AI Commander Capture And Fun Improvement Plan

AI Commander — Capture & Fun Improvement Plan

Source-verified 2026-06-21 against the AI-commander development branch (deploy/2026-06-12-aicom-experital, HEAD b0975da9c). Paths relative to Missions/[55-2hc]warfarev2_073v48co.chernarus/ unless noted. Arma 2 OA 1.64. Produced by a 25-agent analysis fleet (8 subsystem readers → synthesis → adversarial verification against the live SQF). Constant values are branch-specific: this development branch carries WFBE_C_AI_COMMANDER_TEAMS_TARGET = 4, whereas live master halved it to 2 (B36) for headless-client FPS — read every number below as "on the experital branch".

This page is a design proposal, not a behavior reference. For what the commander does today see AI Commander Execution Loop Reference, AI Commander Autonomy Audit, and AI Commander Tunable Constants Reference. For the capture mechanic see Towns, Camps and Capture Atlas.

2026-06-24 branch update: current origin/claude/b74.2-aicom@21b62b04 implements several adjacent quick wins, but remains source Chernarus branch evidence only: no GitHub PR route was found, the merge-base/current stable is origin/master@f8a76de34, git diff --check origin/master..origin/claude/b74.2-aicom is clean, and the branch has no Missions_Vanilla payload. The full current payload is 29 source Chernarus files / +424 / -38. The old documented head d472da6a advanced by four source Chernarus files / +100 / -14: b9515ef57 decouples GUER stipend/tier launch at Init_Server.sqf:808-814 and seeds/rebroadcasts WFBE_GUER_VEHICLE_TIER at Server_GuerStipend.sqf:35,44-46,57-63,95, while 21b62b04 adds join/connect ACK failovers at Init_Client.sqf:654-681,693-714 and retry-budget recovery at Server_OnPlayerConnected.sqf:22-56. Earlier branch commits add the pop-tier AICOM cap model, WFBE_C_AICOM_CAMP_STALL_PASSES = 3, camp range 11.5, small AA trims in Squad_RU.sqf / Squad_USMC.sqf, marker-feed recovery, carrier air-shop/respawn hooks and stats fast-follow writers. Treat this as candidate evidence routed through AI Commander Autonomy Audit, not as a replacement for Hetzner smoke or maintained Vanilla propagation.

How "capture" actually works (the binary bar)

A town is not a flag you stand on. Server/FSM/server_town.sqf runs a per-town loop that, every ~5 s, scans nearEntities inside WFBE_C_TOWNS_CAPTURE_RANGE = 40 m (unitsBelowHeight 10), counts the attacker's presence, and drains the town's supplyValue by roughly round(presence × rate) per cycle — but only while the defender's _activeEnemies == 0 inside that same 40 m circle. Ownership flips when supplyValue falls below 1. Supply regenerates on a separate 60 s tick whenever no enemy is present.

So "good at capturing" is three simultaneous engine demands, and missing any one captures nothing:

1. Arrive — get armed bodies inside the 40 m ring (movement/execution).
2. Clear — kill the last defender and hold every camp so the active-enemy gate opens and the drain multiplier stays high (force composition).
3. Stay — remain long enough to drain supply faster than the 60 s regen restores it (concentration + holding).

A commander that masses force but parks its trucks 50 m out, or clears 90 % and leaves one hidden defender, makes zero progress. This binary bar is why most of the fixes below are about presence inside the ring, not about buying more units.

Why the AI under-captures (thesis) — Updated with 2026-07 GROK Research Cascade Analysis

Four compounding failures strip the AI of effective assault force before teams ever satisfy the bar. The 2026-07 GROK-AICOM-RESEARCH (Wave1+Wave2 fleets, INIT-CASCADE, AX01-AX10) shows these are not isolated but a failure cascade of 10 adjacent systems that systematically prevent "good" A13/A14/A15 decisions from producing physical success (movement → contact → capture contribution). See outputs/INIT-OSI-UTES-MATRIX.md (OSI metric), INIT-ADJACENT-CASCADE-OFFENSIVE-FAILURE.md, AX01-TOWN-CAMP-CAPTURE-PIPELINE.md, and related AX deeps for full traces at SHA 914cc31c.

