@sanity/ailf 1.0.0 → 2.0.0

package/dist/pipeline/compiler/mode-handlers/literacy/assertions.js

@@ -0,0 +1,118 @@
+/**
+ * Assertion resolution for literacy tasks.
+ *
+ * Handles rubric template resolution, doc-coverage auto-generation,
+ * and baseline assertion filtering.
+ */
+// ---------------------------------------------------------------------------
+// Assertion resolution
+// ---------------------------------------------------------------------------
+export function resolveAssertions(task, options, warnings) {
+    const assertions = [];
+    for (const a of task.assertions ?? []) {
+        if (a.type === "llm-rubric" && "template" in a) {
+            const resolved = resolveTemplatedAssertion(a, options?.rubricConfig, options?.graderProvider, warnings);
+            if (resolved)
+                assertions.push(resolved);
+        }
+        else {
+            assertions.push({
+                type: a.type,
+                ...("value" in a ? { value: a.value } : {}),
+                ...(typeof a.weight === "number"
+                    ? { weight: a.weight }
+                    : {}),
+                ...(a.type === "llm-rubric" && options?.graderProvider
+                    ? { provider: options.graderProvider }
+                    : {}),
+            });
+        }
+    }
+    // Doc-coverage auto-generation
+    if (task.docCoverage) {
+        const docCoverageAssertion = buildDocCoverageAssertion(options?.rubricConfig, options?.graderProvider);
+        if (docCoverageAssertion)
+            assertions.push(docCoverageAssertion);
+    }
+    return assertions;
+}
+// ---------------------------------------------------------------------------
+// Rubric template resolution
+// ---------------------------------------------------------------------------
+function resolveTemplatedAssertion(a, rubricConfig, graderProvider, warnings) {
+    if (!rubricConfig) {
+        warnings.push(`No rubric config — template "${a.template}" cannot be resolved`);
+        return null;
+    }
+    const template = rubricConfig.templates[a.template];
+    if (!template) {
+        warnings.push(`Unknown rubric template: "${a.template}"`);
+        return null;
+    }
+    const scaleText = template.scale.map((s) => `- ${s}`).join("\n");
+    const criteriaText = a.criteria.map((c) => `- ${c}`).join("\n");
+    const rubricValue = `${template.header}\n${scaleText}\n\n` +
+        `${template.criteria_label ?? "Check for:"}\n${criteriaText}\n\n` +
+        `Return ONLY a JSON object: {"score": <number>, "reason": "<explanation>"}`;
+    return {
+        type: "llm-rubric",
+        value: rubricValue,
+        ...(graderProvider ? { provider: graderProvider } : {}),
+        ...(template.dimension
+            ? { metadata: { dimension: template.dimension, maxScore: 100 } }
+            : {}),
+    };
+}
+// ---------------------------------------------------------------------------
+// Doc-coverage assertion
+// ---------------------------------------------------------------------------
+function buildDocCoverageAssertion(rubricConfig, graderProvider) {
+    if (!rubricConfig?.templates["doc-coverage"])
+        return null;
+    const template = rubricConfig.templates["doc-coverage"];
+    const scaleText = template.scale.map((s) => `- ${s}`).join("\n");
+    const rubricValue = `${template.header}\n${scaleText}\n\n` +
+        `Return ONLY a JSON object: {"score": <number>, "reason": "<explanation>"}`;
+    return {
+        type: "llm-rubric",
+        value: rubricValue,
+        ...(graderProvider ? { provider: graderProvider } : {}),
+        ...(template.dimension
+            ? { metadata: { dimension: template.dimension, maxScore: 100 } }
+            : {}),
+    };
+}
+// ---------------------------------------------------------------------------
+// Baseline assertion filtering
+// ---------------------------------------------------------------------------
+/**
+ * Build baseline assertions matching the legacy expand-tasks behavior.
+ *
+ * - "full": all assertions carried over
+ * - "abbreviated": only first llm-rubric with shortened prompt
+ * - "none": no assertions
+ */
+export function buildBaselineAssertions(goldAssertions, rubricMode) {
+    const mode = rubricMode ?? "full";
+    if (mode === "none")
+        return [];
+    if (mode === "full")
+        return [...goldAssertions];
+    // Abbreviated: keep first llm-rubric as summary, skip rest
+    const abbreviated = [];
+    let foundFirst = false;
+    for (const a of goldAssertions) {
+        if (a.type === "llm-rubric") {
+            if (!foundFirst) {
+                foundFirst = true;
+                abbreviated.push({
+                    type: "llm-rubric",
+                    value: "Score task completion from 0 to 100 (same criteria as above).\n" +
+                        'Return ONLY a JSON object: {"score": <number>, "reason": "<explanation>"}',
+                    ...(a.provider ? { provider: a.provider } : {}),
+                });
+            }
+        }
+    }
+    return abbreviated;
+}

