@seanhogg/builderforce-memory 2026.6.34 → 2026.6.39

package/src/workflow/steps.ts

@@ -16,13 +16,19 @@ import {
   BPETokenizer,
   EvermindModelPackage,
   exportEvermind,
+  importEvermind,
   benchmarkText,
   type ExportFormat,
   type BenchmarkReport,
+  type ImportOptions,
 } from "@seanhogg/builderforce-memory-engine";
 import { EvermindCognition } from "../cognition/index.js";
 import { hybridRetrieve, chunkText } from "../retrieval/index.js";
 import { MemoryStore } from "../memory/MemoryStore.js";
+import { OpenAIBridge } from "../bridges/OpenAIBridge.js";
+import { AnthropicBridge } from "../bridges/AnthropicBridge.js";
+import type { TransformerBridge } from "../bridges/TransformerBridge.js";
+import { analyzeCode, runJsCases, type CodeCase } from "./code-eval.js";
 import type { StackStep } from "../diagnostics/stack-diagnostic.js";
 import type { StepFactory, StepTypeInfo, WorkflowStepConfig } from "./types.js";
 
@@ -47,20 +53,185 @@ function pBool(cfg: WorkflowStepConfig, key: string, def: boolean): boolean {
   const v = cfg.params?.[key];
   return typeof v === "boolean" ? v : def;
 }
+/** Read a string[] param (non-strings dropped). */
+function pStrArr(cfg: WorkflowStepConfig, key: string): string[] {
+  const v = cfg.params?.[key];
+  return Array.isArray(v) ? v.filter((x): x is string => typeof x === "string") : [];
+}
+
+/** Default teacher instruction for code distillation — code only, no prose/fences. */
+const CODE_TEACHER_SYSTEM =
+  "You are an expert programmer. Respond with correct, idiomatic code that fulfils the request. " +
+  "Output code only — no explanation, no markdown fences.";
+
+/**
+ * A (prompt → teacher completion) training exemplar produced by distillation.
+ * `cases` (when present) execution-verify the completion; `teacher` records
+ * which panel member produced the surviving answer (attribution/telemetry).
+ */
+interface DistillPair {
+  prompt: string;
+  completion: string;
+  cases?: CodeCase[];
+  teacher?: string;
+}
+
+/** One coding task to distil: a prompt plus the tests its answer must pass. */
+interface DistillTask {
+  prompt: string;
+  cases?: CodeCase[];
+}
+
+/** A single teacher's answer to a prompt, before selection. */
+interface Candidate {
+  teacher: string;
+  completion: string;
+}
+
+/** A configured teacher in the panel (its display label + bridge). */
+interface Teacher {
+  label: string;
+  bridge: TransformerBridge;
+}
+
+/** One teacher's config — a panel entry, or the legacy single-teacher shape. */
+interface TeacherSpec {
+  provider?: string;
+  apiKey?: string;
+  model?: string;
+  baseUrl?: string;
+  systemPrompt?: string;
+  maxTokens?: number;
+}
+
+/**
+ * Strip a wrapping ```lang … ``` markdown fence some teachers add despite the
+ * "code only, no fences" instruction — otherwise the fence fails every code
+ * gate and the exemplar is wrongly dropped.
+ */
+export function stripCodeFences(text: string): string {
+  const t = text.trim();
+  const m = /^```[\w-]*\r?\n([\s\S]*?)\r?\n```$/.exec(t);
+  return m ? m[1]!.trim() : t;
+}
+
+/** Read `cases` off a param blob, dropping malformed entries. */
+function normalizeCases(raw: unknown): CodeCase[] {
+  return Array.isArray(raw)
+    ? raw.filter((c): c is CodeCase => typeof (c as { call?: unknown })?.call === "string")
+    : [];
+}
+
+/**
+ * Best-of-N selection across a teacher panel. When the task carries execution
+ * `cases`, keep only fully-passing candidates and return the first — teachers
+ * are listed strongest-first (e.g. Opus), so the strongest correct answer wins
+ * and a wrong teacher answer is dropped rather than trained on. With no cases,
+ * take the first non-empty candidate (teacher-order priority).
+ */
+export function selectBestCandidate(candidates: Candidate[], cases: CodeCase[]): Candidate | null {
+  if (candidates.length === 0) return null;
+  if (cases.length === 0) return candidates[0]!;
+  for (const c of candidates) {
+    if (runJsCases(c.completion, cases).passRate === 1) return c;
+  }
+  return null;
+}
+
+/**
