Stigmergy means coordination through the environment: no termite knows the plan, each reacts to marks left by others. Here every termite is an agent on a grid of soil and open tunnel cells (side view of the soil, one cell is about 1 cm).
Food patches emit scent that diffuses through the soil and decays: dc/dt = D laplacian(c) - k c; because diffusion is fast compared with digging, the field is taken at steady state, laplacian(c) = c / lambda^2 with lambda = sqrt(D/k) = 30 cm, solved by successive over-relaxation with c = 1 on food. A searching termite picks a neighbouring cell with weight w = exp(b (ln c_n - ln c_here)), a biased random walk up the scent gradient (b is the sensing strength). Choosing a soil cell means trying to dig: it succeeds with probability p = g r (1 + 2 B/(B+0.1)), where the geometry factor g is 1 at a tunnel face, up to 0.012 beside a busy tunnel (slow widening by traffic) and 0 elsewhere, so tunnels stay roughly one cell wide where B is the local build pheromone left by recent diggers, a positive feedback that makes tunnels branch and then thicken.
A termite that reaches food carries it home by descending the distance-to-nest field, laying a trail pheromone that evaporates at rate rho and diffuses a little. Searchers prefer cells with trail T/(T+0.3) times the trail pull, so successful routes are reinforced and unused ones fade, as in ant and termite foraging models.
Numerical check: the same diffusion stepper is run from a point source and the spatial variance is compared with the exact Gaussian result sigma^2 = 2 D t per axis, plus total-mass conservation. The "path / straight" readout compares the carved nest-to-food route (breadth-first search over open cells) to the straight-line distance.
Simplifications: 2D side view, discrete cells and ticks, no moisture/temperature, no castes (workers only), quasi-static scent, no real mound ventilation. It illustrates the mechanism, not a particular species.