diff --git a/src/EntityBehavior.cpp b/src/EntityBehavior.cpp
index 4f8b423..98371f4 100644
--- a/src/EntityBehavior.cpp
+++ b/src/EntityBehavior.cpp
@@ -2,9 +2,22 @@
 #include "UIEntity.h"
 #include "UIGrid.h"
 #include "UIGridPathfinding.h"
+#include "PathProvider.h"
 #include <random>
 #include <algorithm>
 
+// Out-of-line reset lives here so the header can forward-declare PathProvider.
+void EntityBehavior::reset() {
+    type = BehaviorType::IDLE;
+    waypoints.clear();
+    current_waypoint_index = 0;
+    patrol_direction = 1;
+    current_path.clear();
+    path_step_index = 0;
+    sleep_turns_remaining = 0;
+    path_provider.reset();
+}
+
 // Thread-local random engine for behavior randomness
 static thread_local std::mt19937 rng{std::random_device{}()};
 
@@ -226,78 +239,28 @@ static BehaviorOutput executeSleep(UIEntity& entity, UIGrid& grid) {
     return {BehaviorResult::NO_ACTION, {}};
 }
 
-static BehaviorOutput executeSeek(UIEntity& entity, UIGrid& grid) {
+// SEEK and FLEE share one implementation now: both delegate to the active
+// PathProvider. FLEE differs only in which map is stored in the provider -
+// DijkstraProvider over an inverted DijkstraMap descends away from the threat,
+// which matches the old max-distance-neighbor behavior.
+static BehaviorOutput executeProviderStep(UIEntity& entity, UIGrid& grid) {
     auto& behavior = entity.behavior;
-
-    if (!behavior.dijkstra_map) {
+    if (!behavior.path_provider) {
         return {BehaviorResult::NO_ACTION, {}};
     }
 
-    // Use Dijkstra map to find the lowest-distance neighbor (moving toward target)
     int cx = entity.cell_position.x;
     int cy = entity.cell_position.y;
-    float best_dist = std::numeric_limits<float>::max();
-    sf::Vector2i best_cell = {cx, cy};
-    bool found = false;
+    bool ok = false;
+    sf::Vector2i next = behavior.path_provider->nextStep({cx, cy}, grid, &ok);
 
-    sf::Vector2i dirs[] = {{0, -1}, {0, 1}, {-1, 0}, {1, 0},
-                           {-1, -1}, {1, -1}, {-1, 1}, {1, 1}};
-
-    for (auto& dir : dirs) {
-        int nx = cx + dir.x;
-        int ny = cy + dir.y;
-        if (!isCellWalkable(grid, nx, ny)) continue;
-
-        float dist = behavior.dijkstra_map->getDistance(nx, ny);
-        if (dist >= 0 && dist < best_dist) {
-            best_dist = dist;
-            best_cell = {nx, ny};
-            found = true;
-        }
-    }
-
-    if (!found || (best_cell.x == cx && best_cell.y == cy)) {
+    if (!ok) {
         return {BehaviorResult::BLOCKED, {cx, cy}};
     }
-
-    return {BehaviorResult::MOVED, best_cell};
-}
-
-static BehaviorOutput executeFlee(UIEntity& entity, UIGrid& grid) {
-    auto& behavior = entity.behavior;
-
-    if (!behavior.dijkstra_map) {
-        return {BehaviorResult::NO_ACTION, {}};
-    }
-
-    // Use Dijkstra map to find the highest-distance neighbor (fleeing from target)
-    int cx = entity.cell_position.x;
-    int cy = entity.cell_position.y;
-    float best_dist = -1.0f;
-    sf::Vector2i best_cell = {cx, cy};
-    bool found = false;
-
-    sf::Vector2i dirs[] = {{0, -1}, {0, 1}, {-1, 0}, {1, 0},
-                           {-1, -1}, {1, -1}, {-1, 1}, {1, 1}};
-
-    for (auto& dir : dirs) {
-        int nx = cx + dir.x;
-        int ny = cy + dir.y;
-        if (!isCellWalkable(grid, nx, ny)) continue;
-
-        float dist = behavior.dijkstra_map->getDistance(nx, ny);
-        if (dist >= 0 && dist > best_dist) {
-            best_dist = dist;
-            best_cell = {nx, ny};
-            found = true;
-        }
-    }
-
-    if (!found || (best_cell.x == cx && best_cell.y == cy)) {
+    if (next.x == cx && next.y == cy) {
         return {BehaviorResult::BLOCKED, {cx, cy}};
     }
-
-    return {BehaviorResult::MOVED, best_cell};
+    return {BehaviorResult::MOVED, next};
 }
 
 // =============================================================================
@@ -314,8 +277,8 @@ BehaviorOutput executeBehavior(UIEntity& entity, UIGrid& grid) {
         case BehaviorType::PATROL:   return executePatrol(entity, grid);
         case BehaviorType::LOOP:     return executeLoop(entity, grid);
         case BehaviorType::SLEEP:    return executeSleep(entity, grid);
-        case BehaviorType::SEEK:     return executeSeek(entity, grid);
-        case BehaviorType::FLEE:     return executeFlee(entity, grid);
+        case BehaviorType::SEEK:     return executeProviderStep(entity, grid);
+        case BehaviorType::FLEE:     return executeProviderStep(entity, grid);
     }
     return {BehaviorResult::NO_ACTION, {}};
 }
diff --git a/src/EntityBehavior.h b/src/EntityBehavior.h
index 8432683..5cf219a 100644
--- a/src/EntityBehavior.h
+++ b/src/EntityBehavior.h
@@ -8,6 +8,7 @@
 class UIEntity;
 class UIGrid;
 class DijkstraMap;
+class PathProvider;
 
 // =============================================================================
 // BehaviorType - matches Python mcrfpy.Behavior enum values
@@ -60,19 +61,11 @@ struct EntityBehavior {
     // Sleep data
     int sleep_turns_remaining = 0;
 
-    // Dijkstra map (for SEEK/FLEE)
-    std::shared_ptr<DijkstraMap> dijkstra_map;
+    // SEEK/FLEE pathfinding strategy (#315). Nullptr means NO_ACTION.
+    std::unique_ptr<PathProvider> path_provider;
 
-    void reset() {
-        type = BehaviorType::IDLE;
-        waypoints.clear();
-        current_waypoint_index = 0;
-        patrol_direction = 1;
-        current_path.clear();
-        path_step_index = 0;
-        sleep_turns_remaining = 0;
-        dijkstra_map = nullptr;
-    }
+    // Defined in EntityBehavior.cpp to avoid needing the full PathProvider type here.
+    void reset();
 };
 
 // =============================================================================
diff --git a/src/McRFPy_API.cpp b/src/McRFPy_API.cpp
index 95a8d2e..e944be5 100644
--- a/src/McRFPy_API.cpp
+++ b/src/McRFPy_API.cpp
@@ -17,6 +17,7 @@
 #include "PyMouseButton.h"
 #include "PyInputState.h"
 #include "PyPerspective.h"
+#include "PyHeuristic.h"
 #include "PyBehavior.h"
 #include "PyTrigger.h"
 #include "UIGridView.h"
@@ -793,6 +794,12 @@ PyObject* PyInit_mcrfpy()
         PyErr_Clear();
     }
 