+ * Build one teacher bridge from its spec. Any external LLM can teach Evermind:
+ *   • provider "openai" (default) — the BuilderForce gateway, OpenRouter, OpenAI,
+ *     or any OpenAI-compatible server (Ollama / vLLM / LM Studio), by `baseUrl`.
+ *   • provider "anthropic" — the Anthropic Messages API directly (Opus etc).
+ * Returns null when the entry lacks the credentials it needs.
+ */
+function buildTeacher(spec: TeacherSpec, defSystem: string, defMaxTokens: number): Teacher | null {
+  const provider = (spec.provider ?? "openai").toLowerCase();
+  const apiKey = spec.apiKey ?? "";
+  const model = spec.model ?? "";
+  const systemPrompt = spec.systemPrompt ?? defSystem;
+  const maxTokens = spec.maxTokens ?? defMaxTokens;
+  if (provider === "anthropic") {
+    if (!apiKey) return null;
+    return { label: `anthropic:${model || "default"}`, bridge: new AnthropicBridge({ apiKey, systemPrompt, maxTokens, ...(model ? { model } : {}) }) };
+  }
+  // openai-compatible (gateway / OpenRouter / OpenAI / local) — needs a baseUrl.
+  const baseUrl = spec.baseUrl ?? "";
+  if (!baseUrl) return null;
+  return { label: `${provider}:${model || "default"}`, bridge: new OpenAIBridge({ apiKey, baseUrl, systemPrompt, maxTokens, ...(model ? { model } : {}) }) };
+}
+
+/**
+ * Build the teacher panel from step params. `teachers: TeacherSpec[]` runs
+ * several models as a best-of-N panel (list the strongest FIRST — e.g. Opus —
+ * so it wins ties); the legacy single-teacher shape (top-level
+ * provider/baseUrl/apiKey/model) is honoured as a one-entry panel. Entries
+ * without credentials are skipped, so an offline `pairs`-only run yields [].
+ */
+function makeTeachers(cfg: WorkflowStepConfig): Teacher[] {
+  const defSystem = pStr(cfg, "systemPrompt", CODE_TEACHER_SYSTEM);
+  const defMaxTokens = pNum(cfg, "maxTokens", 512);
+  const raw = cfg.params?.teachers;
+  const specs: TeacherSpec[] =
+    Array.isArray(raw) && raw.length > 0
+      ? (raw as TeacherSpec[])
+      : [{ provider: pStr(cfg, "provider", "openai"), apiKey: pStr(cfg, "apiKey", ""), model: pStr(cfg, "model", ""), baseUrl: pStr(cfg, "baseUrl", "") }];
+  const out: Teacher[] = [];
+  for (const s of specs) {
+    const t = buildTeacher(s ?? {}, defSystem, defMaxTokens);
+    if (t) out.push(t);
+  }
+  return out;
+}
+
+/** Read the coding tasks to distil: rich `tasks[{prompt,cases}]` or plain `prompts[]`. */
+function readTasks(cfg: WorkflowStepConfig): DistillTask[] {
+  const raw = cfg.params?.tasks;
+  if (Array.isArray(raw)) {
+    const out: DistillTask[] = [];
+    for (const t of raw) {
+      const prompt = (t as DistillTask)?.prompt;
+      if (typeof prompt !== "string") continue;
+      const cases = normalizeCases((t as DistillTask).cases);
+      out.push(cases.length ? { prompt, cases } : { prompt });
+    }
+    return out;
+  }
+  return pStrArr(cfg, "prompts").map((prompt) => ({ prompt }));
+}
 
 /** Byte length of an emitted export file (text or binary). */
 function fileBytes(data: Uint8Array | string): number {
   return typeof data === "string" ? new TextEncoder().encode(data).length : data.length;
 }
 
+/** Coerce a workflow-param weight blob (bytes, ArrayBuffer, number[], or base64) to Uint8Array. */
+function toBytes(value: unknown, base64?: string): Uint8Array | null {
+  if (value instanceof Uint8Array) return value;
+  if (value instanceof ArrayBuffer) return new Uint8Array(value);
+  if (Array.isArray(value) && value.every((n) => typeof n === "number")) return Uint8Array.from(value as number[]);
+  if (typeof base64 === "string" && base64) {
+    if (typeof atob !== "function") throw new Error("import-model: base64 decoding needs atob (browser/Worker/Node 16+)");
+    const bin = atob(base64);
+    const out = new Uint8Array(bin.length);
+    for (let i = 0; i < bin.length; i++) out[i] = bin.charCodeAt(i);
+    return out;
+  }
+  return null;
+}
+
 function mkStep(cfg: WorkflowStepConfig, layer: string, defaultLabel: string, run: StackStep["run"]): StackStep {
   return { id: cfg.id, layer, label: cfg.label ?? defaultLabel, run };
 }
 
