@mcptoolshop/style-dataset-lab 3.0.1 → 3.1.0

package/README.zh.md

@@ -8,14 +8,16 @@
 
 <p align="center">
   <a href="https://www.npmjs.com/package/@mcptoolshop/style-dataset-lab"><img src="https://img.shields.io/npm/v/@mcptoolshop/style-dataset-lab" alt="npm"></a>
+  <a href="https://github.com/mcp-tool-shop-org/style-dataset-lab/actions/workflows/ci.yml"><img src="https://github.com/mcp-tool-shop-org/style-dataset-lab/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
+  <a href="https://codecov.io/gh/mcp-tool-shop-org/style-dataset-lab"><img src="https://codecov.io/gh/mcp-tool-shop-org/style-dataset-lab/branch/main/graph/badge.svg" alt="codecov"></a>
   <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="MIT License"></a>
 </p>
 
-编写您的视觉规则。生成艺术作品。根据这些规则评估每张图像。将结果作为版本控制的、可审计的训练数据发布。
+编写您的视觉规则。生成图像。根据这些规则评估每张图像。将结果作为版本控制的、可追溯的训练数据发布，然后将训练好的模型应用于实际的生产流程，并将最佳输出反馈回您的数据集。
 
 Style Dataset Lab 将您关于艺术风格的描述与实际用于训练的数据集联系起来。您定义一套规范，包括轮廓规则、调色板限制、材质语言，以及对您的项目而言重要的任何内容。流水线生成候选作品，根据这些规则对其进行评分，并将通过的成果打包成可重复的数据集，其中每个记录都解释了它被包含的原因。
 
-循环闭合：训练一个模型，生成新的输出，根据相同的标准对其进行评分，并将通过的作品重新导入。数据集不断增长，规则始终有效。
+然后，生产工作台接管：从项目基础信息中提取生成指令，通过 ComfyUI 运行这些指令，评估输出结果，批量生成表情图和环境板，选择最佳结果，并将它们作为新的候选对象重新导入。循环完成：生成、选择、审查、改进。
 
 ## 流水线
 
@@ -28,25 +30,37 @@ sdlab generate inputs/prompts/wave1.json --project my-project
 sdlab curate <id> approved "Strong silhouette, correct faction palette"
 
 # Bind approved work to constitution rules
-sdlab bind --project my-project
+# (`sdlab bind` is a shorter alias for `canon-bind`)
+sdlab canon-bind --project my-project
 