+    // Add Heuristic enum class for A* heuristic selection (#315)
+    PyObject* heuristic_class = PyHeuristic::create_enum_class(m);
+    if (!heuristic_class) {
+        PyErr_Clear();
+    }
+
     // Add Alignment enum class for automatic child positioning
     PyObject* alignment_class = PyAlignment::create_enum_class(m);
     if (!alignment_class) {
diff --git a/src/PathProvider.cpp b/src/PathProvider.cpp
new file mode 100644
index 0000000..d697b09
--- /dev/null
+++ b/src/PathProvider.cpp
@@ -0,0 +1,65 @@
+#include "PathProvider.h"
+#include "UIGrid.h"
+#include "UIGridPathfinding.h"
+#include "UIGridPoint.h"
+
+static bool cellWalkable(UIGrid& grid, int x, int y) {
+    if (x < 0 || x >= grid.grid_w || y < 0 || y >= grid.grid_h) return false;
+    return grid.at(x, y).walkable;
+}
+
+// -----------------------------------------------------------------------------
+// DijkstraProvider
+// -----------------------------------------------------------------------------
+DijkstraProvider::DijkstraProvider(std::shared_ptr<DijkstraMap> map)
+    : map_(std::move(map)) {}
+
+sf::Vector2i DijkstraProvider::nextStep(sf::Vector2i from, UIGrid& /*grid*/, bool* ok) {
+    if (!map_) {
+        if (ok) *ok = false;
+        return {-1, -1};
+    }
+    bool valid = false;
+    sf::Vector2i step = map_->descentStep(from.x, from.y, &valid);
+    if (ok) *ok = valid;
+    return step;
+}
+
+// -----------------------------------------------------------------------------
+// AStarProvider
+// -----------------------------------------------------------------------------
+AStarProvider::AStarProvider(std::vector<sf::Vector2i> path)
+    : path_(std::move(path)) {}
+
+sf::Vector2i AStarProvider::nextStep(sf::Vector2i /*from*/, UIGrid& /*grid*/, bool* ok) {
+    if (index_ >= path_.size()) {
+        if (ok) *ok = false;
+        return {-1, -1};
+    }
+    if (ok) *ok = true;
+    return path_[index_++];
+}
+
+// -----------------------------------------------------------------------------
+// TargetProvider
+// -----------------------------------------------------------------------------
+TargetProvider::TargetProvider(sf::Vector2i target)
+    : target_(target) {}
+
+sf::Vector2i TargetProvider::nextStep(sf::Vector2i from, UIGrid& grid, bool* ok) {
+    int dx = target_.x - from.x;
+    int dy = target_.y - from.y;
+    if (dx == 0 && dy == 0) {
+        if (ok) *ok = false;
+        return {-1, -1};  // already at target
+    }
+    int sx = (dx > 0) ? 1 : ((dx < 0) ? -1 : 0);
+    int sy = (dy > 0) ? 1 : ((dy < 0) ? -1 : 0);
+    sf::Vector2i step{from.x + sx, from.y + sy};
+    if (!cellWalkable(grid, step.x, step.y)) {
+        if (ok) *ok = false;
+        return {-1, -1};
+    }
+    if (ok) *ok = true;
+    return step;
+}
diff --git a/src/PathProvider.h b/src/PathProvider.h
new file mode 100644
index 0000000..9ccc5c9
--- /dev/null
+++ b/src/PathProvider.h
@@ -0,0 +1,64 @@
+#pragma once
+#include "Common.h"
+#include <memory>
+#include <vector>
+
+class UIGrid;
+class DijkstraMap;
+
+// =============================================================================
+// PathProvider (#315) - strategy interface for "what's my next cell?"
+//
+// EntityBehavior's SEEK/FLEE execute step() by asking the active PathProvider
+// for a single cell. Three concrete providers satisfy every pathfinding shape
+// the engine exposes today.
+// =============================================================================
+class PathProvider {
+public:
+    virtual ~PathProvider() = default;
+
+    // Return the next cell to step to. Sets *ok=true on a valid step, false
+    // otherwise. The provider is responsible for walkability checks -
+    // DijkstraProvider relies on the TCOD map used at compute time,
+    // TargetProvider re-queries the live grid, AStarProvider trusts the
+    // pre-computed path.
+    virtual sf::Vector2i nextStep(sf::Vector2i from, UIGrid& grid, bool* ok) = 0;
+
+    // Hint for providers that hold iteration state (currently only A*).
+    virtual void reset() {}
+};
+
+// Descend a precomputed DijkstraMap. For SEEK, pass the map as-is; for FLEE,
+// pass DijkstraMap.inverted().
+class DijkstraProvider : public PathProvider {
+public:
+    explicit DijkstraProvider(std::shared_ptr<DijkstraMap> map);
+    sf::Vector2i nextStep(sf::Vector2i from, UIGrid& grid, bool* ok) override;
+
+private:
+    std::shared_ptr<DijkstraMap> map_;
+};
+
+// Replay a pre-computed A* path.
+class AStarProvider : public PathProvider {
+public:
+    explicit AStarProvider(std::vector<sf::Vector2i> path);
+    sf::Vector2i nextStep(sf::Vector2i from, UIGrid& grid, bool* ok) override;
+    void reset() override { index_ = 0; }
+
+private:
+    std::vector<sf::Vector2i> path_;
+    size_t index_ = 0;
+};
+
+// Take a single Chebyshev step toward a fixed target cell. Used when full
+// pathfinding is overkill (e.g. an adjacent target). Returns no-step if the
+// straight-line neighbor is blocked - callers can treat that as BLOCKED.
+class TargetProvider : public PathProvider {
+public:
+    explicit TargetProvider(sf::Vector2i target);
+    sf::Vector2i nextStep(sf::Vector2i from, UIGrid& grid, bool* ok) override;
+
+private:
+    sf::Vector2i target_;
+};
diff --git a/src/PyHeuristic.cpp b/src/PyHeuristic.cpp
new file mode 100644
index 0000000..cf7872c
--- /dev/null
+++ b/src/PyHeuristic.cpp
@@ -0,0 +1,142 @@
+#include "PyHeuristic.h"
+#include <cstring>
+#include <sstream>
+
+PyObject* PyHeuristic::heuristic_enum_class = nullptr;
+
+struct HeuristicEntry {
+    const char* name;
+    int value;
+};
+
+static const HeuristicEntry heuristic_table[] = {
+    {"EUCLIDEAN", 0},
+    {"MANHATTAN", 1},
+    {"CHEBYSHEV", 2},
+    {"DIAGONAL",  3},
+    {"ZERO",      4},
+};
+
+static const int NUM_HEURISTIC_ENTRIES =
+    sizeof(heuristic_table) / sizeof(heuristic_table[0]);
+
+PyObject* PyHeuristic::create_enum_class(PyObject* module) {
+    std::ostringstream code;
+    code << "from enum import IntEnum\n\n";
+    code << "class Heuristic(IntEnum):\n";
+    code << "    \"\"\"Built-in A* heuristic function selector.\n";
+    code << "    \n";
+    code << "    Values:\n";
+    code << "        EUCLIDEAN: sqrt((dx)^2 + (dy)^2). Admissible, default.\n";
+    code << "        MANHATTAN: |dx| + |dy|. Admissible on 4-connected grids.\n";
+    code << "        CHEBYSHEV: max(|dx|, |dy|). Admissible on 8-connected (diag=1).\n";
+    code << "        DIAGONAL: Octile distance. Admissible on 8-connected (diag=sqrt(2)).\n";
+    code << "        ZERO: Always returns 0. A* degenerates to Dijkstra.\n";
+    code << "    \"\"\"\n";
+    for (int i = 0; i < NUM_HEURISTIC_ENTRIES; i++) {
+        code << "    " << heuristic_table[i].name
+             << " = " << heuristic_table[i].value << "\n";
+    }
+    code << "\n";
+    code << "Heuristic.__hash__ = lambda self: hash(int(self))\n";
+    code << "Heuristic.__repr__ = lambda self: f\"{type(self).__name__}.{self.name}\"\n";
+    code << "Heuristic.__str__ = lambda self: self.name\n";
+
+    std::string code_str = code.str();
+
+    PyObject* globals = PyDict_New();
+    if (!globals) return NULL;
+    PyDict_SetItemString(globals, "__builtins__", PyEval_GetBuiltins());
+
+    PyObject* locals = PyDict_New();
+    if (!locals) { Py_DECREF(globals); return NULL; }
+
+    PyObject* result = PyRun_String(code_str.c_str(), Py_file_input, globals, locals);
+    if (!result) {
+        Py_DECREF(globals);
+        Py_DECREF(locals);
+        return NULL;
+    }
+    Py_DECREF(result);
+
+    PyObject* enum_class = PyDict_GetItemString(locals, "Heuristic");
+    if (!enum_class) {
+        PyErr_SetString(PyExc_RuntimeError, "Failed to create Heuristic enum class");
+        Py_DECREF(globals);
+        Py_DECREF(locals);
+        return NULL;
+    }
+    Py_INCREF(enum_class);
+
+    heuristic_enum_class = enum_class;
+    Py_INCREF(heuristic_enum_class);
+
+    if (PyModule_AddObject(module, "Heuristic", enum_class) < 0) {
+        Py_DECREF(enum_class);
+        Py_DECREF(globals);
+        Py_DECREF(locals);
+        heuristic_enum_class = nullptr;
+        return NULL;
+    }
+
+    Py_DECREF(globals);
+    Py_DECREF(locals);
+    return enum_class;
+}
+
+int PyHeuristic::from_arg(PyObject* arg, int* out_value) {
+    if (heuristic_enum_class && PyObject_IsInstance(arg, heuristic_enum_class)) {
+        PyObject* value = PyObject_GetAttrString(arg, "value");
+        if (!value) return 0;
+        long val = PyLong_AsLong(value);
+        Py_DECREF(value);
+        if (val == -1 && PyErr_Occurred()) return 0;
+        if (val < 0 || val >= NUM_HEURISTIC_VALUES) {
+            PyErr_Format(PyExc_ValueError, "Invalid Heuristic value: %ld", val);
+            return 0;
+        }
+        *out_value = static_cast<int>(val);
+        return 1;
+    }
+
+    if (PyLong_Check(arg)) {
+        long val = PyLong_AsLong(arg);
+        if (val == -1 && PyErr_Occurred()) return 0;
+        if (val < 0 || val >= NUM_HEURISTIC_VALUES) {
+            PyErr_Format(PyExc_ValueError,
+                "Invalid Heuristic value: %ld. Must be 0..4.", val);
+            return 0;
+        }
+        *out_value = static_cast<int>(val);
+        return 1;
+    }
+
+    if (PyUnicode_Check(arg)) {
+        const char* name = PyUnicode_AsUTF8(arg);
+        if (!name) return 0;
+        for (int i = 0; i < NUM_HEURISTIC_ENTRIES; i++) {
+            if (strcmp(name, heuristic_table[i].name) == 0) {
+                *out_value = heuristic_table[i].value;
+                return 1;
+            }
+        }
+        PyErr_Format(PyExc_ValueError,
+            "Unknown Heuristic: '%s'. Use EUCLIDEAN, MANHATTAN, CHEBYSHEV, DIAGONAL, or ZERO.", name);
+        return 0;
+    }
+
+    PyErr_SetString(PyExc_TypeError,
+        "Heuristic must be mcrfpy.Heuristic enum member, string, or int");
+    return 0;
+}
+
+TCOD_heuristic_func_t PyHeuristic::get_function(int heuristic_value) {
+    switch (heuristic_value) {
+        case EUCLIDEAN: return TCOD_heuristic_euclidean;
+        case MANHATTAN: return TCOD_heuristic_manhattan;
+        case CHEBYSHEV: return TCOD_heuristic_chebyshev;
+        case DIAGONAL:  return TCOD_heuristic_diagonal;
+        case ZERO:      return TCOD_heuristic_zero;
+        default:        return nullptr;
+    }
+}
diff --git a/src/PyHeuristic.h b/src/PyHeuristic.h
new file mode 100644
index 0000000..2227095
--- /dev/null
+++ b/src/PyHeuristic.h
@@ -0,0 +1,39 @@
+#pragma once
+#include "Common.h"
+#include "Python.h"
+#include <libtcod.h>
+#include <cstdint>
+
+// Module-level Heuristic enum class (created at runtime using Python's IntEnum)
+// Stored as a module attribute: mcrfpy.Heuristic
+//
+// Values:
+//   EUCLIDEAN = 0   (admissible, default, slowest-optimal)
+//   MANHATTAN = 1   (admissible on 4-connected)
+//   CHEBYSHEV = 2   (admissible on 8-connected, diag cost 1)
+//   DIAGONAL  = 3   (octile, admissible on 8-connected, diag cost sqrt(2))
+//   ZERO      = 4   (A* degenerates to Dijkstra)
+class PyHeuristic {
+public:
+    // Create the Heuristic enum class and add to module.
+    static PyObject* create_enum_class(PyObject* module);
+
+    // Helper to extract a Heuristic value from a Python arg.
+    // Accepts Heuristic enum member, string (enum name), or int 0..4.
+    // Returns 1 on success, 0 on error (with exception set).
+    static int from_arg(PyObject* arg, int* out_value);
+
+    // Returns the libtcod built-in heuristic function pointer for a given value.
+    // Returns nullptr if value is invalid.
+    static TCOD_heuristic_func_t get_function(int heuristic_value);
+
+    // Cached reference to the Heuristic enum class for fast type checking.
+    static PyObject* heuristic_enum_class;
+
+    static const int NUM_HEURISTIC_VALUES = 5;
+    static const int EUCLIDEAN = 0;
+    static const int MANHATTAN = 1;
+    static const int CHEBYSHEV = 2;
+    static const int DIAGONAL  = 3;
+    static const int ZERO      = 4;
+};
diff --git a/src/UIEntity.cpp b/src/UIEntity.cpp
index a4173d1..af53756 100644
--- a/src/UIEntity.cpp
+++ b/src/UIEntity.cpp
@@ -2,6 +2,7 @@
 #include "UIGrid.h"
 #include "UIGridView.h"  // #252: Entity.grid accepts GridView
 #include "UIGridPathfinding.h"
+#include "PathProvider.h"
 #include "McRFPy_API.h"
 #include <algorithm>
 #include <cstring>
@@ -1512,14 +1513,16 @@ int UIEntity::set_sight_radius(PyUIEntityObject* self, PyObject* value, void* cl
 }
 
 PyObject* UIEntity::py_set_behavior(PyUIEntityObject* self, PyObject* args, PyObject* kwds) {
-    static const char* kwlist[] = {"type", "waypoints", "turns", "path", nullptr};
+    static const char* kwlist[] = {"type", "waypoints", "turns", "path", "pathfinder", nullptr};
     int type_val = 0;
     PyObject* waypoints_obj = nullptr;
     int turns = 0;
     PyObject* path_obj = nullptr;
+    PyObject* pathfinder_obj = nullptr;
 
