yaml-flow 2.4.0 → 2.5.0

package/dist/index.cjs

@@ -1,5 +1,7 @@
 'use strict';
 
+var child_process = require('child_process');
+
 // src/step-machine/reducer.ts
 function applyStepResult(flow, state, stepName, stepResult) {
   const stepConfig = flow.steps[stepName];
@@ -3029,6 +3031,257 @@ function getDownstream(live, nodeName) {
   return { nodeName, nodes, tokens: [...tokenSet] };
 }
 
+// src/inference/core.ts
+var DEFAULT_THRESHOLD = 0.5;
+var DEFAULT_SYSTEM_PROMPT = `You are a workflow completion analyzer. Given a graph of tasks with their current states, evidence, and inference hints, determine which tasks appear to be completed based on the available evidence.
+
+For each task you analyze, provide a JSON response. Be conservative \u2014 only mark tasks as completed when the evidence strongly supports it.`;
+function buildInferencePrompt(live, options = {}) {
+  const { scope, context, systemPrompt } = options;
+  const graphTasks = getAllTasks(live.config);
+  const { state } = live;
+  const candidates = getAnalyzableCandidates(live, scope);
+  if (candidates.length === 0) {
+    return "";
+  }
+  const lines = [];
+  lines.push(systemPrompt || DEFAULT_SYSTEM_PROMPT);
+  lines.push("");
+  lines.push("## Graph State");
+  lines.push("");
+  lines.push(`Available tokens: ${state.availableOutputs.length > 0 ? state.availableOutputs.join(", ") : "(none)"}`);
+  lines.push("");
+  const completedTasks = Object.entries(state.tasks).filter(([_, ts]) => ts.status === "completed").map(([name]) => name);
+  if (completedTasks.length > 0) {
+    lines.push(`Completed tasks: ${completedTasks.join(", ")}`);
+    lines.push("");
+  }
+  lines.push("## Tasks to Analyze");
+  lines.push("");
+  for (const taskName of candidates) {
+    const taskConfig = graphTasks[taskName];
+    const taskState = state.tasks[taskName];
+    lines.push(`### ${taskName}`);
+    if (taskConfig.description) {
+      lines.push(`Description: ${taskConfig.description}`);
+    }
+    const requires = getRequires(taskConfig);
+    const provides = getProvides(taskConfig);
+    if (requires.length > 0) lines.push(`Requires: ${requires.join(", ")}`);
+    if (provides.length > 0) lines.push(`Provides: ${provides.join(", ")}`);
+    lines.push(`Current status: ${taskState?.status || "not-started"}`);
+    const hints = taskConfig.inference;
+    if (hints) {
+      if (hints.criteria) lines.push(`Completion criteria: ${hints.criteria}`);
+      if (hints.keywords?.length) lines.push(`Keywords: ${hints.keywords.join(", ")}`);
+      if (hints.suggestedChecks?.length) lines.push(`Suggested checks: ${hints.suggestedChecks.join("; ")}`);
+    }
+    lines.push("");
+  }
+  if (context) {
+    lines.push("## Additional Context / Evidence");
+    lines.push("");
+    lines.push(context);
+    lines.push("");
+  }
+  lines.push("## Response Format");
+  lines.push("");
+  lines.push("Respond with a JSON array of objects, one per task you have evidence for:");
+  lines.push("```json");
+  lines.push("[");
+  lines.push("  {");
+  lines.push('    "taskName": "task-name",');
+  lines.push('    "confidence": 0.0 to 1.0,');
+  lines.push('    "reasoning": "explanation of why you believe this task is complete or not"');
+  lines.push("  }");
+  lines.push("]");
+  lines.push("```");
+  lines.push("");
+  lines.push("Rules:");
+  lines.push('- Only include tasks from the "Tasks to Analyze" section');
+  lines.push("- confidence 0.0 = no evidence of completion, 1.0 = certain it is complete");
+  lines.push("- If you have no evidence for a task, omit it from the array");
+  lines.push("- Be conservative \u2014 require clear evidence before high confidence");
+  lines.push("- Respond ONLY with the JSON array, no additional text");
