yxc20098 commited on
Commit
f9013d6
·
1 Parent(s): 82ef6bf

feat(scenario): combat-pincer-coordination — synchronized two-prong attack (SC2 / military pincer anchor)

Browse files

Wave-6 Group N coordination pack — SC2 multi-prong / two-prong attack
timing, military pincer movement doctrine, envelopment from multiple
angles, synchronisation of dispersed forces.

Two 3-tank squads (3× 2tnk each) start on the SAME west edge at
OPPOSING latitudes (north y=8, south y=32). A central enemy cluster
sits at the mid-map (around (50,20)). The win predicate REQUIRES both
squads to engage simultaneously:

- units_in_region_gte:{x:50,y:20,radius:8,n:4} — a single squad has
only 3 tanks so single-squad-A cannot satisfy the region clause
regardless of attrition (structural pincer teeth).
- units_lost_lte:2 — sequenced (A first, B late) is shredded by the
cluster's anti-armour mass (2× 2tnk + 2× e3 + 4× e1) and loses ≥3
tanks before B arrives → busts the cap (timing teeth).
- within_ticks:4500 / after_ticks:4501 (max_tick=4863 with
max_turns:54 — deadline bites, no DRAW degeneracy) — stall hits
the deadline.

Difficulty axis (one new controlled variable per tier):
- easy: lighter cluster (4× e1 + 2× e3, no enemy armour; kill bar 6).
Sequenced may squeak by here per the inert-easy-teeth
convention; stall + single-squad still LOSE.
- medium: full cluster (4× e1 + 2× e3 + 2× 2tnk = 8 enemies; kill
bar 8). Sequenced now LOSES on the attrition cap.
- hard: +2 seed-driven agent spawn_point groups (group 0:
(5,8)/(5,32); group 1: (3,5)/(3,35)) so a memorised
single-cell opening cannot generalise; cluster symmetric
across y=20 so both spawns face the same pincer decision.

Engine guardrails (per CLAUDE.md):
- Enemy cluster anchored at (50,19) / (50,21) — avoids the (50,20)
silent-no-place footgun; win region (50,20, r=8) still covers
both halves.
- Far persistent enemy fact at (120,4) keeps the episode alive past
full cluster elimination so within_ticks / region predicates
evaluate on the terminal frame (engine auto-done mitigation).
- All agent actors declare spawn_point on hard (CLAUDE.md: 'if ANY
agent actor declares spawn_point, every agent actor WITHOUT
spawn_point is filtered OUT'); two full layouts listed explicitly.

Scripted-policy validation (tests/test_combat_pincer_coordination.py
covers all 4 seeds per level on every policy):
- stall → LOSS on every level/seed (clock).
- single-squad-A → LOSS on every level/seed (region n=4).
- sequenced-A-then-B-late → LOSS on medium/hard every seed (cap).
- intended pincer-sync → WIN on every level/seed (0 losses).

Model smoke (Qwen/Qwen3.6-Plus, medium, seed 1):
combat-pincer-coordination:medium:public#1 loss comp=0.2187
(expected discrimination — pack is meant to detect synchronization)

Registered in tests/test_hard_tier.py::UPGRADED.

openra_bench/scenarios/packs/combat-pincer-coordination.yaml ADDED
@@ -0,0 +1,299 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # combat-pincer-coordination.yaml
2
+ #
3
+ # ACTION capability — synchronised two-prong pincer attack. Two armoured
4
+ # squads start on the SAME west edge but at OPPOSING latitudes
5
+ # (north y≈8 and south y≈32). A central enemy cluster sits at the mid-
6
+ # map. The win predicate requires BOTH squads to engage simultaneously:
7
+ # `units_in_region_gte:{n:4}` cannot be satisfied by a single squad
8
+ # (only 3 tanks per squad — even arriving intact is 1 short of n=4),
9
+ # and `units_lost_lte:2` makes a serialised "Squad A engages first,
10
+ # Squad B follows" attempt fail because A is shredded by the cluster
11
+ # before B arrives.
12
+ #
13
+ # Real-world anchor: military pincer movement / SC2 two-prong attack /
14
+ # envelopment from two directions. The capability is the timing
15
+ # coordination — both prongs must commit at the same moment so the
16
+ # defender cannot focus on either prong in detail.
17
+ #
18
+ # Engine notes (per CLAUDE.md):
19
+ # - Enemy cluster anchored at (50,19) / (50,21) — CLAUDE.md warns
20
+ # that (50,20) silently fails to place enemy clusters; the off-by-
21
+ # one cells are confirmed working. The win region centre at
22
+ # (50,20) r=8 still covers both halves of the cluster (and both
23
+ # ingress tanks once they engage the cluster face).
24
+ # - Far persistent enemy `fact` at (120,4) keeps the episode alive
25
+ # past full enemy elimination so within_ticks / region predicates
26
+ # evaluate on the terminal frame (engine auto-`done` would
27
+ # otherwise collapse to DRAW the instant the last MustBeDestroyed
28
+ # enemy actor falls).
29
+ # - `units_killed_gte: 8` matches the central-cluster enemy count
30
+ # on medium/hard (the far sentinel fact at (120,4) is OUT of all
31
+ # squad routes so it survives and never counts toward kills);
32
+ # easy uses a lighter 6-enemy cluster with kill bar 6.