-    if (!PyArg_ParseTupleAndKeywords(args, kwds, "i|OiO", const_cast<char**>(kwlist),
-                                     &type_val, &waypoints_obj, &turns, &path_obj)) {
+    if (!PyArg_ParseTupleAndKeywords(args, kwds, "i|OiOO", const_cast<char**>(kwlist),
+                                     &type_val, &waypoints_obj, &turns, &path_obj,
+                                     &pathfinder_obj)) {
         return NULL;
     }
 
@@ -1582,6 +1585,35 @@ PyObject* UIEntity::py_set_behavior(PyUIEntityObject* self, PyObject* args, PyOb
         behavior.sleep_turns_remaining = turns;
     }
 
+    // Parse pathfinder (#315): DijkstraMap, AStarPath, or (x, y) target tuple.
+    if (pathfinder_obj && pathfinder_obj != Py_None) {
+        if (PyObject_IsInstance(pathfinder_obj, (PyObject*)&mcrfpydef::PyDijkstraMapType)) {
+            auto* dmap = (PyDijkstraMapObject*)pathfinder_obj;
+            if (!dmap->data) {
+                PyErr_SetString(PyExc_RuntimeError, "pathfinder: DijkstraMap is invalid");
+                return NULL;
+            }
+            behavior.path_provider = std::make_unique<DijkstraProvider>(dmap->data);
+        } else if (PyObject_IsInstance(pathfinder_obj, (PyObject*)&mcrfpydef::PyAStarPathType)) {
+            auto* apath = (PyAStarPathObject*)pathfinder_obj;
+            // Copy remaining steps - the provider owns its own iteration state.
+            std::vector<sf::Vector2i> steps(
+                apath->path.begin() + apath->current_index,
+                apath->path.end());
+            behavior.path_provider = std::make_unique<AStarProvider>(std::move(steps));
+        } else if (PyTuple_Check(pathfinder_obj) && PyTuple_Size(pathfinder_obj) == 2) {
+            long tx = PyLong_AsLong(PyTuple_GetItem(pathfinder_obj, 0));
+            long ty = PyLong_AsLong(PyTuple_GetItem(pathfinder_obj, 1));
+            if (PyErr_Occurred()) return NULL;
+            behavior.path_provider = std::make_unique<TargetProvider>(
+                sf::Vector2i(static_cast<int>(tx), static_cast<int>(ty)));
+        } else {
+            PyErr_SetString(PyExc_TypeError,
+                "pathfinder must be a DijkstraMap, AStarPath, or (x, y) tuple");
+            return NULL;
+        }
+    }
+
     Py_RETURN_NONE;
 }
 
diff --git a/src/UIGridPathfinding.cpp b/src/UIGridPathfinding.cpp
index 7a19813..59baa6f 100644
--- a/src/UIGridPathfinding.cpp
+++ b/src/UIGridPathfinding.cpp
@@ -5,49 +5,72 @@
 #include "McRFPy_API.h"
 #include "PyHeightMap.h"
 #include "PyPositionHelper.h"
+#include "PyHeuristic.h"
+#include "PyDiscreteMap.h"
 
 //=============================================================================
 // DijkstraMap Implementation
 //=============================================================================
 
 DijkstraMap::DijkstraMap(TCODMap* map, int root_x, int root_y, float diag_cost)
-    : tcod_map(map)
+    : tcod_dijkstra(nullptr)
+    , tcod_map(map)
     , root(root_x, root_y)
     , diagonal_cost(diag_cost)
     , map_width(map ? map->getWidth() : 0)
     , map_height(map ? map->getHeight() : 0)
 {
-    tcod_dijkstra = new TCODDijkstra(tcod_map, diagonal_cost);
-    tcod_dijkstra->compute(root_x, root_y);  // Compute immediately at creation
+    roots.push_back(sf::Vector2i(root_x, root_y));
+    if (tcod_map) {
+        tcod_dijkstra = TCOD_dijkstra_new(tcod_map->data, diagonal_cost);
+        TCOD_dijkstra_compute(tcod_dijkstra, root_x, root_y);
+    }
+}
+
+DijkstraMap::DijkstraMap(TCODMap* map, const std::vector<sf::Vector2i>& roots_in, float diag_cost)
+    : tcod_dijkstra(nullptr)
+    , tcod_map(map)
+    , root(roots_in.empty() ? sf::Vector2i(-1, -1) : roots_in.front())
+    , roots(roots_in)
+    , diagonal_cost(diag_cost)
+    , map_width(map ? map->getWidth() : 0)
+    , map_height(map ? map->getHeight() : 0)
+{
+    if (!tcod_map || roots.empty()) return;
+
+    tcod_dijkstra = TCOD_dijkstra_new(tcod_map->data, diagonal_cost);
+
+    if (roots.size() == 1) {
+        TCOD_dijkstra_compute(tcod_dijkstra, roots[0].x, roots[0].y);
+    } else {
+        std::vector<int> xs, ys;
+        xs.reserve(roots.size());
+        ys.reserve(roots.size());
+        for (auto& r : roots) { xs.push_back(r.x); ys.push_back(r.y); }
+        TCOD_dijkstra_compute_multi(tcod_dijkstra,
+            static_cast<int>(roots.size()), xs.data(), ys.data());
+    }
 }
 
 DijkstraMap::~DijkstraMap() {
     if (tcod_dijkstra) {
-        delete tcod_dijkstra;
+        TCOD_dijkstra_delete(tcod_dijkstra);
         tcod_dijkstra = nullptr;
     }
 }
 
 float DijkstraMap::getDistance(int x, int y) const {
     if (!tcod_dijkstra) return -1.0f;
-    return tcod_dijkstra->getDistance(x, y);
-}
-
-int DijkstraMap::getWidth() const {
-    return map_width;
-}
-
-int DijkstraMap::getHeight() const {
-    return map_height;
+    return TCOD_dijkstra_get_distance(tcod_dijkstra, x, y);
 }
 
 std::vector<sf::Vector2i> DijkstraMap::getPathFrom(int x, int y) const {
     std::vector<sf::Vector2i> path;
     if (!tcod_dijkstra) return path;
 
-    if (tcod_dijkstra->setPath(x, y)) {
+    if (TCOD_dijkstra_path_set(tcod_dijkstra, x, y)) {
         int px, py;
-        while (tcod_dijkstra->walk(&px, &py)) {
+        while (TCOD_dijkstra_path_walk(tcod_dijkstra, &px, &py)) {
             path.push_back(sf::Vector2i(px, py));
         }
     }
@@ -60,22 +83,47 @@ sf::Vector2i DijkstraMap::stepFrom(int x, int y, bool* valid) const {
         return sf::Vector2i(-1, -1);
     }
 
-    if (!tcod_dijkstra->setPath(x, y)) {
+    if (!TCOD_dijkstra_path_set(tcod_dijkstra, x, y)) {
         if (valid) *valid = false;
         return sf::Vector2i(-1, -1);
     }
 
     int px, py;
-    if (tcod_dijkstra->walk(&px, &py)) {
+    if (TCOD_dijkstra_path_walk(tcod_dijkstra, &px, &py)) {
         if (valid) *valid = true;
         return sf::Vector2i(px, py);
     }
 
-    // At root or no path
     if (valid) *valid = false;
     return sf::Vector2i(-1, -1);
 }
 
+void DijkstraMap::invertInPlace() {
+    if (tcod_dijkstra) {
+        TCOD_dijkstra_invert(tcod_dijkstra);
+    }
+}
+
+std::shared_ptr<DijkstraMap> DijkstraMap::inverted() const {
+    // Recompute from the stored roots, then invert. This preserves the invariant that
+    // the original's distance field is unchanged.
+    auto copy = std::make_shared<DijkstraMap>(tcod_map, roots, diagonal_cost);
+    copy->invertInPlace();
+    return copy;
+}
+
+sf::Vector2i DijkstraMap::descentStep(int x, int y, bool* valid) const {
+    if (!tcod_dijkstra) {
+        if (valid) *valid = false;
+        return sf::Vector2i(-1, -1);
+    }
+    int out_x = -1, out_y = -1;
+    bool ok = TCOD_dijkstra_get_descent(tcod_dijkstra, x, y, &out_x, &out_y);
+    if (valid) *valid = ok;
+    if (!ok) return sf::Vector2i(-1, -1);
+    return sf::Vector2i(out_x, out_y);
+}
+
 //=============================================================================
 // Helper Functions
 //=============================================================================
@@ -405,6 +453,48 @@ PyObject* UIGridPathfinding::DijkstraMap_get_root(PyDijkstraMapObject* self, voi
     return PyVector(sf::Vector2f(static_cast<float>(root.x), static_cast<float>(root.y))).pyObject();
 }
 
+PyObject* UIGridPathfinding::DijkstraMap_invert(PyDijkstraMapObject* self, PyObject* args) {
+    if (!self->data) {
+        PyErr_SetString(PyExc_RuntimeError, "DijkstraMap is invalid");
+        return nullptr;
+    }
+    auto new_map = self->data->inverted();
+    if (!new_map) {
+        PyErr_SetString(PyExc_RuntimeError, "invert() failed");
+        return nullptr;
+    }
+    PyDijkstraMapObject* result = (PyDijkstraMapObject*)mcrfpydef::PyDijkstraMapType.tp_alloc(
+        &mcrfpydef::PyDijkstraMapType, 0);
+    if (!result) return nullptr;
+    new (&result->data) std::shared_ptr<DijkstraMap>(new_map);
+    return (PyObject*)result;
+}
+
+PyObject* UIGridPathfinding::DijkstraMap_descent_step(PyDijkstraMapObject* self, PyObject* args, PyObject* kwds) {
+    static const char* kwlist[] = {"pos", nullptr};
+    PyObject* pos_obj = nullptr;
+    if (!PyArg_ParseTupleAndKeywords(args, kwds, "O", const_cast<char**>(kwlist), &pos_obj)) {
+        return nullptr;
+    }
+    if (!self->data) {
+        PyErr_SetString(PyExc_RuntimeError, "DijkstraMap is invalid");
+        return nullptr;
+    }
+    int x, y;
+    if (!ExtractPosition(pos_obj, &x, &y, nullptr, "pos")) {
+        return nullptr;
+    }
+    if (!dijkstra_bounds_check(self->data.get(), x, y)) return nullptr;
+
+    bool valid = false;
+    sf::Vector2i step = self->data->descentStep(x, y, &valid);
+    if (!valid) {
+        Py_RETURN_NONE;
+    }
+    return PyVector(sf::Vector2f(static_cast<float>(step.x),
+                                 static_cast<float>(step.y))).pyObject();
+}
+
 PyObject* UIGridPathfinding::DijkstraMap_to_heightmap(PyDijkstraMapObject* self, PyObject* args, PyObject* kwds) {
     static const char* kwlist[] = {"size", "unreachable", nullptr};
     PyObject* size_obj = nullptr;
@@ -520,14 +610,18 @@ static void restoreCollisionLabel(GridData* grid,
 }
 
 PyObject* UIGridPathfinding::Grid_find_path(PyUIGridObject* self, PyObject* args, PyObject* kwds) {
-    static const char* kwlist[] = {"start", "end", "diagonal_cost", "collide", NULL};
+    static const char* kwlist[] = {"start", "end", "diagonal_cost", "collide",
+                                   "heuristic", "weight", NULL};
     PyObject* start_obj = NULL;
     PyObject* end_obj = NULL;
     float diagonal_cost = 1.41f;
     const char* collide_label = NULL;
+    PyObject* heuristic_obj = NULL;
+    float heuristic_weight = 1.0f;
 