1. Teams arrive undermanned or too late. Live telemetry and AX02/03 quant show ecology (garrisons/patrols/playerAI/GUER) + admission caps steal 50-80%+ of group budget before AICOM fields bodies (A15 breadth collapse). AX04 scheduler skew turns nominal 60s decisions into 90-180s+ effective age (stale targets by arrival). Root causes now include Produce sideAI tier caps, global group pressure (A44), and Send bursts amplifying AX04.

2. Full teams still register no pressure (AX01 capture pipeline). Arrival detection is leader-centric (RunCommanderTeam.sqf L1569-1574, 250m ARRIVE gate). Capture requires dismounted "Man" units in camp range (~10m) + depot hold (~40m) for presence drain. Mounted teams, no force-dismount, or _captureDone latch (L288/985/2259) prevent the phase. AX01 details: townorder set at AssignTowns L802+ (often no-broadcast on HC path) → townOrderArr nil at Run L1910 (nearest fallback). BEGIN_CAPTURE (L1890) never fires or the sweep fails. Adjacent: GC (AX06, server_groupsGC re-adopt strips orders mid-march), respawn (preserves group but drops wfbe_aicom*), non-AICOM interference.

3. Forward teams can't be reinforced or protected. AX02 admission + AX03 ecology + AX06 GC/Service as hidden writers (ServiceTick clears townorder/order on detours; GC 60s re-adopt writes last-writer on aicom state) erase or starve the dispatch. Delegation collapse (prior perf 85-93% server) means "HC" teams often execute server-side without locality for orders.

4. Captured towns are naked + post-capture dead-air (A19). Post-flip re-task fires locally but the next purposeful dispatch waits on scheduler gates + strips + ecology. Dead-air 60-180s+ common; low exploitation_rate. Garrison off, statics delayed. Humans flip back in the window.

New unifying metric (GROK research): Offensive Success Index (OSI) = stage_survival product (departed/disp × arrived/departed × contacted/arrived × capture_start/contacted) × tempo_norm × effective_bodies_factor × pressure_adjust × (1-staleness). See INIT-OSI-UTES-MATRIX.md for formula, Utes matrix, events (DECISION_TARGET, BEHAVIOR_ARRIVAL with mounted/locality, OFFENSE_PRESSURE, CAPTURE_START reason, STAGE_SURVIVAL), and gates (OSI >0.55-0.6 for active offensive). Utes fast-bench (OSI-UTES-EXPERIMENT-HARNESS.md + RUN-PLAN.md) makes the cascade falsifiable with AX injection.

See the dedicated AICOM 2026-07 GROK Research Integration page for the consolidated pipeline traces, AX matrix, root causes, and experiment specs.

This binary bar + cascade is why most "improvements" fail without fixing the adjacent pipeline (AX01 capture, AX04 scheduler, AX06 GC/Service, AX02/03 ecology/admission, locality/non-broadcast).

Detailed AX01 Capture Pipeline (from AX01-TOWN-CAMP-CAPTURE-PIPELINE.md):

1. Assignment (AI_Commander_AssignTowns.sqf L802+): _team setVariable ["wfbe_aicom_townorder", [_target, _dispT0, pos]]; Often no broadcast on HC path (L802 2-arg form). dispatch_open + ASSAULT_DISPATCH logged.

2. Arrival (Common_RunCommanderTeam.sqf ~L1569-1574): Leader-distance gate (~250 m or WFBE_C_AICOM_ASSAULT_ARRIVE_RADIUS). _arrived = true; ARRIVAL_GATE log. _captureDone reset on new order.

3. Capture phase entry (L1890): if (_arrived && !_captureDone && mode=="towns-target") → BEGIN_CAPTURE. Resolve townObj: townOrderArr = _team getVariable "wfbe_aicom_townorder" (L1910). Nil fallback → nearest town (wrong target risk under multi-town or HC).

4. Physical capture requirements: Force dismount footInf. Camp sweep to unheldCamps at ~10 m MOVE. Then depot hold + SAD inside 40 m. Actual town flip is presence-based in server_town (40 m nearEntities, no activeEnemies, supply drain).