33
+ # - Enemy stance:3 (AttackAnything) so the cluster fires on any
34
+ # tank in range from the instant of contact — no bot_type needed.
35
+ # For the squad-A-then-B-late timing teeth, the enemy 2× 2tnk +
36
+ # 2× e3 anti-armour mass MUST be sufficient to shred a 3-tank
37
+ # squad arriving alone (~12s to clear with 3 attackers = ~800
38
+ # damage absorbed = 2+ tanks lost on the lone squad). With both
39
+ # squads synchronised the same cluster falls in ~6s and the
40
+ # 6-tank force loses ≤1 tank — well inside the
41
+ # `units_lost_lte: 2` cap and with ≥4 tanks (in fact ≥5) inside
42
+ # the central region.
43
+ # - Tick budget: 4500 within / 4501 fail. With max_turns 54 the max
44
+ # reachable tick is 93 + 90·53 = 4863 → the `after_ticks` clause
45
+ # fires before the run ends → no DRAW degeneracy.
46
+ #
47
+ # Discrimination (the four-script bar; scripted-policy validation in
48
+ # tests/test_combat_pincer_coordination.py covers all 4 seeds per
49
+ # level on every policy):
50
+ # • stall (Command.observe only) — agent units never engage → kill
51
+ # bar unmet → after_ticks LOSS.
52
+ # • single-squad-A — Squad B held back; A walks alone into the
53
+ # cluster. Even if A clears the cluster, only 3 tanks (max) can
54
+ # occupy the objective region while n=4 is required → LOSS. On
55
+ # medium/hard A is also shredded by the cluster's anti-armour →
56
+ # units_lost_lte:2 busts as a second teeth.
57
+ # • sequenced-A-then-B-late — A advances first, takes attrition,
58
+ # B follows after a long delay. On medium/hard A loses 3 of 3
59
+ # tanks before B arrives → units_lost_lte:2 busts → LOSS. On
60
+ # easy with no enemy armour A may clear cleanly and B walks in →
61
+ # this may squeak by (acceptable per SCENARIO_REVIEW_CHECKLIST
62
+ # inert-easy-teeth convention — easy is the bare-skill tier).
63
+ # • intended pincer-sync — both squads attack_move the cluster on
64
+ # turn 1; their staging latitudes (y=8, y=32) put them
65
+ # equidistant (~45 cells each) from the cluster so a naive
66
+ # simultaneous launch arrives together; mass DPS clears the
67
+ # cluster before either side bleeds → WIN.
68
+ #
69
+ # Validate:
70
+ # cd /Users/berta/Projects/OpenRA-Bench && \
71
+ # python3 -m openra_bench.scenarios.validate \
72
+ # openra_bench/scenarios/packs/combat-pincer-coordination.yaml
73
+
74
+ meta:
75
+ id: combat-pincer-coordination
76
+ title: 'Pincer Attack — Two Squads Strike Simultaneously From Two Sides'
77
+ capability: action
78
+ real_world_meaning: >
79
+ Two armoured squads start at OPPOSING latitudes on the west edge
80
+ (one to the north, one to the south) and must converge on a
81
+ central enemy cluster simultaneously, hitting the defender from
82
+ two sides at once. Sending a single squad alone fails on two
83
+ counts: only 3 tanks (not the required 4) can occupy the
84
+ objective region, and the lone squad is shredded by the cluster's
85
+ mass anti-armour before clearing it. Sending both squads but not
86
+ synchronised (one held back, the other commits first) lets the
87
+ cluster focus on the lead squad and destroy it before the trailing
88
+ squad arrives, busting the attrition cap. Only a true
89
+ simultaneous two-prong commit clears the cluster cleanly and
90
+ leaves the joint force standing on the objective.
91
+ robotics_analogue: >
92
+ Synchronous two-team pincer attack on a contested objective: each
93
+ team approaches on a different bearing and both must arrive within
94
+ a common window so the defending agents cannot focus on either
95
+ team in detail. Serialising one team behind the other lets the
96
+ defender concentrate fire and destroy the lead team before
97
+ reinforcement; only the joint simultaneous commit overwhelms
98
+ the defence and preserves the strike force.
99
+ benchmark_anchor:
100
+ - "SC2 multi-prong / two-prong attack timing"
101
+ - "military pincer movement doctrine"
102
+ - "envelopment from multiple angles"
103
+ - "synchronisation of dispersed forces"
104
+ author: openra-bench
105
+
106
+ base_map: rush-hour-arena
107
+
108
+ base:
109
+ agent: {faction: allies, cash: 0}
110
+ enemy: {faction: soviet, cash: 0, bot_type: ''}
111
+ tools: [move_units, attack_unit, attack_move, stop]
112
+ planning: true
113
+ # React the instant contact happens or a tank is lost — pincer
114
+ # micro is reactive; without these interrupts the sync policy only
115
+ # re-targets on the next turn boundary and any gap is fatal to the
116
+ # already-thin attrition cap.
117
+ interrupts:
118
+ enemy_unit_spotted: true
119
+ own_unit_destroyed: true
120
+ termination: {max_ticks: 6000}
121
+ actors: [] # every level supplies its own actor list via overrides.
