diff --git a/package-lock.json b/package-lock.json
index 7e6eccba..013ae1e2 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -25,6 +25,9 @@
         "@commitlint/cli": "^21.1.0",
         "@commitlint/config-conventional": "^21.2.0",
         "@eslint/js": "^10.0.1",
+        "@langchain/anthropic": "^1.5.1",
+        "@langchain/core": "^1.2.2",
+        "@langchain/ollama": "^1.3.0",
         "@types/better-sqlite3": "^7.6.12",
         "@types/node": "^25.9.3",
         "@vitest/coverage-v8": "^4.1.10",
@@ -34,6 +37,7 @@
         "eslint-config-prettier": "^10.1.8",
         "fast-check": "^4.8.0",
         "husky": "^9.1.7",
+        "langchain": "^1.5.3",
         "lint-staged": "^17.0.8",
         "playwright": "^1.61.1",
         "prettier": "^3.9.4",
@@ -48,6 +52,28 @@
         "node": ">=20"
       }
     },
+    "node_modules/@anthropic-ai/sdk": {
+      "version": "0.103.0",
+      "resolved": "https://registry.npmjs.org/@anthropic-ai/sdk/-/sdk-0.103.0.tgz",
+      "integrity": "sha512-1uG7RNgoHTUxzOXqSCODKt0UTVlxWiHk/2Tt2/uQJiPW7XzBeKVuJyd3Aw6T3LPyvZV/jDTnPLX7SaM70WLLjA==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "json-schema-to-ts": "^3.1.1",
+        "standardwebhooks": "^1.0.0"
+      },
+      "bin": {
+        "anthropic-ai-sdk": "bin/cli"
+      },
+      "peerDependencies": {
+        "zod": "^3.25.0 || ^4.0.0"
+      },
+      "peerDependenciesMeta": {
+        "zod": {
+          "optional": true
+        }
+      }
+    },
     "node_modules/@babel/code-frame": {
       "version": "7.29.7",
       "resolved": "https://registry.npmjs.org/@babel/code-frame/-/code-frame-7.29.7.tgz",
@@ -106,6 +132,16 @@
         "node": ">=6.0.0"
       }
     },
+    "node_modules/@babel/runtime": {
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/runtime/-/runtime-7.29.7.tgz",
+      "integrity": "sha512-Nq8OhGWiZIZGV6hLHoyAKLLcJihP/xFeBMGJoUrxTX2psI8dCifzLhZISFb+VWS3wFMRDmCGw5R+dOySCqPLhw==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
     "node_modules/@babel/types": {
       "version": "7.29.0",
       "resolved": "https://registry.npmjs.org/@babel/types/-/types-7.29.0.tgz",
@@ -180,6 +216,13 @@
         "keyv": "^5.6.0"
       }
     },
+    "node_modules/@cfworker/json-schema": {
+      "version": "4.1.1",
+      "resolved": "https://registry.npmjs.org/@cfworker/json-schema/-/json-schema-4.1.1.tgz",
+      "integrity": "sha512-gAmrUZSGtKc3AiBL71iNWxDsyUC5uMaKKGdvzYsBoTW/xi42JQHl7eKV2OYzCUqvc+D2RCcf7EXY2iCyFIk6og==",
+      "devOptional": true,
+      "license": "MIT"
+    },
     "node_modules/@clack/core": {
       "version": "1.2.0",
       "resolved": "https://registry.npmjs.org/@clack/core/-/core-1.2.0.tgz",
@@ -1467,9 +1510,6 @@
       "cpu": [
         "arm"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "LGPL-3.0-or-later",
       "optional": true,
       "os": [
@@ -1486,9 +1526,6 @@
       "cpu": [
         "arm64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "LGPL-3.0-or-later",
       "optional": true,
       "os": [
@@ -1505,9 +1542,6 @@
       "cpu": [
         "ppc64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "LGPL-3.0-or-later",
       "optional": true,
       "os": [
@@ -1524,9 +1558,6 @@
       "cpu": [
         "riscv64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "LGPL-3.0-or-later",
       "optional": true,
       "os": [
@@ -1543,9 +1574,6 @@
       "cpu": [
         "s390x"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "LGPL-3.0-or-later",
       "optional": true,
       "os": [
@@ -1562,9 +1590,6 @@
       "cpu": [
         "x64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "LGPL-3.0-or-later",
       "optional": true,
       "os": [
@@ -1581,9 +1606,6 @@
       "cpu": [
         "arm64"
       ],
-      "libc": [
-        "musl"
-      ],
       "license": "LGPL-3.0-or-later",
       "optional": true,
       "os": [
@@ -1600,9 +1622,6 @@
       "cpu": [
         "x64"
       ],
-      "libc": [
-        "musl"
-      ],
       "license": "LGPL-3.0-or-later",
       "optional": true,
       "os": [
@@ -1619,9 +1638,6 @@
       "cpu": [
         "arm"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "Apache-2.0",
       "optional": true,
       "os": [
@@ -1644,9 +1660,6 @@
       "cpu": [
         "arm64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "Apache-2.0",
       "optional": true,
       "os": [
@@ -1669,9 +1682,6 @@
       "cpu": [
         "ppc64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "Apache-2.0",
       "optional": true,
       "os": [
@@ -1694,9 +1704,6 @@
       "cpu": [
         "riscv64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "Apache-2.0",
       "optional": true,
       "os": [
@@ -1719,9 +1726,6 @@
       "cpu": [
         "s390x"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "Apache-2.0",
       "optional": true,
       "os": [
@@ -1744,9 +1748,6 @@
       "cpu": [
         "x64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "Apache-2.0",
       "optional": true,
       "os": [
@@ -1769,9 +1770,6 @@
       "cpu": [
         "arm64"
       ],
-      "libc": [
-        "musl"
-      ],
       "license": "Apache-2.0",
       "optional": true,
       "os": [
@@ -1794,9 +1792,6 @@
       "cpu": [
         "x64"
       ],
-      "libc": [
-        "musl"
-      ],
       "license": "Apache-2.0",
       "optional": true,
       "os": [
@@ -1956,6 +1951,169 @@
       "dev": true,
       "license": "MIT"
     },
+    "node_modules/@langchain/anthropic": {
+      "version": "1.5.1",
+      "resolved": "https://registry.npmjs.org/@langchain/anthropic/-/anthropic-1.5.1.tgz",
+      "integrity": "sha512-j92zCCd5BFH3rHMRzc2wBmSKDoVpinof1oh8aFiAz9TWbSOc4tGU4n6bqwy/wP0GH1uO96zZHLGCHBMPgrxTNw==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@anthropic-ai/sdk": "^0.103.0",
+        "zod": "^3.25.76 || ^4"
+      },
+      "engines": {
+        "node": ">=20"
+      },
+      "peerDependencies": {
+        "@langchain/core": "^1.2.1"
+      }
+    },
+    "node_modules/@langchain/core": {
+      "version": "1.2.2",
+      "resolved": "https://registry.npmjs.org/@langchain/core/-/core-1.2.2.tgz",
+      "integrity": "sha512-KfjEOT6sCg0vvItagfEtGpmrGoLMGfma4Affb5BGEqPmS2YR3AxW54pABSkhQlzCehTB+0BnLquAe1lGF4J9zQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@cfworker/json-schema": "^4.0.2",
+        "@standard-schema/spec": "^1.1.0",
