Show the Bots How to Move

The Bots Had Almost Nowhere to Go

On DM2, none of 63 floor points had a walking link. A height mix-up made ordinary floors look disconnected, leaving bots to improvise between spawns, items and teleporters.

Testing from standing height restores 272 walking links on that map. Route checks also use short runs of the player's own movement, so stairs, ramps and small obstacles no longer turn every corridor into a dead end. Jump-pad routes are flown to their real landing positions.

In the release's bot tests, Search & Destroy on wbomb2 gets a planted and armed bomb in three of four rounds. A DM12 Duel goes from 0–0 to 5–4. These are specific test matches, not a claim that every map is solved.

Record the Trick

The new training_demo command starts a recorded warmup session. Perform the wall jump or high jump, stop the recording, and the bot learner finds the attempts.

It replays each trick with the game's movement, finds a run-up that works from standing, and saves the routes that land reliably. Bots load those scripts into their navigation graph.

DM2's Red Armor becomes reachable through the wall jump above the teleporter. Actually choosing that armor in a live match is still the next step, in v0.9.28.

The Editor Can Take Instructions Too

SHAGGED gains a JSON command line and an MCP server, using the same map document the editor opens. Scripts can create rooms, corridors, stairs, teleporters, jump pads and textured brushes, then compile and check the result.

The first fresh-agent trial built a three-area map. Its upper room was sealed off, and the attempted repair leaked. The reports caught both. The run ended at a usage limit before the map was finished; the useful result is that the tools reported the failures rather than calling it playable.

Measure Before Promising

Offline timedemos gain repeatable workloads, CPU/GPU diagnostics and profiling tools. Off-screen player bodies are culled, medium-distance poses update at 30 Hz, and gib fragments use collision brushes instead of render-triangle tests.

The ten-player benchmark reports lower median frame times on both renderers. Those are workload measurements, not a universal FPS promise.

WebGPU weather moves to GPU particles. Render bundles remain opt-in and are not yet faster; browser-route and paired-hardware benchmark acceptance are pending, and fixed-work throughput mode is not implemented.

A Little More Bite

Head hits rise from 10% to 25% extra damage, except for the Riotgun. Leg hits retain their 25% reduction. The machinegun gets spread, tracer and kick work, and Amphi stops carrying the Quad into the following round.