-    if (!PyArg_ParseTupleAndKeywords(args, kwds, "OO|fz", const_cast<char**>(kwlist),
-                                     &start_obj, &end_obj, &diagonal_cost, &collide_label)) {
+    if (!PyArg_ParseTupleAndKeywords(args, kwds, "OO|fzOf", const_cast<char**>(kwlist),
+                                     &start_obj, &end_obj, &diagonal_cost, &collide_label,
+                                     &heuristic_obj, &heuristic_weight)) {
         return NULL;
     }
 
@@ -536,6 +630,11 @@ PyObject* UIGridPathfinding::Grid_find_path(PyUIGridObject* self, PyObject* args
         return NULL;
     }
 
+    if (heuristic_weight <= 0.0f) {
+        PyErr_SetString(PyExc_ValueError, "weight must be positive");
+        return NULL;
+    }
+
     int x1, y1, x2, y2;
     if (!ExtractPosition(start_obj, &x1, &y1, self->data.get(), "start")) {
         return NULL;
@@ -551,51 +650,159 @@ PyObject* UIGridPathfinding::Grid_find_path(PyUIGridObject* self, PyObject* args
         return NULL;
     }
 
+    // Resolve heuristic selection before any allocations so we fail fast on bad args.
+    TCOD_heuristic_func_t heuristic_func = nullptr;
+    if (heuristic_obj && heuristic_obj != Py_None) {
+        int hval = 0;
+        if (!PyHeuristic::from_arg(heuristic_obj, &hval)) {
+            return NULL;
+        }
+        heuristic_func = PyHeuristic::get_function(hval);
+    }
+
     // Mark-and-restore: temporarily block cells occupied by entities with collide label
     std::string label_str = collide_label ? collide_label : "";
     auto restore_list = markCollisionLabel(self->data.get(), label_str);
 
-    // Compute path using temporary TCODPath
-    TCODPath tcod_path(self->data->getTCODMap(), diagonal_cost);
-    bool found = tcod_path.compute(x1, y1, x2, y2);
+    TCODMap* tcmap = self->data->getTCODMap();
+
+    // Build path handle. Use C API so we can set the heuristic/weight when requested.
+    TCOD_path_t tcod_path = TCOD_path_new_using_map(tcmap->data, diagonal_cost);
+    if (heuristic_func || heuristic_weight != 1.0f) {
+        // Passing null heuristic_func keeps the default (Euclidean) while still allowing
+        // weight override; non-null installs the chosen built-in.
+        TCOD_path_set_heuristic(tcod_path, heuristic_func, heuristic_weight);
+    }
+
+    bool found = TCOD_path_compute(tcod_path, x1, y1, x2, y2);
 
     // Restore walkability before returning
     restoreCollisionLabel(self->data.get(), restore_list);
 
     if (!found) {
-        Py_RETURN_NONE;  // No path exists
+        TCOD_path_delete(tcod_path);
+        Py_RETURN_NONE;
     }
 
-    // Create AStarPath result object
     PyAStarPathObject* result = (PyAStarPathObject*)mcrfpydef::PyAStarPathType.tp_alloc(
         &mcrfpydef::PyAStarPathType, 0);
-    if (!result) return NULL;
+    if (!result) {
+        TCOD_path_delete(tcod_path);
+        return NULL;
+    }
 
-    // Initialize
     new (&result->path) std::vector<sf::Vector2i>();
     result->current_index = 0;
     result->origin = sf::Vector2i(x1, y1);
     result->destination = sf::Vector2i(x2, y2);
 
-    // Copy path data
-    result->path.reserve(tcod_path.size());
-    for (int i = 0; i < tcod_path.size(); i++) {
+    int size = TCOD_path_size(tcod_path);
+    result->path.reserve(size);
+    for (int i = 0; i < size; i++) {
         int px, py;
-        tcod_path.get(i, &px, &py);
+        TCOD_path_get(tcod_path, i, &px, &py);
         result->path.push_back(sf::Vector2i(px, py));
     }
 
+    TCOD_path_delete(tcod_path);
     return (PyObject*)result;
 }
 
+// Collect roots from a Python object, which may be:
+//   - a single (x,y) (tuple/list/Vector/Entity)
+//   - a list/iterable of (x,y) positions
+//   - a DiscreteMap mask (non-zero cells become roots)
+// Returns true on success; populates `out_roots` and `out_mask_used`.
+// If a DiscreteMap mask is used, caller should prefer the masked C path.
+static bool collectRoots(PyObject* root_obj, UIGrid* grid,
+                         std::vector<sf::Vector2i>* out_roots,
+                         PyDiscreteMapObject** out_mask)
+{
+    out_roots->clear();
+    if (out_mask) *out_mask = nullptr;
+
+    // DiscreteMap mask path
+    if (PyObject_IsInstance(root_obj, (PyObject*)&mcrfpydef::PyDiscreteMapType)) {
+        auto* dmap = (PyDiscreteMapObject*)root_obj;
+        if (!dmap->data) {
+            PyErr_SetString(PyExc_RuntimeError, "DiscreteMap is invalid");
+            return false;
+        }
+        if (dmap->data->width() != grid->grid_w || dmap->data->height() != grid->grid_h) {
+            PyErr_Format(PyExc_ValueError,
+                "DiscreteMap size (%dx%d) does not match grid size (%dx%d)",
+                dmap->data->width(), dmap->data->height(),
+                grid->grid_w, grid->grid_h);
+            return false;
+        }
+        if (out_mask) *out_mask = dmap;
+        return true;
+    }
+
+    // Single position path (Vector / Entity / (x,y) tuple): ExtractPosition accepts these.
+    int x = 0, y = 0;
+    if (UIGridPathfinding::ExtractPosition(root_obj, &x, &y, grid, "root")) {
+        if (x < 0 || x >= grid->grid_w || y < 0 || y >= grid->grid_h) {
+            PyErr_SetString(PyExc_ValueError, "Root position out of grid bounds");
+            return false;
+        }
+        out_roots->push_back(sf::Vector2i(x, y));
+        return true;
+    }
+    // ExtractPosition set an error - clear it only if we still have an iterable to try.
+    if (!PyErr_ExceptionMatches(PyExc_TypeError)) {
+        return false;
+    }
+    PyErr_Clear();
+
+    // List/iterable of positions
+    if (PySequence_Check(root_obj) || PyIter_Check(root_obj)) {
+        PyObject* iter = PyObject_GetIter(root_obj);
+        if (!iter) {
+            PyErr_SetString(PyExc_TypeError,
+                "roots must be (x,y), a sequence of (x,y), or a DiscreteMap mask");
+            return false;
+        }
+        PyObject* item;
+        while ((item = PyIter_Next(iter)) != NULL) {
+            int rx = 0, ry = 0;
+            if (!UIGridPathfinding::ExtractPosition(item, &rx, &ry, grid, "root")) {
+                Py_DECREF(item);
+                Py_DECREF(iter);
+                return false;
+            }
+            Py_DECREF(item);
+            if (rx < 0 || rx >= grid->grid_w || ry < 0 || ry >= grid->grid_h) {
+                Py_DECREF(iter);
+                PyErr_Format(PyExc_ValueError,
+                    "Root (%d,%d) out of grid bounds", rx, ry);
+                return false;
+            }
+            out_roots->push_back(sf::Vector2i(rx, ry));
+        }
+        Py_DECREF(iter);
+        if (PyErr_Occurred()) return false;
+        if (out_roots->empty()) {
+            PyErr_SetString(PyExc_ValueError, "roots sequence is empty");
+            return false;
+        }
+        return true;
+    }
+
+    PyErr_SetString(PyExc_TypeError,
+        "roots must be (x,y), a sequence of (x,y), or a DiscreteMap mask");
+    return false;
+}
+
 PyObject* UIGridPathfinding::Grid_get_dijkstra_map(PyUIGridObject* self, PyObject* args, PyObject* kwds) {
-    static const char* kwlist[] = {"root", "diagonal_cost", "collide", NULL};
+    static const char* kwlist[] = {"root", "diagonal_cost", "collide", "roots", NULL};
     PyObject* root_obj = NULL;
+    PyObject* roots_obj = NULL;
     float diagonal_cost = 1.41f;
     const char* collide_label = NULL;
 
-    if (!PyArg_ParseTupleAndKeywords(args, kwds, "O|fz", const_cast<char**>(kwlist),
-                                     &root_obj, &diagonal_cost, &collide_label)) {
+    if (!PyArg_ParseTupleAndKeywords(args, kwds, "|OfzO", const_cast<char**>(kwlist),
+                                     &root_obj, &diagonal_cost, &collide_label, &roots_obj)) {
         return NULL;
     }
 
@@ -604,50 +811,82 @@ PyObject* UIGridPathfinding::Grid_get_dijkstra_map(PyUIGridObject* self, PyObjec
         return NULL;
     }
 
-    int root_x, root_y;
-    if (!ExtractPosition(root_obj, &root_x, &root_y, self->data.get(), "root")) {
+    // Accept either `root=` (back-compat, also accepts multi-input now) or `roots=`.
+    PyObject* input_obj = roots_obj ? roots_obj : root_obj;
+    if (!input_obj) {
+        PyErr_SetString(PyExc_TypeError,
+            "get_dijkstra_map() requires 'root' or 'roots' argument");
+        return NULL;
+    }
+    if (roots_obj && root_obj) {
+        PyErr_SetString(PyExc_TypeError,
+            "get_dijkstra_map(): pass 'root' or 'roots', not both");
         return NULL;
     }
 
-    // Bounds check
-    if (root_x < 0 || root_x >= self->data->grid_w || root_y < 0 || root_y >= self->data->grid_h) {
-        PyErr_SetString(PyExc_ValueError, "Root position out of grid bounds");
+    std::vector<sf::Vector2i> roots;
+    PyDiscreteMapObject* mask_obj = nullptr;
+    if (!collectRoots(input_obj, self->data.get(), &roots, &mask_obj)) {
         return NULL;
     }
 
     std::string label_str = collide_label ? collide_label : "";
-    auto key = std::make_tuple(root_x, root_y, label_str);
 
-    // Check cache
-    auto it = self->data->dijkstra_maps.find(key);
-    if (it != self->data->dijkstra_maps.end()) {
-        // Check diagonal cost matches
-        if (std::abs(it->second->getDiagonalCost() - diagonal_cost) < 0.001f) {
-            // Return existing
-            PyDijkstraMapObject* result = (PyDijkstraMapObject*)mcrfpydef::PyDijkstraMapType.tp_alloc(
-                &mcrfpydef::PyDijkstraMapType, 0);
-            if (!result) return NULL;
-            new (&result->data) std::shared_ptr<DijkstraMap>(it->second);
-            return (PyObject*)result;
+    // Cache path for the common single-root case (preserves prior behavior).
+    if (!mask_obj && roots.size() == 1) {
+        auto key = std::make_tuple(roots[0].x, roots[0].y, label_str);
+        auto it = self->data->dijkstra_maps.find(key);
+        if (it != self->data->dijkstra_maps.end()) {
+            if (std::abs(it->second->getDiagonalCost() - diagonal_cost) < 0.001f) {
+                PyDijkstraMapObject* result = (PyDijkstraMapObject*)mcrfpydef::PyDijkstraMapType.tp_alloc(
+                    &mcrfpydef::PyDijkstraMapType, 0);
+                if (!result) return NULL;
+                new (&result->data) std::shared_ptr<DijkstraMap>(it->second);
+                return (PyObject*)result;
+            }
+            self->data->dijkstra_maps.erase(it);
         }
-        // Different diagonal cost - remove old one
-        self->data->dijkstra_maps.erase(it);
     }
 
