Development log · 2026-06-08

The Combat Outpost, made real

The COP was geometrically correct and dramatically inert — a cluster of b-huts on a tan patch, defended by eight identical nubs, garrisoned by men who loitered, with a fire plan no soldier would recognise and a wire attack you could only watch as a cutscene. This pass rebuilt every aspect of it toward one bar: a skeptical infantryman looks at it and says "holy shit — that's a COP." It was driven by a six-specialist design survey, a synthesis judge, and seven atomic commits, each verified.

3 → 1blind heuristics replaced by one terrain spine

0 → 1Mk19 on the wall (sited on 94%-dead ground)

dead → livefob.hesco: now written & read

cutscene → the watcha wire assault now demands a decision

0 / 9copaudit sealed pockets — connectivity untouched

The shape of the problem

The outpost data model — CopLayout in terrain.ts — was clean and the movement audit (copaudit) was already 100% green across nine seeds: egress unblocked, 97% perimeter ring, zero sealed pockets, zero unreachable posts. So this was never a bug hunt. It was the harder job: taking a system that works and making it convincing — defensible, lived-in, alive.

A six-agent survey (one specialist each on defensive doctrine, visual fidelity, garrison life, facilities, gameplay, and atmosphere), grounded in the real code and in ATP 3-21.8, fed a synthesis judge. Its verdict was a Law-6 unification — three robust mechanisms instead of a pile of features:

One terrain-geometry spine feeds the render silhouette, the mortar pit, the crew-served siting, and the final protective fire. One render env seam makes the base the warm, lit, human place in a dark valley. One reused fire-request loop turns a wire attack into the commander's watch.

1 · The fire plan, sited by the terrain

Before, the perimeter was eight identical positions {id,cx,cy,facing,tower}; crew-served guns were handed out by raw array index (m240→fps[0]), the tower flag was fp % 3, the "mortar pit" was the flagpole, and there was no Mk19 at all. A range card a soldier would laugh at.

Now a deterministic line-of-sight sweep runs over the baked elevation inside buildCop(). For every position it marches its outward sector cell-by-cell, tracking the rising horizon, and measures two things: the avenue score — how far down the sector the gun can graze before terrain masks the ground — and the dead-space fraction. Then it sites the weapons the way a platoon would:

the M2 .50 cal (1830 m) takes the longest open avenue;
the M240B (1100 m) takes the next;
the Mk19 grenade launcher — whose 40 mm plunges — takes the worst dead-space sector, exactly where direct grazing fire fails and only high-angle fire reaches.

Bar chart of each fighting position's avenue reach and dead space, coloured by assigned weapon — The data that drives the siting (seed valley-2533). The .50 lands on the 922 m avenue with the lowest dead space; the Mk19 lands on the 195 m position that is **94% dead ground** — the defilade nothing else can touch. No hand-placement: pure terrain geometry, identical on every replay.

Top-down tactical diagram of the COP showing each crew-served sector of fire as a wedge — The same fire plan as a tactical overlay. Wedge length is scaled to avenue reach — you can see the .50's long arc to the south-west, the M240 covering south, and the Mk19's short, steep sector to the north registered onto the **FPF** (the most dangerous, most-concealed avenue). The mortar pit sits in rear defilade.

Why it's honest: the dead-space numbers are high (58–94%). That's not a bug — it's the Korengal. A COP on a knob in that terrain has brutal dead ground close in; it is literally why Restrepo and Keating were so dangerous. The sweep found the real masks, and the Mk19 was placed to answer them.

2 · A strongpoint you can see

The wire used to be a flat tan tint in the baked terrain. The art to fix it already existed in the sprite manifest — hesco-straight, antenna-array, jersey-barrier, ico-mortar — and was never drawn. drawCop() now paints the fortified silhouette: a real HESCO bastion wall of gabion segments, a comms mast on the skyline, the mortar pit, a serpentine ECP of staggered T-walls, drifting burn-pit smoke, a flag that streams downwind, and the crew-served sectors of fire colour-coded by weapon.

The COP before the overhaul — a cluster of buildings on a tan patch — **Before.** Functional, but the perimeter doesn't read as a wall and the fire plan is invisible.

The COP after — a recognisable fortified outpost with a HESCO wall and sectors — **After.** The bastion ring reads as a wall; sectors, mortar pit, antenna, ECP and motor pool all legible — a recognisable outpost.

3 · The one warm place in a dark valley

At night the COP used to go dark exactly like empty hillside. The fix is the second unifying idea: the drawCop signature was widened to take a render-only env struct ({night, windX, windY, tNow}, mirroring the weather view-struct and never re-entering the sim). An additive lighter pass then emits warm window glow (the TOC, aid station and chow hall brightest — always manned), cold floodlight pools over the gate and LZ, and slow-blinking red tower beacons — all gated on darkness so a daytime frame pays nothing.

The COP at 02:00 — warm lit windows and floodlights against a dark valley — 02:00. The outpost is the one warm, human place in a cold valley — exactly the feeling of returning to the wire from a night patrol. Render-only; the sim clock is untouched.