-/** Encode a corpus into next-token training sequences (one per sentence). */
+/**
+ * Encode a corpus into next-token training sequences. Split on sentence
+ * boundaries (prose) OR blank lines (code blocks / distilled prompt→code pairs),
+ * so both natural-language and code corpora yield multiple trainable sequences.
+ */
 function corpusToSequences(corpus: string, tok: BPETokenizer): number[][] {
   return corpus
-    .split(/(?<=\.)\s+/)
+    .split(/(?<=\.)\s+|\n\s*\n/)
     .map((s) => tok.encode(s.trim()))
     .filter((ids) => ids.length >= 2);
 }
@@ -175,16 +346,38 @@ export const BUILTIN_STEPS: Record<string, Registered> = {
 
   // ── LLM-creation pipeline (BUILD) — produces a custom .evermind model ────────
   "train-tokenizer": {
-    info: { type: "train-tokenizer", layer: "BUILD", label: "Train tokenizer", description: "Learn a BPE vocab from your corpus (params: corpus, numMerges)" },
+    info: {
+      type: "train-tokenizer",
+      layer: "BUILD",
+      label: "Train tokenizer",
+      description:
+        "Build a BPE vocab — train fresh from your corpus, OR import merges from a proven code tokenizer (params: corpus, numMerges, hfTokenizer, hfTokenizerUrl)",
+    },
     factory: (cfg) =>
       mkStep(cfg, "BUILD", "Train tokenizer (BPE)", async (ctx) => {
-        const corpus = pStr(cfg, "corpus", KNOWLEDGE.repeat(4));
-        const numMerges = pNum(cfg, "numMerges", 100);
         const tok = new BPETokenizer();
+        // Two paths, both first-class (this is an IDE choice):
+        //   • import-merges: seed from a Hugging Face tokenizer.json (code vocab on day one)
+        //   • train-fresh:   learn a vocab from the corpus
+        const hfTokenizer = cfg.params?.hfTokenizer as Record<string, unknown> | undefined;
+        const hfTokenizerUrl = pStr(cfg, "hfTokenizerUrl", "");
+        if (hfTokenizer && typeof hfTokenizer === "object") {
+          tok.loadHuggingFace(hfTokenizer);
+          ctx.bag.tokenizer = tok;
+          return `imported vocab ${tok.vocabSize} (Hugging Face merges)`;
+        }
+        if (hfTokenizerUrl) {
+          await tok.loadHuggingFaceUrl(hfTokenizerUrl);
+          ctx.bag.tokenizer = tok;
+          return `imported vocab ${tok.vocabSize} from ${hfTokenizerUrl}`;
+        }
+        // train-fresh — default to the distilled corpus from a prior step when present.
+        const corpus = pStr(cfg, "corpus", (ctx.bag.corpus as string) ?? KNOWLEDGE.repeat(4));
+        const numMerges = pNum(cfg, "numMerges", 100);
         tok.train(corpus, { numMerges });
         ctx.bag.tokenizer = tok;
         ctx.bag.corpus = corpus;
-        return `vocab ${tok.vocabSize} (${numMerges} merges)`;
+        return `trained vocab ${tok.vocabSize} (${numMerges} merges)`;
       }),
   },
   "train-model": {
@@ -415,4 +608,172 @@ export const BUILTIN_STEPS: Record<string, Registered> = {
         return `exported ${format}: ${result.files.length} file(s), ${total} bytes, ${result.paramCount} params`;
       }),
   },
+
+  "import-model": {
+    info: {
+      type: "import-model",
+      layer: "BUILD",
+      label: "Import model (warm-start)",
+      description:
+        "Warm-start an EvermindLM from a .safetensors checkpoint instead of random init — round-trips our own exports, or maps a foreign SSM checkpoint via renameMap (params: safetensors, safetensorsBase64, topK, renameMap)",
+    },
+    factory: (cfg) =>
+      mkStep(cfg, "BUILD", "Import model (warm-start)", async (ctx) => {
+        const bytes = toBytes(cfg.params?.safetensors, pStr(cfg, "safetensorsBase64", ""));
+        if (!bytes) {
+          throw new Error("import-model: provide 'safetensors' (bytes/ArrayBuffer/number[]) or 'safetensorsBase64'");
+        }
+        const opts: ImportOptions = {};
+        const topK = pNum(cfg, "topK", 0);
+        if (topK > 0) opts.topK = topK;
+        // renameMap is the JSON-friendly form of ImportOptions.rename — used to
