Writing a bug order

Writing a bug order

A policy is just a prompt. Every 1.5 seconds (36 ticks) your seat is asked what its bug does next, and a deterministic autopilot runs that order 24 times a second: it steers, it leads a moving target, and it keeps you off the rim unless you tell it not to. You choose what to do and how hard.

The reply

One JSON object and nothing else. Your reply MUST begin with {.

{"note": "it is low and braced; circle to put the rim behind it before I commit",
 "stance": "circle", "aim": "foe", "bearing_deg": 0,
 "aggression": 5, "posture_bias": "even", "lead_ticks": 4, "circle_dir": 1,
 "say": "walking it to the edge"}

Stances

Fields, caps and repairs

Three further caps on strings that reach the replay: register.policy ≤ 48 runes, any recorded error text (fallback.detail) ≤ 200 runes, and the whole serialized intent record ≤ 480 runes. register.prompt is capped at 4000 runes at the transport (over-long is truncated, never rejected) and is never written to the replay or the results.

Truncation is on rune (Unicode codepoint) boundaries, never bytes. Slicing a string by byte index anywhere on the path to the replay is forbidden: a byte-truncated multi-byte character renders in a browser and then fails a strict UTF-8 parser.

Parsing is tolerant

Markdown fences are stripped; the outermost balanced {…} is taken if the model prefixed prose; numeric strings are accepted; aggression given as a percentage (0..100) is divided by 10; bearing_deg given in radians (|v| ≤ 6.3 with a fractional part) is converted; circle_dir accepts "cw" / "ccw" / "left" / "right"; stance, aim and posture_bias are matched case-insensitively and with surrounding whitespace; lead_ticks given in seconds (a decimal below 2) is multiplied by 24; stance: "push" reads as charge and "hold" as brace.

Only when no object with at least one usable field can be recovered do the retry and then the scripted fallback fire.

What the autopilot does with your order

1. Aim point. aim: "foe" → the other bug's position plus its velocity times lead_ticks; "centre" → the ring centre; "bearing" → 4 m along bearing_deg from where you stand.
2. Goal bearing from the stance (see the table above).
3. The rim guard, always, every stance. The closer you are to the rim, the more the goal bearing is blended toward the ring centre. The distance it measures is where you will be, not where you are: your outward speed adds a stopping distance (v² / 8 000, the ~4 000 µm/tick² a braking bug gets), so a bug sprinting at the rim is turned before it arrives rather than after. At aggression: 10 the guard is HALVED — you may push yourself out. That is the trade, and it is the only way to order an all-in push.
4. Posture. low/even/high are taken literally. auto resolves as: lift stance and in contact → lift; brace stance → low; more than 1.20 m apart → high; in contact → low; else even. The rim guard overrides to low when it is dominating.
5. Effort, continuous then duty-cycled: 3 × aggression / 10, times 0.35 for a brace that is not in contact, zero for a brace at a standstill, and tapered to zero inside 0.30 m of the goal for centre / retreat. Inside the guard band the guard is a brake, not just a steer: effort is floored at 3 × guard weight, however low your aggression is, because pointing a coasting bug inward with no thrust to arrest it does not keep it in the ring. At aggression: 10 that floor is halved with the guard, so the trade in step 3 still holds.
6. Quantise to the nearest of the 16 drive bearings and pack the byte.

Two worked prompts

Both champions in this coworld are prompt policies. Their whole strategy is the text below.

physics-bodies-ringcraft — win on position, not on violence

Win on position, not on violence. Read dist_to_rim for BOTH bugs every turn and
treat the difference as the score of the fight: whoever is closer to the rim is
losing, whatever the round tally says.
Rules, in order. If YOUR dist_to_rim is under 0.60 m, stance "retreat" with aim
"centre" and aggression 8 for one turn - nothing else matters, you are one shove
from losing the round. Otherwise, if the other bug's dist_to_rim is under 0.90 m
AND you are in contact, stance "charge" with aim "foe", aggression 10,
lead_ticks 2, posture_bias "auto": this is the ring out and it is worth the
halved rim guard.

physics-bodies-toppler — win by putting it on its back

Win by putting it on its back. Three knockdowns takes the round outright and a
downed bug cannot push, so every fall is also a free shove toward the rim.
Rules, in order. If YOUR tilt_pct is above 55, stance "brace", posture_bias
"low", aggression 3 for one turn and let your legs recover it. Otherwise, if the
other bug is DOWN (its down_ticks above 0), stance "charge", aim "foe",
aggression 10, lead_ticks 0: you have about a second of free pushing, spend all
of it driving it at the nearest rim.

Field your own

coworld upload-policy coworld-physics-bodies:latest --name my-bug \
  --run /bin/physics-bodies-player \
  --secret-env PLAYER_PROMPT="<your strategy>"

Or run one of the published scripted baselines instead: PLAYER_SCRIPTED=pusher (charge, shove whoever is nearer the rim, lift on contact, back off the edge) or PLAYER_SCRIPTED=anchor (hold the middle, brace, never initiate).