4 · The watch — the hardest part of command

The design's soul is "you watch." Yet a complex attack on the COP itself was a passive cutscene: the garrison has no maneuver task, so the call-for-fire system never fired for it. Now tickCopDefense() watches the wire. When fighters mass against it while the garrison stands to, the TOC raises a Final Protective Fire request — and the clock auto-pauses until you CLEAR HOT or DENY.

The COP under assault: stand-to in red, and a Final Protective Fire request awaiting approval — The watch, live. Left panel flips to **⚠ STAND-TO — manning the wire**; the centre demands a decision: *"6 fighters in the wire — FPF, danger close. Requesting 60mm on grid 177-361."* The mortar fires *from the pit*, onto the assault as it forms. The FPF is danger-close by doctrine — and only the commander can accept that risk.

It reuses the squad fire-request loop verbatim — one pending request, a 45-second cooldown, and it never auto-fires — so it obeys the ROE law: combat is 100% AI, but the human clears the fires. Verified headlessly on valley-2533: a quiet COP raises nothing; six fighters at the wire raise the FPF; approving it fires the mortar from the pit, expends four rounds (110→106), and kills an attacker.

5 · A garrison that lives

The COP was manned but lifeless — off-duty men loitered, sentries stared one way, stand-to only ever reacted to contact, and fob.hesco (a 0–100 fortification stat) was written at birth and read nowhere. Three additions, all deterministic, wired into tickGarrison:

Work details — "a COP is never finished." A rotating slice of the off-duty riflemen improve the wire; while they work, fob.hesco climbs (60.95→62.56 over two game-hours in test). The dead stat finally has a writer — and the squad panel a reader: a COP integrity bar.
Dawn/dusk stand-to (BMNT/EENT) — the whole garrison mans the wire at first and last light, the ritual every infantryman knows and the enemy favours, not just a reaction to a threat.
Sweeping sentries — a guard now sweeps his gaze across his fighting position's sector (from the Phase-1 limits) instead of staring one fixed bearing.

The squad readout showing the garrison activity and COP integrity bar — The garrison is now legible. At a glance: *9/9 effective · on guard — watching their sectors · COP integrity 60%*. Sample a midday COP and you find 7 on guard, 2 MG crews, 11 men on work detail, the medic in the aid station, and the leaders in the TOC — each shown for what it is.

Verification

Every commit kept the standing checks green and was proved before it landed.

Check	Result
copaudit (9 seeds)	clean — 0 sealed pockets, 0 unreachable posts, 97% ring, 0 wire-pins (connectivity untouched by the mortar pit / new positions)
smoke (serialize round-trip)	SMOKE OK — no NaN, save/load identical; `cop` is seed-rebuilt, so zero migration
balance (stall check)	8/8 deployments saw a firefight, 0 stranded, no stalls (KIA 0.13)
determinism	bit-identical — the LOS sweep and the sentry sweep reproduce exactly across two same-seed runs
tsc / build	clean / green across all seven commits
the watch (live + headless)	fires — request raised, auto-pause, approve → mortar from the pit, attacker down

How it was built

This was an orchestration, not a solo sprint. A workflow fanned out six independent specialists — each pinned to a senior persona, the real file:line facts, and ATP 3-21.8 — to survey one axis of the COP and return scored, doctrine-cited proposals. A seventh judge deduped them into a partitioned, dependency-ordered build plan (data spine first; render, world, and garrison behind it, one writer per hot file). Then seven atomic commits implemented it, each with its own headless probe and live capture, on a main shared with other agents committing in parallel.

survey ─┬─ defense ┐ ├─ visual │ ├─ garrison ├─► judge ─► partitioned plan ─► 7 atomic commits ─► verify ─► this report ├─ facilities│ ├─ gameplay │ └─ atmosphere┘

What we did not do (restraint, logged)

No terrain-stamped ECP chicane. A real serpentine would stamp impassable stubs into the gate apron — and risk pinching the proven ≥3-cell egress corridor that copaudit guards. The chicane is render-only; the apron is untouched.
fob.hesco doesn't yet feed combat. It is now written (work details) and read (the integrity bar), but coupling it to the FPF threshold or the enemy's effectiveness changes firefight balance and needs held-out tuning — deferred deliberately rather than shipped unproven.
The Mk19 isn't crewed full-time. Only two machine-gunners exist; the Mk19 sits emplaced and ready, manned at stand-to — realistic, not a stub to paper over.

Engineering record: docs/progress/2026-06-08-cop-overhaul/ (baselines, after-numbers, live + diagram assets). Probes & tools: scripts/{cop-shot,cop-diagram}.{mjs,ts}, scripts/copaudit.ts. Touched: terrain.ts (the spine), world.ts + helpers.ts (mortar origin + crew siting + the watch), draw.ts + WorldView.tsx (the silhouette + atmosphere), garrison.ts (the living garrison), DeployScreen.tsx (legibility). Built and verified by an autonomous agent; numbers are quoted as measured.