 # Freeze a versioned dataset
 sdlab snapshot create --project my-project
 sdlab split build
 sdlab export build
 
-# Build a training package and close the loop
+# Build a training package
 sdlab training-manifest create --profile character-style-lora
 sdlab training-package build
-sdlab eval-run create && sdlab eval-run score <id> --outputs results.jsonl
-sdlab reingest generated --source ./outputs --manifest <id>
+
+# Compile a production brief and run it
+sdlab brief compile --workflow character-portrait-set --subject kael_maren
+sdlab run generate --brief brief_2026-04-16_001
+
+# Critique, refine, batch-produce
+sdlab critique --run run_2026-04-16_001
+sdlab refine --run run_2026-04-16_001 --pick 001.png
+sdlab batch generate --mode expression-sheet --subject kael_maren
+
+# Select the best outputs and bring them back
+sdlab select --run run_2026-04-16_001 --approve 001.png,003.png
+sdlab reingest selected --selection selection_2026-04-16_001
 ```
 
-那个最后的命令至关重要。生成的输出会经过与所有其他内容相同的审查流程。循环闭合。
+这个最后的命令至关重要。选定的输出会经过与所有其他内容相同的审查流程。数据集不断增长，规则始终有效。
 
 ## 它产生的内容
 
-七个版本控制的、具有校验和的成果。每个成果都链接到其前身，因此您可以追溯任何训练记录，了解它是根据哪个规则被批准的。
+包含七个数据集文件和一套完整的生产工作台。每个文件都链接到其前驱文件，因此您可以追溯任何训练记录，直到找到最初批准它的规则。
 
 | 成果 | 它的定义 |
 |----------|-----------|
@@ -58,6 +72,17 @@ sdlab reingest generated --source ./outputs --manifest <id>
 | **Eval scorecard** | 每个任务的通过/失败结果，基于生成的输出与评估包的比较。 |
 | **Implementation pack** | 提示示例、已知的失败案例、连续性测试以及重新导入指南。 |
 
+生产工作台包含以下功能：
+
+| 表面功能 | 功能描述 |
+|---------|-------------|
+| **Compiled brief** | 从工作流程配置文件和项目基础信息中获取确定性的生成指令。 |
+| **Run** | 冻结的执行文件：包含指令、种子、ComfyUI 输出和清单。 |
+| **Critique** | 对生成结果进行结构化的多维度评估，与标准进行比较。 |
+| **Batch** | 协调的多通道生产（表情图、环境板、人物轮廓图）。 |
+| **Selection** | 创意决策文件：记录选择了哪些输出，原因以及它们的来源。 |
+| **Re-ingest** | 选定的输出作为包含完整生成信息的候选记录返回。 |
+
 ## 存在的理由
 
 训练数据是任何视觉人工智能流水线中最具价值的资源。但大多数训练数据只是一个包含图像的文件夹，没有任何历史记录、评估记录，以及与它应该遵循的风格规则的联系。
@@ -94,6 +119,12 @@ projects/my-project/
   splits/                Train/val/test partitions
   exports/               Versioned export packages
   training/              Profiles, manifests, packages, eval runs, implementations
+  workflows/             Workflow profiles + batch mode definitions
+  briefs/                Compiled generation briefs
+  runs/                  Execution artifacts (brief + outputs + manifest)
+  batches/               Coordinated multi-slot productions
+  selections/            Chosen outputs with reasons and provenance
+  inbox/generated/       Re-ingested images awaiting review
 ```
 
 ## 信任属性
@@ -130,13 +161,73 @@ npm install -g @mcptoolshop/style-dataset-lab
 
 需要 Node.js 20+ 以及安装在本地主机 8188 端口上的 [ComfyUI](https://github.com/comfyanonymous/ComfyUI) 才能进行生成。
 
+### 尝试不使用 ComfyUI
+
+您可以使用捆绑的 Star Freight 项目，探索完整的非生成功能，例如检查、整理、快照、分割和导出，而无需安装 ComfyUI 或下载任何 SDXL 模型文件。
+
+```bash
+# Scaffold a fresh project (no ComfyUI needed)
+sdlab init test --domain game-art
+
+# Run the canonical health check (no ComfyUI needed)
+sdlab project doctor --project test
+
+# Dry-run a snapshot against the bundled Star Freight corpus
+sdlab snapshot create --dry-run --project star-freight
+```
+
+`sdlab project doctor` 命令可以验证每个项目的配置（宪法、流程、评分标准、术语），并在不访问 GPU 的情况下报告资格。任何会修改生成状态的命令都支持 `--dry-run` 参数，以便在执行前预览效果。
+
+如果您忘记了 `--project` 参数，CLI 会自动选择 `projects/` 目录下找到的第一个项目，并打印一条警告。要强制指定项目，请明确使用 `--project` 参数。
+
+### 恢复中断的运行
+
+可以恢复长时间的生成运行，而无需重新执行已完成的工作：
+
+```bash
+# Skip subjects whose record + image are already on disk.
+# Seeds are preserved — resumed runs are bit-identical to fresh ones.
+sdlab generate inputs/prompts/wave1.json --project my-project --resume
+
+# Re-run only failed/missing slots in an existing batch.
+# Inherits mode/subject/theme from the prior manifest.
+sdlab batch generate --resume batch_2026-04-22_001 --project my-project
+```
+
+这两个命令都有效，因为每个通道在完成时都会原子性地写入其清单条目，即使在运行过程中崩溃，也不会损坏部分状态。
+
+## 故障排除
+
+常见的错误模式和解决方法：
+
+**`ECONNREFUSED 127.0.0.1:8188` 错误，出现在任何 `sdlab generate` / `sdlab run generate` / `sdlab batch generate` 命令中**
+ComfyUI 未运行。启动 ComfyUI (`python main.py --listen 127.0.0.1 --port 8188`)，并使用 `curl http://127.0.0.1:8188/system_stats` 确认其是否正在运行。要指定不同的主机/端口，请设置 `COMFY_URL=http://host:port`。
+
+**`missing checkpoint` / `LoRA weight not found` 错误**
+您的工作流程配置文件中指定了 ComfyUI 的 `models/checkpoints/` 或 `models/loras/` 文件夹中不存在的模型文件。打开 `projects/<project>/workflows/profiles/<profile>.json`，找到 `checkpoint` 或 `lora` 字段，然后要么下载指定的模型文件，要么将其替换为您已有的模型文件。重新运行 `sdlab project doctor --project <project>` 以确认修复。
+
+**`sdlab project doctor` 错误**
+Doctor 返回结构化的错误代码。常见的错误代码包括：
+- `E_PROJECT_NOT_FOUND`：项目目录不存在于 `projects/` 目录下。检查拼写。
+- `E_CONFIG_INVALID`：五个 JSON 配置文件之一未能通过模式验证。`hint` 字段会指出哪个文件和字段存在问题。
+- `E_RECORD_DRIFT`：记录的配置指纹与其来源不再匹配。根据提示进行重新整理或重新绑定。
+
+**`未指定 --project，将使用 <name> 作为默认项目`**
+这是一个提示信息。为了选择正确的项目并消除此提示，请明确使用 `--project <name>` 参数。
+
+**绘画风格 / 显存不足问题**
+请参考 `docs/internal/HANDOFF.md` 中的关于绘画风格降噪的调整说明。 简而言之：降低降噪强度，减少批次大小，或者在工作流程配置文件中切换到较小的模型文件。
+
+**报告错误**
+请在 https://github.com/mcp-tool-shop-org/style-dataset-lab/issues 上提交问题，并提供您的 sdlab 版本（`sdlab --version`）、Node 版本（`node -v`）、完整的命令以及结构化的错误输出。 错误报告模板会预先填充字段。
+
 ## 安全性
 