package/dist/pipeline/compiler/mode-handlers/literacy/compiler.d.ts

@@ -0,0 +1,14 @@
+/**
+ * Literacy task compilation — core compilation logic.
+ *
+ * Produces the same structure as the legacy expand-tasks.ts path:
+ * - Gold entry with with-docs prompt and canonical doc context
+ * - Baseline entry with without-docs prompt and empty docs
+ * - Rubric assertions with structured dimension metadata
+ */
+import type { LiteracyTaskDefinition } from "../../../../_vendor/ailf-core/index.d.ts";
+import type { LiteracyCompileOptions, LiteracyCompileResult } from "./types.js";
+/**
+ * Compile a literacy task into Promptfoo configuration.
+ */
+export declare function compileLiteracyTask(task: LiteracyTaskDefinition, options?: LiteracyCompileOptions): LiteracyCompileResult;

package/dist/pipeline/compiler/mode-handlers/literacy/compiler.js

@@ -0,0 +1,105 @@
+/**
+ * Literacy task compilation — core compilation logic.
+ *
+ * Produces the same structure as the legacy expand-tasks.ts path:
+ * - Gold entry with with-docs prompt and canonical doc context
+ * - Baseline entry with without-docs prompt and empty docs
+ * - Rubric assertions with structured dimension metadata
+ */
+import { LiteracyVariant, } from "../../../normalize-mode.js";
+import { buildBaselineAssertions, resolveAssertions } from "./assertions.js";
+import { LITERACY_PROMPT_TEMPLATES } from "./prompts.js";
+import { validateLiteracyTask } from "./validation.js";
+/**
+ * Compile a literacy task into Promptfoo configuration.
+ */
+export function compileLiteracyTask(task, options) {
+    const warnings = [];
+    const evalMode = options?.evalMode ?? LiteracyVariant.STANDARD;
+    // Validation
+    for (const err of validateLiteracyTask(task)) {
+        warnings.push(`Literacy task "${task.id}": ${err.field} — ${err.message}`);
+    }
+    const providers = buildProviders(options);
+    const prompts = buildPrompts(evalMode);
+    const tests = buildTestCases(task, evalMode, options, warnings);
+    return { providers, tests, prompts, warnings };
+}
+// ---------------------------------------------------------------------------
+// Provider assembly
+// ---------------------------------------------------------------------------
+function buildProviders(options) {
+    if (options?.models && options.models.length > 0) {
+        return options.models.map((m) => ({
+            id: m.id,
+            label: m.label,
+            config: m.config,
+        }));
+    }
+    return [];
+}
+// ---------------------------------------------------------------------------
+// Prompt assembly
+// ---------------------------------------------------------------------------
+function templateToPromptfoo(pt) {
+    return { id: pt.id, label: pt.label, raw: pt.template };
+}
+function buildPrompts(evalMode) {
+    if (evalMode === "agentic") {
+        return [templateToPromptfoo(LITERACY_PROMPT_TEMPLATES["agentic"])];
+    }
+    return [
+        templateToPromptfoo(LITERACY_PROMPT_TEMPLATES["with-docs"]),
+        templateToPromptfoo(LITERACY_PROMPT_TEMPLATES["without-docs"]),
+    ];
+}
+// ---------------------------------------------------------------------------
+// Test case assembly
+// ---------------------------------------------------------------------------
+function buildTestCases(task, evalMode, options, warnings) {
+    const tests = [];
+    const promptText = task.prompt?.text ?? task.prompt?.template ?? "";
+    const contextDocs = task.context?.docs ?? [];
+    const taskArea = task.area ?? "";
+    const taskTitle = task.title;
+    const promptVars = task.prompt?.vars ?? {};
+    const hasDocs = contextDocs.length > 0;
+    const docsVar = hasDocs ? `file://contexts/canonical/${task.id}.md` : "";
+    const assertions = resolveAssertions(task, options, warnings);
+    // Gold entry — canonical docs injected
+    const goldVars = {
+        task: promptText,
+        docs: docsVar,
+        __featureArea: taskArea,
+        ...promptVars,
+    };
+    tests.push({
+        description: `${taskTitle} (gold)`,
+        vars: goldVars,
+        ...(evalMode === LiteracyVariant.STANDARD
+            ? { prompts: ["with-docs"] }
+            : {}),
+        ...(assertions.length > 0 ? { assert: assertions } : {}),
+    });
+    // Baseline entry — no docs (floor measurement)
+    if (evalMode !== "agentic") {
+        const baselineEnabled = task.baseline?.enabled !== false;
+        if (baselineEnabled) {
+            const baselineAssertions = buildBaselineAssertions(assertions, task.baseline?.rubric);
+            tests.push({
+                description: `${taskTitle} (baseline)`,
+                vars: {
+                    task: promptText,
+                    docs: "",
+                    __featureArea: taskArea,
+                    ...promptVars,
+                },
+                prompts: ["without-docs"],
+                ...(baselineAssertions.length > 0
+                    ? { assert: baselineAssertions }
+                    : {}),
+            });
+        }
+    }
+    return tests;
+}

