameplay recorder finish its WebM-to-MP4 conversion.
One frame for video and stills
The gameplay recorder moves to a phone-sized frame and captures full-screen marketing stills during the same run as its videos.
A scaffold for the workbench
Migrating the devcontainer from a bespoke compose stack to a standardized single-container scaffold supporting four distinct AI assistant CLIs.
The server that ran somewhere else
Wiring up Sentry hosted Model Context Protocol server inside the devcontainer, routing remote HTTPS queries through the egress firewall, and keeping access tokens out of version control.
Three aliases, one posture
Fixing a mount-namespace bug that broke git worktrees across autonomous CLI aliases, dropping phantom sandboxes, and persisting workspace ownership across container rebuilds.
The game was still there
A player replaying the Daily Challenge who switched tabs mid-game came back to the previous attempt's recap instead of the game still running underneath it.
Gone by the next restart
A prototyping rule edited straight into an installed third-party skill file looked shipped, until a devcontainer restart quietly reverted it, for two reasons the agent had missed.
The cap that was already gone
A month after the PvP lives system replaced the 40-word duel cap and the single-collision ending, the glossary and architecture docs still described the old rules.