 仅在本地运行。没有遥测数据，没有分析，没有外部请求。图像保留在您的 GPU 和文件系统中。
 
 ## 许可证
 
-MIT
+MIT 许可证
 
 ---
 

package/bin/sdlab.js

@@ -113,7 +113,7 @@ Flags:
   --verbose            Show per-check detail
   --quiet              Only print PASS / FAIL summary`,
 
-  'generate': `sdlab generate [<pack-path>] --project <name> [--dry-run]
+  'generate': `sdlab generate [<pack-path>] --project <name> [--dry-run] [--resume]
 
 Generate candidates from a prompt pack (JSON file under
 projects/<name>/inputs/prompts/).
@@ -124,6 +124,10 @@ Positional:
 Flags:
   --project <name>     Project to operate on
   --dry-run            Preview work without contacting ComfyUI
+  --resume             Skip subjects whose record + image are already on disk.
+                       Seeds remain stable — skipped slots still advance the
+                       seed counter so downstream outputs are bit-identical to
+                       a fresh run.
 
 Env:
   COMFY_URL            ComfyUI HTTP endpoint (default: http://127.0.0.1:8188)
@@ -131,15 +135,19 @@ Env:
 Precedence: CLI flag > env COMFY_URL > project defaults > built-in default.`,
 
   'batch generate': `sdlab batch generate --mode <id> [--subject <id>] [--theme <label>] --project <name>
+       sdlab batch generate --resume <batch_id> --project <name>
 
 Execute a batch production mode (expression sheet, silhouette pack,
 environment board, etc).
 
 Flags:
-  --mode <id>          Batch mode id (required)
+  --mode <id>          Batch mode id (required unless --resume is given)
   --subject <id>       Subject for subject-driven modes
   --theme <label>      Theme label for environment modes
   --asset <id>         Training asset reference
+  --resume <id>        Resume an existing batch — re-runs only the failed or
+                       missing slots, keeps the same batch_id, and inherits
+                       mode/subject/theme/asset from the prior manifest.
   --project <name>     Project to operate on
   --dry-run            Prepare batch without submitting to ComfyUI
   --json               Output manifest as JSON`,

package/lib/adapters/comfyui-runner.js

@@ -11,6 +11,7 @@ import { comfyHealth, submitAndWait, downloadImage } from '../comfyui.js';
 import { buildWorkflowGraph, resolveTemplateId } from './comfyui-workflows.js';
 import { buildSeedPlan, prepareRunDir, saveRunManifest } from '../runtime-runs.js';
 import { loadProjectConfig } from '../config.js';
+import { pickOutputImage } from '../comfyui-output.js';
 import { log, info, warn } from '../log.js';
 import { runtimeError } from '../errors.js';
 