122
+
123
+ levels:
124
+ # ── EASY ─────────────────────────────────────────────────────────
125
+ # Bare pincer skill: lighter cluster (4× e1 + 2× e3, no enemy
126
+ # armour; 6 enemies total, kill bar 6). Stall still LOSES on the
127
+ # clock; single-squad still LOSES on the region predicate (3 tanks
128
+ # cannot satisfy n=4). Sequenced may squeak by on this lighter
129
+ # cluster — that is the SCENARIO_REVIEW_CHECKLIST "inert anti-
130
+ # cheat teeth acceptable on easy" convention: the bare-skill tier
131
+ # only enforces stall-loses and intended-wins; the sequence-vs-
132
+ # sync delta lives on medium and hard.
133
+ easy:
134
+ description: >
135
+ You command TWO squads of medium tanks (3× 2tnk each, 6 tanks
136
+ total). Squad A stages at the NORTH-WEST corner (5,8); Squad B
137
+ stages at the SOUTH-WEST corner (5,32). A central enemy cluster
138
+ sits at the mid-map (around (50,20)): 4 rifle infantry (e1) +
139
+ 2 anti-tank rocket soldiers (e3). Both squads must converge on
140
+ the cluster SIMULTANEOUSLY from the north and the south
141
+ (a pincer). Win when you have killed 6 enemies AND at least 4
142
+ of your 6 tanks are inside an 8-cell radius of (50,20) AND no
143
+ more than 2 tanks are lost, before tick 4500. Sending only one
144
+ squad cannot satisfy the 4-in-region clause. A stall loses on
145
+ the clock.
146
+ overrides:
147
+ actors:
148
+ # Squad A (NORTH-WEST) — 3 medium tanks at (5,8).
149
+ - {type: 2tnk, owner: agent, position: [5, 8], stance: 1, count: 3}
150
+ # Squad B (SOUTH-WEST) — 3 medium tanks at (5,32).
151
+ - {type: 2tnk, owner: agent, position: [5, 32], stance: 1, count: 3}
152
+ # Enemy cluster — anchored at (50,19) / (50,21) (CLAUDE.md
153
+ # notes (50,20) silently fails to place enemy clusters; the
154
+ # off-by-one cells work). The win region (50,20, r=8) still
155
+ # covers both halves.
156
+ - {type: e1, owner: enemy, position: [50, 19], stance: 3, count: 2}
157
+ - {type: e1, owner: enemy, position: [50, 21], stance: 3, count: 2}
158
+ - {type: e3, owner: enemy, position: [50, 19], stance: 3}
159
+ - {type: e3, owner: enemy, position: [50, 21], stance: 3}
160
+ # Far persistent enemy marker (sentinel) — out of every
161
+ # squad route, keeps the episode alive past full cluster
162
+ # elimination so within_ticks / region predicates evaluate
163
+ # on the terminal frame (no DRAW collapse).
164
+ - {type: fact, owner: enemy, position: [120, 4]}
165
+ win_condition:
166
+ all_of:
167
+ - {units_killed_gte: 6}
168
+ - {units_in_region_gte: {x: 50, y: 20, radius: 8, n: 4}}
169
+ - {units_lost_lte: 2}
170
+ - {within_ticks: 4500}
171
+ fail_condition:
172
+ any_of:
173
+ - {after_ticks: 4501}
174
+ - {not: {units_lost_lte: 2}}
175
+ - {not: {own_units_gte: 1}}
176
+ max_turns: 54
177
+
178
+ # ── MEDIUM ───────────────────────────────────────────────────────
179
+ # +1 controlled variable: full enemy cluster — 4× e1 + 2× e3 +
180
+ # 2× 2tnk (8 enemies total; kill bar = 8). The enemy armour gives
181
+ # the cluster enough sustained anti-armour DPS to shred a lone
182
+ # 3-tank squad (~12s to clear, ~800 damage absorbed) → A-only or
183
+ # A-then-B-late busts units_lost_lte:2. The 6-tank synchronised
184
+ # commit halves the clear time and the per-tank damage share, so
185
+ # the sync wins comfortably (validated: 0 losses on every seed).
186
+ medium:
187
+ description: >
188
+ You command TWO squads of medium tanks (3× 2tnk each, 6 tanks
189
+ total). Squad A stages at the NORTH-WEST corner (5,8); Squad B
190
+ stages at the SOUTH-WEST corner (5,32). A central enemy cluster
191
+ sits at the mid-map (around (50,20)): 4 rifle infantry (e1) +
192
+ 2 anti-tank rocket soldiers (e3, Dragon launcher) + 2 medium
193
+ tanks (2tnk) — the enemy armour gives the cluster enough
194
+ sustained anti-armour DPS to shred a lone 3-tank squad before
195
+ it can clear the cluster. Both squads must converge
196
+ SIMULTANEOUSLY (a pincer): attack-moving with both squads at
197
+ once on turn 1 doubles the strike force and halves the per-tank
198
+ damage share, clearing the cluster before either side bleeds
199
+ out. Sending only one squad cannot satisfy the 4-in-region
200
+ clause; sending A first and B late lets the cluster shred A
201
+ before B arrives. Win when you have killed 8 enemies AND at
202
+ least 4 of your 6 tanks are inside an 8-cell radius of (50,20)
203
+ AND no more than 2 tanks are lost, before tick 4500.
