superlab 0.1.48 → 0.1.50

package/bin/superlab.cjs

@@ -39,7 +39,8 @@ Usage:
   superlab init [--target <dir>] [--platform codex|claude|both|all] [--lang en|zh] [--force]
   superlab install [--target <dir>] [--platform codex|claude|both|all] [--lang en|zh] [--force]
   superlab paper attach-template --path <dir> [--target <dir>]
-  superlab auto start [--target <dir>] [--objective <text>] [--campaign-kind <kind>] [--allowed-stages <csv>] [--autonomy-level <L1|L2|L3>]
+  superlab auto start [--target <dir>] [--campaign-kind <kind>] [--allowed-stages <csv>] [--autonomy-level <L1|L2|L3>|--l1|--l2|--l3] [--objective <text>|<objective>]
+  superlab auto experiment [--target <dir>] [--l1|--l2|--l3] <objective>
   superlab auto status [--target <dir>]
   superlab auto stop [--target <dir>]
   superlab update [--target <dir>]
@@ -219,7 +220,7 @@ function parsePaperArgs(argv) {
 
 function parseAutoArgs(argv) {
   const [action, ...rest] = argv;
-  if (!["start", "status", "stop"].includes(action || "")) {
+  if (!["start", "status", "stop", "experiment"].includes(action || "")) {
     throw new Error(`Unknown auto action: ${action || "(missing)"}`);
   }
   const options = {
@@ -230,13 +231,14 @@ function parseAutoArgs(argv) {
     requestedAllowedStages: "",
     requestedAutonomyLevel: "",
   };
+  const positionalObjectiveParts = [];
 
   for (let index = 0; index < rest.length; index += 1) {
     const value = rest[index];
     if (value === "--target") {
       options.targetDir = path.resolve(rest[index + 1]);
       index += 1;
-    } else if (action === "start" && value === "--objective") {
+    } else if ((action === "start" || action === "experiment") && value === "--objective") {
       options.requestedObjective = rest[index + 1] || "";
       index += 1;
     } else if (action === "start" && value === "--campaign-kind") {
@@ -245,14 +247,25 @@ function parseAutoArgs(argv) {
     } else if (action === "start" && value === "--allowed-stages") {
       options.requestedAllowedStages = rest[index + 1] || "";
       index += 1;
-    } else if (action === "start" && value === "--autonomy-level") {
+    } else if ((action === "start" || action === "experiment") && value === "--autonomy-level") {
       options.requestedAutonomyLevel = (rest[index + 1] || "").trim();
       index += 1;
+    } else if ((action === "start" || action === "experiment") && ["--l1", "--l2", "--l3"].includes(value)) {
+      options.requestedAutonomyLevel = value.slice(2).toUpperCase();
+    } else if (action === "start" || action === "experiment") {
+      positionalObjectiveParts.push(value);
     } else {
       throw new Error(`Unknown option: ${value}`);
     }
   }
 
+  if (!options.requestedObjective && positionalObjectiveParts.length > 0) {
+    options.requestedObjective = positionalObjectiveParts.join(" ").trim();
+  }
+  if (action === "experiment" && !options.requestedCampaignKind) {
+    options.requestedCampaignKind = "experiment-loop";
+  }
+
   return options;
 }
 
@@ -990,7 +1003,7 @@ async function main() {
       printAutoStatus(options);
       return;
     }
-    if (options.action === "start") {
+    if (options.action === "start" || options.action === "experiment") {
       const result = await startAutoMode({
         targetDir: options.targetDir,
         requestedContract: {

package/lib/auto_contracts.cjs

@@ -23,6 +23,12 @@ const AUTO_LEVEL_STAGE_ENVELOPES = {
 };
 const VALID_APPROVAL_STATUSES = new Set(["draft", "approved"]);
 const VALID_TERMINAL_GOAL_TYPES = new Set(["rounds", "metric-threshold", "task-completion"]);
+const CAMPAIGN_KIND_DEFAULT_STAGES = {
+  planning: ["idea", "data", "framing", "spec", "review", "report"],
+  spec: ["idea", "data", "framing", "spec", "review", "report"],
+  "experiment-loop": ["run", "iterate", "review", "report"],
+  "report-polish": ["review", "report", "write"],
+};
 const FROZEN_CORE_ALIASES = {
   mission: [path.join(".lab", "context", "mission.md")],
   framing: [
@@ -82,19 +88,23 @@ function inferCampaignKind({ campaignKind = "", allowedStages = [] }) {
 }
 