+        // translate a foreign SSM checkpoint's tensor names to our canonical names.
+        const renameMap = cfg.params?.renameMap as Record<string, string> | undefined;
+        if (renameMap && typeof renameMap === "object") {
+          opts.rename = (name) => (name in renameMap ? renameMap[name]! : name);
+        }
+        const model = importEvermind(bytes, opts);
+        ctx.bag.model = model;
+        const params = model.parameters().reduce((n, p) => n + p.data.length, 0);
+        const c = model.config;
+        return `imported ${params} params (dModel ${c.dModel}, ${c.numLayers}L, ${c.numExperts}E) from ${bytes.length} bytes`;
+      }),
+  },
+
+  // ── Teaching Evermind to code — multi-teacher distillation + exec gates ───────
+  "distill-corpus": {
+    info: {
+      type: "distill-corpus",
+      layer: "BUILD",
+      label: "Distill from teacher panel",
+      description:
+        "Build a training corpus from a panel of teacher LLMs (Opus + any other model — the gateway, OpenRouter, OpenAI, a local server, Anthropic). Per task, queries every teacher, keeps only completions that PASS the task's execution cases (best-of-N, strongest teacher first), tags each with its teacher, and ACCUMULATES so stacked steps add up. Feeds the corpus to train-model (params: teachers[{provider,baseUrl,apiKey,model}], tasks[{prompt,cases}], prompts, provider, baseUrl, apiKey, model, systemPrompt, maxTokens, temperature, pairs)",
+    },
+    factory: (cfg) =>
+      mkStep(cfg, "BUILD", "Distill corpus from teacher panel", async (ctx) => {
+        // Accumulate across steps so multiple teacher panels stack instead of
+        // overwriting a prior distill step's exemplars.
+        const pairs: DistillPair[] = Array.isArray(ctx.bag.distillPairs) ? (ctx.bag.distillPairs as DistillPair[]) : [];
+        const startCount = pairs.length;
+        let offlineKept = 0;
+        let verified = 0;
+        let dropped = 0;
+
+        // Offline exemplars (pre-captured) — a single trusted candidate each,
+        // execution-verified when the pair carries `cases`.
+        const offline = cfg.params?.pairs;
+        if (Array.isArray(offline)) {
+          for (const p of offline) {
+            const prompt = (p as DistillPair)?.prompt;
+            const completion = (p as DistillPair)?.completion;
+            if (typeof prompt !== "string" || typeof completion !== "string") continue;
+            const cases = normalizeCases((p as DistillPair).cases);
+            const chosen = selectBestCandidate([{ teacher: "offline", completion: stripCodeFences(completion) }], cases);
+            if (!chosen) { dropped++; continue; }
+            const pair: DistillPair = { prompt, completion: chosen.completion, teacher: "offline" };
+            if (cases.length) pair.cases = cases;
+            pairs.push(pair);
+            offlineKept++;
+          }
+        }
+
+        // Live teacher panel — Opus and/or any other models, best-of-N + exec-filtered.
+        const teachers = makeTeachers(cfg);
+        const tasks = readTasks(cfg);
+        if (teachers.length > 0 && tasks.length > 0) {
+          const temperature = pNum(cfg, "temperature", 0.2);
+          // Dedupe identical (teacher, prompt) calls within this run so a repeated
+          // prompt is not billed twice.
+          const memo = new Map<string, string>();
+          for (const task of tasks) {
+            const candidates: Candidate[] = [];
+            for (const t of teachers) {
+              const key = `${t.label}${task.prompt}`;
+              let completion = memo.get(key);
+              if (completion === undefined) {
+                completion = stripCodeFences(await t.bridge.generate(task.prompt, { temperature }));
+                memo.set(key, completion);
+              }
+              if (completion.trim()) candidates.push({ teacher: t.label, completion });
+            }
+            const chosen = selectBestCandidate(candidates, task.cases ?? []);