204
+ overrides:
205
+ actors:
206
+ - {type: 2tnk, owner: agent, position: [5, 8], stance: 1, count: 3}
207
+ - {type: 2tnk, owner: agent, position: [5, 32], stance: 1, count: 3}
208
+ # 4× e1 + 2× e3 + 2× 2tnk = 8 enemies. Avoid the (50,20) cell
209
+ # (CLAUDE.md silent-no-place footgun); split between (50,19)
210
+ # and (50,21). The enemy 2tnks are the load-bearing anti-
211
+ # armour that shreds a lone 3-tank squad arriving alone.
212
+ - {type: e1, owner: enemy, position: [50, 19], stance: 3, count: 2}
213
+ - {type: e1, owner: enemy, position: [50, 21], stance: 3, count: 2}
214
+ - {type: e3, owner: enemy, position: [50, 19], stance: 3}
215
+ - {type: e3, owner: enemy, position: [50, 21], stance: 3}
216
+ - {type: 2tnk, owner: enemy, position: [50, 19], stance: 3}
217
+ - {type: 2tnk, owner: enemy, position: [50, 21], stance: 3}
218
+ - {type: fact, owner: enemy, position: [120, 4]}
219
+ win_condition:
220
+ all_of:
221
+ - {units_killed_gte: 8}
222
+ - {units_in_region_gte: {x: 50, y: 20, radius: 8, n: 4}}
223
+ - {units_lost_lte: 2}
224
+ - {within_ticks: 4500}
225
+ fail_condition:
226
+ any_of:
227
+ - {after_ticks: 4501}
228
+ - {not: {units_lost_lte: 2}}
229
+ - {not: {own_units_gte: 1}}
230
+ max_turns: 54
231
+
232
+ # ── HARD ─────────────────────────────────────────────────────────
233
+ # +1 more controlled variable: TWO seed-driven agent spawn groups
234
+ # with DIFFERENT cell coordinates. Group 0 stages Squad A NORTH
235
+ # (5,8) / Squad B SOUTH (5,32). Group 1 SHIFTS the squads to (8,6)
236
+ # / (8,34) — same west-edge / opposing-latitude pincer shape but
237
+ # the exact start cells differ across seeds so a memorised single-
238
+ # cell opening cannot generalise. Both groups remain approximately
239
+ # equidistant from the central cluster (~46-47 cells) so the naive
240
+ # simultaneous launch still arrives together. Per CLAUDE.md, "if
241
+ # ANY agent actor declares spawn_point, every agent actor WITHOUT
242
+ # spawn_point is filtered OUT", so the two full squad layouts are
243
+ # listed explicitly. The defender cluster is symmetric across y=20
244
+ # so both spawns face the same simultaneous-pincer decision.
245
+ hard:
246
+ description: >
247
+ You command TWO squads of medium tanks (3× 2tnk each, 6 tanks
248
+ total) staged on the west edge — one squad to the NORTH, one
249
+ to the SOUTH (the exact staging cells vary by seed; treat the
250
+ two squads symmetrically). A central enemy cluster
251
+ sits at the mid-map (around (50,20)): 4 rifle infantry (e1) +
252
+ 2 anti-tank rocket soldiers (e3, Dragon launcher) + 2 medium
253
+ tanks (2tnk) — the enemy armour gives the cluster enough
254
+ sustained anti-armour DPS to shred a lone 3-tank squad. Both
255
+ squads must converge SIMULTANEOUSLY (a pincer): attack-moving
256
+ with both squads at once on turn 1 clears the cluster before
257
+ either side bleeds out. Win when you have killed 8 enemies AND
258
+ at least 4 of your 6 tanks are inside an 8-cell radius of
259
+ (50,20) AND no more than 2 tanks are lost, before tick 4500.
260
+ overrides:
261
+ actors:
262
+ # spawn_point 0 — Squad A NORTH (5,8), Squad B SOUTH (5,32).
263
+ - {type: 2tnk, owner: agent, position: [5, 8], stance: 1, count: 3, spawn_point: 0}
264
+ - {type: 2tnk, owner: agent, position: [5, 32], stance: 1, count: 3, spawn_point: 0}
265
+ # spawn_point 1 — SHIFTED layout: NORTH squad at (3, 5) and
266
+ # SOUTH squad at (3, 35). Both remain on the west edge and
267
+ # approximately equidistant (~49-50 cells) from the central
268
+ # cluster at (50,20); the slight extra distance vs group 0
269
+ # keeps the synchronised launch viable while ensuring the
270
+ # squad arriving alone takes ≥ the same attrition (still
271
+ # loses ≥3 of 3 tanks on the sequenced policy). The exact
272
+ # start cells DIFFER from group 0, satisfying the hard-tier
273
+ # invariant (≥2 distinct seed-driven starts) and ensuring a
274
+ # memorised single-cell opening cannot generalise across
275
+ # seeds.
276
+ - {type: 2tnk, owner: agent, position: [3, 5], stance: 1, count: 3, spawn_point: 1}
277
+ - {type: 2tnk, owner: agent, position: [3, 35], stance: 1, count: 3, spawn_point: 1}
278
+ # Central cluster — 4× e1 + 2× e3 + 2× 2tnk = 8 enemies,
279
+ # symmetric across y=20 so both spawn groups face the same
280
+ # pincer decision.