 function normalizeRequestedAutoContract(requested = {}) {
-  const allowedStages = Array.isArray(requested.allowedStages)
+  const rawAllowedStages = Array.isArray(requested.allowedStages)
     ? requested.allowedStages.map((stage) => String(stage).trim().toLowerCase()).filter(Boolean)
     : normalizeList(requested.allowedStages || "").map((stage) => stage.toLowerCase());
+  const normalizedCampaignKind =
+    isMeaningful(requested.campaignKind || "") || rawAllowedStages.length > 0
+      ? inferCampaignKind({
+          campaignKind: requested.campaignKind || "",
+          allowedStages: rawAllowedStages,
+        })
+      : "";
+  const allowedStages = rawAllowedStages.length > 0
+    ? rawAllowedStages
+    : [...(CAMPAIGN_KIND_DEFAULT_STAGES[normalizedCampaignKind] || [])];
   return {
     objective: (requested.objective || "").trim(),
     autonomyLevel: normalizeScalar(requested.autonomyLevel || ""),
-    campaignKind:
-      isMeaningful(requested.campaignKind || "") || allowedStages.length > 0
-        ? inferCampaignKind({
-            campaignKind: requested.campaignKind || "",
-            allowedStages,
-          })
-        : "",
+    campaignKind: normalizedCampaignKind,
     allowedStages,
   };
 }
@@ -539,6 +549,7 @@ module.exports = {
   REVIEW_CONTEXT_FILES,
   VALID_APPROVAL_STATUSES,
   VALID_TERMINAL_GOAL_TYPES,
+  CAMPAIGN_KIND_DEFAULT_STAGES,
   classifyAutoContractFit,
   changedSnapshotPaths,
   detectFrozenCoreChanges,

package/lib/i18n.cjs

@@ -207,7 +207,7 @@ const ZH_SKILL_FILES = {
 - 本轮假设
 - 具体代码或配置改动
 - 标准化评估摘要
-- 书面 iteration report
+- 写入 \`.lab/iterations/\` 的书面 iteration report
 - 写入 \`results_root\` 的持久 run 输出
 - continue 或 stop 决策
 - 风险连续两轮升高时触发 diagnostic mode
@@ -235,6 +235,7 @@ const ZH_SKILL_FILES = {
 - 不要修改指标定义、baseline 语义或对比方法实现，除非评估协议已经记录了来源和与原始实现的偏差。
 - 如果要调整 ladder、样本量或升格 gate，必须继续锚定到带来源的评估协议，而不是靠聊天临时判断。
 - durable findings 和证据边界变化应写回 canonical context，再刷新派生的 \`state.md\` 快照；当前执行进度只写 \`workflow-state.md\`。
+- \`.lab/iterations/*.md\` 属于对外可读的 workflow artifact。如果 workflow language 是中文，iteration report、本轮摘要、阻塞点和 continue-or-stop 决策都应使用中文，除非文件路径、代码标识符或字面指标名必须保持原样。
 
 ## 交互约束
 
@@ -887,13 +888,13 @@ const ZH_SKILL_FILES = {
 - [ ] 产出 final report。
 `,
   [path.join(".lab", ".managed", "templates", "iteration-report.md")]:
-`# Iteration Report
+`# 迭代报告
 
 ## 轮次
 
-- Iteration number:
-- Run id:
-- Completion Promise:
+- 迭代编号:
+- 运行 id:
+- 完成承诺:
 
 ## 假设
 
@@ -905,27 +906,27 @@ const ZH_SKILL_FILES = {
 
 ## 评估摘要
 
-- Dataset:
-- Split:
-- Baseline:
-- Variant:
-- Primary metric:
-- Secondary metrics:
-- Failure count:
+- 数据集:
+- 划分:
+- 基线:
+- 变体:
+- 主指标:
+- 次级指标:
+- 失败次数:
 
 ## 专家批评
 
-- Main concerns:
-- Methodology concerns:
-- Interpretation concerns:
+- 主要关注点:
+- 方法论关注点:
+- 解释风险:
 - 是否进入 diagnostic mode:
 
 ## 决策
 
-- Continue or stop:
-- Next action:
-- Top blockers:
-- Next best actions:
+- 继续还是停止:
+- 下一步动作:
+- 主要阻塞点:
+- 次优后续动作:
 `,
   [path.join(".lab", ".managed", "templates", "review-checklist.md")]:
 `# Review Checklist
@@ -2854,6 +2855,7 @@ ZH_CONTENT[path.join(".codex", "skills", "lab", "stages", "auto.md")] = `# \`/la
 - 如果用户显式调用 \`/lab:auto\` 或 \`/lab-auto\`，就保持在 auto 执行路径里；只要请求仍在已批准 execution envelope 内，即使目标听起来像 feature selection、baseline selection、离散化或 candidate sweep，也不要重新路由到 brainstorming 或 spec review。
 - 如果用户同时提了论文层、实验 phase 和自治级别，先用一句话重述：objective、自治级别、terminal goal、scope、allowed modifications。
 - 如果 workflow language 是中文，摘要、清单条目、任务标签和进度更新都应使用中文，除非文件路径、代码标识符或字面指标名必须保持原样。
+- 对 auto 写出的 \`.lab/iterations/*.md\` 也应用同样的 workflow-language 规则，包括本轮摘要、阻塞点、diagnostic notes 和 continue-or-stop 决策。
 - 当循环进入 \`report\` 时，要主动给出用户可读的白话总结，解释主指标、次级指标和主表作用；不要等用户额外发一句“解释这些指标”。
 - 当循环即将进入 \`write\`，且 \`paper_template_root\` 为空时：
   - 如果 \`paper_template_decision\` 是 \`unconfirmed\`，必须先追问一次：继续使用默认 scaffold，还是先接入模板目录