@@ -81,7 +82,7 @@ export async function executeRun({ projectRoot, projectName, brief, baseSeed, dr
 
     try {
       // Build graph
-      const { nodes } = buildWorkflowGraph({
+      const { nodes, saveNodeId } = buildWorkflowGraph({
         brief,
         projectDefaults: defaults,
         seed,
@@ -94,26 +95,20 @@ export async function executeRun({ projectRoot, projectName, brief, baseSeed, dr
 
       const elapsed = Date.now() - startMs;
 
-      // Extract output image
-      const nodeOutputs = result.outputs || {};
-      let imageFile = null;
-      let imageSubfolder = '';
-      for (const nodeOut of Object.values(nodeOutputs)) {
-        if (nodeOut.images && nodeOut.images.length > 0) {
-          imageFile = nodeOut.images[0].filename;
-          imageSubfolder = nodeOut.images[0].subfolder || '';
-          break;
-        }
-      }
+      // Extract output image — prefer the explicit save node from the
+      // workflow builder (or `expected_outputs.save_node_id` from the
+      // brief/profile), falling back to the highest numeric node id.
+      const preferNodeId = brief.expected_outputs?.save_node_id ?? saveNodeId;
+      const picked = pickOutputImage(result.outputs, { preferNodeId });
 
-      if (!imageFile) {
+      if (!picked) {
         warn(`  No output image for seed ${seed}`);
         outputs.push({ index: i, seed, status: 'no_output', elapsed_ms: elapsed });
         continue;
       }
 
       // Download and save
-      const imgData = await downloadImage(imageFile, imageSubfolder, comfyUrl);
+      const imgData = await downloadImage(picked.filename, picked.subfolder, comfyUrl);
       const destFilename = `${imageIndex}.png`;
       await writeFile(join(runDir, 'outputs', destFilename), imgData);
 
@@ -124,7 +119,8 @@ export async function executeRun({ projectRoot, projectName, brief, baseSeed, dr
         filename: destFilename,
         bytes: imgData.length,
         elapsed_ms: elapsed,
-        comfy_filename: imageFile,
+        comfy_filename: picked.filename,
+        comfy_node_id: picked.nodeId,
       });
 
       log(`    ✓ ${destFilename} (${imgData.length} bytes, ${(elapsed / 1000).toFixed(1)}s)`);

package/lib/batch-runs.js

@@ -85,6 +85,57 @@ export async function createBatchDir(projectRoot) {
   return { batchId, batchDir };
 }
 