+            if (!chosen) { dropped++; continue; }
+            const pair: DistillPair = { prompt: task.prompt, completion: chosen.completion, teacher: chosen.teacher };
+            if (task.cases?.length) pair.cases = task.cases;
+            pairs.push(pair);
+            verified++;
+          }
+        }
+
+        if (pairs.length === 0) {
+          throw new Error(
+            "distill-corpus produced no exemplars — set teachers (provider + apiKey + prompts/tasks; baseUrl for openai-compatible) or supply offline 'pairs'",
+          );
+        }
+
+        const corpus = pairs.map((p) => `${p.prompt}\n${p.completion}`).join("\n\n");
+        ctx.bag.corpus = corpus;
+        ctx.bag.distillPairs = pairs;
+
+        const added = pairs.length - startCount;
+        const parts: string[] = [];
+        if (offlineKept) parts.push(`${offlineKept} offline`);
+        if (verified) parts.push(`${verified} verified via ${teachers.length} teacher(s): ${teachers.map((t) => t.label).join(", ")}`);
+        if (dropped) parts.push(`${dropped} dropped`);
+        return `distilled ${added} exemplar(s) (${pairs.length} total, ${corpus.length} chars)${parts.length ? ` — ${parts.join(", ")}` : ""}`;
+      }),
+  },
+  "code-parse-check": {
+    info: {
+      type: "code-parse-check",
+      layer: "BUILD",
+      label: "Code parse check",
+      description:
+        "Validate generated output is structurally valid code — balanced delimiters/strings, optional JS parse (params: code, language, minScore)",
+    },
+    factory: (cfg) =>
+      mkStep(cfg, "BUILD", "Code structure / parse gate", async (ctx) => {
+        const code = pStr(cfg, "code", (ctx.bag.sample as string) ?? "");
+        if (!code.trim()) throw new Error("no code to check — run 'evaluate'/'generate-check' first or pass a 'code' param");
+        const language = pStr(cfg, "language", "js");
+        const analysis = analyzeCode(code, language);
+        ctx.bag.codeParse = analysis;
+        const minScore = pNum(cfg, "minScore", 0); // only enforced when configured
+        if (minScore > 0 && analysis.score < minScore) {
+          throw new Error(
+            `code structure score ${analysis.score.toFixed(2)} < min ${minScore} (unbalanced delimiters/strings — output is not valid ${language})`,
+          );
+        }
+        return `structure ${(analysis.score * 100) | 0}%${language === "js" ? `, js-parse ${analysis.jsParse ? "ok" : "fail"}` : ""}`;
+      }),
+  },
+  "code-eval": {
+    info: {
+      type: "code-eval",
+      layer: "BUILD",
+      label: "Code eval (reward)",
+      description:
+        "Execution-grounded reward: run the generated JS against test cases and score the pass-rate (params: code, cases, minPassRate)",
+    },
+    factory: (cfg) =>
+      mkStep(cfg, "BUILD", "Code execution reward", async (ctx) => {
+        const code = pStr(cfg, "code", (ctx.bag.sample as string) ?? "");
+        if (!code.trim()) throw new Error("no code to evaluate — run 'evaluate'/'generate-check' first or pass a 'code' param");
+        const rawCases = cfg.params?.cases;
+        const cases = Array.isArray(rawCases)
+          ? rawCases.filter((c): c is { call: string; expect: unknown } => typeof (c as { call?: unknown })?.call === "string")
+          : [];
+        const result = runJsCases(code, cases);
+        ctx.bag.codeEval = result;
+        const minPassRate = pNum(cfg, "minPassRate", 0); // only enforced when configured
+        if (minPassRate > 0 && result.passRate < minPassRate) {
+          throw new Error(
+            `code-eval pass-rate ${(result.passRate * 100) | 0}% < min ${(minPassRate * 100) | 0}%${result.error ? ` (${result.error})` : ""}`,
+          );
+        }
+        return result.total > 0
+          ? `${result.passed}/${result.total} cases passed (${(result.passRate * 100) | 0}%)`
+          : "no cases — execution reward skipped";
+      }),
+  },
 };