+        "js-tiktoken": "^1.0.12",
+        "langsmith": ">=0.5.0 <1.0.0",
+        "mustache": "^4.2.0",
+        "p-queue": "^6.6.2",
+        "zod": "^3.25.76 || ^4"
+      },
+      "engines": {
+        "node": ">=20"
+      }
+    },
+    "node_modules/@langchain/core/node_modules/eventemitter3": {
+      "version": "4.0.7",
+      "resolved": "https://registry.npmjs.org/eventemitter3/-/eventemitter3-4.0.7.tgz",
+      "integrity": "sha512-8guHBZCwKnFhYdHr2ysuRWErTwhoN2X8XELRlrRwpmfeY2jjuUN4taQMsULKUVo1K4DvZl+0pgfyoysHxvmvEw==",
+      "dev": true,
+      "license": "MIT"
+    },
+    "node_modules/@langchain/core/node_modules/p-queue": {
+      "version": "6.6.2",
+      "resolved": "https://registry.npmjs.org/p-queue/-/p-queue-6.6.2.tgz",
+      "integrity": "sha512-RwFpb72c/BhQLEXIZ5K2e+AhgNVmIejGlTgiB9MzZ0e93GRvqZ7uSi0dvRF7/XIXDeNkra2fNHBxTyPDGySpjQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "eventemitter3": "^4.0.4",
+        "p-timeout": "^3.2.0"
+      },
+      "engines": {
+        "node": ">=8"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
+    "node_modules/@langchain/core/node_modules/p-timeout": {
+      "version": "3.2.0",
+      "resolved": "https://registry.npmjs.org/p-timeout/-/p-timeout-3.2.0.tgz",
+      "integrity": "sha512-rhIwUycgwwKcP9yTOOFK/AKsAopjjCakVqLHePO3CC6Mir1Z99xT+R63jZxAT5lFZLa2inS5h+ZS2GvR99/FBg==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "p-finally": "^1.0.0"
+      },
+      "engines": {
+        "node": ">=8"
+      }
+    },
+    "node_modules/@langchain/langgraph": {
+      "version": "1.4.7",
+      "resolved": "https://registry.npmjs.org/@langchain/langgraph/-/langgraph-1.4.7.tgz",
+      "integrity": "sha512-2tcyf3QGC7v89kqSxMCtRvzg/3L/4yHtOaWC49A8KieCciWJs7LGaxHoPB6QRxXyUgyR+Zg9Q1ss/XJIE+JuSQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@langchain/langgraph-checkpoint": "^1.1.3",
+        "@langchain/langgraph-sdk": "~1.9.25",
+        "@langchain/protocol": "^0.0.18",
+        "@standard-schema/spec": "1.1.0"
+      },
+      "engines": {
+        "node": ">=18"
+      },
+      "peerDependencies": {
+        "@langchain/core": "^1.1.48",
+        "zod": "^3.25.32 || ^4.2.0"
+      }
+    },
+    "node_modules/@langchain/langgraph-checkpoint": {
+      "version": "1.1.3",
+      "resolved": "https://registry.npmjs.org/@langchain/langgraph-checkpoint/-/langgraph-checkpoint-1.1.3.tgz",
+      "integrity": "sha512-wgzdQNeEsdw1e+4lvlj0tdq/RYR/k1vPin10g0ymGoehZDDgd9nvIllGXSXN4TFgF9sf5qQP/KTkOcLfeseIhA==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=18"
+      },
+      "peerDependencies": {
+        "@langchain/core": "^1.1.48"
+      }
+    },
+    "node_modules/@langchain/langgraph-sdk": {
+      "version": "1.9.25",
+      "resolved": "https://registry.npmjs.org/@langchain/langgraph-sdk/-/langgraph-sdk-1.9.25.tgz",
+      "integrity": "sha512-mRKW8zyQUaHox+HirRFMRrPqOvNbQI3xeXDt6kkk4PbBg77V92bsO1WzUVNrmJ81zCkvxyOrWSK8D6ioCj0a8A==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@langchain/protocol": "^0.0.18",
+        "@types/json-schema": "^7.0.15",
+        "p-queue": "^9.0.1",
+        "p-retry": "^7.1.1"
+      },
+      "peerDependencies": {
+        "@langchain/core": "^1.1.48",
+        "react": "^18 || ^19",
+        "react-dom": "^18 || ^19",
+        "svelte": "^4.0.0 || ^5.0.0",
+        "vue": "^3.0.0"
+      },
+      "peerDependenciesMeta": {
+        "react": {
+          "optional": true
+        },
+        "react-dom": {
+          "optional": true
+        },
+        "svelte": {
+          "optional": true
+        },
+        "vue": {
+          "optional": true
+        }
+      }
+    },
+    "node_modules/@langchain/ollama": {
+      "version": "1.3.0",
+      "resolved": "https://registry.npmjs.org/@langchain/ollama/-/ollama-1.3.0.tgz",
+      "integrity": "sha512-uzLkZVTXN0SI5zAoJbJzL4y/QN7gbmJ1txBgFfqJ5M0Qf+Yu2Zu8q98L3jIrfM+akWMaNmRU4R4APhcfQmeIsw==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "ollama": "^0.6.3"
+      },
+      "engines": {
+        "node": ">=20"
+      },
+      "peerDependencies": {
+        "@langchain/core": "^1.0.0"
+      }
+    },
+    "node_modules/@langchain/protocol": {
+      "version": "0.0.18",
+      "resolved": "https://registry.npmjs.org/@langchain/protocol/-/protocol-0.0.18.tgz",
+      "integrity": "sha512-XW1egQtPfsGI41w2AMZNFZrUIwFSQHTjVMZs0OaTpCAvht/QLoaPN8FQcsysMVypOhupG28J29yOorrc70otBQ==",
+      "dev": true,
+      "license": "MIT"
+    },
     "node_modules/@linear/sdk": {
       "version": "88.0.0",
       "resolved": "https://registry.npmjs.org/@linear/sdk/-/sdk-88.0.0.tgz",
@@ -2210,9 +2368,6 @@
         "arm64"
       ],
       "dev": true,
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -2230,9 +2385,6 @@
         "arm64"
       ],
       "dev": true,
-      "libc": [
-        "musl"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -2250,9 +2402,6 @@
         "ppc64"
       ],
       "dev": true,
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -2270,9 +2419,6 @@
         "s390x"
       ],
       "dev": true,
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -2290,9 +2436,6 @@
         "x64"
       ],
       "dev": true,
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -2310,9 +2453,6 @@
         "x64"
       ],
       "dev": true,
-      "libc": [
-        "musl"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -2428,6 +2568,13 @@
         "url": "https://ko-fi.com/dangreen"
       }
     },
+    "node_modules/@stablelib/base64": {
+      "version": "1.0.1",
+      "resolved": "https://registry.npmjs.org/@stablelib/base64/-/base64-1.0.1.tgz",
+      "integrity": "sha512-1bnPQqSxSuc3Ii6MhBysoWCg58j97aUjuCSZrGSmDxNqtytIi0k8utUenAwTZN4V5mXXYGsVUI9zeBqy+jBOSQ==",
+      "dev": true,
+      "license": "MIT"
+    },
     "node_modules/@standard-schema/spec": {
       "version": "1.1.0",
       "resolved": "https://registry.npmjs.org/@standard-schema/spec/-/spec-1.1.0.tgz",
@@ -4510,6 +4657,13 @@
       "integrity": "sha512-W+KJc2dmILlPplD/H4K9l9LcAHAfPtP6BY84uVLXQ6Evcz9Lcg33Y2z1IVblT6xdY54PXYVHEv+0Wpq8Io6zkA==",
       "license": "MIT"
     },
+    "node_modules/fast-sha256": {
+      "version": "1.3.0",
+      "resolved": "https://registry.npmjs.org/fast-sha256/-/fast-sha256-1.3.0.tgz",
+      "integrity": "sha512-n11RGP/lrWEFI/bWdygLxhI+pVeo1ZYIVwvvPkW7azl/rOy+F3HYRZ2K5zeE9mmkhQppyv9sQFx0JM9UabnpPQ==",