-    // Mark-and-restore: temporarily block cells with collide label
     auto restore_list = markCollisionLabel(self->data.get(), label_str);
 
-    // Create new DijkstraMap
-    auto dijkstra = std::make_shared<DijkstraMap>(
-        self->data->getTCODMap(), root_x, root_y, diagonal_cost);
+    std::shared_ptr<DijkstraMap> dijkstra;
+    TCODMap* tcmap = self->data->getTCODMap();
+
+    if (mask_obj) {
+        // Translate mask -> explicit root list, then drive compute_multi. Distance-only
+        // results are identical to compute_masked; this keeps DijkstraMap's invariant
+        // that it always holds exactly one computed TCOD_Dijkstra handle.
+        std::vector<sf::Vector2i> mask_roots;
+        const uint8_t* buf = mask_obj->data->data();
+        int w = mask_obj->data->width();
+        int h = mask_obj->data->height();
+        mask_roots.reserve(static_cast<size_t>(w) * 4);  // heuristic
+        for (int y = 0; y < h; y++) {
+            for (int x = 0; x < w; x++) {
+                if (buf[y * w + x] != 0) {
+                    mask_roots.push_back(sf::Vector2i(x, y));
+                }
+            }
+        }
+        if (mask_roots.empty()) {
+            restoreCollisionLabel(self->data.get(), restore_list);
+            PyErr_SetString(PyExc_ValueError, "DiscreteMap mask has no non-zero cells");
+            return NULL;
+        }
+        dijkstra = std::make_shared<DijkstraMap>(tcmap, mask_roots, diagonal_cost);
+    } else {
+        dijkstra = std::make_shared<DijkstraMap>(tcmap, roots, diagonal_cost);
+    }
 
-    // Restore walkability
     restoreCollisionLabel(self->data.get(), restore_list);
 
-    // Cache it
-    self->data->dijkstra_maps[key] = dijkstra;
+    // Cache only single-root case
+    if (!mask_obj && roots.size() == 1) {
+        auto key = std::make_tuple(roots[0].x, roots[0].y, label_str);
+        self->data->dijkstra_maps[key] = dijkstra;
+    }
 
-    // Return Python wrapper
     PyDijkstraMapObject* result = (PyDijkstraMapObject*)mcrfpydef::PyDijkstraMapType.tp_alloc(
         &mcrfpydef::PyDijkstraMapType, 0);
     if (!result) return NULL;
@@ -800,6 +1039,27 @@ PyMethodDef PyDijkstraMap_methods[] = {
      "Returns:\n"
      "    HeightMap with distance values as heights."},
 
+    {"invert", (PyCFunction)UIGridPathfinding::DijkstraMap_invert, METH_NOARGS,
+     "invert() -> DijkstraMap\n\n"
+     "Return a NEW DijkstraMap whose distance field is the safety field.\n\n"
+     "Cells near a root become high values and cells far from any root become\n"
+     "low values. Combined with step_from or descent_step, this gives flee\n"
+     "behavior: descend the inverted map to move away from the original roots.\n\n"
+     "The original DijkstraMap is unchanged.\n\n"
+     "Returns:\n"
+     "    New DijkstraMap with inverted distances."},
+
+    {"descent_step", (PyCFunction)UIGridPathfinding::DijkstraMap_descent_step, METH_VARARGS | METH_KEYWORDS,
+     "descent_step(pos) -> Vector | None\n\n"
+     "Get the adjacent cell with the lowest distance (steepest descent).\n\n"
+     "Unlike step_from (which follows the path set by path_from), descent_step\n"
+     "always returns the best neighbor in a single hop. Useful for AI that\n"
+     "reacts to the current distance field rather than following a fixed path.\n\n"
+     "Args:\n"
+     "    pos: Current position as Vector, Entity, or (x, y) tuple.\n\n"
+     "Returns:\n"
+     "    Next position as Vector, or None if pos is a local minimum or off-grid."},
+
     {NULL}
 };
 
diff --git a/src/UIGridPathfinding.h b/src/UIGridPathfinding.h
index 89610c8..28e0b92 100644
--- a/src/UIGridPathfinding.h
+++ b/src/UIGridPathfinding.h
@@ -28,10 +28,15 @@ struct PyAStarPathObject {
 
 class DijkstraMap {
 public:
+    // Single-root construction (back-compat).
     DijkstraMap(TCODMap* map, int root_x, int root_y, float diagonal_cost);
+
+    // Multi-root construction (#315). roots must be non-empty.
+    DijkstraMap(TCODMap* map, const std::vector<sf::Vector2i>& roots, float diagonal_cost);
+
     ~DijkstraMap();
 
-    // Non-copyable (owns TCODDijkstra)
+    // Non-copyable (owns TCOD_Dijkstra)
     DijkstraMap(const DijkstraMap&) = delete;
     DijkstraMap& operator=(const DijkstraMap&) = delete;
 
@@ -40,18 +45,35 @@ public:
     std::vector<sf::Vector2i> getPathFrom(int x, int y) const;
     sf::Vector2i stepFrom(int x, int y, bool* valid = nullptr) const;
 
+    // Phase B: FLEE primitives (#315)
+    // invertInPlace() mutates this map's distance field. Prefer inverted() in new code —
+    // the Python surface exposes the non-mutating form to keep maps immutable after
+    // creation.
+    void invertInPlace();
+    // Returns a freshly computed DijkstraMap with the same roots and diagonal_cost,
+    // then inverts its distance field. The caller owns the returned shared_ptr.
+    std::shared_ptr<DijkstraMap> inverted() const;
+    // descent_step returns the next cell along steepest descent, or (-1,-1) + valid=false.
+    sf::Vector2i descentStep(int x, int y, bool* valid = nullptr) const;
+
     // Accessors
-    sf::Vector2i getRoot() const { return root; }
+    sf::Vector2i getRoot() const { return root; }  // First root for multi-root
+    const std::vector<sf::Vector2i>& getRoots() const { return roots; }
+    bool isMultiRoot() const { return roots.size() > 1; }
     float getDiagonalCost() const { return diagonal_cost; }
-    int getWidth() const;
-    int getHeight() const;
+    int getWidth() const { return map_width; }
+    int getHeight() const { return map_height; }
+
+    // Raw C handle, for internal use in new constructor paths (e.g. from_invert).
+    TCOD_Dijkstra* getHandle() const { return tcod_dijkstra; }
 
 private:
-    TCODDijkstra* tcod_dijkstra;  // Owned by this object
-    TCODMap* tcod_map;            // Borrowed from Grid
+    TCOD_Dijkstra* tcod_dijkstra;  // Owned
+    TCODMap* tcod_map;             // Borrowed from Grid
     sf::Vector2i root;
+    std::vector<sf::Vector2i> roots;
     float diagonal_cost;
-    int map_width;                // Cached from TCODMap at construction
+    int map_width;
     int map_height;
 };
 
@@ -110,6 +132,8 @@ namespace UIGridPathfinding {
     PyObject* DijkstraMap_path_from(PyDijkstraMapObject* self, PyObject* args, PyObject* kwds);
     PyObject* DijkstraMap_step_from(PyDijkstraMapObject* self, PyObject* args, PyObject* kwds);
     PyObject* DijkstraMap_to_heightmap(PyDijkstraMapObject* self, PyObject* args, PyObject* kwds);
+    PyObject* DijkstraMap_invert(PyDijkstraMapObject* self, PyObject* args);
+    PyObject* DijkstraMap_descent_step(PyDijkstraMapObject* self, PyObject* args, PyObject* kwds);
 
