@oisincoveney/pipeline 2.3.1 → 2.5.0

package/dist/schedule/passes/candidates.js

@@ -0,0 +1,42 @@
+//#region src/schedule/passes/candidates.ts
+/**
+* PIPE-83.7: best-of-N candidate generation. When config.best_of_n is enabled
+* with n > 1, each agent node whose id carries a configured category (e.g.
+* "green") is expanded into a kind:parallel node holding N candidate children
+* (each a full copy with a fresh id and no inter-candidate deps). The wrapper
+* keeps the original id + upstream needs, so downstream consumers and the
+* PIPE-83.9 selector see a single dependency. Default off / n=1 is identity, so
+* generated schedules and the PIPE-57 goldens are unchanged.
+*/
+function expandBestOfNCandidates(config, artifact) {
+	const bestOfN = config.best_of_n;
+	if (!bestOfN?.enabled || bestOfN.n <= 1) return artifact;
+	return {
+		...artifact,
+		workflows: Object.fromEntries(Object.entries(artifact.workflows).map(([id, workflow]) => [id, {
+			...workflow,
+			nodes: workflow.nodes.flatMap((node) => expandNode(node, bestOfN.categories, bestOfN.n))
+		}]))
+	};
+}
+function expandNode(node, categories, n) {
+	if (node.kind !== "agent" || !categories.some((category) => node.id.includes(category))) return [node];
+	const candidatesId = `${node.id}--candidates`;
+	return [{
+		id: candidatesId,
+		kind: "parallel",
+		nodes: Array.from({ length: n }, (_, index) => ({
+			...node,
+			id: `${node.id}--c${index + 1}`,
+			needs: []
+		})),
+		...node.needs ? { needs: node.needs } : {}
+	}, {
+		builtin: "select-candidate",
+		id: node.id,
+		kind: "builtin",
+		needs: [candidatesId]
+	}];
+}
+//#endregion
+export { expandBestOfNCandidates };

package/dist/schedule/passes/index.js

@@ -1,6 +1,7 @@
 //#region src/schedule/passes/index.ts
 const SCHEDULE_PASS_ORDER = [
 	"coverage",
+	"candidates",
 	"models",
 	"ids",
 	"references"

package/dist/schedule/passes/models.js

@@ -25,12 +25,16 @@ function applyNodeCatalogModelsToParallelNode(node, templates) {
 	};
 }
 function applyNodeCatalogModelsToAgentNode(node, templates) {
-	if (node.models?.length) return node;
 	const template = nodeCatalogTemplateFor(node, templates);
-	return template ? {
+	if (!template) return node;
+	return {
 		...node,
-		models: template.models
-	} : node;
+		category: node.category ?? template.category,
+		models: nodeModelsOrCatalog(node, template)
+	};
+}
+function nodeModelsOrCatalog(node, template) {
+	return node.models?.length ? node.models : template.models;
 }
 function nodeCatalogTemplateFor(node, templates) {
 	return templates[node.id] ?? Object.values(templates).find((candidate) => node.id.includes(candidate.category)) ?? Object.values(templates).find((candidate) => candidate.profile === node.profile);

package/dist/schedule/prompts.js

@@ -44,6 +44,9 @@ function plannerPrompt(entrypointId, task, baseline, config, planningContext) {
 		"Scheduler node catalog:",
 		schedulerCatalogPrompt(config, entrypointId),
 		"",
+		"Token budget:",
+		tokenBudgetPrompt(config),
+		"",
 		"Gate recipes:",
 		"- Prefer preserving valid gates from the baseline workflows instead of recreating them.",
 		"- RED/test coverage may use changed_files gates on test-writing nodes. A changed_files gate must include a changed_files object with allow and/or require_any glob arrays.",
@@ -84,6 +87,18 @@ function plannerRepairPrompt(inputs) {
 		stringify(inputs.baseline)
 	].join("\n");
 }
+function tokenBudgetPrompt(config) {
+	const budget = config.token_budget;
+	const windows = Object.entries(budget.model_context_windows);
+	const fanOut = Object.entries(budget.fan_out_width.by_category);
+	return [
+		`- Keep each node's assembled context under ${budget.max_context_pct}% of its model's context window; prefer the smallest-tier model whose window comfortably holds the node within that cap.`,
+		`- Assume ${budget.default_context_window} tokens of context window for a model with no declared window.`,
+		windows.length > 0 ? `- Known model context windows: ${windows.map(([id, size]) => `${id}=${size}`).join(", ")}.` : void 0,
+		`- Do not exceed the per-category fan-out width (max concurrent same-category nodes). Default width: ${budget.fan_out_width.default}.`,
+		fanOut.length > 0 ? `- Category fan-out caps: ${fanOut.map(([category, width]) => `${category}=${width}`).join(", ")}.` : void 0
+	].filter((line) => Boolean(line)).join("\n");
+}
 function allowedProfilePromptLine(config, id) {
 	const profile = config.profiles[id];
 	const runner = config.runners[profile.runner];

package/dist/token-estimator.js

@@ -0,0 +1,23 @@
+import { getEncoding } from "js-tiktoken";
+//#region src/token-estimator.ts
+/**
+* Token estimation for node sizing. Uses the `o200k_base` BPE as a
+* model-agnostic heuristic — NOT a guarantee of any specific model's tokenizer.
+*
+* This is a cross-model ESTIMATE, not a billing-accurate count: the pipeline
+* routes nodes across OpenAI/Kimi/Qwen models whose exact tokenizers differ (and
+* are not all known here), so treat the value as a sizing heuristic for
+* budget/routing decisions only. For exact counts on Anthropic runners, use the
+* Anthropic `count_tokens` API.
+*/
+let encoder;
+function encoding() {
+	encoder ??= getEncoding("o200k_base");
+	return encoder;
+}
+function estimateTokens(text) {
+	if (text.length === 0) return 0;
+	return encoding().encode(text).length;
+}
+//#endregion
+export { estimateTokens };

package/package.json

@@ -9,6 +9,7 @@
     "execa": "^9.5.2",
     "git-url-parse": "^16.1.0",
     "gray-matter": "^4.0.3",
+    "js-tiktoken": "^1.0.21",
     "jsonc-parser": "^3.3.1",
     "ky": "^2.0.2",
     "micromatch": "^4.0.8",
@@ -120,7 +121,7 @@
     "prepack": "bun run build:cli"
   },
   "type": "module",
-  "version": "2.3.1",
+  "version": "2.5.0",
   "description": "Config-driven multi-agent pipeline runner for repository work",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",