+      "dev": true,
+      "license": "Unlicense"
+    },
     "node_modules/fast-string-truncated-width": {
       "version": "1.2.1",
       "resolved": "https://registry.npmjs.org/fast-string-truncated-width/-/fast-string-truncated-width-1.2.1.tgz",
@@ -5251,6 +5405,19 @@
         "node": ">=0.10.0"
       }
     },
+    "node_modules/is-network-error": {
+      "version": "1.3.2",
+      "resolved": "https://registry.npmjs.org/is-network-error/-/is-network-error-1.3.2.tgz",
+      "integrity": "sha512-PhBY86zaxNZUuWP6h13Vu5oFe0XY6/UlKzQnYFELzGVHygP3MxmvTfYSG7GN3aIab/iWudSMgjSnG9Dq+nHrgA==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=16"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
     "node_modules/is-obj": {
       "version": "2.0.0",
       "resolved": "https://registry.npmjs.org/is-obj/-/is-obj-2.0.0.tgz",
@@ -5353,6 +5520,16 @@
         "node": ">=10"
       }
     },
+    "node_modules/js-tiktoken": {
+      "version": "1.0.21",
+      "resolved": "https://registry.npmjs.org/js-tiktoken/-/js-tiktoken-1.0.21.tgz",
+      "integrity": "sha512-biOj/6M5qdgx5TKjDnFT1ymSpM5tbd3ylwDtrQvFQSu0Z7bBYko2dF+W/aUkXUPuk6IVpRxk/3Q2sHOzGlS36g==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "base64-js": "^1.5.1"
+      }
+    },
     "node_modules/js-tokens": {
       "version": "10.0.0",
       "resolved": "https://registry.npmjs.org/js-tokens/-/js-tokens-10.0.0.tgz",
@@ -5406,6 +5583,20 @@
       "dev": true,
       "license": "MIT"
     },
+    "node_modules/json-schema-to-ts": {
+      "version": "3.1.1",
+      "resolved": "https://registry.npmjs.org/json-schema-to-ts/-/json-schema-to-ts-3.1.1.tgz",
+      "integrity": "sha512-+DWg8jCJG2TEnpy7kOm/7/AxaYoaRbjVB4LFZLySZlWn8exGs3A4OLJR966cVvU26N7X9TWxl+Jsw7dzAqKT6g==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@babel/runtime": "^7.18.3",
+        "ts-algebra": "^2.0.0"
+      },
+      "engines": {
+        "node": ">=16"
+      }
+    },
     "node_modules/json-schema-traverse": {
       "version": "1.0.0",
       "resolved": "https://registry.npmjs.org/json-schema-traverse/-/json-schema-traverse-1.0.0.tgz",
@@ -5456,6 +5647,96 @@
         "@keyv/serialize": "^1.1.1"
       }
     },
+    "node_modules/langchain": {
+      "version": "1.5.3",
+      "resolved": "https://registry.npmjs.org/langchain/-/langchain-1.5.3.tgz",
+      "integrity": "sha512-YRYPy1xPq4CgKe0NdPoscYHnKBW5cViEyWiZsUKQfN3xL1X5DpGlIpaBQg1A/t6o9ePKE7J5b7nqrfLz8t7/Rw==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@langchain/langgraph": "^1.4.7",
+        "@langchain/langgraph-checkpoint": "^1.1.3",
+        "langsmith": ">=0.5.0 <1.0.0",
+        "zod": "^3.25.76 || ^4"
+      },
+      "engines": {
+        "node": ">=20"
+      },
+      "peerDependencies": {
+        "@langchain/core": "^1.2.1"
+      }
+    },
+    "node_modules/langsmith": {
+      "version": "0.8.2",
+      "resolved": "https://registry.npmjs.org/langsmith/-/langsmith-0.8.2.tgz",
+      "integrity": "sha512-WX85yuL9vHAFIXJ4Nde/2VsjZT5o6FzX617uKDGtcgMojPe+IM6lgiZeVagv/egC79bQS9HPKVD/UPe/6zGFPg==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "p-queue": "6.6.2"
+      },
+      "peerDependencies": {
+        "@opentelemetry/api": "*",
+        "@opentelemetry/exporter-trace-otlp-proto": "*",
+        "@opentelemetry/sdk-trace-base": "*",
+        "openai": "*",
+        "ws": ">=7"
+      },
+      "peerDependenciesMeta": {
+        "@opentelemetry/api": {
+          "optional": true
+        },
+        "@opentelemetry/exporter-trace-otlp-proto": {
+          "optional": true
+        },
+        "@opentelemetry/sdk-trace-base": {
+          "optional": true
+        },
+        "openai": {
+          "optional": true
+        },
+        "ws": {
+          "optional": true
+        }
+      }
+    },
+    "node_modules/langsmith/node_modules/eventemitter3": {
+      "version": "4.0.7",
+      "resolved": "https://registry.npmjs.org/eventemitter3/-/eventemitter3-4.0.7.tgz",
+      "integrity": "sha512-8guHBZCwKnFhYdHr2ysuRWErTwhoN2X8XELRlrRwpmfeY2jjuUN4taQMsULKUVo1K4DvZl+0pgfyoysHxvmvEw==",
+      "dev": true,
+      "license": "MIT"
+    },
+    "node_modules/langsmith/node_modules/p-queue": {
+      "version": "6.6.2",
+      "resolved": "https://registry.npmjs.org/p-queue/-/p-queue-6.6.2.tgz",
+      "integrity": "sha512-RwFpb72c/BhQLEXIZ5K2e+AhgNVmIejGlTgiB9MzZ0e93GRvqZ7uSi0dvRF7/XIXDeNkra2fNHBxTyPDGySpjQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "eventemitter3": "^4.0.4",
+        "p-timeout": "^3.2.0"
+      },
+      "engines": {
+        "node": ">=8"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
+    "node_modules/langsmith/node_modules/p-timeout": {
+      "version": "3.2.0",
+      "resolved": "https://registry.npmjs.org/p-timeout/-/p-timeout-3.2.0.tgz",
+      "integrity": "sha512-rhIwUycgwwKcP9yTOOFK/AKsAopjjCakVqLHePO3CC6Mir1Z99xT+R63jZxAT5lFZLa2inS5h+ZS2GvR99/FBg==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "p-finally": "^1.0.0"
+      },
+      "engines": {
+        "node": ">=8"
+      }
+    },
     "node_modules/levn": {
       "version": "0.4.1",
       "resolved": "https://registry.npmjs.org/levn/-/levn-0.4.1.tgz",
@@ -5624,9 +5905,6 @@
         "arm64"
       ],
       "dev": true,
-      "libc": [
-        "glibc"
-      ],
       "license": "MPL-2.0",
       "optional": true,
       "os": [
@@ -5648,9 +5926,6 @@
         "arm64"
       ],
       "dev": true,
-      "libc": [
-        "musl"
-      ],
       "license": "MPL-2.0",
       "optional": true,
       "os": [
@@ -5672,9 +5947,6 @@
         "x64"
       ],
       "dev": true,
-      "libc": [
-        "glibc"
-      ],
       "license": "MPL-2.0",
       "optional": true,
       "os": [
@@ -5696,9 +5968,6 @@
         "x64"
       ],
       "dev": true,
-      "libc": [
-        "musl"
-      ],
       "license": "MPL-2.0",
       "optional": true,
       "os": [
@@ -6243,6 +6512,16 @@
         "node": ">=18"
       }
     },
+    "node_modules/mustache": {
+      "version": "4.2.0",
+      "resolved": "https://registry.npmjs.org/mustache/-/mustache-4.2.0.tgz",
+      "integrity": "sha512-71ippSywq5Yb7/tVYyGbkBggbU8H3u5Rz56fH60jGFgr8uHwxs+aSKeqmluIVzM0m0kB7xQjKS6qPfd0b2ZoqQ==",
+      "dev": true,
+      "license": "MIT",
+      "bin": {
+        "mustache": "bin/mustache"
+      }