     // Properties
     PyObject* DijkstraMap_get_root(PyDijkstraMapObject* self, void* closure);
diff --git a/tests/benchmarks/baseline/phase5_2/dijkstra_bench.json b/tests/benchmarks/baseline/phase5_2/dijkstra_bench.json
index 018361d..74521b1 100644
--- a/tests/benchmarks/baseline/phase5_2/dijkstra_bench.json
+++ b/tests/benchmarks/baseline/phase5_2/dijkstra_bench.json
@@ -4,82 +4,82 @@
       "grid": "100x100",
       "kind": "multi_root",
       "roots": 1,
-      "mean_ms": 0.7709094556048512
+      "mean_ms": 0.6948539987206459
     },
     {
       "grid": "100x100",
       "kind": "multi_root",
       "roots": 2,
-      "mean_ms": 0.7632468361407518
+      "mean_ms": 0.8002225775271654
     },
     {
       "grid": "100x100",
       "kind": "multi_root",
       "roots": 5,
-      "mean_ms": 1.200081780552864
+      "mean_ms": 1.1821302119642496
     },
     {
       "grid": "100x100",
       "kind": "multi_root",
       "roots": 20,
-      "mean_ms": 2.137616788968444
+      "mean_ms": 2.0206935703754425
     },
     {
       "grid": "100x100",
       "kind": "mask",
       "roots": 500,
-      "mean_ms": 30.424197972752154
+      "mean_ms": 24.073211615905166
     },
     {
       "grid": "100x100",
       "kind": "invert",
-      "mean_ms": 1.0323396185413003
+      "mean_ms": 0.7887090090662241
     },
     {
       "grid": "100x100",
       "kind": "descent_step_per_call",
-      "mean_us": 0.4075700417160988,
+      "mean_us": 0.3262499812990427,
       "valid_per_trial": 100
     },
     {
       "grid": "500x500",
       "kind": "multi_root",
       "roots": 1,
-      "mean_ms": 26.075413217768073
+      "mean_ms": 20.39959062822163
     },
     {
       "grid": "500x500",
       "kind": "multi_root",
       "roots": 2,
-      "mean_ms": 25.83242394030094
+      "mean_ms": 22.81180394347757
     },
     {
       "grid": "500x500",
       "kind": "multi_root",
       "roots": 5,
-      "mean_ms": 33.73005616012961
+      "mean_ms": 29.194996040314436
     },
     {
       "grid": "500x500",
       "kind": "multi_root",
       "roots": 20,
-      "mean_ms": 78.58918677084148
+      "mean_ms": 67.84450197592378
     },
     {
       "grid": "500x500",
       "kind": "mask",
       "roots": 12500,
-      "mean_ms": 18658.679087948985
+      "mean_ms": 19711.241053813137
     },
     {
       "grid": "500x500",
       "kind": "invert",
-      "mean_ms": 25.918347598053515
+      "mean_ms": 27.875673002563417
     },
     {
       "grid": "500x500",
       "kind": "descent_step_per_call",
-      "mean_us": 0.3717193566262722,
+      "mean_us": 0.3495373670011759,
       "valid_per_trial": 2500
     }
   ]
diff --git a/tests/benchmarks/baseline/phase5_2/grid_step_bench.json b/tests/benchmarks/baseline/phase5_2/grid_step_bench.json
index 8565bb3..e7c96fe 100644
--- a/tests/benchmarks/baseline/phase5_2/grid_step_bench.json
+++ b/tests/benchmarks/baseline/phase5_2/grid_step_bench.json
@@ -2,8 +2,8 @@
   "grid": "100x100",
   "entities": 100,
   "rounds": 1000,
-  "total_sec": 0.07824495097156614,
-  "mean_round_ms": 0.07824495097156614,
-  "p95_round_ms": 0.1227830071002245,
-  "per_entity_step_us": 0.7824495097156614
+  "total_sec": 0.06852402002550662,
+  "mean_round_ms": 0.06852402002550662,
+  "p95_round_ms": 0.09302806574851274,
+  "per_entity_step_us": 0.6852402002550662
 }
\ No newline at end of file
diff --git a/tests/benchmarks/baseline/phase5_2/gridview_render_bench.json b/tests/benchmarks/baseline/phase5_2/gridview_render_bench.json
index 757efb2..b58bca5 100644
--- a/tests/benchmarks/baseline/phase5_2/gridview_render_bench.json
+++ b/tests/benchmarks/baseline/phase5_2/gridview_render_bench.json
@@ -4,31 +4,31 @@
       "views": 1,
       "frames": 60,
       "warmup_frames": 5,
-      "total_sec": 4.710599604761228,
-      "mean_frame_ms": 78.50999341268714,
-      "p95_frame_ms": 90.47448402270675,
-      "implied_fps": 12.737231994702995,
-      "per_view_frame_ms": 78.50999341268714
+      "total_sec": 4.731390134897083,
+      "mean_frame_ms": 78.85650224828473,
+      "p95_frame_ms": 86.92639297805727,
+      "implied_fps": 12.681262438593032,
+      "per_view_frame_ms": 78.85650224828473
     },
     {
       "views": 2,
       "frames": 60,
       "warmup_frames": 5,
-      "total_sec": 4.6525509969796985,
-      "mean_frame_ms": 77.54251661632831,
-      "p95_frame_ms": 91.92000096663833,
-      "implied_fps": 12.896150958678424,
-      "per_view_frame_ms": 38.77125830816416
+      "total_sec": 4.975782633875497,
+      "mean_frame_ms": 82.92971056459162,
+      "p95_frame_ms": 91.40842803753912,
+      "implied_fps": 12.058404559619538,
+      "per_view_frame_ms": 41.46485528229581
     },
     {
       "views": 4,
       "frames": 60,
       "warmup_frames": 5,
-      "total_sec": 4.727940998389386,
-      "mean_frame_ms": 78.7990166398231,
-      "p95_frame_ms": 99.88687501754612,
-      "implied_fps": 12.690513697281654,
-      "per_view_frame_ms": 19.699754159955774
+      "total_sec": 4.861755113233812,
+      "mean_frame_ms": 81.0292518872302,
+      "p95_frame_ms": 90.46144201420248,
+      "implied_fps": 12.341222172354708,
+      "per_view_frame_ms": 20.25731297180755
     }
   ],
   "config": {
diff --git a/tests/demo/screens/pathfinding_demo.py b/tests/demo/screens/pathfinding_demo.py
new file mode 100644
index 0000000..adfce1a
--- /dev/null
+++ b/tests/demo/screens/pathfinding_demo.py
@@ -0,0 +1,290 @@
+"""pathfinding_demo.py - Visual demo of the #315 pathfinding primitives.
+
+Three panels side-by-side on one scene:
+  [Panel 1] A* with selectable heuristic. Keys 1-5 cycle EUCLIDEAN, MANHATTAN,
+            CHEBYSHEV, DIAGONAL, ZERO. Q/W bump the weight by 0.25.
+  [Panel 2] Dijkstra flood from a cursor-controlled root. Arrow keys move the
+            cursor; the distance field re-renders as a blue gradient.
+  [Panel 3] Multi-root FLEE: three "guard" entities flee from a shared set of
+            threats, using an inverted multi-root Dijkstra map. Animated one
+            step per frame tick. Press T to drop a new threat on the panel.
+
+Also exercises: DijkstraMap.invert(), DijkstraMap.descent_step(),
+mcrfpy.Heuristic, Grid.find_path(heuristic=, weight=), Grid.get_dijkstra_map(
+roots=...).
+"""
+import mcrfpy
+import sys
+
+
+GRID_W, GRID_H = 20, 20
+CELL_PX = 14
+PANEL_W_PX = GRID_W * CELL_PX
+GAP = 20
+
+SCREEN_W = 3 * PANEL_W_PX + 4 * GAP
+SCREEN_H = GRID_H * CELL_PX + 140
+
+
+scene = mcrfpy.Scene("pathfinding_demo")
+
+bg = mcrfpy.Frame(pos=(0, 0), size=(SCREEN_W, SCREEN_H),
+                  fill_color=mcrfpy.Color(18, 18, 26))
+scene.children.append(bg)
+
+title = mcrfpy.Caption(text="Pathfinding Next-Gen Demo (#315)", pos=(GAP, 10))
+title.fill_color = mcrfpy.Color(240, 240, 255)
+scene.children.append(title)
+
+
+def make_open_grid(x_off):
+    g = mcrfpy.Grid(grid_size=(GRID_W, GRID_H),
+                    pos=(x_off, 50),
+                    size=(PANEL_W_PX, GRID_H * CELL_PX))
+    for y in range(GRID_H):
+        for x in range(GRID_W):
+            c = g.at(x, y)
+            wall = (x in (0, GRID_W - 1)) or (y in (0, GRID_H - 1))
+            c.walkable = not wall
+            c.transparent = not wall
+    # A small wedge of walls in the middle for visual interest.
+    for y in range(4, 12):
+        g.at(GRID_W // 2, y).walkable = False
+        g.at(GRID_W // 2, y).transparent = False
+    scene.children.append(g)
+    return g
+
+
+# =============================================================================
+# Panel 1: A* with heuristic switching
+# =============================================================================
+g1 = make_open_grid(GAP)
+astar_layer = mcrfpy.ColorLayer(z_index=1, name="astar_path")
+g1.add_layer(astar_layer)
+
+HEURISTICS = [
+    (mcrfpy.Heuristic.EUCLIDEAN, "EUCLIDEAN"),
+    (mcrfpy.Heuristic.MANHATTAN, "MANHATTAN"),
+    (mcrfpy.Heuristic.CHEBYSHEV, "CHEBYSHEV"),
+    (mcrfpy.Heuristic.DIAGONAL,  "DIAGONAL"),
+    (mcrfpy.Heuristic.ZERO,      "ZERO"),
+]
+state_astar = {"hidx": 0, "weight": 1.0,
+               "start": (2, 10), "end": (GRID_W - 3, 10)}
+
+cap_astar = mcrfpy.Caption(text="A* heuristic: EUCLIDEAN  weight=1.00",
+                           pos=(GAP, 50 + GRID_H * CELL_PX + 6))
+cap_astar.fill_color = mcrfpy.Color(180, 220, 255)
+scene.children.append(cap_astar)
+
+
+def redraw_astar():
+    h, hname = HEURISTICS[state_astar["hidx"]]
+    astar_layer.fill(mcrfpy.Color(0, 0, 0, 0))
+    p = g1.find_path(state_astar["start"], state_astar["end"],
+                     heuristic=h, weight=state_astar["weight"])
+    n_steps = 0
+    if p is not None:
+        for step in p:
+            astar_layer.set((int(step.x), int(step.y)),
+                            mcrfpy.Color(255, 220, 80, 220))
+            n_steps += 1
+    # Start/end markers.
+    astar_layer.set(state_astar["start"], mcrfpy.Color(80, 255, 120, 255))
+    astar_layer.set(state_astar["end"], mcrfpy.Color(255, 90, 90, 255))
+    cap_astar.text = (f"A* heuristic: {hname}  "
+                      f"weight={state_astar['weight']:.2f}  steps={n_steps}")
+
+
+# =============================================================================
+# Panel 2: Dijkstra flood
+# =============================================================================
+g2 = make_open_grid(GAP * 2 + PANEL_W_PX)
+dij_layer = mcrfpy.ColorLayer(z_index=1, name="dij_flood")
+g2.add_layer(dij_layer)
+
+state_dij = {"root": (GRID_W // 2 - 3, GRID_H // 2)}
+
+cap_dij = mcrfpy.Caption(text="Dijkstra flood (arrows move root)",
+                         pos=(GAP * 2 + PANEL_W_PX, 50 + GRID_H * CELL_PX + 6))
+cap_dij.fill_color = mcrfpy.Color(180, 255, 220)
+scene.children.append(cap_dij)
+
+
+def redraw_dijkstra():
+    dij_layer.fill(mcrfpy.Color(0, 0, 0, 0))
+    root = state_dij["root"]
+    if not g2.at(root[0], root[1]).walkable:
+        cap_dij.text = "Dijkstra root on wall - move with arrows"
+        return
+    dmap = g2.get_dijkstra_map(root)
+    # Sample distances to find the maximum for normalization.
+    max_dist = 0.0
+    for y in range(1, GRID_H - 1):
+        for x in range(1, GRID_W - 1):
+            d = dmap.distance((x, y))
+            if d is not None and d > max_dist:
+                max_dist = d
+    if max_dist <= 0:
+        return
+    for y in range(1, GRID_H - 1):
+        for x in range(1, GRID_W - 1):
+            d = dmap.distance((x, y))
+            if d is None:
+                continue
+            t = min(1.0, d / max_dist)
+            # Cool gradient: near-root = bright, far = dark.
+            r = int(40 * t + 80 * (1 - t))
+            gc = int(160 * (1 - t) + 40 * t)
+            bc = int(240 * (1 - t) + 80 * t)
+            dij_layer.set((x, y), mcrfpy.Color(r, gc, bc, 180))
+    dij_layer.set(root, mcrfpy.Color(255, 255, 90, 255))
+    cap_dij.text = f"Dijkstra flood max={max_dist:.1f}  root={root}"
+
+
+# =============================================================================
+# Panel 3: Multi-root FLEE
+# =============================================================================
+g3 = make_open_grid(GAP * 3 + 2 * PANEL_W_PX)
+flee_layer = mcrfpy.ColorLayer(z_index=1, name="flee_layer")
+g3.add_layer(flee_layer)
+
+state_flee = {
+    "threats": [(3, 3), (GRID_W - 4, GRID_H - 4)],
+    "guards": [(10, 6), (10, 10), (10, 14)],
+    "safety": None,
+    "threat_map": None,
+}
+
+cap_flee = mcrfpy.Caption(text="Multi-root FLEE (T adds threat, R resets)",
+                          pos=(GAP * 3 + 2 * PANEL_W_PX,
+                               50 + GRID_H * CELL_PX + 6))
+cap_flee.fill_color = mcrfpy.Color(255, 190, 190)
+scene.children.append(cap_flee)
+
+
+def recompute_flee():
+    state_flee["threat_map"] = g3.get_dijkstra_map(roots=state_flee["threats"])
+    state_flee["safety"] = state_flee["threat_map"].invert()
+
+
+def redraw_flee():
+    flee_layer.fill(mcrfpy.Color(0, 0, 0, 0))
+    # Threats: red.
+    for t in state_flee["threats"]:
+        flee_layer.set(t, mcrfpy.Color(255, 60, 60, 255))
+    # Guards: green.
+    for gd in state_flee["guards"]:
+        flee_layer.set(gd, mcrfpy.Color(80, 255, 120, 255))
+    cap_flee.text = (f"FLEE: {len(state_flee['threats'])} threats  "
+                     f"{len(state_flee['guards'])} guards")
+
+
+def step_guards():
+    if state_flee["safety"] is None:
+        return
+    new_guards = []
+    for gd in state_flee["guards"]:
+        nxt = state_flee["safety"].descent_step(gd)
+        if nxt is None:
+            new_guards.append(gd)
+        else:
+            candidate = (int(nxt.x), int(nxt.y))
+            # Avoid landing on another guard or a threat.
+            if candidate in new_guards or candidate in state_flee["threats"]:
+                new_guards.append(gd)
+            else:
+                new_guards.append(candidate)
+    state_flee["guards"] = new_guards
+    redraw_flee()
+
+
+# =============================================================================
+# Key handling
+# =============================================================================
+instructions = [
+    "Panel 1 (A*):   [1-5] heuristic   [Q/W] weight -/+",
+    "Panel 2 (Dij):  [Arrow keys] move root",
+    "Panel 3 (FLEE): [T] add threat   [R] reset   (guards auto-step)",
+    "[ESC] quit",
+]
+for i, text in enumerate(instructions):
+    c = mcrfpy.Caption(text=text, pos=(GAP, SCREEN_H - 90 + i * 22))
+    c.fill_color = mcrfpy.Color(180, 180, 200)
+    scene.children.append(c)
+
+
+def on_key(key, state):
+    if state != mcrfpy.InputState.PRESSED:
+        return
+
+    # Heuristic switching
+    for i, digit in enumerate([mcrfpy.Key.Num1, mcrfpy.Key.Num2, mcrfpy.Key.Num3,
+                               mcrfpy.Key.Num4, mcrfpy.Key.Num5]):
+        if key == digit:
+            state_astar["hidx"] = i
+            redraw_astar()
+            return
+
+    if key == mcrfpy.Key.Q:
+        state_astar["weight"] = max(0.25, state_astar["weight"] - 0.25)
+        redraw_astar()
+        return
+    if key == mcrfpy.Key.W:
+        state_astar["weight"] = min(5.0, state_astar["weight"] + 0.25)
+        redraw_astar()
+        return
+
+    # Dijkstra root movement
+    rx, ry = state_dij["root"]
+    moved = False
+    if key == mcrfpy.Key.LEFT:
+        rx = max(1, rx - 1); moved = True
+    elif key == mcrfpy.Key.RIGHT:
+        rx = min(GRID_W - 2, rx + 1); moved = True
+    elif key == mcrfpy.Key.UP:
+        ry = max(1, ry - 1); moved = True
+    elif key == mcrfpy.Key.DOWN:
+        ry = min(GRID_H - 2, ry + 1); moved = True
+    if moved:
+        state_dij["root"] = (rx, ry)
+        redraw_dijkstra()
+
+    # FLEE panel: drop threat at a random walkable cell.
+    if key == mcrfpy.Key.T:
+        import random
+        for _ in range(20):
+            p = (random.randint(1, GRID_W - 2), random.randint(1, GRID_H - 2))
+            if g3.at(p[0], p[1]).walkable and p not in state_flee["threats"]:
+                state_flee["threats"].append(p)
+                break
+        recompute_flee()
+        redraw_flee()
+
+    if key == mcrfpy.Key.R:
+        state_flee["threats"] = [(3, 3), (GRID_W - 4, GRID_H - 4)]
+        state_flee["guards"] = [(10, 6), (10, 10), (10, 14)]
+        recompute_flee()
+        redraw_flee()
+
+    if key == mcrfpy.Key.ESCAPE:
+        mcrfpy.exit()
+
+
+scene.on_key = on_key
+
+# Step FLEE guards on a timer (6 ticks/sec is slow enough to watch).
+def tick(timer, runtime):
+    step_guards()
+
+
+# Initial render
+redraw_astar()
+redraw_dijkstra()
+recompute_flee()
+redraw_flee()
+
+mcrfpy.Timer("flee_tick", tick, 160)
+mcrfpy.current_scene = scene
+
+print("pathfinding_demo loaded - see on-screen instructions.")
diff --git a/tests/regression/issue_315_path_provider_test.py b/tests/regression/issue_315_path_provider_test.py
new file mode 100644
index 0000000..a5b153d
--- /dev/null
+++ b/tests/regression/issue_315_path_provider_test.py
@@ -0,0 +1,150 @@
+"""Phase C regression: PathProvider abstraction drives SEEK/FLEE via set_behavior.
+
+Verifies that each of the three provider shapes (DijkstraMap, AStarPath, and
+plain target tuple) produces a valid SEEK step, and that swapping pathfinder
+mid-run changes the next step.
+"""
+import mcrfpy
+import sys
+
+
+def make_open_grid(w=20, h=20):
+    scene = mcrfpy.Scene("issue315")
+    mcrfpy.current_scene = scene
+    grid = mcrfpy.Grid(grid_size=(w, h))
+    scene.children.append(grid)
+    for y in range(h):
+        for x in range(w):
+            c = grid.at(x, y)
+            # Walkable interior, walls on the border.
+            if x == 0 or y == 0 or x == w - 1 or y == h - 1:
+                c.walkable = False
+                c.transparent = False
+            else:
+                c.walkable = True
+                c.transparent = True
+    return grid
+
+
+def closer_to(p, goal, start):
+    """True when p is closer than start to goal (Chebyshev)."""
+    def d(a, b): return max(abs(a[0] - b[0]), abs(a[1] - b[1]))
+    return d(p, goal) < d(start, goal)
+
+
+def test_dijkstra_provider():
+    grid = make_open_grid()
+    e = mcrfpy.Entity((5, 5), grid=grid)
+    e.move_speed = 0
+    goal = (15, 15)
+    dmap = grid.get_dijkstra_map(goal)
+    e.set_behavior(int(mcrfpy.Behavior.SEEK), pathfinder=dmap)
+
+    start = (e.cell_x, e.cell_y)
+    grid.step()
+    ended = (e.cell_x, e.cell_y)
+    assert ended != start, "DijkstraProvider SEEK must move the entity"
+    assert closer_to(ended, goal, start), f"moved away from goal: {start} -> {ended}"
+    print("  DijkstraProvider SEEK: OK")
+
+
+def test_astar_provider():
+    grid = make_open_grid()
+    e = mcrfpy.Entity((5, 5), grid=grid)
+    e.move_speed = 0
+    goal = (10, 10)
+    path = grid.find_path((5, 5), goal)
+    assert path is not None
+    e.set_behavior(int(mcrfpy.Behavior.SEEK), pathfinder=path)
+
+    start = (e.cell_x, e.cell_y)
+    grid.step()
+    ended = (e.cell_x, e.cell_y)
+    assert ended != start, "AStarProvider SEEK must move the entity"
+    assert closer_to(ended, goal, start), f"moved away from goal: {start} -> {ended}"
+    print("  AStarProvider SEEK: OK")
+
+
+def test_target_provider():
+    grid = make_open_grid()
+    e = mcrfpy.Entity((5, 5), grid=grid)
+    e.move_speed = 0
+    goal = (10, 10)
+    e.set_behavior(int(mcrfpy.Behavior.SEEK), pathfinder=goal)
+
+    start = (e.cell_x, e.cell_y)
+    grid.step()
+    ended = (e.cell_x, e.cell_y)
+    assert ended != start, "TargetProvider SEEK must move the entity"
+    assert closer_to(ended, goal, start), f"moved away from target: {start} -> {ended}"
+    print("  TargetProvider SEEK: OK")
+
+
+def test_flee_via_inverted_dijkstra():
+    grid = make_open_grid()
+    e = mcrfpy.Entity((10, 10), grid=grid)
+    e.move_speed = 0
+    threat = (12, 10)
+    threat_map = grid.get_dijkstra_map(threat)
+    safety_map = threat_map.invert()
+    e.set_behavior(int(mcrfpy.Behavior.FLEE), pathfinder=safety_map)
+
+    start = (e.cell_x, e.cell_y)
+    start_dist = threat_map.distance(start)
+    grid.step()
+    ended = (e.cell_x, e.cell_y)
+    assert ended != start, "FLEE must move the entity"
+    new_dist = threat_map.distance(ended)
+    assert new_dist >= start_dist, (
+        f"FLEE should not get closer: d={start_dist} -> {new_dist}")
+    print("  FLEE via inverted DijkstraMap: OK")
+
+
+def test_provider_swap_midrun():
+    """Swapping pathfinder mid-run changes the next step."""
+    grid = make_open_grid()
+    e = mcrfpy.Entity((10, 10), grid=grid)
+    e.move_speed = 0
+
+    # First: seek (5, 10) - should move left.
+    e.set_behavior(int(mcrfpy.Behavior.SEEK),
+                   pathfinder=grid.get_dijkstra_map((5, 10)))
+    grid.step()
+    after_first = (e.cell_x, e.cell_y)
+    assert after_first[0] < 10, f"expected leftward step, got {after_first}"
+
+    # Swap pathfinder: now seek (15, 10) - should move right from here.
+    e.set_behavior(int(mcrfpy.Behavior.SEEK),
+                   pathfinder=grid.get_dijkstra_map((15, 10)))
+    grid.step()
+    after_second = (e.cell_x, e.cell_y)
+    assert after_second[0] > after_first[0], (
+        f"expected rightward step after swap, got {after_first} -> {after_second}")
+    print("  Mid-run provider swap: OK")
+
+
+def test_invalid_pathfinder_raises():
+    grid = make_open_grid()
+    e = mcrfpy.Entity((5, 5), grid=grid)
+    try:
+        e.set_behavior(int(mcrfpy.Behavior.SEEK), pathfinder="not a pathfinder")
+    except TypeError:
+        pass
+    else:
+        raise AssertionError("expected TypeError for bad pathfinder argument")
+    print("  Invalid pathfinder rejected: OK")
+
+
+def main():
+    test_dijkstra_provider()
+    test_astar_provider()
+    test_target_provider()
+    test_flee_via_inverted_dijkstra()
+    test_provider_swap_midrun()
+    test_invalid_pathfinder_raises()
+    print("PASS")
+
+
+if __name__ == "__main__":