5. Latch (L2259): _captureDone = true only on town flip. Many early exits strand the team.

Adjacent cascade impacts (high stage_survival killers):

• AX02/03: ecology + admission caps steal bodies before AICOM (CONC=6 fist starved; SPEARHEAD_MAX=2).

• AX04 + AX11: 60 s strat + sleeps + Send compile bursts → effective decision age 90-180 s+; stale by arrival.

• AX06: GC (60 s BASE re-adopt) + ServiceTick (Run L2502) + respawn rewrite or clear townorder/order mid-pipeline. Pre-arrival survival ~0.5-0.7.

• HC locality + delegation collapse (prior 85-93 % server): teams often execute without correct order state.

• _captureDone + CapLock + camp guards prevent churn but create zombie orders when any prior stage fails.

Offensive Success Index (OSI) (from INIT-OSI-UTES-MATRIX.md):

OSI = (stage_survival product) × tempo_norm × effective_bodies_factor × pressure_adjust × (1 − staleness_penalty)

Stage survivals (example chain):

• departed / dispatched

• arrived / departed

• contacted / arrived

• capture_start / contacted

Events / telemetry (A51 extensions):

• DECISION_TARGET

• BEHAVIOR_ARRIVAL (with mounted/locality flags)

• OFFENSE_PRESSURE

• CAPTURE_START (reason)

• STAGE_FAILURE

• SCHED_LOAD, DECISION_AGE

Gate: sustained OSI > 0.55–0.6 indicates "active offensive" (purposeful sustained contact + capture contribution).

Utes matrix (controlled 0-player, 4-6 towns, equal work):

• Baseline

• +AX02 (admission starve)

• +AX03 (ecology pollution)

• +AX04 (scheduler skew / bursts)

• +AX06 (GC hidden writer)

• Mixed + HC variants

• Columns: OSI components + raw KPIs (ASSAULT_DISPATCH vs BEGIN_CAPTURE counts, capture contribution, dead-air time)

See full spec in INIT-OSI-UTES-MATRIX.md and harness docs for instrumentation points and run plan.

Root Causes (from 65-MASTER-SYNTHESIS + FINAL-INDEX):

1. Capture pipeline mismatch (AX01) — leader arrival vs dismounted camp presence + nil townorder.

2. Scheduler timebase skew (AX04/11) — nominal intervals become stale in practice.

3. Hidden writers (AX06 GC + Service + AX09/12 respawn) strip orders.

4. Admission + ecology starvation (AX02/03) — fist never fields enough bodies.

5. Post-capture dead-air + weak re-eval (A19) + no immediate continuation stamp.

Safest High-ROI Improvements (from A61 catalog + synthesis, offensive-preserving, Utes-testable):

• Broadcast + protect townorder / dispatch_id (AX01 + HC parity).

• Admission headroom for concentrated fist (AX02).

• GC/Service guards or seq stamps on aicom_* state (AX06).

• Post-capture re-eval / continuation stamp + reduced dead-air (A19).

• Full A51 telemetry + OSI consumer/analyzer.

All changes must be measured by OSI lift + stage survival gains on Utes (no regression in offensive tempo).

Why fighting it is less fun (thesis)

The AI is opaque and clockwork-predictable at the same time — the worst combination. Players can't read its intent (posture, HQ-strike launch, wildcard draws, relief dispatch are all RPT-only), so there's no counterplay signal; yet its decisions are fully deterministic (same two spearhead towns, same coin-flip doctrine, same 120 s reassignment cadence, same truck-into-the-depot wave). Relief arrives before the attacker is even in position (detection-triggered, not combat-triggered), but the AI never retreats a bleeding team, never feints, never varies tempo, never relocates its HQ, and never reacts to its own base being demolished. The result feels simultaneously unfair (omniscient reaction) and easy (exploit the one predictable axis), collapsing the tension arc that makes combined-arms PvE compelling.

(The 2026-07 research adds: low legibility from missing A51 events and A49/A50 gaps; player obstruction (A48) on dismount/camps further kills presence; map/terrain (A47) amplifies the pipeline failures differently per map.)