+    },
     "node_modules/nanoid": {
       "version": "3.3.15",
       "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.15.tgz",
@@ -6378,6 +6657,16 @@
       ],
       "license": "MIT"
     },
+    "node_modules/ollama": {
+      "version": "0.6.3",
+      "resolved": "https://registry.npmjs.org/ollama/-/ollama-0.6.3.tgz",
+      "integrity": "sha512-KEWEhIqE5wtfzEIZbDCLH51VFZ6Z3ZSa6sIOg/E/tBV8S51flyqBOXi+bRxlOYKDf8i327zG9eSTb8IJxvm3Zg==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "whatwg-fetch": "^3.6.20"
+      }
+    },
     "node_modules/on-exit-leak-free": {
       "version": "2.1.2",
       "resolved": "https://registry.npmjs.org/on-exit-leak-free/-/on-exit-leak-free-2.1.2.tgz",
@@ -6442,6 +6731,16 @@
         "node": ">= 0.8.0"
       }
     },
+    "node_modules/p-finally": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/p-finally/-/p-finally-1.0.0.tgz",
+      "integrity": "sha512-LICb2p9CB7FS+0eR1oqWnHhp0FljGLZCWBE9aix0Uye9W8LTQPwMTYVGWQWIw9RdQiDg4+epXQODwIYJtSJaow==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=4"
+      }
+    },
     "node_modules/p-queue": {
       "version": "9.1.1",
       "resolved": "https://registry.npmjs.org/p-queue/-/p-queue-9.1.1.tgz",
@@ -6459,6 +6758,22 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
+    "node_modules/p-retry": {
+      "version": "7.1.1",
+      "resolved": "https://registry.npmjs.org/p-retry/-/p-retry-7.1.1.tgz",
+      "integrity": "sha512-J5ApzjyRkkf601HpEeykoiCvzHQjWxPAHhyjFcEUP2SWq0+35NKh8TLhpLw+Dkq5TZBFvUM6UigdE9hIVYTl5w==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "is-network-error": "^1.1.0"
+      },
+      "engines": {
+        "node": ">=20"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
     "node_modules/p-timeout": {
       "version": "7.0.1",
       "resolved": "https://registry.npmjs.org/p-timeout/-/p-timeout-7.0.1.tgz",
@@ -7504,6 +7819,17 @@
       "dev": true,
       "license": "MIT"
     },
+    "node_modules/standardwebhooks": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/standardwebhooks/-/standardwebhooks-1.0.0.tgz",
+      "integrity": "sha512-BbHGOQK9olHPMvQNHWul6MYlrRTAOKn03rOe4A8O3CLWhNf4YHBqq2HJKKC+sfqpxiBY52pNeesD6jIiLDz8jg==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@stablelib/base64": "^1.0.0",
+        "fast-sha256": "^1.3.0"
+      }
+    },
     "node_modules/statuses": {
       "version": "2.0.2",
       "resolved": "https://registry.npmjs.org/statuses/-/statuses-2.0.2.tgz",
@@ -7784,6 +8110,13 @@
         "node": ">=14.17"
       }
     },
+    "node_modules/ts-algebra": {
+      "version": "2.0.0",
+      "resolved": "https://registry.npmjs.org/ts-algebra/-/ts-algebra-2.0.0.tgz",
+      "integrity": "sha512-FPAhNPFMrkwz76P7cdjdmiShwMynZYN6SgOujD1urY4oNm80Ou9oMdmbR45LotcKOXoy7wSmHkRFE6Mxbrhefw==",
+      "dev": true,
+      "license": "MIT"
+    },
     "node_modules/ts-api-utils": {
       "version": "2.5.0",
       "resolved": "https://registry.npmjs.org/ts-api-utils/-/ts-api-utils-2.5.0.tgz",
@@ -8151,6 +8484,13 @@
       "dev": true,
       "license": "MIT"
     },
+    "node_modules/whatwg-fetch": {
+      "version": "3.6.20",
+      "resolved": "https://registry.npmjs.org/whatwg-fetch/-/whatwg-fetch-3.6.20.tgz",
+      "integrity": "sha512-EqhiFU6daOA8kpjOWTL0olhVOF3i7OrFzSYiGsEMB8GcXS+RrzauAERX65xMeNWVqxA6HXH2m69Z9LaKKdisfg==",
+      "dev": true,
+      "license": "MIT"
+    },
     "node_modules/which": {
       "version": "2.0.2",
       "resolved": "https://registry.npmjs.org/which/-/which-2.0.2.tgz",
diff --git a/package.json b/package.json
index 14d55177..815dbb80 100644
--- a/package.json
+++ b/package.json
@@ -58,6 +58,9 @@
     "@commitlint/cli": "^21.1.0",
     "@commitlint/config-conventional": "^21.2.0",
     "@eslint/js": "^10.0.1",
+    "@langchain/anthropic": "^1.5.1",
+    "@langchain/core": "^1.2.2",
+    "@langchain/ollama": "^1.3.0",
     "@types/better-sqlite3": "^7.6.12",
     "@types/node": "^25.9.3",
     "@vitest/coverage-v8": "^4.1.10",
@@ -67,6 +70,7 @@
     "eslint-config-prettier": "^10.1.8",
     "fast-check": "^4.8.0",
     "husky": "^9.1.7",
+    "langchain": "^1.5.3",
     "lint-staged": "^17.0.8",
     "playwright": "^1.61.1",
     "prettier": "^3.9.4",
diff --git a/scripts/spikes/lia396_provider_override_walking_skeleton.md b/scripts/spikes/lia396_provider_override_walking_skeleton.md
new file mode 100644
index 00000000..1cc88a8e
--- /dev/null
+++ b/scripts/spikes/lia396_provider_override_walking_skeleton.md
@@ -0,0 +1,169 @@
+# Spike: per-request provider overrides through LangChain's `wrapModelCall` middleware (LIA-396 / A3)
+
+**Date:** 2026-07-14 · **Verdict: per-AC breakdown below (no blanket PASS)**
+— selection and provider-attribution logging proven live and deterministic;
+the subagent/Ollama leg is fully live with zero mocking; the main-agent/
+Claude leg's live network execution is explicitly deferred to A4
+(LIA-397, credential-proxy) since `ANTHROPIC_API_KEY` is absent here.
+
+## Question
+
+Can LangChain JS's real `wrapModelCall` middleware contract — the third
+member of the middleware family after `createAgent` (A1) and `wrapToolCall`
+(A2) — route different chat model providers to different requests **within
+one logical session**? This is A3 of the Linear-tracked MA milestone
+(LIA-394..400), continuing from A1 (LIA-394, PR #1031) and A2 (LIA-395, PR
+#1032, both still open/unmerged).
+
+LIA-396 lists five acceptance criteria:
+
+1. `wrapModelCall` selects a Claude model for the main-agent request.
+2. `wrapModelCall` selects a local Ollama model for a subagent request.
+3. Both requests execute within one logical session.
+4. Logs or test assertions identify which provider handled each request.
+5. The proof can be reused by the A7 benchmark harness.
+
+## Method
+
+### Reuse target — unmodified
+
+`wrapModelCall`'s real contract, confirmed by reading
+`node_modules/langchain/dist/agents/middleware/types.d.ts`:
+`createMiddleware({ name, wrapModelCall: async (request: ModelRequest,
+handler: WrapModelCallHandler) => Promise<AIMessage | Command> })`, passed
+to `createAgent({ middleware: [...] })`. The official JSDoc example shows
+`request.model` being swapped and re-dispatched via
+`handler({...request, model: fallbackModel})` — this spike's
+`createProviderRoutingMiddleware` uses exactly that documented pattern, not
+a novel mechanism.
