JEV-Star
Explore Jev macro actions and micromanagement in reproducible StarCraft II experiments.
Added to Jevfast
How it uses Jev
JEV-Star sends a structured StarCraft II observation and the currently available Protoss macro actions to Jev as one choice question. Jev selects one legal action. The bot validates the returned choice against the offered action IDs, then executes it if it remains available. In constrained planner mode, Astra can filter candidates; in advisory mode, its plan is context for Jev.
What Jev decides
Source-defined Protoss action test requestExample answers · not a recorded Jev response · Source ↗
Question 1 · next macro action
YOUR APP
INSTRUCTION
Choose one action from the currently available action criteria that advances the Protoss mission.
STATE
The repository test supplies state {minerals:50} with exactly two legal candidate actions:0 TRAIN PROBE and71 EMPTY ACTION. This is a mocked source-owned request example, not a recorded game result.
JEV · CHOICE
- 0
- 71
App workflow
- Observe the game
The bot builds a structured snapshot with current game information, visible enemy counts, army intent, navigation, supply forecast, base threat, ready idle producers, and recent outcomes.
- Generate legal choices
The bot checks game abilities, resources, prerequisites, cooldowns, and executor state. It removes unavailable actions and records why they are blocked. A constrained Astra plan may further filter candidates.
- Submit asynchronously
When the scheduler is ready and more than one action is available, it deep-copies the snapshot and choices and starts a Jev request. The game loop does not wait for the HTTP request.
- Validate the response
The client checks the returned choice ID and probability distribution against the submitted criteria. The scheduler later consumes the completed result.
- Recheck and execute
The bot recalculates available actions and discards a stale or newly unavailable action. Otherwise, its local executor performs the selected game command.
Why it is interesting
The project uses Jev as a real-time game controller. A non-blocking scheduler submits snapshots at a configured interval while the game continues, then the bot checks the returned action against fresh game state before execution. The repository also compares Jev with random and planner-assisted configurations.