+  return lines.join("\n");
+}
+async function inferCompletions(live, adapter, options = {}) {
+  options.threshold ?? DEFAULT_THRESHOLD;
+  const analyzedNodes = getAnalyzableCandidates(live, options.scope);
+  if (analyzedNodes.length === 0) {
+    return { suggestions: [], promptUsed: "", rawResponse: "", analyzedNodes: [] };
+  }
+  const prompt = buildInferencePrompt(live, options);
+  const rawResponse = await adapter.analyze(prompt);
+  const suggestions = parseInferenceResponse(rawResponse, analyzedNodes);
+  return {
+    suggestions,
+    promptUsed: prompt,
+    rawResponse,
+    analyzedNodes
+  };
+}
+function applyInferences(live, result, threshold = DEFAULT_THRESHOLD) {
+  let current = live;
+  for (const suggestion of result.suggestions) {
+    if (suggestion.confidence < threshold) continue;
+    const taskState = current.state.tasks[suggestion.taskName];
+    if (!taskState) continue;
+    if (taskState.status === "completed" || taskState.status === "running") continue;
+    const now = (/* @__PURE__ */ new Date()).toISOString();
+    current = applyEvent(current, {
+      type: "task-started",
+      taskName: suggestion.taskName,
+      timestamp: now
+    });
+    current = applyEvent(current, {
+      type: "task-completed",
+      taskName: suggestion.taskName,
+      timestamp: now,
+      result: "llm-inferred"
+    });
+  }
+  return current;
+}
+async function inferAndApply(live, adapter, options = {}) {
+  const threshold = options.threshold ?? DEFAULT_THRESHOLD;
+  const inference = await inferCompletions(live, adapter, options);
+  const updated = applyInferences(live, inference, threshold);
+  const applied = inference.suggestions.filter((s) => s.confidence >= threshold);
+  const skipped = inference.suggestions.filter((s) => s.confidence < threshold);
+  return {
+    live: updated,
+    inference,
+    applied,
+    skipped
+  };
+}
+function getAnalyzableCandidates(live, scope) {
+  const graphTasks = getAllTasks(live.config);
+  const { state } = live;
+  const candidates = [];
+  for (const [name, config] of Object.entries(graphTasks)) {
+    const taskState = state.tasks[name];
+    if (taskState?.status === "completed" || taskState?.status === "running") continue;
+    if (scope) {
+      if (scope.includes(name)) candidates.push(name);
+    } else {
+      if (config.inference?.autoDetectable) candidates.push(name);
+    }
+  }
+  return candidates;
+}
+function parseInferenceResponse(rawResponse, validNodes, _threshold) {
+  const validSet = new Set(validNodes);
+  try {
+    const jsonStr = extractJson(rawResponse);
+    if (!jsonStr) return [];
+    const parsed = JSON.parse(jsonStr);
+    if (!Array.isArray(parsed)) return [];
+    const suggestions = [];
+    for (const item of parsed) {
+      if (!item || typeof item !== "object") continue;
+      if (typeof item.taskName !== "string") continue;
+      if (typeof item.confidence !== "number") continue;
+      if (!validSet.has(item.taskName)) continue;
+      const confidence = Math.max(0, Math.min(1, item.confidence));
+      suggestions.push({
+        taskName: item.taskName,
+        confidence,
+        reasoning: typeof item.reasoning === "string" ? item.reasoning : "",
+        detectionMethod: "llm-inferred"
+      });
+    }
+    return suggestions;
+  } catch {
+    return [];
+  }
+}
+function extractJson(text) {
+  if (!text || typeof text !== "string") return null;
+  const trimmed = text.trim();
+  const fenceMatch = trimmed.match(/```(?:json)?\s*\n?([\s\S]*?)\n?\s*```/);
+  if (fenceMatch) return fenceMatch[1].trim();
+  const firstBracket = trimmed.indexOf("[");
+  const lastBracket = trimmed.lastIndexOf("]");
+  if (firstBracket !== -1 && lastBracket > firstBracket) {