+
+### `createProviderRoutingMiddleware` — the reusable primitive
+
+A real `wrapModelCall` middleware that reads `request.runtime.context.role`,
+resolves a model instance via an injectable `resolveModel` Strategy, swaps
+`request.model`, and records every selection to an inspectable `log:
+ProviderRoutingRecord[]` (`requestIndex`, `role`, `providerClass`, `modelId`,
+`executed`). `resolveModel`/`executeOverride` are Strategy-pattern seams —
+interchangeable per-role algorithms, matching A2's `invokeGate`
+injectable-seam precedent. `executeOverride` is an optional second Strategy
+seam that, when it returns a message, short-circuits the real network call
+for a given role — used ONLY to route around the disclosed
+`ANTHROPIC_API_KEY` absence, never silently, and always recorded truthfully
+in the log (`executed: false`).
+
+### "One logical session" — one top-level `agent.invoke()`, one nested `agent.invoke()`
+
+Modeled the deus-native two-tier architecture directly: a `delegate_to_
+subagent` tool whose `execute()` constructs and invokes a *second*,
+nested `createAgent`, sharing the *same* `wrapModelCall` middleware instance
+as the outer agent, disambiguated by `request.runtime.context.role`
+(`'main'` vs `'sub'`). This is call-stack/session continuity — both legs
+occur within one real `mainAgent.invoke()` call tree, one shared middleware
+instance, one shared routing log. Two independent top-level `.invoke()`
+calls with no nesting would not satisfy this — there would be no shared
+call stack and nothing forcing the two calls to be causally related within
+one run.
+
+### Provider selection — both real, live, in the same call stack
+
+`defaultMainSubResolver()`: role `'main'` → a real `ChatAnthropic` instance
+(placeholder API key when `ANTHROPIC_API_KEY` is absent — construction
+throws without *some* key present, confirmed empirically); role `'sub'` →
+a real `ChatOllama` instance pointed at local Ollama (`OLLAMA_URL`, reusing
+the repo's existing env-var convention). Both providers expose the model id
+on the same `.model` field, confirmed by construction — uniform
+provider-attribution logging with no per-provider branching.
+
+### Driving the loop without a Claude credential
+
+`makeScriptedMainStub()`: a *stateful*, call-count-aware `executeOverride`
+for role `'main'` only — its first invocation returns a `tool_calls`-bearing
+`AIMessage` naming `delegate_to_subagent`, every subsequent invocation
+returns a plain final `AIMessage`. A stateless single-canned-response stub
+would never actually invoke the tool, so the `'sub'` leg would never fire —
+this was caught and fixed during planning, before implementation. The
+`'sub'`/Ollama leg has no override — it always hits the real local Ollama
+server for real.
+
+## What was built
+
+- `scripts/spikes/lia396_provider_override_walking_skeleton.ts` —
+  `createProviderRoutingMiddleware` (the reusable primitive),
+  `defaultMainSubResolver`, `makeScriptedMainStub`,
+  `makeDelegateToSubagentTool`, and a `main()` direct-execution demo.
+- `scripts/spikes/lia396_provider_override_walking_skeleton.test.ts` — 7
+  tests (6 passing, 1 conditionally skipped):
+  - Selection: `'main'` role resolves a real `ChatAnthropic` instance with
+    the right model id; `'sub'` role resolves a real `ChatOllama` instance;
+    an unrecognized role passes the original `request.model` through
+    unchanged.
+  - Routing log: entries recorded in order with correct
+    `requestIndex`/`role`/`providerClass`/`modelId`; `executed: false` on
+    the `executeOverride` short-circuit path vs `executed: true` on the
+    default real-handler path.
+  - **The integration test** — one real `mainAgent.invoke()` call, real
+    nested `subAgent.invoke()`, real Ollama network call: asserts exactly
+    one `'sub'` log entry (`ChatOllama`, `executed: true`), at least two
+    `'main'` entries (`ChatAnthropic`, `executed: false`) straddling the
+    `'sub'` entry in `requestIndex` order (the live `main → sub → main`
+    trace), and that the subagent's actual returned content is real
+    (non-empty, not the canned stub string) — proving the Ollama round-trip
+    genuinely happened, not just that the mechanism fired.
+  - Optional live full-Claude smoke test, gated on `ANTHROPIC_API_KEY`
+    presence (same disposition as A1) — skipped in this environment.
+- `vitest.config.ts` — same `scripts/spikes/**/*.test.ts` glob addition
+  A1/A2 made (independently, since neither is merged yet).
+- Direct-execution demo output (`npx tsx
+  scripts/spikes/lia396_provider_override_walking_skeleton.ts`), matching
+  the automated assertions exactly:
+  ```
+  ── Transcript ──
+  [human] Delegate to your subagent to find out its favorite fruit, then report back.
+  [ai:model]
+  [tool:delegate_to_subagent] {"subagent_answer":"Apple"}
+  [ai:model] Delegated to the subagent and got an answer back.
+
+  ── Routing log ──
+  #0 role=main provider=ChatAnthropic model=claude-opus-4-8 executed=false
+  #1 role=sub provider=ChatOllama model=gemma4:e2b executed=true
+  #2 role=main provider=ChatAnthropic model=claude-opus-4-8 executed=false
+  ```
+
+## Verdict — per acceptance criterion
+
+- **AC1** "wrapModelCall selects a Claude model for the main-agent request"
+  — **PASS**. A real `ChatAnthropic` instance is genuinely constructed and
+  assigned to `request.model` for the `'main'` role, live, in the same
+  session run as the `'sub'` leg.
+- **AC2** "wrapModelCall selects a local Ollama model for a subagent
+  request" — **PASS**. A real `ChatOllama` instance is genuinely
+  constructed, assigned, AND executed against local Ollama
+  (`gemma4:e2b`) — zero mocking on this leg.
+- **AC3** "Both requests execute within one logical session" — **PARTIAL**.
+  Session/call-stack nesting: yes — one top-level `agent.invoke()`, one
+  nested `agent.invoke()` inside `delegate_to_subagent`, sharing one
+  middleware instance and one routing log, live-verified `main → sub →
+  main` trace. Live network *execution* of both legs: no — the `'main'`/
+  Claude leg is stubbed via the disclosed `executeOverride` seam because
+  `ANTHROPIC_API_KEY` is absent; only the `'sub'`/Ollama leg genuinely
+  executes over the network. Full literal AC3 satisfaction (both legs
+  actually executing) is A4's job (LIA-397, credential-proxy), not A3's —
+  stated explicitly here rather than folded into a blanket pass.
+- **AC4** "Logs or test assertions identify which provider handled each
+  request" — **PASS**. `ProviderRoutingRecord[]` entries carry
+  `role`/`providerClass`/`modelId`/`executed` per request.
+- **AC5** "The proof can be reused by the A7 benchmark harness" —
+  **PASS**. `createProviderRoutingMiddleware` is a standalone, parameterized
+  export with no hardcoded `'main'`/`'sub'` coupling — that logic lives only
+  in the separately-exported, explicitly-example-labeled
+  `defaultMainSubResolver()`. A7 supplies its own resolver per benchmark
+  matrix cell.