Why fighting it is less fun (thesis)

The AI is opaque and clockwork-predictable at the same time — the worst combination. Players can't read its intent (posture, HQ-strike launch, wildcard draws, relief dispatch are all RPT-only), so there's no counterplay signal; yet its decisions are fully deterministic (same two spearhead towns, same coin-flip doctrine, same 120 s reassignment cadence, same truck-into-the-depot wave). Relief arrives before the attacker is even in position (detection-triggered, not combat-triggered), but the AI never retreats a bleeding team, never feints, never varies tempo, never relocates its HQ, and never reacts to its own base being demolished. The result feels simultaneously unfair (omniscient reaction) and easy (exploit the one predictable axis), collapsing the tension arc that makes combined-arms PvE compelling.

Design constraints the plan respects

• Capture power competes directly with server FPS. The team caps (TEAMS_TARGET, TOTAL_AI_MAX = 60), WFBE_C_TOWNS_ACTIVE_MAX = 12, and SPEARHEAD_TOWNS_MAX = 2 are FPS/headless-client levers, not game-design choices. Any "field more force" change is also a frame-rate bet and belongs behind a Hetzner playtest gate.
• The 144-groups/side engine hard cap bounds everything (see Server Group GC, Cap Warning and Zombie Reaper). Empty founded teams and over-garrisoning both burn the budget.
• Two deliberate owner vetoes (do not flip without sign-off): AI artillery is hard-locked off (WFBE_C_AI_COMMANDER_ARTILLERY = 0, Steff 2026-06-13) and the mobile garrison is off by default (Owner call 2026-06-11, "everything goes to the front").

Already fixed (do not re-propose)

• The distTgt = 8122 piecemeal bug — teams cherry-picking the enemy's rich rear 8 km away instead of the front — is fixed by the coherent-front scorer in AI_Commander_Strategy.sqf (distance-to-front divisor dropped 150 → 50, WFBE_C_AICOM_DISTANCE_DIVISOR). AICOMDBG confirms front-adjacent targeting.

Recommendations

Every item below survived adversarial verification against the live code. The verifier rejected three synthesis items as already-shipped or factually wrong and corrected several others (noted inline). Effort/risk use the standard scale; "Status" distinguishes new work from already-built branches and owner-gated flips.

Tier 0 — Merge the four in-flight branches (highest ROI, playtest-pending)

The biggest wins are already coded on feature branches and only need a Hetzner playtest + merge. This is the single highest-leverage action.

Branch	Fixes	Axis	Verified note
claude/aicom-deficitfill-0618	Under-strength teams refill to 8+ (_floorN raises the produce target; adds a FILL-TO-FLOOR pad). Directly targets the unitsPerTeam 5.4 / captures-0 root cause.	capture	Supersedes the synthesis "deficit-fill floor" idea, which was a no-op against the deployed file. Merge the branch, don't re-patch.
claude/aicom-punchy	Kill-rewards credited per kill (wired in RequestOnUnitKilled.sqf, was missing outside the W12 flag); SUPPLY_RESERVE 8000→1000; BOOTSTRAP_SUPPLY 50→120; bigger team count; ASSAULT_HOLD/ASSAULT_SAD extracted; time-curve income.	both	Breaks the supply-starvation spiral and keeps the late-game treasury solvent. FPS bet (more teams / higher TOTAL_AI_MAX) — gate on playtest.
claude/aicom-freewins	Garrison stance locked to RED/AWARE/NORMAL (was a random roll that left defenders passive ~25 % of spawns); W21 VBIED water/forest re-roll; W17 supply convoy targets the front, not the rear HQ town.	both	4 small behavior fixes, low risk. Replicate the garrison-stance fix across all map copies of Common_WaypointPatrolTown.sqf.
claude/aicom-light-team-redo	Light/motorized teams dismount at 400 m standoff, hull supports by fire, infantry clear on foot, survivors remount after the flip. Ends the infanticidal truck-into-the-ring assault.	both	Supersedes the synthesis "light-team dismount" idea (whose sketch contained the A3-only select {} filter and vectorAdd — both fatal in A2). Use the branch.