package/package-assets/shared/skills/lab/SKILL.md

@@ -28,6 +28,7 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - Use `.lab/context/eval-protocol.md` as the shared evaluation contract for run, iterate, auto, and report stages, including metric glossary and experiment ladder semantics.
 - Treat evaluation semantics as source-backed once evaluation planning starts: metrics, benchmark gates, baseline behavior, comparison implementations, and deviations should come from recorded sources, not memory.
 - Workflow artifacts should follow the installed workflow language.
+- Iteration artifacts under `.lab/iterations/` are workflow artifacts and should follow the installed workflow language.
 - Final paper output should default to LaTeX, and its manuscript language should be decided separately from the workflow language.
 - Separate sourced facts from model-generated hypotheses.
 - Preserve failed runs, failed ideas, and limitations.
@@ -174,6 +175,7 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - Keep metric definitions, baseline behavior, and comparison implementations anchored to the source-backed evaluation protocol before changing thresholds, gates, or ladder transitions.
 - Switch to diagnostic mode if risk increases for two consecutive rounds.
 - Write round reports with `.lab/.managed/templates/iteration-report.md`.
+- Write round reports under `.lab/iterations/` with `.lab/.managed/templates/iteration-report.md`, and keep those iteration artifacts in the workflow language.
 - Update `.lab/context/workflow-state.md`, `.lab/context/decisions.md`, `.lab/context/evidence-index.md`, `.lab/context/open-questions.md`, and `.lab/context/eval-protocol.md` each round as needed, then refresh derived views.
 - Keep `.lab/context/eval-protocol.md` synchronized with accepted ladder changes, benchmark scope, and source-backed implementation deviations.
 - Stop at threshold success or iteration cap, and record blockers plus next-best actions when the campaign ends without success.

package/package-assets/shared/skills/lab/stages/auto.md

@@ -166,6 +166,7 @@
   - allowed modifications
 - Then ask at most one clarifying question if a blocking field is still missing.
 - If `.lab/config/workflow.json` sets the workflow language to Chinese, write summaries, options, checklist items, task labels, and progress updates in Chinese unless a file path, code identifier, or literal metric name must remain unchanged.
+- Apply the same workflow-language rule to auto-written iteration artifacts under `.lab/iterations/`, including round summaries, blocker lists, diagnostic notes, and continue-or-stop decisions.
 - When the loop reaches `report`, apply the same workflow-language rule to `report.md` and the managed `main-tables.md` artifact.
 - When the loop reaches `report`, proactively deliver a user-facing plain-language summary of the selected metrics, what they mean, what the tables prove, and what remains unproven. Do not wait for a separate user request asking for interpretation.
 - When the loop is about to enter `write` and `paper_template_root` is empty:

package/package-assets/shared/skills/lab/stages/iterate.md

@@ -37,7 +37,7 @@ Declare and keep fixed:
 - round hypothesis
 - concrete code or config changes
 - normalized evaluation summary
-- written iteration report
+- written iteration report under `.lab/iterations/`
 - durable run outputs under `results_root`
 - continue or stop decision
 - diagnostic mode trigger when risk increases for two consecutive rounds
@@ -70,6 +70,7 @@ If the loop stops without success, record:
 - Re-run the `Sanity and Alternative-Explanation Checks` whenever a round produces anomaly signals, suspiciously unchanged results, impl/result mismatches, or other outcomes that could still have simpler explanations.
 - If a round reveals leakage risk, invalid scale comparisons, unsupported metric semantics, or overstated claim boundaries, treat that as a fatal methodological flaw instead of a normal failed iteration.
 - If anomaly signals remain unresolved after implementation reality checks and at least one cross-check, switch to diagnostic mode instead of continuing as if the interpretation were settled.
+- Keep `.lab/iterations/*.md` as reader-facing workflow artifacts. If `.lab/config/workflow.json` sets the workflow language to Chinese, write the iteration report, round summary, blocker list, and continue-or-stop decision in Chinese unless a file path, code identifier, or literal metric name must remain unchanged.
 
 ## Interaction Contract
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlab",
-  "version": "0.1.48",
+  "version": "0.1.50",
   "description": "Strict /lab research workflow installer for Codex and Claude",
   "keywords": [
     "codex",