281
+ - {type: e1, owner: enemy, position: [50, 19], stance: 3, count: 2}
282
+ - {type: e1, owner: enemy, position: [50, 21], stance: 3, count: 2}
283
+ - {type: e3, owner: enemy, position: [50, 19], stance: 3}
284
+ - {type: e3, owner: enemy, position: [50, 21], stance: 3}
285
+ - {type: 2tnk, owner: enemy, position: [50, 19], stance: 3}
286
+ - {type: 2tnk, owner: enemy, position: [50, 21], stance: 3}
287
+ - {type: fact, owner: enemy, position: [120, 4]}
288
+ win_condition:
289
+ all_of:
290
+ - {units_killed_gte: 8}
291
+ - {units_in_region_gte: {x: 50, y: 20, radius: 8, n: 4}}
292
+ - {units_lost_lte: 2}
293
+ - {within_ticks: 4500}
294
+ fail_condition:
295
+ any_of:
296
+ - {after_ticks: 4501}
297
+ - {not: {units_lost_lte: 2}}
298
+ - {not: {own_units_gte: 1}}
299
+ max_turns: 54
tests/test_combat_pincer_coordination.py ADDED
@@ -0,0 +1,364 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """combat-pincer-coordination — synchronised two-prong pincer attack.
2
+
3
+ The bar: TWO 3-tank squads start on the SAME west edge at OPPOSING
4
+ latitudes (north y=8, south y=32). A central enemy cluster sits at
5
+ the mid-map (around (50,20)). The win predicate REQUIRES both squads
6
+ to engage simultaneously:
7
+
8
+ • `units_in_region_gte:{x:50,y:20,radius:8,n:4}` — at least 4 tanks
9
+ must be inside the central region; a single squad has only 3 tanks
10
+ so single-squad-A cannot satisfy this clause regardless of
11
+ attrition.
12
+ • `units_lost_lte:2` — sequenced (A first, B late) is shredded by
13
+ the cluster's anti-armour mass and loses ≥3 tanks before B
14
+ arrives → busts the cap.
15
+ • `within_ticks:4500` — stall hits the deadline → after_ticks LOSS.
16
+
17
+ Discrimination (the four-script bar; the scripted runs cover all
18
+ seeds 1–4 per level on every policy):
19
+
20
+ • stall — every level/seed must LOSE (clock).
21
+ • single-squad-A — every level/seed must LOSE (region n=4).
22
+ • sequenced-A-then-B-late — medium/hard must LOSE (attrition cap
23
+ busted by lone Squad A); easy is the
24
+ bare-skill tier where this may squeak
25
+ by (SCENARIO_REVIEW_CHECKLIST inert-
26
+ easy-teeth convention).
27
+ • intended pincer-sync — every level/seed must WIN.
28
+
29
+ Validation is scripted (no model / network)."""
30
+ from __future__ import annotations
31
+
32
+ from pathlib import Path
33
+
34
+ import pytest
35
+
36
+ pytest.importorskip("openra_rl_training", reason="Rust env wheel not installed")
37
+ from openra_bench.scenarios import load_pack
38
+ from openra_bench.scenarios.loader import compile_level
39
+ from openra_bench.scenarios.win_conditions import WinContext, evaluate
40
+
41
+ PACKS = Path(__file__).parent.parent / "openra_bench" / "scenarios" / "packs"
42
+ PACK_PATH = PACKS / "combat-pincer-coordination.yaml"
43
+
44
+
45
+ # ── unit-level predicate / metadata checks (no engine) ──────────────
46
+
47
+
48
+ def test_pack_compiles_and_meta_fields_populated():
49
+ pack = load_pack(PACK_PATH)
50
+ assert pack.meta.id == "combat-pincer-coordination"
51
+ assert pack.meta.capability == "action"
52
+ assert pack.meta.real_world_meaning, "real_world_meaning required"
53
+ assert pack.meta.robotics_analogue, "robotics_analogue required"
54
+ anchors = pack.meta.benchmark_anchor
55
+ assert isinstance(anchors, list) and len(anchors) == 4, (
56
+ f"benchmark_anchor must list 4 anchors, got {anchors!r}"
57
+ )
58
+ joined = " ".join(anchors).lower()
59
+ for needle in ("sc2", "pincer", "envelopment", "synchronisation"):
60
+ assert needle in joined, f"missing anchor keyword: {needle}"
61
+ for lvl in ("easy", "medium", "hard"):
62
+ c = compile_level(pack, lvl)
63
+ assert c.map_supported
64
+ assert c.win_condition is not None
65
+ assert c.fail_condition is not None
66
+
67
+
68
+ def _ctx(*, units=(), tick=1000, kills=0, lost=0):
69
+ """Synthesize a WinContext for predicate-level checks."""
70
+ import types
71
+
72
+ sig = types.SimpleNamespace(
73
+ game_tick=tick,
74
+ units_killed=kills,
75
+ units_lost=lost,
76
+ cash=0,
77
+ resources=0,
78
+ power_provided=0,
79
+ power_drained=0,
80
+ own_buildings=[],
81
+ own_building_types=set(),
82
+ enemies_seen_ids=set(),
83
+ enemy_buildings_seen_ids=set(),
84
+ explored_percent=0.0,
85
+ then_progress={},
86
+ seq_progress={},
87
+ )
88
+ return WinContext(
89
+ signals=sig,
90
+ render_state={"units_summary": list(units)},
91
+ )
92
+
93
+
94
+ def _tanks_at(n, x=50, y=20):
95
+ return [
96
+ {"cell_x": x, "cell_y": y, "type": "2tnk", "id": str(2000 + i)}
97
+ for i in range(n)
98
+ ]
99
+
100
+
101
+ def test_predicates_per_level():
102
+ """Region n=4 + kill bar + attrition + deadline all load-bearing."""