Tier 1 — New capture quick-wins (verified A2-safe, not yet built)

#	Change	Files	Effort	Risk	Verifier correction
C1	Arc fallback. When both flanking paths are unsafe the arc builder hits a bare exitWith {} and the team gets zero waypoints and freezes. Replace with a direct AIMoveTo … "SAD" to town centre.	Server/Functions/Server_AI_SetTownAttackPath.sqf:49	trivial	low	The freeze is via the AssignTowns sticky-order guard (it won't re-dispatch a team that already has a valid target), not wfbe_exec_sig as first claimed. The final-SAD jitter is already tight on this branch — verify before changing it.
C2	Mop-up despawn scan → all hostiles. The post-capture squad counts only resistance units, so in WEST-vs-EAST it despawns instantly and leaves the town open. Count every side hostile to the new owner.	Server/FSM/server_town.sqf (mop-up block)	trivial	low	Use the existing _townRange and entity-class list (not a hardcoded 200 m / Man-only), keep the vehicle-crew inner loop, and declare the new var in the Private[] array.
C3	Arc shortcut + hop depth. A hardcoded random 100 < 30 early-exitWith gives ~1 in 3 dispatches only a single flanking hop; a second random 100 < 50 per-hop exit dominates path quality. Parameterise the 30 % to 0 and raise the hop ceiling.	Server/Functions/Server_AI_SetTownAttackPath.sqf:41,62; Init_CommonConstants.sqf	trivial	low	The shortcut does not reduce the team to one waypoint (depot WPs are always appended after) — it skips the flanking variety. Address the 50 % per-hop exit in the same pass or the effect is small.
C4	Forward reinforcement reach. WFBE_C_AICOM_FWD_REINFORCE_RANGE was already raised to 900 m (Init_CommonConstants.sqf:544, FILL-FIX 2026-06-18). This item is done; verify deep-spearhead refill behavior in soak.	Init_CommonConstants.sqf:544	small	low	Keep the owned-town gate — it is intentional (front-line resupply, not a bug). Do not remove it as first proposed; that would let teams resupply next to enemy towns.

Tier 2 — Fun quick-wins (telegraph + counterplay)

The unifying theme: make the AI's intent legible and give players a reaction window. Most of the "unfair" feeling is opacity, and most of the "easy" feeling is determinism.

#	Change	Files	Effort	Risk	Verifier correction
F1	Broadcast posture + HQ-strike. Surface PRESS/DEFEND/HOLD/HQ_STRIKE transitions and the HQ-strike launch as side radio/subtitle messages (fire-on-change). Turns invisible strategy into dramatic, reactable beats.	AI_Commander_Strategy.sqf; Server_SideMessage.sqf; Client/kb/hq.bikb; constants	small	low	Requires all four steps or it silently no-ops: add the Server_SideMessage switch case, register the sentence classes in hq.bikb, gate on a new WFBE_C_AICOM_ANNOUNCE, and note enableSentences = false ⇒ subtitle text only, no VO. The "600 m trigger" in the draft was invented.
F2	Relief deliberation delay. Today relief dispatches the instant wfbe_active fires (detection, ~600 m), before the attacker is in position — the AI "reads your mind". Buffer threatened towns and only commit if still contested on the next strategy tick (WFBE_C_AICOM_RELIEF_DELAY = 60). Enables the feint: poke a town, draw the reliever, hit the real objective.	AI_Commander_Strategy.sqf; constants	small	low	Verified accurate and A2-safe — but the draft showed only the consumer loop. Must also add the producer that inserts [town, time] tuples, or the buffer is always empty.
F3	Spread the capture force. The depot hold uses WEDGE, so one grenade/mortar wipes a 150 s capture attempt. Alternate LINE/STAG COLUMN at the hold.	Common_RunCommanderTeam.sqf:637	trivial	low	Scope to the depot-hold line only — the road-march arrival handoff (different lines) is a separate posture; don't change both in one edit.
F4	Break the clockwork camp dwell. A fixed sleep 45 per camp lets defenders time the AI to the second. Randomise it (sleep (35 + random 20)).	Common_RunCommanderTeam.sqf:551	trivial	low	Do not use the proposed "exit when the camp flips" poll here — it causes a zero-dwell regression on already-held camps and abandons freshly-flipped camps. Randomised dwell is the safe form; the flip-gated loop already exists in the separate camp-first phase.

Tier 3 — Capture levers behind an owner veto (sign-off required)

Both are one-line flips with the supporting code already in place, but each reverses an explicit, dated owner decision. Do not ship without Steff's go. Note: the isNil-guard form does not work here — both constants are bare unconditional assignments, so the value must be changed at the assignment line.

