The people in Act II of my point-and-click adventure used to stand in one place. Between the evening of November 15 and the next noon they started to walk: Caterina between her counter and a stack of bales, Luka to a barrel, Peter to his fishing nets, the boatman Zorzi along the quay. The feature had 24 green tests. Then a playtest found that Martin, the hero, walked straight through all four of them.

Animation in two rows. First playtest: Martin crosses the warehouse, walks straight through Caterina standing at her counter and on behind the counter. After the first fix: the same click takes him past her in front of the counter

It took five more rounds on November 16 to get from "through" to "past". The game is the one from the four-day prototype and the finished first act; it is built in Godot by Claude Code sessions, and I decide.

Who tests and who fixes

One session writes the code. A playtest is done by a subagent with an empty context: it plays the build, records, measures, and writes a report. Its brief ends with "Fix nothing, commit nothing, publish nothing", and every round of the report opens by confirming it. A tester that may not repair what it finds has no reason to find less.

The feature had numbered criteria. Two of them carried this story:

  • Criterion 9: Martin and a person never pass through each other.
  • Criterion 2: feet do not slide on the floor.

The first playtest

It found three severe faults. Two matter here. Martin's floor ran 0 to 20 px from the places where people stand, so his path crossed them. And Peter slid sideways at about 80 px per second, because he was drawn from the back while his path ran diagonally.

Round 1: the tests were green at 0.36 m

The first fix cut holes into Martin's floor around every place of work and added 8 replay tests built from the report's own clicks. Green again.

For the re-test the tester changed its unit. Screen pixels do not say whether two bodies overlap, because the rooms have depth: along Peter's path alone the drawing scale goes from 1.00 to 0.81. So a script converted both pairs of feet to floor coordinates through the room's camera and logged the distance in metres for every frame. The tester picked the limit itself and wrote down why: a body is about half a metre wide, so under 0.5 m two bodies are inside each other.

It measured 0.36 to 0.47 m beside all four people. The report explains the green tests in one sentence: the fixer's limit was 0.6 of Martin's reach to the side, and that is exactly where the new floor ends. The test checked that Martin stays on the floor the fix had just drawn.

Two rows of three game frames: Martin beside Caterina at her counter, beside her at the stack of bales, and beside Zorzi on the quay. In round 1 the feet are 0.36 to 0.47 metres apart and the bodies overlap by a third to a half. In round 2 the same three places measure 0.53 to 0.64 metres and only the shoulders touch

The round also found a fault that the fix had created. A person did not leave their place for 60 seconds and more while Martin stood beside it. The cause, read from the code: a person counted Martin as "in the way" if he was anywhere in the half-plane in front of the first leg of their path. The new floor made that first leg point straight at the camera, so a Martin standing to the side and a little lower was always ahead.

Rounds 2 and 3: a fix that moved the bug

The second fix widened the holes to about 0.59 m and made a person look for Martin only on the stretch they are about to walk. The tester measured 0.53 to 0.65 m: shoulders touch. One spot was still under the limit, the right side of Luka's barrel at 0.48 to 0.51 m, where Martin half covered him. Round 3 closed it at 0.55 m.

Chart of the closest distance between feet beside seven places, in metres, with a dashed line at 0.5 m. Round 1 is under the line everywhere, from 0.36 to 0.46. Round 2 is at or over it, 0.50 to 0.64, except the right side of Luka's barrel at 0.48, which round 3 moves to 0.55

And Peter started to twitch. With Martin standing left of the nets, every 4.3 seconds Peter turned, took a 2.5 px step and jumped back, 11 to 14 times in 45 to 60 seconds. The check on setting out and the check while walking disagreed by a few pixels.

The third fix used a slightly smaller box while walking. The twitch was gone. Now Peter left his nets every 22 to 25 seconds, walked three seconds to the corner of his path, stood there for four, and went back. The tester called it "changed, not fixed" and named the cause: the box was measured at Peter's own depth, and 32 px closer to the camera the same box is bigger.

The fourth fix is one argument (comment shortened):

func _px(px: float, y := NAN) -> float: # a distance tuned in Act I's screen px; in a camera room the same in metres if owner and owner.get("camera"): return px / owner.SCREEN_PX_PER_M * owner.px_per_m(position.y if is_nan(y) else y) return px 
var row: float = (owner as Node2D).to_local(point).y if owner is Node2D else NAN # at that point's depth if absf(him.x - point.x) < _px(60.0, row) * size and absf(him.y - point.y) < _px(40.0, row) * size: return true 

The box is now measured at the depth of each point on the path. In round 4 Peter stayed at his nets for 65 and 61 seconds and did 14 work actions. The test for this case fails without the fix, which the first replay tests could not say.

Rounds 4 and 5: a deterministic game needs varied timing

Round 4 left one stand-off. Martin passed behind Caterina as she turned to her place, both waited, and after 2.4 seconds his click was dropped.

For round 5 the tester ran the same tour with three random seeds and reported that the three runs were almost identical, because the game runs at a fixed 30 frames per second. So it added six variants that shift the moment of the meeting along her path (at 0.1, 0.3 and 0.45 of the way, and three stops at the stack). In 18 meetings nobody passed through anybody, the closest was 0.53 m, and no click was dropped.

The same report has a list of its own invalid runs: a click that hit the stack and not the floor, a hint that swallowed a click, a path that started a fight by accident. It also notes that the recorded runs played 1.2 to 1.4 times faster than the clock, and that one rule ("nobody sets out within a second of another") counted clock seconds. I trust the 0.53 m more because of that list.

What is still wrong

Criterion 2 holds with a remainder. Peter now walks across and then into the depth, one drawn view per leg, and the sideways slide is gone. In the back view the standing shoe still creeps about 3 px per step, which at normal size is hard to see.

Shoulders touch beside places at 0.53 to 0.57 m. Martin stalls for one frame once or twice per meeting. The tester also found an old fault outside the feature: in Act I Martin walks through the smith, and has since before any of this. The second agent that reviews each candidate accepted these as limits of this version. My own test is next, and I am the only one of the three who will play it with a mouse.

Wrapping up

The 24 tests and the 8 replays answered the question their author asked. The tester asked a different one, in metres, with a limit it could defend, and it was not allowed to fix anything. Three of the five rounds ended with a fault that the previous fix had made or moved, and each of those was found because somebody replayed the same clicks and measured again.