103
+ pack = load_pack(PACK_PATH)
104
+ expectations = {
105
+ # (kill_bar, region_n, lost_cap, within, after_fail)
106
+ "easy": (6, 4, 2, 4500, 4501),
107
+ "medium": (8, 4, 2, 4500, 4501),
108
+ "hard": (8, 4, 2, 4500, 4501),
109
+ }
110
+ for lvl, (kb, n, cap, w, af) in expectations.items():
111
+ c = compile_level(pack, lvl)
112
+ # Intended win: ≥n tanks in region, ≥kb kills, ≤cap lost, in time.
113
+ ctx_win = _ctx(units=_tanks_at(n + 1), tick=w - 100, kills=kb, lost=cap)
114
+ assert evaluate(c.win_condition, ctx_win), f"{lvl}: intended should WIN"
115
+ # Region clause unmet (only 3 tanks at objective) → not a win
116
+ # — this is the load-bearing pincer teeth (single squad has only 3 tanks).
117
+ ctx_3 = _ctx(units=_tanks_at(3), tick=w - 100, kills=kb, lost=0)
118
+ assert not evaluate(c.win_condition, ctx_3), (
119
+ f"{lvl}: 3 tanks in region < n={n} must not satisfy win"
120
+ )
121
+ # Kill bar unmet → not a win.
122
+ ctx_low_kills = _ctx(units=_tanks_at(n + 1), tick=w - 100, kills=kb - 1, lost=0)
123
+ assert not evaluate(c.win_condition, ctx_low_kills), (
124
+ f"{lvl}: kills < {kb} must not satisfy win"
125
+ )
126
+ # Attrition cap busted → not a win + fails.
127
+ ctx_lost = _ctx(units=_tanks_at(n + 1), tick=w - 100, kills=kb, lost=cap + 1)
128
+ assert not evaluate(c.win_condition, ctx_lost)
129
+ assert evaluate(c.fail_condition, ctx_lost), f"{lvl}: lost>{cap} must FAIL"
130
+ # Timeout busted → fails.
131
+ ctx_late = _ctx(units=_tanks_at(n + 1), tick=af + 1, kills=kb, lost=0)
132
+ assert evaluate(c.fail_condition, ctx_late), f"{lvl}: tick>after must FAIL"
133
+ # Force-wipe → fails.
134
+ ctx_wipe = _ctx(units=[], tick=w - 100, kills=kb, lost=6)
135
+ assert evaluate(c.fail_condition, ctx_wipe)
136
+
137
+
138
+ def test_timeout_reachable_inside_max_turns():
139
+ """No DRAW degeneracy: after_ticks fail trigger must be reachable
140
+ within max_turns (engine advances ~90 ticks per turn ⇒ max tick
141
+ ≈ 93 + 90·(max_turns-1))."""
142
+ pack = load_pack(PACK_PATH)
143
+ for lvl in ("easy", "medium", "hard"):
144
+ c = compile_level(pack, lvl)
145
+ max_tick = 93 + 90 * (c.max_turns - 1)
146
+ # After-ticks fail clause is 4501 on every level here.
147
+ assert 4501 <= max_tick, (
148
+ f"{lvl}: after_ticks 4501 > max reachable tick {max_tick} "
149
+ f"(max_turns={c.max_turns}); deadline never bites"
150
+ )
151
+
152
+
153
+ def test_two_squads_on_every_level():
154
+ """Two 3-tank squads at opposing latitudes (north y≈8 vs south
155
+ y≈32); the pincer mechanism requires both prongs."""
156
+ pack = load_pack(PACK_PATH)
157
+ for lvl in ("easy", "medium", "hard"):
158
+ c = compile_level(pack, lvl)
159
+ agent_actors = [a for a in c.scenario.actors if a.owner == "agent"]
160
+ # All agent units are medium tanks (2tnk).
161
+ for a in agent_actors:
162
+ assert a.type == "2tnk", (
163
+ f"{lvl}: agent must be 2tnk (medium tanks); got {a.type}"
164
+ )
165
+ # Two latitudes represented (north y<20 and south y>20).
166
+ latitudes = {("N" if a.position[1] < 20 else "S") for a in agent_actors}
167
+ assert latitudes == {"N", "S"}, (
168
+ f"{lvl}: agent squads must span both latitudes; got {latitudes}"
169
+ )
170
+
171
+
172
+ def test_enemy_cluster_avoids_silent_no_place_cell():
173
+ """CLAUDE.md: (50,20) silently fails to place enemy clusters.
174
+ Cluster must anchor on the off-by-one cells (50,19) / (50,21)."""