+/**
+ * Compute which prior slot_ids are eligible to skip on resume.
+ * Truth source: status === 'ok'. For pre-status manifests we infer success
+ * from a truthy selected_output. Returns a Map<slot_id, prior_result>.
+ */
+export function buildCompletedSlotMap(priorResults) {
+  const completed = new Map();
+  if (!Array.isArray(priorResults)) return completed;
+  for (const r of priorResults) {
+    if (!r || !r.slot_id) continue;
+    const ok = r.status === 'ok' || (r.status === undefined && !!r.selected_output);
+    if (ok) completed.set(r.slot_id, r);
+  }
+  return completed;
+}
+
+/**
+ * Resume an existing batch directory. Loads its prior manifest so the caller
+ * can replay only the incomplete slots while keeping the same batch_id.
+ *
+ * Throws BATCH_NOT_FOUND if the batch dir or manifest is missing.
+ * Throws BATCH_NO_PROGRESS if the manifest predates the incremental-checkpoint
+ * format (no `slots` array) — those batches can't be partially resumed.
+ */
+export function openBatchDirForResume(projectRoot, batchId) {
+  const batchesDir = getBatchesDir(projectRoot);
+  const batchDir = join(batchesDir, batchId);
+  const manifestPath = join(batchDir, 'manifest.json');
+
+  if (!existsSync(manifestPath)) {
+    throw inputError(
+      'BATCH_NOT_FOUND',
+      `Batch "${batchId}" not found at ${manifestPath}`,
+      'Run `sdlab batch list` to see available batch IDs.'
+    );
+  }
+
+  const manifest = JSON.parse(readFileSync(manifestPath, 'utf-8'));
+  const priorResults = Array.isArray(manifest.slots) ? manifest.slots : null;
+
+  if (!priorResults) {
+    throw inputError(
+      'BATCH_NO_PROGRESS',
+      `Batch "${batchId}" has no slot-level progress to resume from`,
+      'This batch was created before incremental checkpointing — re-run from scratch instead.'
+    );
+  }
+
+  return { batchId, batchDir, priorResults, priorManifest: manifest };
+}
+
 /**
  * Execute all slot briefs as coordinated runs.
  *
@@ -95,6 +146,7 @@ export async function createBatchDir(projectRoot) {
  * @param {Object} opts.mode — batch mode definition
  * @param {Object[]} opts.slotBriefs — one brief per slot
  * @param {boolean} [opts.dryRun]
+ * @param {Object[]} [opts.priorResults] — previously-completed slot results to carry forward (resume)
  * @returns {Promise<Object[]>} — array of { slot, brief, manifest }
  */
 export async function executeBatchRuns({
@@ -104,6 +156,7 @@ export async function executeBatchRuns({
   mode,
   slotBriefs,
   dryRun = false,
+  priorResults = [],
 }) {
   // PB-005: fail fast on length mismatch instead of crashing deep in the loop.
   if (!Array.isArray(mode?.variant_plan) || mode.variant_plan.length !== slotBriefs.length) {
@@ -115,6 +168,8 @@ export async function executeBatchRuns({
     );
   }
 
+  const completed = buildCompletedSlotMap(priorResults);
+
   const results = [];
   // PB-005: incremental batch-progress manifest — written after every slot
   // so a crash never orphans the batch dir without a record of what ran.
@@ -125,6 +180,13 @@ export async function executeBatchRuns({
     const brief = slotBriefs[i];
     const slotLabel = slot.label || slot.slot_id;
 
+    if (completed.has(slot.slot_id)) {
+      const prior = completed.get(slot.slot_id);
+      info(`[${i + 1}/${slotBriefs.length}] Slot: ${slotLabel} \x1b[2m(resumed — already complete)\x1b[0m`);
+      results.push(prior);
+      continue;
+    }
+
     info(`[${i + 1}/${slotBriefs.length}] Slot: ${slotLabel}`);
 