+
+## Scope of this spike
+
+Read-only proof-of-concept; nothing wired into production. No modifications
+to any reused mechanism. Not merged to `main` — pushed as a PR for review
+only. When A4 (LIA-397) lands, the same `createProviderRoutingMiddleware`
+runs with no `executeOverride` on the `'main'` leg and a real Anthropic call
+happens live — nothing about this design needs to change for that
+transition; `executeOverride` simply becomes unused for that role.
diff --git a/scripts/spikes/lia396_provider_override_walking_skeleton.test.ts b/scripts/spikes/lia396_provider_override_walking_skeleton.test.ts
new file mode 100644
index 00000000..69253446
--- /dev/null
+++ b/scripts/spikes/lia396_provider_override_walking_skeleton.test.ts
@@ -0,0 +1,288 @@
+import { ChatAnthropic } from '@langchain/anthropic';
+import { AIMessage } from '@langchain/core/messages';
+import { ChatOllama } from '@langchain/ollama';
+import { createAgent, type ModelRequest } from 'langchain';
+import { describe, expect, it, vi } from 'vitest';
+
+import {
+  CLAUDE_MODEL_ID,
+  OLLAMA_BASE_URL,
+  OLLAMA_SUB_MODEL_ID,
+  createProviderRoutingMiddleware,
+  defaultMainSubResolver,
+  makeDelegateToSubagentTool,
+  makeScriptedMainStub,
+} from './lia396_provider_override_walking_skeleton.js';
+
+/**
+ * The integration test below requires a real, reachable local Ollama server
+ * with OLLAMA_SUB_MODEL_ID actually pulled — infrastructure this repo's
+ * default test suite cannot assume (CI, other contributors' machines, or
+ * even a machine with Ollama running but a different model set). Self-gates
+ * via a live probe (daemon reachable AND the specific model present in
+ * /api/tags, short timeout) rather than a manual opt-in env var, mirroring
+ * the ANTHROPIC_API_KEY-presence gate used for the Claude smoke test below
+ * — the suite adapts to what's actually available instead of failing with a
+ * missing-model error on a machine that has Ollama but not this model.
+ */
+async function isOllamaSubModelAvailable(): Promise<boolean> {
+  try {
+    const response = await fetch(`${OLLAMA_BASE_URL}/api/tags`, {
+      signal: AbortSignal.timeout(1000),
+    });
+    if (!response.ok) return false;
+    const body = (await response.json()) as {
+      models?: Array<{ name?: string; model?: string }>;
+    };
+    return (body.models ?? []).some(
+      (m) => m.name === OLLAMA_SUB_MODEL_ID || m.model === OLLAMA_SUB_MODEL_ID,
+    );
+  } catch {
+    return false;
+  }
+}
+
+const OLLAMA_AVAILABLE = await isOllamaSubModelAvailable();
+
+/**
+ * Minimal fake ModelRequest for direct wrapModelCall invocation — only the
+ * fields the middleware actually reads (request.model, request.runtime
+ * .context.role) need to be present, mirroring A2's own precedent
+ * (lia395_warden_veto_walking_skeleton.test.ts:251-255) of casting a
+ * minimal object literal rather than fully satisfying ModelRequest's shape.
+ */
+function fakeRequest(
+  role: string | undefined,
+  model: ModelRequest['model'],
+): ModelRequest {
+  return {
+    model,
+    messages: [],
+    runtime: { context: role === undefined ? undefined : { role } },
+  } as unknown as ModelRequest;
+}
+
+describe('createProviderRoutingMiddleware — selection', () => {
+  it("selects a real ChatAnthropic instance for role 'main'", async () => {
+    const { middleware } = createProviderRoutingMiddleware({
+      resolveModel: defaultMainSubResolver(),
+    });
+    const canned = new AIMessage({ content: 'ok' });
+    const handler = vi.fn().mockResolvedValue(canned);
+
+    await middleware.wrapModelCall!(
+      fakeRequest('main', new ChatOllama({ model: OLLAMA_SUB_MODEL_ID })),
+      handler,
+    );
+
+    expect(handler).toHaveBeenCalledTimes(1);
+    const passedRequest = handler.mock.calls[0]?.[0] as ModelRequest;
+    expect(passedRequest.model).toBeInstanceOf(ChatAnthropic);
+    expect((passedRequest.model as ChatAnthropic).model).toBe(CLAUDE_MODEL_ID);
+  });
+
+  it("selects a real ChatOllama instance for role 'sub'", async () => {
+    const { middleware } = createProviderRoutingMiddleware({
+      resolveModel: defaultMainSubResolver(),
+    });
+    const canned = new AIMessage({ content: 'ok' });
+    const handler = vi.fn().mockResolvedValue(canned);
+
+    await middleware.wrapModelCall!(
+      fakeRequest(
+        'sub',
+        new ChatAnthropic({ model: CLAUDE_MODEL_ID, apiKey: 'placeholder' }),
+      ),
+      handler,
+    );
+
+    expect(handler).toHaveBeenCalledTimes(1);
+    const passedRequest = handler.mock.calls[0]?.[0] as ModelRequest;
+    expect(passedRequest.model).toBeInstanceOf(ChatOllama);
+    expect((passedRequest.model as ChatOllama).model).toBe(OLLAMA_SUB_MODEL_ID);
+  });
+
+  it('passes the original request.model through for an unrecognized role', async () => {
+    const { middleware } = createProviderRoutingMiddleware({
+      resolveModel: defaultMainSubResolver(),
+    });
+    const original = new ChatOllama({ model: OLLAMA_SUB_MODEL_ID });
+    const handler = vi.fn().mockResolvedValue(new AIMessage({ content: 'ok' }));
+
+    await middleware.wrapModelCall!(
+      fakeRequest('mystery-role', original),
+      handler,
+    );
+
+    const passedRequest = handler.mock.calls[0]?.[0] as ModelRequest;
+    expect(passedRequest.model).toBe(original);
+  });
+});
+
+describe('createProviderRoutingMiddleware — routing log', () => {
+  it('records requestIndex/role/providerClass/modelId/executed in order', async () => {
+    const { middleware, log } = createProviderRoutingMiddleware({
+      resolveModel: defaultMainSubResolver(),
+    });
+    const handler = vi.fn().mockResolvedValue(new AIMessage({ content: 'ok' }));
+
+    await middleware.wrapModelCall!(
+      fakeRequest('main', new ChatOllama({ model: OLLAMA_SUB_MODEL_ID })),
+      handler,
+    );
+    await middleware.wrapModelCall!(
+      fakeRequest('sub', new ChatOllama({ model: OLLAMA_SUB_MODEL_ID })),
+      handler,
+    );
+
+    expect(log).toHaveLength(2);
+    expect(log[0]).toMatchObject({
+      requestIndex: 0,
+      role: 'main',
+      providerClass: 'ChatAnthropic',
+      modelId: CLAUDE_MODEL_ID,
+      executed: true,
+    });
+    expect(log[1]).toMatchObject({
+      requestIndex: 1,
+      role: 'sub',
+      providerClass: 'ChatOllama',
+      modelId: OLLAMA_SUB_MODEL_ID,
+      executed: true,
+    });
+  });
+
+  it('records executed:false when executeOverride short-circuits the call', async () => {
+    const { middleware, log } = createProviderRoutingMiddleware({
+      resolveModel: defaultMainSubResolver(),
+      executeOverride: (role) =>
+        role === 'main' ? new AIMessage({ content: 'stubbed' }) : undefined,
+    });
+    const handler = vi.fn().mockResolvedValue(new AIMessage({ content: 'ok' }));
+
+    const result = await middleware.wrapModelCall!(
+      fakeRequest('main', new ChatOllama({ model: OLLAMA_SUB_MODEL_ID })),
+      handler,
+    );
+
+    expect(handler).not.toHaveBeenCalled();
+    expect((result as AIMessage).content).toBe('stubbed');
+    expect(log[0]).toMatchObject({ role: 'main', executed: false });