package/dist/pipeline/compiler/mode-handlers/literacy/index.d.ts

@@ -0,0 +1,11 @@
+/**
+ * Literacy mode handler — compiles LiteracyTaskDefinition into Promptfoo config.
+ *
+ * @see docs/exec-plans/architecture-overhaul/phase-7-migrate-literacy.md
+ */
+import type { ModeHandler } from "../../../../_vendor/ailf-core/index.d.ts";
+export { LITERACY_PROMPT_TEMPLATES } from "./prompts.js";
+export { validateLiteracyTask, type LiteracyValidationError, } from "./validation.js";
+export { compileLiteracyTask } from "./compiler.js";
+export type { LiteracyCompileOptions, LiteracyCompileResult, RubricConfig, } from "./types.js";
+export declare const handler: ModeHandler;

package/dist/pipeline/compiler/mode-handlers/literacy/index.js

@@ -0,0 +1,38 @@
+/**
+ * Literacy mode handler — compiles LiteracyTaskDefinition into Promptfoo config.
+ *
+ * @see docs/exec-plans/architecture-overhaul/phase-7-migrate-literacy.md
+ */
+import { compileLiteracyTask } from "./compiler.js";
+import { LITERACY_PROMPT_TEMPLATES } from "./prompts.js";
+// Re-export public API
+export { LITERACY_PROMPT_TEMPLATES } from "./prompts.js";
+export { validateLiteracyTask, } from "./validation.js";
+export { compileLiteracyTask } from "./compiler.js";
+// ---------------------------------------------------------------------------
+// ModeHandler adapter — wraps compileLiteracyTask for registry dispatch
+// ---------------------------------------------------------------------------
+export const handler = {
+    getPrompts() {
+        return LITERACY_PROMPT_TEMPLATES;
+    },
+    compileTask(task, ctx) {
+        if (task.mode !== "literacy") {
+            throw new Error(`Literacy handler received task with mode "${task.mode}" — expected "literacy"`);
+        }
+        const result = compileLiteracyTask(task, {
+            graderProvider: ctx.graderProvider,
+            rootDir: ctx.rootDir,
+            models: ctx.models,
+            rubricConfig: ctx.rubricConfig,
+            evalMode: ctx
+                .evalMode,
+        });
+        return {
+            providers: result.providers,
+            tests: result.tests,
+            prompts: result.prompts,
+            warnings: result.warnings,
+        };
+    },
+};