175
+ pack = load_pack(PACK_PATH)
176
+ for lvl in ("easy", "medium", "hard"):
177
+ c = compile_level(pack, lvl)
178
+ cluster = [
179
+ a for a in c.scenario.actors
180
+ if a.owner == "enemy" and a.type in ("e1", "e3", "2tnk")
181
+ ]
182
+ assert cluster, f"{lvl}: no central cluster declared"
183
+ for a in cluster:
184
+ assert tuple(a.position) != (50, 20), (
185
+ f"{lvl}: enemy cluster member at (50,20) — CLAUDE.md "
186
+ f"silent-no-place footgun"
187
+ )
188
+ assert a.position[0] == 50 and a.position[1] in (19, 21), (
189
+ f"{lvl}: cluster member off-anchor; got {a.position}"
190
+ )
191
+
192
+
193
+ def test_persistent_far_enemy_marker_present():
194
+ """Engine auto-`done` mitigation: a far unarmed enemy `fact` keeps
195
+ the episode alive past full cluster elimination so within_ticks /
196
+ region predicates evaluate on the terminal frame (no DRAW collapse)."""
197
+ pack = load_pack(PACK_PATH)
198
+ for lvl in ("easy", "medium", "hard"):
199
+ c = compile_level(pack, lvl)
200
+ far_facts = [
201
+ a for a in c.scenario.actors
202
+ if a.owner == "enemy" and a.type == "fact"
203
+ and a.position[0] >= 100 # far east of the cluster
204
+ ]
205
+ assert far_facts, f"{lvl}: needs a far persistent enemy fact sentinel"
206
+
207
+
208
+ def test_hard_has_two_spawn_point_groups():
209
+ """Hard-tier curation: ≥2 distinct agent spawn_point groups so the
210
+ seed round-robins which squad ID is north vs south."""
211
+ c = compile_level(load_pack(PACK_PATH), "hard")
212
+ groups = {
213
+ (a.spawn_point if a.spawn_point is not None else 0)
214
+ for a in c.scenario.actors
215
+ if a.owner == "agent"
216
+ }
217
+ assert len(groups) >= 2, f"hard needs ≥2 spawn_point groups, got {groups}"
218
+
219
+
220
+ # ── engine-driven scripted policies ─────────────────────────────────
221
+
222
+
223
+ def _split_squads(rs):
224
+ """Return (north_ids, south_ids) — split agent 2tnk units by
225
+ latitude. The agent's two staging latitudes are y≈8 (north) and
226
+ y≈32 (south); the cluster sits at y=20."""
227
+ n_ids, s_ids = [], []
228
+ for u in (rs.get("units_summary") or []):
229
+ if str(u.get("type", "")).lower() != "2tnk":
230
+ continue
231
+ if int(u["cell_y"]) < 20:
232
+ n_ids.append(str(u["id"]))
233
+ else:
234
+ s_ids.append(str(u["id"]))
235
+ return n_ids, s_ids
236
+
237
+
238
+ def _stall(rs, Command):
239
+ """Pure observe — agent units never engage → kill bar unmet AND
240
+ region clause unmet → after_ticks LOSS."""
241
+ return [Command.observe()]
242
+
243
+
244
+ def _single_squad_a(rs, Command):
245
+ """Send only the NORTH squad; hold the SOUTH squad. Even if A
246
+ clears the cluster on its own, only 3 tanks can be in the
247
+ objective region → `units_in_region_gte:{n:4}` cannot be
248
+ satisfied → LOSS. On medium/hard the cluster's enemy 2tnks also
249
+ shred A → units_lost_lte:2 busts as a second teeth."""
250
+ a, b = _split_squads(rs)
251
+ cmds = []
252
+ if a:
253
+ cmds.append(Command.attack_move(a, 50, 20))
254
+ if b:
255
+ cmds.append(Command.stop(b))
256
+ return cmds or [Command.observe()]
257
+
258
+
259
+ def _make_sequenced_a_then_b_late(delay_turns=20):
260
+ """Squad A commits first; Squad B holds for `delay_turns` then
261
+ advances. The cluster shreds A in detail before B arrives →
262
+ units_lost_lte:2 busts on medium/hard. On easy with no enemy
263
+ armour A may clear the cluster cleanly and B walks in → may WIN
264
+ (acceptable per the inert-easy-teeth convention)."""
265
+ state = {"turn": 0}
266
+
267
+ def policy(rs, Command):
268
+ state["turn"] += 1
269
+ a, b = _split_squads(rs)
270
+ cmds = []
271
+ if a:
272
+ cmds.append(Command.attack_move(a, 50, 20))
273
+ if b:
274
+ if state["turn"] >= delay_turns:
275
+ cmds.append(Command.attack_move(b, 50, 20))
276
+ else:
277
+ cmds.append(Command.stop(b))
278
+ return cmds or [Command.observe()]
279
+
280
+ return policy
281
+
282
+
283
+ def _intended_pincer_sync(rs, Command):
284
+ """Both squads attack_move the cluster on turn 1. The two squads
285
+ are equidistant from the cluster (~45 cells each) so a naive
286
+ simultaneous launch arrives together; mass DPS clears the
287
+ cluster in ~6s before either side bleeds out."""