     // Save brief into batch/briefs/

package/lib/comfyui-output.js

@@ -0,0 +1,70 @@
+/**
+ * Pick the canonical output image from a ComfyUI run's `outputs` map.
+ *
+ * ComfyUI workflows can contain more than one SaveImage node (e.g. a
+ * preview save + a final save, or multiple intermediate saves). Picking
+ * the first via `Object.values()` is non-deterministic across builds —
+ * Node guarantees insertion order, but the order in which ComfyUI
+ * populates `outputs` depends on its execution scheduler.
+ *
+ * Selection precedence:
+ *   1. `preferNodeId` — explicit hint from the workflow builder
+ *      (`buildWorkflowGraph` returns `{ saveNodeId }`) or from a
+ *      workflow profile's `expected_outputs.save_node_id`.
+ *   2. Highest numeric node id among nodes that emitted images.
+ *      This matches the typical convention where the final save node
+ *      is appended last, and is deterministic regardless of map order.
+ *   3. First node in iteration order — last-resort fallback for the
+ *      pathological case where node ids aren't numeric.
+ *
+ * @param {Object} nodeOutputs - the `outputs` field from a ComfyUI prompt result
+ * @param {Object} [opts]
+ * @param {string|number|null} [opts.preferNodeId] - hint to prefer
+ * @returns {{ filename: string, subfolder: string, nodeId: string } | null}
+ */
+export function pickOutputImage(nodeOutputs, { preferNodeId = null } = {}) {
+  if (!nodeOutputs || typeof nodeOutputs !== 'object') return null;
+
+  const candidates = [];
+  for (const [nodeId, nodeOut] of Object.entries(nodeOutputs)) {
+    if (nodeOut?.images?.length > 0) {
+      candidates.push({ nodeId, image: nodeOut.images[0] });
+    }
+  }
+  if (candidates.length === 0) return null;
+
+  // 1. Honor explicit hint when the matching node actually produced an image.
+  if (preferNodeId !== null && preferNodeId !== undefined) {
+    const wanted = String(preferNodeId);
+    const match = candidates.find(c => c.nodeId === wanted);
+    if (match) {
+      return {
+        filename: match.image.filename,
+        subfolder: match.image.subfolder || '',
+        nodeId: match.nodeId,
+      };
+    }
+  }
+
+  // 2. Highest numeric node id wins (typical "final save" convention).
+  const numeric = candidates
+    .map(c => ({ ...c, n: Number(c.nodeId) }))
+    .filter(c => Number.isFinite(c.n));
+  if (numeric.length > 0) {
+    numeric.sort((a, b) => b.n - a.n);
+    const winner = numeric[0];
+    return {
+      filename: winner.image.filename,
+      subfolder: winner.image.subfolder || '',
+      nodeId: winner.nodeId,
+    };
+  }
+
+  // 3. Last resort: first iteration entry (non-numeric ids).
+  const first = candidates[0];
+  return {
+    filename: first.image.filename,
+    subfolder: first.image.subfolder || '',
+    nodeId: first.nodeId,
+  };
+}

package/lib/training-adapters.js

@@ -0,0 +1,79 @@
+/**
+ * Training adapter registry.
+ *
+ * Single source of truth for training-package adapters. Each adapter
+ * lives in `lib/adapters/<id>.js` and exports `buildPackage(opts)`.
+ *
+ * To add a new adapter (e.g. `kohya-lora`, `onedtrainer`):
+ *   1. Add `lib/adapters/<id>.js` exporting `buildPackage`.
+ *   2. Add an entry to ADAPTER_REGISTRY below with a one-line description.
+ *
+ * Why explicit registration: the available adapter list shows up in error
+ * messages, training-profile validation, and (eventually) CLI completion.
+ * A static map keeps those views consistent and makes new adapters a
+ * one-line surface change rather than a directory-listing side effect.
+ */
+
+import { inputError } from './errors.js';
+
+export const ADAPTER_REGISTRY = Object.freeze({
+  'generic-image-caption': {
+    id: 'generic-image-caption',
+    description: 'Image folders + caption JSONL — broadest trainer compatibility',
+    module: './adapters/generic-image-caption.js',
+  },
+  'diffusers-lora': {
+    id: 'diffusers-lora',
+    description: 'Diffusers fine-tuning layout: image + caption .txt sidecars + per-partition JSONL',
+    module: './adapters/diffusers-lora.js',
+  },
+});
+
+/**
+ * Names of all registered adapters.
+ * @returns {string[]}
+ */
+export function listAdapters() {
+  return Object.keys(ADAPTER_REGISTRY);
+}
+
+/**
+ * Whether `id` names a registered adapter.
+ * @param {string} id
+ * @returns {boolean}
+ */
+export function isRegisteredAdapter(id) {
+  return Object.prototype.hasOwnProperty.call(ADAPTER_REGISTRY, id);
+}
+
+/**
+ * Resolve and dynamically import an adapter module.
+ *
+ * Throws `ADAPTER_NOT_REGISTERED` (input error, exit 1) for unknown ids
+ * — distinct from `MODULE_LOAD_FAILED` (runtime, exit 2) for the rare
+ * case where the file vanished after registration.
+ *
+ * @param {string} id
+ * @returns {Promise<{ buildPackage: Function }>}
+ */
+export async function loadAdapter(id) {
+  if (!isRegisteredAdapter(id)) {
+    throw inputError(
+      'ADAPTER_NOT_REGISTERED',
+      `Adapter "${id}" is not registered. Available: ${listAdapters().join(', ')}`,
+      'Pick one of the registered adapters or add a new one to lib/training-adapters.js.'
+    );
+  }
+  const entry = ADAPTER_REGISTRY[id];
+  try {
+    return await import(entry.module);
+  } catch (err) {
+    // The registry says it should exist but import failed — surface the
+    // underlying error so it doesn't get masked as a typo.
+    throw inputError(
+      'ADAPTER_MODULE_LOAD_FAILED',
+      `Adapter "${id}" is registered but its module failed to load: ${err.message}`,
+      `Check that ${entry.module} exists and exports buildPackage().`
+    );
+  }
+}