package/dist/pipeline/compiler/mode-handlers/literacy/prompts.d.ts

@@ -0,0 +1,9 @@
+/**
+ * Canonical prompt templates for literacy-mode evaluations.
+ *
+ * These are the source-of-truth templates. Previously lived in
+ * config/prompts.ts as global templates; now handler-owned so
+ * non-literacy modes can define their own prompts without collision.
+ */
+import type { PromptTemplate } from "../../../../_vendor/ailf-core/index.d.ts";
+export declare const LITERACY_PROMPT_TEMPLATES: Record<string, PromptTemplate>;

package/dist/pipeline/compiler/mode-handlers/literacy/prompts.js

@@ -0,0 +1,74 @@
+/**
+ * Canonical prompt templates for literacy-mode evaluations.
+ *
+ * These are the source-of-truth templates. Previously lived in
+ * config/prompts.ts as global templates; now handler-owned so
+ * non-literacy modes can define their own prompts without collision.
+ */
+export const LITERACY_PROMPT_TEMPLATES = {
+    "with-docs": {
+        id: "with-docs",
+        label: "With Documentation",
+        template: `You are an expert Sanity.io developer. Use the following documentation to help implement the task.
+
+## Sanity Documentation
+{{docs}}
+
+## Task
+{{task}}
+
+## Requirements
+
+1. Use ONLY the APIs and patterns shown in the documentation
+2. Provide a complete, working implementation
+3. Include all necessary imports
+4. Follow Sanity best practices as documented
+
+Provide your implementation:
+`,
+        variables: ["docs", "task"],
+    },
+    "without-docs": {
+        id: "without-docs",
+        label: "Baseline (No Docs)",
+        template: `You are an expert Sanity.io developer.
+
+## Task
+{{task}}
+
+## Requirements
+
+1. Provide a complete, working implementation
+2. Include all necessary imports
+3. Follow Sanity best practices
+
+Provide your implementation:
+`,
+        variables: ["task"],
+    },
+    agentic: {
+        id: "agentic",
+        label: "Agentic (self-retrieval)",
+        template: `You are an expert developer helping implement a Sanity.io feature.
+You have access to web search and page fetching tools.
+
+IMPORTANT: Before writing any code, search for and read the relevant
+Sanity.io documentation to ensure you are using the latest APIs and
+best practices. Do not rely on memory alone.
+
+## Task
+{{task}}
+
+## Requirements
+
+1. Search for relevant Sanity documentation before implementing
+2. Use ONLY the APIs and patterns from the current official docs
+3. Provide a complete, working implementation
+4. Include all necessary imports
+5. Follow Sanity best practices as documented
+
+Provide your implementation:
+`,
+        variables: ["task"],
+    },
+};

package/dist/pipeline/compiler/mode-handlers/literacy/types.d.ts

@@ -0,0 +1,41 @@
+/**
+ * Shared types for the literacy mode handler.
+ */
+import type { PromptfooPrompt, PromptfooProvider, PromptfooTestCase } from "../../promptfoo-compiler.js";
+/** Options for compiling a literacy task */
+export interface LiteracyCompileOptions {
+    /** Grader provider for LLM-graded assertions */
+    graderProvider?: string;
+    /** Root directory (for resolving file:// doc paths) */
+    rootDir?: string;
+    /** Evaluation sub-mode — controls which entries are generated */
+    evalMode?: import("../../../normalize-mode.js").LiteracyEvalSubMode;
+    /** Model providers to include */
+    models?: {
+        id: string;
+        label: string;
+        config?: Record<string, unknown>;
+    }[];
+    /** Rubric config (templates, weights, profiles) — loaded from rubrics config */
+    rubricConfig?: RubricConfig;
+}
+/** Minimal rubric config needed by the handler */
+export interface RubricConfig {
+    templates: Record<string, {
+        dimension?: string;
+        header: string;
+        scale: string[];
+        criteria_label?: string;
+    }>;
+}
+/** Result of compiling a single literacy task */
+export interface LiteracyCompileResult {
+    /** Promptfoo provider configs */
+    providers: PromptfooProvider[];
+    /** Compiled test cases (gold + optional baseline) */
+    tests: PromptfooTestCase[];
+    /** Prompts for evaluation */
+    prompts: PromptfooPrompt[];
+    /** Warnings generated during compilation */
+    warnings: string[];
+}