+    return trimmed.slice(firstBracket, lastBracket + 1);
+  }
+  if (trimmed.startsWith("[")) return trimmed;
+  return null;
+}
+function createCliAdapter(opts) {
+  const timeout = opts.timeout ?? 6e4;
+  return {
+    analyze: (prompt) => {
+      return new Promise((resolve, reject) => {
+        const args = opts.args(prompt);
+        const child = child_process.execFile(
+          opts.command,
+          opts.stdin ? opts.args("") : args,
+          {
+            timeout,
+            cwd: opts.cwd,
+            env: opts.env ? { ...process.env, ...opts.env } : void 0,
+            maxBuffer: 10 * 1024 * 1024
+            // 10MB
+          },
+          (error, stdout, stderr) => {
+            if (error) {
+              reject(new Error(
+                `CLI adapter failed: ${opts.command} exited with ${error.code ?? "error"}` + (stderr ? `
+stderr: ${stderr.slice(0, 500)}` : "") + `
+${error.message}`
+              ));
+            } else {
+              resolve(stdout);
+            }
+          }
+        );
+        if (opts.stdin && child.stdin) {
+          child.stdin.write(prompt);
+          child.stdin.end();
+        }
+      });
+    }
+  };
+}
+function createHttpAdapter(opts) {
+  const timeout = opts.timeout ?? 6e4;
+  return {
+    analyze: async (prompt) => {
+      const body = opts.buildBody ? opts.buildBody(prompt) : { prompt };
+      const controller = new AbortController();
+      const timer = setTimeout(() => controller.abort(), timeout);
+      try {
+        const response = await fetch(opts.url, {
+          method: "POST",
+          headers: {
+            "Content-Type": "application/json",
+            ...opts.headers ?? {}
+          },
+          body: JSON.stringify(body),
+          signal: controller.signal
+        });
+        if (!response.ok) {
+          const text = await response.text().catch(() => "");
+          throw new Error(`HTTP ${response.status}: ${text.slice(0, 500)}`);
+        }
+        const json = await response.json();
+        if (opts.extractResponse) {
+          return opts.extractResponse(json);
+        }
+        if (typeof json.response === "string") return json.response;
+        if (typeof json.text === "string") return json.text;
+        if (typeof json.content === "string") return json.content;
+        return JSON.stringify(json);
+      } finally {
+        clearTimeout(timer);
+      }
+    }
+  };
+}
+
 exports.COMPLETION_STRATEGIES = COMPLETION_STRATEGIES;
 exports.CONFLICT_STRATEGIES = CONFLICT_STRATEGIES;
 exports.DEFAULTS = DEFAULTS;
@@ -3047,13 +3300,17 @@ exports.addRequires = addRequires;
 exports.apply = apply;
 exports.applyAll = applyAll;
 exports.applyEvent = applyEvent;
+exports.applyInferences = applyInferences;
 exports.applyStepResult = applyStepResult;
 exports.batch = batch;
+exports.buildInferencePrompt = buildInferencePrompt;
 exports.checkCircuitBreaker = checkCircuitBreaker;
 exports.computeAvailableOutputs = computeAvailableOutputs;
 exports.computeStepInput = computeStepInput;
+exports.createCliAdapter = createCliAdapter;
 exports.createDefaultTaskState = createDefaultTaskState;
 exports.createEngine = createStepMachine;
+exports.createHttpAdapter = createHttpAdapter;
 exports.createInitialExecutionState = createInitialExecutionState;
 exports.createInitialState = createInitialState;
 exports.createLiveGraph = createLiveGraph;
@@ -3078,6 +3335,8 @@ exports.getUnreachableTokens = getUnreachableTokens;
 exports.getUpstream = getUpstream;
 exports.graphToMermaid = graphToMermaid;
 exports.hasTask = hasTask;
+exports.inferAndApply = inferAndApply;
+exports.inferCompletions = inferCompletions;
 exports.injectTokens = injectTokens;
 exports.inspect = inspect;
 exports.isExecutionComplete = isExecutionComplete;