package/lib/training-packages.js

@@ -17,6 +17,7 @@ import { loadRecord } from './records.js';
 import { loadProjectConfig, detectLane, detectGroup } from './config.js';
 import { inputError } from './errors.js';
 import { SCHEMA_VERSION, checkManifestVersion } from './snapshot.js';
+import { loadAdapter } from './training-adapters.js';
 
 /**
  * Generate a training package ID: tp-YYYYMMDD-HHMMSS-XXXX
@@ -33,20 +34,6 @@ async function sha256File(filePath) {
   return createHash('sha256').update(data).digest('hex');
 }
 
-/**
- * Load an adapter module by name.
- */
-async function loadAdapter(adapterTarget) {
-  const adapterPath = `./adapters/${adapterTarget}.js`;
-  try {
-    return await import(adapterPath);
-  } catch {
-    throw new Error(
-      `Adapter "${adapterTarget}" not found. Available: generic-image-caption, diffusers-lora`
-    );
-  }
-}
-
 /**
  * Build a training package from a manifest.
  *

package/lib/training-profiles.js

@@ -9,6 +9,7 @@
 import { readFile, readdir, writeFile, mkdir } from 'node:fs/promises';
 import { join } from 'node:path';
 import { existsSync } from 'node:fs';
+import { isRegisteredAdapter, listAdapters } from './training-adapters.js';
 
 const REQUIRED_FIELDS = ['profile_id', 'label', 'target_family', 'asset_type', 'eligible_lanes', 'adapter_targets'];
 
@@ -32,6 +33,15 @@ export function validateProfile(profile) {
   if (profile.adapter_targets?.length === 0) {
     errors.push('adapter_targets must have at least one target');
   }
+  // Reject unregistered adapter targets early — surfaces typos at profile
+  // load time instead of at package-build time.
+  if (Array.isArray(profile.adapter_targets)) {
+    for (const target of profile.adapter_targets) {
+      if (!isRegisteredAdapter(target)) {
+        errors.push(`adapter_targets contains unregistered adapter "${target}". Available: ${listAdapters().join(', ')}`);
+      }
+    }
+  }
   return { valid: errors.length === 0, errors };
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mcptoolshop/style-dataset-lab",
-  "version": "3.0.1",
+  "version": "3.1.0",
   "description": "Canon-aligned dataset production and generation workbench — define visual rules, build versioned training data, compile production briefs, run local workflows, batch-produce, select winners, and re-ingest into your corpus",
   "type": "module",
   "bin": {
@@ -8,6 +8,7 @@
   },
   "scripts": {
     "test": "node --test tests/**/*.test.js",
+    "coverage": "c8 npm test",
     "verify": "node bin/sdlab.js --help && node bin/sdlab.js project doctor --project star-freight",
     "generate": "node scripts/generate.js",
     "curate": "node scripts/curate.js",
@@ -16,6 +17,20 @@
     "painterly": "node scripts/painterly.js",
     "generate:identity": "node scripts/generate-identity.js"
   },
+  "c8": {
+    "include": ["lib/**/*.js"],
+    "exclude": [
+      "tests/**",
+      "scripts/**",
+      "site/**",
+      "projects/**",
+      "templates/**",
+      "**/*.test.js"
+    ],
+    "reporter": ["text", "lcov", "text-summary"],
+    "report-dir": "coverage",
+    "all": true
+  },
   "files": [
     "bin/",
     "lib/",
@@ -58,5 +73,8 @@
     "node": ">=20"
   },
   "author": "mcp-tool-shop",
-  "license": "MIT"
+  "license": "MIT",
+  "devDependencies": {
+    "c8": "^10.1.3"
+  }
 }