package/src/workflow/templates.ts

@@ -69,7 +69,65 @@ export const TRAIN_LLM: WorkflowConfig = {
   ],
 };
 
-export const WORKFLOW_TEMPLATES: WorkflowConfig[] = [AGENTIC_SEVEN_LAYER, TRAIN_LLM];
+/**
+ * Out-of-the-box (prompt → code) exemplars so "Teach Evermind to Code" runs with
+ * no network. Each carries execution `cases` so the same best-of-N / exec-filter
+ * path a live run uses is exercised offline (an exemplar whose code fails its
+ * cases is dropped, never trained on).
+ *
+ * A real run instead sets a teacher PANEL on the `distill-corpus` step —
+ * `teachers: [{provider:"anthropic", model:"claude-opus-4-8", apiKey},
+ *             {provider:"openai", baseUrl:<gateway>, model:<coder>, apiKey}]`
+ * (strongest first) — plus `tasks: [{prompt, cases}]`. Per task every teacher is
+ * queried, only completions that PASS the cases survive, and the strongest
+ * correct answer is kept: Opus is the authority, cheaper models contribute only
+ * where they are verifiably right.
+ */
+const CODE_DISTILL_PAIRS = [
+  { prompt: "Write a JS function add(a, b) that returns their sum.", completion: "function add(a, b) {\n  return a + b;\n}", cases: [{ call: "add(2, 3)", expect: 5 }] },
+  { prompt: "Write a JS function isEven(n) that returns true when n is even.", completion: "function isEven(n) {\n  return n % 2 === 0;\n}", cases: [{ call: "isEven(4)", expect: true }, { call: "isEven(7)", expect: false }] },
+  { prompt: "Write a JS function max(a, b) returning the larger value.", completion: "function max(a, b) {\n  return a > b ? a : b;\n}", cases: [{ call: "max(2, 9)", expect: 9 }] },
+  { prompt: "Write a JS function reverse(s) that reverses a string.", completion: "function reverse(s) {\n  return s.split('').reverse().join('');\n}", cases: [{ call: "reverse('abc')", expect: "cba" }] },
+  { prompt: "Write a JS function clamp(x, lo, hi) bounding x to [lo, hi].", completion: "function clamp(x, lo, hi) {\n  return Math.min(hi, Math.max(lo, x));\n}", cases: [{ call: "clamp(12, 0, 10)", expect: 10 }] },
+  { prompt: "Write a JS function sum(xs) that adds an array of numbers.", completion: "function sum(xs) {\n  return xs.reduce((a, b) => a + b, 0);\n}", cases: [{ call: "sum([1, 2, 3])", expect: 6 }] },
+];
+
+/**
+ * Teach Evermind to Code — the foundation pipeline for the coding skill.
+ *
+ * Distil a PANEL of teacher coders — Opus as the authority plus any cheaper
+ * models (gateway/OpenRouter/local) — into an execution-verified (prompt → code)
+ * corpus (best-of-N: only completions that pass their tests survive), learn a
+ * code tokenizer (train fresh by default; set `hfTokenizerUrl` on the tokenizer
+ * step to import merges from a proven code tokenizer instead), train the model,
+ * and gate it with the standard convergence/benchmark checks PLUS code-specific
+ * ones: structural/parse validity and an execution-grounded test reward.
+ *
+ * The code gates ship report-only (no thresholds) so the template completes and
+ * surfaces metrics; a real run sets `minScore` / `minPassRate` to enforce them.
+ */
+export const TEACH_CODE: WorkflowConfig = {
+  id: "teach-code",
+  name: "Teach Evermind to Code",
+  description:
+    "Distil a teacher coder into a (prompt → code) corpus, train an EvermindLM on it, and validate with code-aware gates: structural parse + an execution test reward. Tokenizer supports both train-fresh and import-merges.",
+  steps: [
+    { id: "distill", type: "distill-corpus", params: { pairs: CODE_DISTILL_PAIRS } },
+    { id: "tok", type: "train-tokenizer", params: { numMerges: 200 } },
+    { id: "data", type: "dataset-quality", params: { minSequences: 3, minWords: 20 } },
+    { id: "model", type: "train-model", params: { epochs: 60, dModel: 24, numLayers: 2, hiddenDim: 32 } },
+    { id: "converge", type: "convergence" },
+    { id: "eval", type: "evaluate", params: { prompt: "function" } },
+    { id: "gen", type: "generate-check", params: { prompt: "function" } },
+    { id: "parse", type: "code-parse-check", params: { language: "js" } },
+    { id: "reward", type: "code-eval", params: { cases: [{ call: "add(2, 3)", expect: 5 }] } },
+    { id: "bench", type: "benchmark" },
+    { id: "pkg", type: "roundtrip", params: { name: "evermind-coder" } },
+    { id: "export", type: "export", params: { format: "huggingface", name: "evermind-coder", version: "1.0.0" } },
+  ],
+};
+
+export const WORKFLOW_TEMPLATES: WorkflowConfig[] = [AGENTIC_SEVEN_LAYER, TRAIN_LLM, TEACH_CODE];
 
 export function getTemplate(id: string): WorkflowConfig | undefined {
   return WORKFLOW_TEMPLATES.find((t) => t.id === id);