package/dist/pipeline/compiler/mode-handlers/literacy/types.js

@@ -0,0 +1,4 @@
+/**
+ * Shared types for the literacy mode handler.
+ */
+export {};

package/dist/pipeline/compiler/mode-handlers/literacy/validation.d.ts

@@ -0,0 +1,12 @@
+/**
+ * Validation for literacy task definitions.
+ */
+import type { LiteracyTaskDefinition } from "../../../../_vendor/ailf-core/index.d.ts";
+export interface LiteracyValidationError {
+    field: string;
+    message: string;
+}
+/**
+ * Validate a literacy task definition.
+ */
+export declare function validateLiteracyTask(task: LiteracyTaskDefinition): LiteracyValidationError[];

package/dist/pipeline/compiler/mode-handlers/literacy/validation.js

@@ -0,0 +1,28 @@
+/**
+ * Validation for literacy task definitions.
+ */
+/**
+ * Validate a literacy task definition.
+ */
+export function validateLiteracyTask(task) {
+    const errors = [];
+    if (!task.id)
+        errors.push({ field: "id", message: "Task ID is required" });
+    if (!task.title) {
+        errors.push({
+            field: "title",
+            message: "Task title is required",
+        });
+    }
+    const promptText = task.prompt?.text ??
+        task.prompt?.template ??
+        task.prompt?.vars?.task ??
+        "";
+    if (!promptText) {
+        errors.push({
+            field: "prompt",
+            message: "Task prompt text is required",
+        });
+    }
+    return errors;
+}

package/dist/pipeline/compiler/mode-handlers/{mcp-assertions.d.ts → mcp-server/assertions.d.ts}

@@ -18,16 +18,8 @@
  *
  * @see docs/exec-plans/architecture-overhaul/phase-3-mcp-server-mode.md
  */
-import type { PromptfooAssertion } from "../assertion-mapper.js";
-/** Context for building MCP assertions */
-export interface MCPAssertionContext {
-    /** Task ID (for error messages) */
-    taskId: string;
-    /** Expected server capabilities */
-    capabilities: string[];
-    /** Grader provider for LLM-graded assertions */
-    graderProvider?: string;
-}
+import type { PromptfooAssertion } from "../../assertion-mapper.js";
+import type { MCPAssertionContext } from "./types.js";
 /** An AILF assertion definition — accepts both core and generalized types */
 interface AssertionInput {
     type: string;

package/dist/pipeline/compiler/mode-handlers/{mcp-assertions.js → mcp-server/assertions.js}