#	Change	Owner decision being reversed
G1	Enable mobile garrison (WFBE_C_AI_COMMANDER_GARRISON = 1) and cut the static-defense delay (WFBE_C_TOWNS_DEFENSE_SPAWN_DELAY 300 → 90 at line 1010 of Init_CommonConstants.sqf). If garrison is on, raise TEAMS_TARGET to keep attacker count, don't lower it.	"Owner call 2026-06-11: OFF by default — everything goes to the front."
G2	Enable AI artillery (WFBE_C_AI_COMMANDER_ARTILLERY = 1 at line 335 of Init_CommonConstants.sqf). The friendly-fire guard (_ownNear == 0) and range check are already implemented; adds a real suppressive-threat dimension.	"Steff 2026-06-13: the AI must NOT be able to use artillery. Forced off so no param can enable it." Likely placed after a team-kill/imbalance incident — confirm the original concern is resolved first.

Bigger bets (design work, not yet built)

These need design before code and are where the AI stops feeling like a script and starts feeling like an opponent:

• Retreat / withdrawal. The AI never gives ground — sub-30 % teams sit and die in place. A fallback-and-regroup behavior adds maneuver and the "exploit a retreat" moment.
• HQ relocation + base-defense reaction. The MHQ never moves (always the start grid) and the AI never reacts to its own Command Center being demolished. Counter-basing is currently a human-only tactic; the AI should both relocate under threat and recall teams when its base is hit.
• Adaptive composition. The type mix is a static [0.65,0.20,0.12,0.03]; the AI never answers an all-air or all-armor human with AA/AT. Read the enemy order of battle and bias the next teams.
• Player agency on the AI side. Players commanded by the AI are spectators — no way to donate, request, or suggest a target (and the donation path currently double-charges, see below). A lightweight "request attack / reinforce here" channel would restore agency.
• Reusable air. Air teams do a one-shot insert then walk for the rest of the match (heli refunded). Let them re-task as air.
• Decisive endgame. HQ-hunt needs both a 1.5× town edge and 1.1× strength, so balanced rounds never end. Loosen so AI-vs-AI and stalemated rounds reach a conclusion.

Discovered bugs (fix regardless of the plan)

• Donation double-debit — donating to the AI commander charges the donor twice (GUI_TransferMenu.sqf client-side debit + RequestAIComDonate.sqf server-side debit; confirmed in the deploy JOURNAL as G6). Discourages the co-op economy entirely.
• Wildcard announce — conflicting reads. Two subsystem readers reported wildcard draws (paradrop, air cavalry, supply surge) are silent to players; the adversarial verifier re-read the file and found a shared announce block (WFBE_CO_FNC_SendToClients … "LocalizeMessage") already present. Quick human check which is true on the deployed build before acting.

Telemetry to confirm any change

The commander already emits a rich diag stream (see AICOM Logging and AICOMSTAT Telemetry). Read these to grade a fix without guessing:

• CMDRSTAT … unitsPerTeam — should climb from ~5.5 toward 8–11 after the deficit-fill / punchy merges.
• AICOMSTAT … FRONT / POSTURE — front coherence and stance transitions.
• AICOMDBG … SPEARHEAD — target choice (town, supply, distance-to-front, on-front?).
• STUCKSTAT — dispatch-vs-arrive ratio; the live baseline was 256 dispatches : 13 arrivals. Arrivals should rise sharply after the arc-fallback (C1) and reach fixes.
• AICOMSTAT … ECONOMY netFunds — should stop going negative in long sessions once kill-rewards land.

Continue Reading

• AI Commander Execution Loop Reference — the supervisor loop and worker cadences these changes touch.
• AI Commander Autonomy Audit — which workers are actually wired vs dead.
• AI Commander Tunable Constants Reference — every constant cited here, by source line and default.
• Towns, Camps and Capture Atlas — the supply-attrition capture mechanic in full.
• AI Commander Logging and AICOMSTAT Telemetry — the diag tags used to grade these fixes.
• Quad AI Commander — the multi-side AI-commander architecture.
• Pending Owner Decisions — where the Tier 3 artillery/garrison flips should be logged for sign-off.