+  });
+});
+
+describe('one logical session — nested delegate_to_subagent (real Ollama call)', () => {
+  const maybeIt = OLLAMA_AVAILABLE ? it : it.skip;
+
+  maybeIt(
+    'routes main -> sub -> main within one agent.invoke() call, sub leg genuinely live',
+    async () => {
+      const { middleware: sharedMiddleware, log } =
+        createProviderRoutingMiddleware({
+          resolveModel: defaultMainSubResolver(),
+          executeOverride: makeScriptedMainStub(),
+        });
+      const placeholderModel = new ChatOllama({ model: OLLAMA_SUB_MODEL_ID });
+      const delegateTool = makeDelegateToSubagentTool(
+        sharedMiddleware,
+        placeholderModel,
+      );
+
+      const mainAgent = createAgent({
+        model: placeholderModel,
+        tools: [delegateTool],
+        middleware: [sharedMiddleware],
+      });
+
+      const result = await mainAgent.invoke(
+        {
+          messages: [
+            {
+              role: 'user',
+              content:
+                'Delegate to your subagent to find out its favorite fruit, then report back.',
+            },
+          ],
+        },
+        { context: { role: 'main' } },
+      );
+
+      const subEntries = log.filter((r) => r.role === 'sub');
+      const mainEntries = log.filter((r) => r.role === 'main');
+
+      expect(subEntries).toHaveLength(1);
+      expect(subEntries[0]).toMatchObject({
+        providerClass: 'ChatOllama',
+        executed: true,
+      });
+
+      expect(mainEntries.length).toBeGreaterThanOrEqual(2);
+      expect(
+        mainEntries.every(
+          (r) => r.providerClass === 'ChatAnthropic' && r.executed === false,
+        ),
+      ).toBe(true);
+
+      const subIndex = subEntries[0]!.requestIndex;
+      expect(mainEntries.some((r) => r.requestIndex < subIndex)).toBe(true);
+      expect(mainEntries.some((r) => r.requestIndex > subIndex)).toBe(true);
+
+      const toolMessage = result.messages.find(
+        (m) => (m as { name?: string }).name === 'delegate_to_subagent',
+      );
+      expect(toolMessage).toBeDefined();
+      const payload = JSON.parse(String(toolMessage!.content)) as {
+        subagent_answer: string;
+      };
+      // Real local Ollama response — non-empty, not the canned main-leg stub string.
+      expect(payload.subagent_answer.length).toBeGreaterThan(0);
+      expect(payload.subagent_answer).not.toBe(
+        'Delegated to the subagent and got an answer back.',
+      );
+    },
+    30_000,
+  );
+});
+
+describe('optional live full-Claude smoke test', () => {
+  const hasCredential = Boolean(process.env.ANTHROPIC_API_KEY);
+  const maybeIt = hasCredential ? it : it.skip;
+
+  maybeIt(
+    'runs the same one-session flow with a real Claude call, no executeOverride',
+    async () => {
+      const { middleware: sharedMiddleware, log } =
+        createProviderRoutingMiddleware({
+          resolveModel: defaultMainSubResolver(),
+        });
+      const placeholderModel = new ChatOllama({ model: OLLAMA_SUB_MODEL_ID });
+      const delegateTool = makeDelegateToSubagentTool(
+        sharedMiddleware,
+        placeholderModel,
+      );
+
+      const mainAgent = createAgent({
+        model: placeholderModel,
+        tools: [delegateTool],
+        middleware: [sharedMiddleware],
+      });
+
+      await mainAgent.invoke(
+        {
+          messages: [
+            {
+              role: 'user',
+              content:
+                'Delegate to your subagent to find out its favorite fruit, then report back.',
+            },
+          ],
+        },
+        { context: { role: 'main' } },
+      );
+
+      const mainEntries = log.filter((r) => r.role === 'main');
+      expect(mainEntries.some((r) => r.executed === true)).toBe(true);
+    },
+    60_000,
+  );
+});
diff --git a/scripts/spikes/lia396_provider_override_walking_skeleton.ts b/scripts/spikes/lia396_provider_override_walking_skeleton.ts
new file mode 100644
index 00000000..d952fc2a
--- /dev/null
+++ b/scripts/spikes/lia396_provider_override_walking_skeleton.ts
@@ -0,0 +1,292 @@
+import { fileURLToPath } from 'node:url';
+import path from 'node:path';
+
+import { ChatAnthropic } from '@langchain/anthropic';
+import { AIMessage } from '@langchain/core/messages';
+import { ChatOllama } from '@langchain/ollama';
+import {
+  createAgent,
+  createMiddleware,
+  tool,
+  type AgentMiddleware,
+  type ModelRequest,
+} from 'langchain';
+
+export const CLAUDE_MODEL_ID = 'claude-opus-4-8'; // same as A1's MODEL_ID
+export const OLLAMA_SUB_MODEL_ID = 'gemma4:e2b';
+// Reuses the repo's existing OLLAMA_URL convention (.env.example:36, used
+// elsewhere for embeddings/judge) rather than introducing a new env var —
+// mapped to @langchain/ollama's baseUrl constructor field internally.
+export const OLLAMA_BASE_URL =
+  process.env.OLLAMA_URL ?? 'http://localhost:11434';
+const ANTHROPIC_PLACEHOLDER_KEY = 'sk-ant-lia396-placeholder-never-sent';
+
+export type ProviderRole = string; // deliberately open (not 'main'|'sub'
+// union) so A7 can add more roles without touching this file.
+
+export interface ProviderRoutingRecord {
+  /** Monotonic, 0-based, across this middleware instance's whole lifetime —
+   *  lets a caller reconstruct request order across a nested session. */
+  requestIndex: number;
+  role: ProviderRole;
+  /** request.model's constructor name AFTER selection, e.g. 'ChatAnthropic'. */
+  providerClass: string;
+  /** Best-effort model id: (model as { model?: string }).model. */
+  modelId: string | undefined;
+  /** true = the real LangChain handler ran (network call happened);
+   *  false = executeOverride short-circuited it (e.g. no live credential). */
+  executed: boolean;
+}
+
+/**
+ * Resolves which model instance should handle a request for a given role.
+ * Return the request's own request.model unchanged to mean "no override for
+ * this role" (passthrough). Strategy-pattern seam (interchangeable per-role
+ * algorithm, matching A2's `invokeGate` injectable-seam precedent) — this IS
+ * the thing under test/reuse; A7 supplies its own resolver per benchmark
+ * matrix cell.
+ */
+export type ProviderResolver = (
+  role: ProviderRole,
+  request: ModelRequest,
+) => ModelRequest['model'];
+
+/**
+ * Optional per-role execution override (Strategy-pattern seam, same family
+ * as ProviderResolver). Return undefined to mean "call the real handler"
+ * (default for every role). Return an AIMessage to short-circuit the real
+ * network call — used ONLY to route around a known, disclosed environmental
+ * gap (e.g. missing credential), never silently.
+ *
+ * IMPORTANT: a caller driving a tool-calling loop from a stub MUST close
+ * over call-count state and script a tool_calls-bearing AIMessage on the
+ * turn that should trigger a tool, and a plain AIMessage on the turn that
+ * should end the loop — same shape as FakeToolCallingModel's scripted
+ * toolCalls array, just applied here (a stateless single-canned-response
+ * stub would never actually invoke a tool, so a nested subagent leg would
+ * never fire).
