Agentic loops, projection caps, and steer regressions

25 May 2026 · Dirt Bowl Rally (offroad)

Progress

Ran an agentic 1-lap arc (tester → director → coder): ten loops plus an extra harness try. Best scripted progress: 59.3% of one lap (tp=2396, loop 6) after advanceTrackS() caps per-frame arc-length jumps (bug-lap-steer-incomplete).

In-repo gate script: games/offroad/scripts/lap-from-grid.mjs (t-regression-matrix).
Docs: doc/offroad-design.md, doc/offroad-workflow.md (t-c7ef).
Progress tracking consolidated into t-loop-1lap-gate (no duplicate /log tables).
README lists offroad as game #1; hub + journal scaffold (t-4b51).

Arc 2 (loops 12–21) is queued; baseline after loop 11 is tp=1819 (45%).

Learnings

Do not auto-steer from the harness. Both DOM KeyboardEvent injection and Puppeteer centerline-following on pure-W segments dropped progress from ~53–59% to ~32–45% (t-f031). The scripted W+A/D timing is part of the test; “fixing” steer in the runner masks physics bugs and makes runs worse.

Pins (fold, hairpin, south wall) are largely solved; remaining gate blockers are projection monotonicity, lap counter on a self-crossing track, and eventually harness/replay that can actually close the circuit.

Failures

Gate never passed — Lap completed (0→1, monotonic tp): false across all loops. Postmortem: t-agentic-1lap-postmortem.

Loop 7–8: pace + auto-steer → tp crashed ~32% (reverted).
Loop 11: Puppeteer steer assist → 45% in 152s (reverted).
Milestones M2–M4 never pass at 45–60s budgets — mid/late track pace or tp jumps still wrong.

Screenshots

Late-lap frame from lap gate script showing truck on dirt track — Lap-gate run — late progress frame (`/tmp/offroad-play/lap-steer-*`).

More frames from 24 May playtest: pins & wall-slide entry.