+    main()
+    sys.exit(0)
diff --git a/tests/unit/dijkstra_flee_test.py b/tests/unit/dijkstra_flee_test.py
new file mode 100644
index 0000000..a8b4836
--- /dev/null
+++ b/tests/unit/dijkstra_flee_test.py
@@ -0,0 +1,85 @@
+"""DijkstraMap.invert() + descent_step() produce FLEE behavior.
+
+Build a Dijkstra map rooted on a threat, invert it, and confirm that walking
+descent steps from a nearby cell strictly increases distance from the threat.
+"""
+import mcrfpy
+import sys
+
+
+def make_grid(w, h):
+    g = mcrfpy.Grid(grid_size=(w, h))
+    for y in range(h):
+        for x in range(w):
+            c = g.at(x, y)
+            c.walkable = True
+            c.transparent = True
+    return g
+
+
+def main():
+    g = make_grid(20, 20)
+    threat = (10, 10)
+
+    threat_map = g.get_dijkstra_map(threat)
+    assert threat_map.distance(threat) == 0.0
+
+    # invert() returns a NEW map; the original is unchanged.
+    safety_map = threat_map.invert()
+    assert safety_map is not threat_map, "invert() should return a new object"
+    assert threat_map.distance(threat) == 0.0, "original map must not be mutated"
+    # After invert, the threat cell itself is a local minimum (low safety),
+    # and cells far from the threat are peaks.
+
+    # descent_step on the safety map from a cell near the threat must move AWAY,
+    # i.e. its distance in the *original* (threat-rooted) map strictly increases.
+    start = (11, 10)
+    start_dist = threat_map.distance(start)
+
+    pos = start
+    for _ in range(5):
+        nxt = safety_map.descent_step(pos)
+        if nxt is None:
+            break
+        nxt_tuple = (int(nxt.x), int(nxt.y))
+        # Must actually move.
+        assert nxt_tuple != pos, f"descent stuck at {pos}"
+        # Must move to a walkable cell inside the grid.
+        assert 0 <= nxt_tuple[0] < 20 and 0 <= nxt_tuple[1] < 20
+        new_dist = threat_map.distance(nxt_tuple)
+        assert new_dist >= start_dist, (
+            f"FLEE descent from {pos} to {nxt_tuple}: threat distance dropped "
+            f"from {start_dist} to {new_dist}")
+        pos = nxt_tuple
+        start_dist = new_dist
+
+    # descent_step on the original (non-inverted) map from a far cell SEEKs the threat.
+    far = (0, 0)
+    nxt = threat_map.descent_step(far)
+    assert nxt is not None
+    nxt_tuple = (int(nxt.x), int(nxt.y))
+    # Closer to the threat than `far`.
+    assert threat_map.distance(nxt_tuple) < threat_map.distance(far), \
+        "descent on threat_map should SEEK the root"
+
+    # descent_step at the root itself has no better neighbor — returns None.
+    at_root = safety_map.descent_step(threat)
+    # Note: at_root might not be None on the inverted map since the threat is a local
+    # minimum of the inverted field — any neighbor has lower (or equal) value. So allow
+    # either None or a valid step. Just ensure we don't crash.
+    _ = at_root
+
+    # Out-of-bounds raises IndexError.
+    try:
+        safety_map.descent_step((999, 999))
+    except IndexError:
+        pass
+    else:
+        raise AssertionError("expected IndexError for out-of-bounds descent_step")
+
+    print("PASS")
+
+
+if __name__ == "__main__":
+    main()
+    sys.exit(0)
diff --git a/tests/unit/find_path_heuristic_test.py b/tests/unit/find_path_heuristic_test.py
new file mode 100644
index 0000000..bfaf51c
--- /dev/null
+++ b/tests/unit/find_path_heuristic_test.py
@@ -0,0 +1,61 @@
+"""Grid.find_path heuristic/weight kwargs produce valid paths across each built-in."""
+import mcrfpy
+import sys
+
+
+def make_open_grid(w, h):
+    g = mcrfpy.Grid(grid_size=(w, h))
+    for y in range(h):
+        for x in range(w):
+            c = g.at(x, y)
+            c.walkable = True
+            c.transparent = True
+    return g
+
+
+def main():
+    g = make_open_grid(30, 30)
+
+    # On an obstacle-free grid every admissible heuristic yields an optimal-length path.
+    # libtcod returns steps (excluding origin), so a diagonal-permitting move from (0,0)
+    # to (20,20) is 20 steps.
+    for h in (mcrfpy.Heuristic.EUCLIDEAN,
+              mcrfpy.Heuristic.MANHATTAN,
+              mcrfpy.Heuristic.CHEBYSHEV,
+              mcrfpy.Heuristic.DIAGONAL,
+              mcrfpy.Heuristic.ZERO):
+        p = g.find_path((0, 0), (20, 20), heuristic=h)
+        assert p is not None, f"no path for {h}"
+        steps = list(p)
+        assert len(steps) == 20, f"heuristic {h} gave {len(steps)} steps, expected 20"
+        # Last step must be the goal.
+        assert (int(steps[-1].x), int(steps[-1].y)) == (20, 20), \
+            f"heuristic {h} did not end at goal"
+
+    # Weighted A* with weight>=1 must still find a path (not necessarily optimal).
+    for w in (1.0, 1.5, 2.0):
+        p = g.find_path((0, 0), (20, 20), heuristic=mcrfpy.Heuristic.EUCLIDEAN, weight=w)
+        assert p is not None, f"no path for weight={w}"
+        steps = list(p)
+        assert len(steps) >= 20, f"weight={w} gave impossibly short path"
+
+    # With an obstacle, the path still reaches the goal.
+    g2 = make_open_grid(30, 30)
+    for y in range(5, 25):
+        g2.at(15, y).walkable = False
+
+    p = g2.find_path((0, 0), (29, 0), heuristic=mcrfpy.Heuristic.MANHATTAN)
+    assert p is not None
+    steps = list(p)
+    assert (int(steps[-1].x), int(steps[-1].y)) == (29, 0)
+    # No step may land on a blocked cell.
+    for s in steps:
+        assert not (int(s.x) == 15 and 5 <= int(s.y) < 25), \
+            f"path stepped through wall at {s}"
+
+    print("PASS")
+
+
+if __name__ == "__main__":
+    main()
+    sys.exit(0)
diff --git a/tests/unit/heuristic_enum_test.py b/tests/unit/heuristic_enum_test.py
new file mode 100644
index 0000000..45b6698
--- /dev/null
+++ b/tests/unit/heuristic_enum_test.py
@@ -0,0 +1,61 @@
+"""mcrfpy.Heuristic enum exists with expected members and accepts several arg forms."""
+import mcrfpy
+import sys
+
+
+def main():
+    assert hasattr(mcrfpy, "Heuristic"), "mcrfpy.Heuristic missing"
+    H = mcrfpy.Heuristic
+
+    expected = {"EUCLIDEAN": 0, "MANHATTAN": 1, "CHEBYSHEV": 2, "DIAGONAL": 3, "ZERO": 4}
+    for name, value in expected.items():
+        assert hasattr(H, name), f"Heuristic.{name} missing"
+        assert int(getattr(H, name)) == value, f"Heuristic.{name} != {value}"
+
+    members = list(H)
+    assert len(members) == 5, f"expected 5 members, got {len(members)}"
+
+    # find_path accepts enum, int, string
+    g = mcrfpy.Grid(grid_size=(20, 20))
+    for y in range(20):
+        for x in range(20):
+            c = g.at(x, y)
+            c.walkable = True
+            c.transparent = True
+
+    for arg in (H.MANHATTAN, 1, "MANHATTAN"):
+        p = g.find_path((0, 0), (10, 10), heuristic=arg)
+        assert p is not None, f"find_path returned None for heuristic={arg!r}"
+        steps = list(p)
+        assert len(steps) > 0, f"empty path for heuristic={arg!r}"
+
+    # Invalid string raises
+    try:
+        g.find_path((0, 0), (10, 10), heuristic="NOT_A_HEURISTIC")
+    except ValueError:
+        pass
+    else:
+        raise AssertionError("expected ValueError for bad heuristic string")
+
+    # Invalid int raises
+    try:
+        g.find_path((0, 0), (10, 10), heuristic=99)
+    except ValueError:
+        pass
+    else:
+        raise AssertionError("expected ValueError for out-of-range heuristic int")
+
+    # Non-positive weight raises
+    try:
+        g.find_path((0, 0), (10, 10), weight=0.0)
+    except ValueError:
+        pass
+    else:
+        raise AssertionError("expected ValueError for non-positive weight")
+
+    print("PASS")
+
+
+if __name__ == "__main__":
+    main()
+    sys.exit(0)
diff --git a/tests/unit/multi_root_dijkstra_test.py b/tests/unit/multi_root_dijkstra_test.py
new file mode 100644
index 0000000..b009f3c
--- /dev/null
+++ b/tests/unit/multi_root_dijkstra_test.py
@@ -0,0 +1,83 @@
+"""Multi-root Dijkstra distance equals min of per-root distances.
+
+Also covers the DiscreteMap-mask root-input form introduced in #315.
+"""
+import mcrfpy
+import sys
+
+
+def make_grid(w, h):
+    g = mcrfpy.Grid(grid_size=(w, h))
+    for y in range(h):
+        for x in range(w):
+            c = g.at(x, y)
+            c.walkable = True
+            c.transparent = True
+    return g
+
+
+def approx(a, b, tol=0.01):
+    return abs(a - b) < tol
+
+
+def main():
+    g = make_grid(20, 20)
+    roots = [(0, 0), (19, 19), (0, 19)]
+
+    multi = g.get_dijkstra_map(roots=roots)
+    singles = [g.get_dijkstra_map(r) for r in roots]
+
+    # Pick sample cells spread across the grid.
+    samples = [(5, 5), (10, 10), (15, 5), (2, 18), (18, 2), (9, 15)]
+    for p in samples:
+        expected = min(s.distance(p) for s in singles)
+        got = multi.distance(p)
+        assert approx(got, expected), (
+            f"multi-root distance at {p} was {got}, expected {expected}")
+
+    # Distance at each root is 0.
+    for r in roots:
+        assert multi.distance(r) == 0.0, f"root {r} distance should be 0"
+
+    # Single-root via roots= also works.
+    d_single = g.get_dijkstra_map(roots=[(5, 5)])
+    d_ref = g.get_dijkstra_map((5, 5))
+    for p in samples:
+        assert approx(d_single.distance(p), d_ref.distance(p)), \
+            f"single-element roots list diverges at {p}"
+
+    # DiscreteMap mask form: mark four corners, compare against explicit roots.
+    dmap = mcrfpy.DiscreteMap((20, 20))
+    corners = [(0, 0), (19, 0), (0, 19), (19, 19)]
+    for x, y in corners:
+        dmap.set(x, y, 1)
+
+    d_mask = g.get_dijkstra_map(roots=dmap)
+    d_corners = g.get_dijkstra_map(roots=corners)
+    for p in samples:
+        assert approx(d_mask.distance(p), d_corners.distance(p)), \
+            f"mask-root diverges from explicit corners at {p}"
+
+    # Empty mask errors out rather than silently returning all-infinity.
+    empty_mask = mcrfpy.DiscreteMap((20, 20))
+    try:
+        g.get_dijkstra_map(roots=empty_mask)
+    except ValueError:
+        pass
+    else:
+        raise AssertionError("expected ValueError on empty DiscreteMap mask")
+
+    # Passing both root and roots raises.
+    try:
+        g.get_dijkstra_map(root=(0, 0), roots=[(1, 1)])
+    except TypeError:
+        pass
+    else:
+        raise AssertionError("expected TypeError when both root and roots supplied")
+
+    print("PASS")
+
+
+if __name__ == "__main__":
+    main()
+    sys.exit(0)