+ */
+export type ExecuteOverride = (
+  role: ProviderRole,
+  request: ModelRequest,
+  model: ModelRequest['model'],
+) => Promise<AIMessage | undefined> | AIMessage | undefined;
+
+export interface CreateProviderRoutingMiddlewareOptions {
+  resolveModel: ProviderResolver;
+  executeOverride?: ExecuteOverride;
+  /** role read when request.runtime.context is missing/has no `.role`. */
+  defaultRole?: ProviderRole;
+}
+
+export interface ProviderRoutingMiddleware {
+  middleware: AgentMiddleware;
+  log: ProviderRoutingRecord[];
+}
+
+/**
+ * Reusable primitive: a real wrapModelCall middleware that selects a model
+ * per-request based on request.runtime.context.role, and records every
+ * selection (and whether it actually executed) to an inspectable log.
+ * A7 (LIA-400) imports this directly for its cross-provider benchmark
+ * matrix — no reimplementation needed.
+ */
+export function createProviderRoutingMiddleware(
+  options: CreateProviderRoutingMiddlewareOptions,
+): ProviderRoutingMiddleware {
+  const log: ProviderRoutingRecord[] = [];
+  let requestIndex = 0;
+
+  const middleware = createMiddleware({
+    name: 'ProviderRouting',
+    wrapModelCall: async (request, handler) => {
+      const role =
+        (request.runtime.context as { role?: ProviderRole } | undefined)
+          ?.role ??
+        options.defaultRole ??
+        'unknown';
+      const model = options.resolveModel(role, request);
+      const index = requestIndex++;
+
+      const override = await options.executeOverride?.(role, request, model);
+      if (override) {
+        log.push({
+          requestIndex: index,
+          role,
+          providerClass: model.constructor.name,
+          modelId: (model as { model?: string }).model,
+          executed: false,
+        });
+        return override;
+      }
+
+      const result = await handler({ ...request, model });
+      log.push({
+        requestIndex: index,
+        role,
+        providerClass: model.constructor.name,
+        modelId: (model as { model?: string }).model,
+        executed: true,
+      });
+      return result;
+    },
+  });
+
+  return { middleware, log };
+}
+
+/**
+ * Default resolver for THIS spike's two roles — 'main' -> ChatAnthropic,
+ * 'sub' -> ChatOllama. Constructs a NEW instance per call. A7 will likely
+ * bring its own resolver covering more providers, so this is documented as
+ * an example, not baked into createProviderRoutingMiddleware itself.
+ */
+export function defaultMainSubResolver(): ProviderResolver {
+  return (role, request) => {
+    if (role === 'main') {
+      const apiKey = process.env.ANTHROPIC_API_KEY ?? ANTHROPIC_PLACEHOLDER_KEY;
+      return new ChatAnthropic({ model: CLAUDE_MODEL_ID, apiKey });
+    }
+    if (role === 'sub') {
+      return new ChatOllama({
+        model: OLLAMA_SUB_MODEL_ID,
+        baseUrl: OLLAMA_BASE_URL,
+      });
+    }
+    return request.model; // passthrough for unrecognized roles
+  };
+}
+
+/**
+ * Builds a stateful executeOverride that short-circuits ONLY the 'main'
+ * role's execution, scripting a tool_calls-bearing AIMessage on its first
+ * invocation (so the ReAct loop actually invokes delegate_to_subagent) and
+ * a plain final AIMessage on every subsequent invocation. The 'sub' role
+ * always returns undefined (no override — real handler, real Ollama call).
+ */
+export function makeScriptedMainStub(): ExecuteOverride {
+  let mainCallCount = 0;
+  return (role) => {
+    if (role !== 'main') return undefined;
+    mainCallCount += 1;
+    if (mainCallCount === 1) {
+      return new AIMessage({
+        content: '',
+        tool_calls: [
+          {
+            name: 'delegate_to_subagent',
+            args: { task: 'What is your favorite fruit? Answer in one word.' },
+            id: 'call_1',
+            type: 'tool_call',
+          },
+        ],
+      });
+    }
+    return new AIMessage({
+      content: 'Delegated to the subagent and got an answer back.',
+    });
+  };
+}
+
+/**
+ * Builds the delegate_to_subagent tool: its execute() nests a SECOND
+ * createAgent + invoke(), sharing the caller-supplied middleware instance,
+ * tagged { role: 'sub' }. This nesting IS "one logical session" (call-stack
+ * continuity) — see the plan's design-question-1 resolution.
+ */
+export function makeDelegateToSubagentTool(
+  sharedMiddleware: AgentMiddleware,
+  placeholderModel: ModelRequest['model'],
+) {
+  return tool(
+    async (args: { task: string }) => {
+      const subAgent = createAgent({
+        model: placeholderModel, // overridden by sharedMiddleware every call
+        middleware: [sharedMiddleware],
+      });
+      const subResult = await subAgent.invoke(
+        { messages: [{ role: 'user', content: args.task }] },
+        { context: { role: 'sub' } },
+      );
+      const last = subResult.messages.at(-1);
+      return JSON.stringify({ subagent_answer: last?.content ?? '' });
+    },
+    {
+      name: 'delegate_to_subagent',
+      description: 'Delegates a sub-task to a local subagent model.',
+      schema: {
+        type: 'object',
+        properties: { task: { type: 'string' } },
+        required: ['task'],
+        additionalProperties: false,
+      },
+    },
+  );
+}
+
+// ── Direct-execution smoke demo ─────────────────────────────────────────
+
+async function runOneSessionDemo(): Promise<void> {
+  const { middleware: sharedMiddleware, log } = createProviderRoutingMiddleware(
+    {
+      resolveModel: defaultMainSubResolver(),
+      executeOverride: makeScriptedMainStub(),
+    },
+  );
+  const placeholderModel = new ChatOllama({
+    model: OLLAMA_SUB_MODEL_ID,
+    baseUrl: OLLAMA_BASE_URL,
+  });
+  const delegateTool = makeDelegateToSubagentTool(
+    sharedMiddleware,
+    placeholderModel,
+  );
+
+  const mainAgent = createAgent({
+    model: placeholderModel, // overridden by sharedMiddleware every call
+    tools: [delegateTool],
+    middleware: [sharedMiddleware],
+  });
+
+  const result = await mainAgent.invoke(
+    {
+      messages: [
+        {
+          role: 'user',
+          content:
+            'Delegate to your subagent to find out its favorite fruit, then report back.',
+        },
+      ],
+    },
+    { context: { role: 'main' } },
+  );
+
+  console.log('\n── Transcript ──');
+  for (const message of result.messages) {
+    const name = (message as { name?: string }).name;
+    console.log(
+      `[${message.getType()}${name ? `:${name}` : ''}] ${String(message.content)}`,
+    );
+  }
+
+  console.log('\n── Routing log ──');
+  for (const record of log) {
+    console.log(
+      `#${record.requestIndex} role=${record.role} provider=${record.providerClass} model=${record.modelId} executed=${record.executed}`,
+    );
+  }
+}
+
+async function main(): Promise<void> {
+  await runOneSessionDemo();
+}
+
+const invokedDirectly =
+  process.argv[1] !== undefined &&
+  fileURLToPath(import.meta.url) === path.resolve(process.argv[1]);
+if (invokedDirectly) {
+  main().catch((err) => {
+    console.error('A3 spike failed:', err instanceof Error ? err.message : err);
+    process.exit(1);
+  });
+}
diff --git a/vitest.config.ts b/vitest.config.ts
index 46ba2703..2ada0c2b 100644
--- a/vitest.config.ts
+++ b/vitest.config.ts
@@ -2,7 +2,11 @@ import { defineConfig } from 'vitest/config';
 
 export default defineConfig({
   test: {
-    include: ['src/**/*.test.ts', 'setup/**/*.test.ts'],
+    include: [
+      'src/**/*.test.ts',
+      'setup/**/*.test.ts',
+      'scripts/spikes/**/*.test.ts',
+    ],
     coverage: {
       provider: 'v8',
       reporter: ['text', 'text-summary', 'lcov'],