288
+ a, b = _split_squads(rs)
289
+ cmds = []
290
+ if a:
291
+ cmds.append(Command.attack_move(a, 50, 20))
292
+ if b:
293
+ cmds.append(Command.attack_move(b, 50, 20))
294
+ return cmds or [Command.observe()]
295
+
296
+
297
+ @pytest.mark.parametrize("level", ["easy", "medium", "hard"])
298
+ @pytest.mark.parametrize("seed", [1, 2, 3, 4])
299
+ def test_stall_loses(level, seed):
300
+ pytest.importorskip("openra_train")
301
+ from openra_bench.eval_core import run_level
302
+
303
+ c = compile_level(load_pack(PACK_PATH), level)
304
+ r = run_level(c, _stall, seed=seed)
305
+ assert r.outcome == "loss", (
306
+ f"{level} seed={seed}: stall must be a real timeout LOSS "
307
+ f"(no engagement → kill bar + region clause unmet), got {r.outcome}"
308
+ )
309
+
310
+
311
+ @pytest.mark.parametrize("level", ["easy", "medium", "hard"])
312
+ @pytest.mark.parametrize("seed", [1, 2, 3, 4])
313
+ def test_single_squad_a_loses(level, seed):
314
+ """Single-squad-A cannot satisfy `units_in_region_gte:{n:4}` —
315
+ only 3 tanks exist in Squad A so the region clause is structurally
316
+ unsatisfiable. On medium/hard the cluster also shreds A as a
317
+ second teeth (units_lost_lte:2 busts)."""
318
+ pytest.importorskip("openra_train")
319
+ from openra_bench.eval_core import run_level
320
+
321
+ c = compile_level(load_pack(PACK_PATH), level)
322
+ r = run_level(c, _single_squad_a, seed=seed)
323
+ assert r.outcome == "loss", (
324
+ f"{level} seed={seed}: single-squad-A must LOSE "
325
+ f"(region n=4 impossible with 3 tanks); got {r.outcome} "
326
+ f"(kills={r.signals.units_killed}, lost={r.signals.units_lost})"
327
+ )
328
+
329
+
330
+ @pytest.mark.parametrize("level", ["medium", "hard"])
331
+ @pytest.mark.parametrize("seed", [1, 2, 3, 4])
332
+ def test_sequenced_a_then_b_late_loses_on_medium_hard(level, seed):
333
+ """A commits first, B holds 20 turns. The cluster's enemy 2tnks
334
+ shred A (lose 3 of 3 tanks) before B arrives → units_lost_lte:2
335
+ busts → LOSS. Easy has no enemy armour so this may squeak by
336
+ there (bare-skill-tier inert-easy-teeth convention)."""
337
+ pytest.importorskip("openra_train")
338
+ from openra_bench.eval_core import run_level
339
+
340
+ c = compile_level(load_pack(PACK_PATH), level)
341
+ r = run_level(c, _make_sequenced_a_then_b_late(20), seed=seed)
342
+ assert r.outcome == "loss", (
343
+ f"{level} seed={seed}: sequenced-A-then-B-late must LOSE "
344
+ f"(A shredded by enemy 2tnks before B arrives); got {r.outcome} "
345
+ f"(kills={r.signals.units_killed}, lost={r.signals.units_lost})"
346
+ )
347
+
348
+
349
+ @pytest.mark.parametrize("level", ["easy", "medium", "hard"])
350
+ @pytest.mark.parametrize("seed", [1, 2, 3, 4])
351
+ def test_intended_pincer_sync_wins(level, seed):
352
+ """Both squads attack_move the cluster on turn 1 — naive
353
+ simultaneous launch with equidistant staging arrives together,
354
+ mass DPS clears the cluster before either side bleeds out."""
355
+ pytest.importorskip("openra_train")
356
+ from openra_bench.eval_core import run_level
357
+
358
+ c = compile_level(load_pack(PACK_PATH), level)
359
+ r = run_level(c, _intended_pincer_sync, seed=seed)
360
+ assert r.outcome == "win", (
361
+ f"{level} seed={seed}: intended pincer-sync should WIN, "
362
+ f"got {r.outcome} after {r.turns} turns "
363
+ f"(kills={r.signals.units_killed}, lost={r.signals.units_lost})"
364
+ )
tests/test_hard_tier.py CHANGED
@@ -539,6 +539,20 @@ UPGRADED = [
539
  # filtered out), so the strike geometry is symmetric but a
540
  # memorised opening cannot generalise.
541
  "rob-unit-loss-recovery",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
542
  ]
543
 
544
  # Consciously NOT spawn-varied, with the reason (keeps the curation
 
539
  # filtered out), so the strike geometry is symmetric but a
540
  # memorised opening cannot generalise.
541
  "rob-unit-loss-recovery",
542
+ # Wave-6 Group N coordination pack — military pincer movement /
543
+ # SC2 two-prong attack / envelopment from two directions. Two
544
+ # 3-tank squads start at OPPOSING west-edge latitudes (north y=8 /
545
+ # south y=32) and must converge on a central enemy cluster
546
+ # SIMULTANEOUSLY: the win predicate requires ≥4 tanks in the
547
+ # central region (single squad has only 3 — structurally
548
+ # impossible) AND ≤2 tanks lost (sequenced A-then-B-late shreds A
549
+ # before B arrives → cap busted). Hard defines two agent
550
+ # spawn_point groups round-robined by seed with DIFFERENT cell
551
+ # coordinates (group 0: (5,8)/(5,32); group 1: (3,5)/(3,35)) so
552
+ # a memorised single-cell opening cannot generalise; the cluster
553
+ # is symmetric across y=20 so both spawns face the same pincer
554
+ # decision.
555
+ "combat-pincer-coordination",
556
  ]
557
 
558
  # Consciously NOT spawn-varied, with the reason (keeps the curation