@@ -83,18 +83,72 @@ function mapMCPAssertion(assertion, context, warnings) {
 // ---------------------------------------------------------------------------
 function buildToolCalledAssertion(assertion, _context) {
     const toolName = String(assertion.value ?? "");
+    // Strategy: check multiple sources for tool call evidence.
+    // 1. context.vars.__toolCalls (structured, if Promptfoo populates it)
+    // 2. Response metadata toolCallLog (from custom mcp-tool-provider)
+    // 3. Response output text (LLM+MCP providers embed tool_use JSON blocks)
     return {
         type: "javascript",
         value: buildJsAssertion(`tool-called: ${toolName}`, `
-      const toolCalls = context.vars.__toolCalls || [];
-      const called = toolCalls.some(tc => tc.name === ${JSON.stringify(toolName)});
+      var toolName = ${JSON.stringify(toolName)};
+
+      // Strategy 1: structured tool calls from Promptfoo
+      var toolCalls = context.vars.__toolCalls || [];
+      if (Array.isArray(toolCalls) && toolCalls.length > 0) {
+        var called = toolCalls.some(function(tc) { return tc.name === toolName; });
+        return {
+          pass: called,
+          score: called ? 1 : 0,
+          reason: called
+            ? 'Tool "' + toolName + '" was called (via __toolCalls)'
+            : 'Expected tool "' + toolName + '" but found: ' + toolCalls.map(function(tc) { return tc.name; }).join(', '),
+        };
+      }
+
+      // Strategy 2: MCP_TOOLS_CALLED summary appended by custom mcp-tool-provider
+      var outputStr = typeof output === 'string' ? output : JSON.stringify(output || '');
+      var summaryMatch = outputStr.match(/<!-- MCP_TOOLS_CALLED: (\\[.*?\\]) -->/);
+      if (summaryMatch) {
+        try {
+          var calledTools = JSON.parse(summaryMatch[1]);
+          var called = calledTools.includes(toolName);
+          var count = calledTools.filter(function(n) { return n === toolName; }).length;
+          return {
+            pass: called,
+            score: called ? 1 : 0,
+            reason: called
+              ? 'Tool "' + toolName + '" was called ' + count + ' time(s)'
+              : 'Expected tool "' + toolName + '" but found: ' + calledTools.join(', '),
+          };
+        } catch (e) { /* fall through to Strategy 3 */ }
+      }
+
+      // Strategy 3: parse output for tool_use blocks (built-in provider fallback)
+      var outputStr = typeof output === 'string' ? output : JSON.stringify(output || '');
+      var toolUsePattern = /"type"\\s*:\\s*"tool_use"[^}]*"name"\\s*:\\s*"([^"]+)"/g;
+      var foundTools = [];
+      var match;
+      while ((match = toolUsePattern.exec(outputStr)) !== null) {
+        foundTools.push(match[1]);
+      }
+      var fnCallPattern = /"function"\\s*:\\s*\\{[^}]*"name"\\s*:\\s*"([^"]+)"/g;
+      while ((match = fnCallPattern.exec(outputStr)) !== null) {
+        foundTools.push(match[1]);
+      }
+      if (foundTools.length === 0 && outputStr.includes(toolName) && outputStr.includes('tool_use')) {
+        foundTools.push(toolName);
+      }
+
+      var called = foundTools.includes(toolName);
       return {
         pass: called,
         score: called ? 1 : 0,
         reason: called
-          ? 'Tool ' + ${JSON.stringify(JSON.stringify(toolName))} + ' was called as expected'
-          : 'Expected tool ' + ${JSON.stringify(JSON.stringify(toolName))} + ' to be called, but it was not. ' +
-            'Tools called: ' + (toolCalls.map(tc => tc.name).join(', ') || 'none'),
+          ? 'Tool "' + toolName + '" was called (detected in output)'
+          : 'Expected tool "' + toolName + '" to be called. ' +
+            (foundTools.length > 0
+              ? 'Tools found in output: ' + foundTools.join(', ')
+              : 'No tool calls detected in output'),
       };`),
         ...(assertion.weight !== undefined ? { weight: assertion.weight } : {}),
     };
@@ -273,5 +327,8 @@ function buildCapabilityAssertion(assertion, _context) {
  * from Promptfoo's assertion runner.
  */
 function buildJsAssertion(label, body) {
-    return `// MCP assertion: ${label}\n(function() {\n${body.trim()}\n})()`;
+    // No IIFE wrapper — Promptfoo wraps the assertion in its own function via
+    // new Function('output', 'context', ...). The body must use `return` at
+    // the top level for the result to reach Promptfoo's validator.
+    return `// MCP assertion: ${label}\n${body.trim()}`;
 }

package/dist/pipeline/compiler/mode-handlers/mcp-server/compiler.d.ts

@@ -0,0 +1,19 @@
+/**
+ * MCP server task compilation — core compiler logic.
+ *
+ * Produces Promptfoo configuration from MCP server task definitions:
+ * 1. A provider config pointing to the MCP server
+ * 2. Test cases with tool-call assertions
+ * 3. Appropriate prompts for the evaluation
+ */
+import type { MCPServerTaskDefinition } from "../../../../_vendor/ailf-core/index.d.ts";
+import type { MCPCompileOptions, MCPCompileResult } from "./types.js";
+/**
+ * Compile an MCP server task definition into Promptfoo configuration.
+ *
+ * This is the core of the MCP mode handler. It produces:
+ * 1. A provider config pointing to the MCP server
+ * 2. Test cases with tool-call assertions
+ * 3. Appropriate prompts for the evaluation
+ */
+export declare function compileMCPTask(task: MCPServerTaskDefinition, options?: MCPCompileOptions): MCPCompileResult;