trip-moments-skill 1.1.1 → 1.2.0

package/.agents/skills/trip-moments/scripts/agent_ui.py

@@ -28,18 +28,51 @@ AGENT_STYLES = r"""
 .ap-compose textarea{width:100%;min-height:65px;max-height:160px;resize:vertical;padding:9px;border:1px solid var(--line-2);border-radius:7px;font-size:12px}
 .ap-compose .row{margin-top:8px}.ap-compose .btn{white-space:normal}
 .ap-next{width:100%;margin-top:8px}
-.workflow-control{padding:13px;border:1px solid var(--line-2);border-radius:9px;background:var(--card);margin-bottom:13px}
-.workflow-control h3{font-size:12px;margin-bottom:8px}
-.workflow-mode{display:grid;grid-template-columns:1fr 1fr;gap:5px}
+.workflow-mode{display:grid;grid-template-columns:1fr 1fr;gap:5px;min-width:0}
 .workflow-mode button{padding:8px 5px;border:1px solid var(--line-2);border-radius:6px;background:var(--paper-hi);cursor:pointer;font-size:11px}
 .workflow-mode button.on{background:var(--navy);border-color:var(--navy);color:#fff}
-.workflow-control .row{margin-top:9px}.workflow-control .row .btn{flex:1}
-.workflow-control .note{margin-top:8px}
-.agent-workflow{border:1px solid #cbd7c8;border-radius:9px;background:var(--card);margin:12px 0;overflow:hidden}
-.agent-workflow>header{padding:10px 12px;border-bottom:1px solid var(--line);font-size:12px;font-weight:600}
-.agent-workflow .ap-stream{padding:10px;max-height:360px;overflow:auto}
-.agent-workflow .ap-status{margin:0 0 8px}
-.agent-workflow .ap-compose{padding:10px}
+/* ---- the per-step action deck -------------------------------------------
+   写要求 → 点按钮 → 看结果 in one card. Before this the request box sat in the
+   middle card, the trigger at the top of the rail and the result at the bottom,
+   and five labels told the user to click 「上方」. */
+.deck{border:1px solid var(--line-2);border-radius:10px;background:var(--card);margin-bottom:13px;overflow:hidden}
+.deck-h{display:flex;align-items:center;gap:8px;padding:10px 12px;border-bottom:1px solid var(--line);
+  background:linear-gradient(120deg,#f7f5ee,#eef2ee)}
+.deck-t{font-size:12px;font-weight:600}.deck-s{font-size:10px;color:var(--sub);text-align:right}
+.deck-ctl{padding:10px 12px 2px}
+.deck-row{display:flex;align-items:center;gap:9px;margin-bottom:8px}
+.deck-k{font-size:10.5px;color:var(--sub);flex:none;width:52px}
+.deck-row>.workflow-mode{flex:1}
+.deck-row select{flex:1;min-width:0;padding:7px;border:1px solid var(--line-2);border-radius:6px;background:#fff;font-size:11px}
+.deck-ctl #wf-mode-note{margin:0 0 10px}
+.deck-write{padding:0 12px}
+.deck-write .pref-grid{margin-top:2px}
+.deck-write .cp-actions{display:flex;gap:6px;flex-wrap:wrap;margin-bottom:8px}
+.deck-write .cp-actions button{font-size:11px;padding:6px 10px;border:1px solid #becdbb;border-radius:6px;
+  background:#f4f8f2;color:#2c4a37;cursor:pointer;transition:all .16s var(--ease)}
+.deck-write .cp-actions button:hover{background:#fff;transform:translateY(-1px)}
+.deck-l{display:block;font-size:10.5px;color:var(--sub);margin:2px 0 5px}
+.deck-write textarea{width:100%;min-height:62px;max-height:170px;resize:vertical;padding:9px;font-size:12px;
+  border:1px solid var(--line-2);border-radius:7px;background:#fff;font-family:inherit}
+.deck-go{display:flex;gap:7px;flex-wrap:wrap;padding:10px 12px 0}
+.deck-go .btn{white-space:normal}
+.deck-go #wf-work{flex:1;min-width:120px}
+.deck-go .note{margin:0}
+#wf-connection{margin:8px 12px 0}
+#wf-connection:empty{display:none}
+.deck-brief{margin:9px 12px 0;border:1px solid var(--line);border-radius:7px;background:var(--paper-hi)}
+.deck-brief>summary{font-size:10.5px;color:var(--sub);padding:8px 10px;cursor:pointer}
+.deck-brief pre{white-space:pre-wrap;overflow-wrap:anywhere;font:10.5px/1.65 var(--mono);color:var(--ink-2);
+  max-height:300px;overflow:auto;padding:0 10px 9px;margin:0}
+.deck-out{padding:11px 12px 12px;border-top:1px solid var(--line);margin-top:11px}
+.deck-out-h{display:flex;align-items:center;gap:8px;font-size:10px;color:var(--sub);
+  text-transform:uppercase;letter-spacing:.08em;margin-bottom:8px}
+.deck-out-h .sp{flex:1}.deck-out-s{font-size:10px;color:var(--sub);text-transform:none;letter-spacing:0}
+.deck-out .ai-out:empty{margin:0}
+.deck-agent{border-top:1px dashed var(--line-2);padding:11px 12px 12px;background:#fbfcfa}
+.deck-agent .ap-stream{max-height:340px;overflow:auto}
+.deck-agent .ap-status{margin:0 0 8px}
+.deck .proposal-card{margin:0 0 9px}
 @media(max-width:720px){.agent-panel{top:8px;bottom:8px;right:8px;width:calc(100vw - 16px)}}
 """
 
@@ -73,57 +106,70 @@ AGENT_SHELL = r"""
 AGENT_SCRIPT = r"""
 const AP = { open: false, data: null, polling: false, timer: null, seq: -1, pending: false, draft: "", answers: {} };
 function useAgentModel() { return S.data?.preferences?.ai_mode === "agent"; }
-function workflowControlsHtml() {
+/* The deck's control rows. No trigger button here any more: the deck owns the
+   写要求 → 点按钮 → 看结果 order, so the button sits under the input box. */
+function deckControlsHtml() {
   const agent = useAgentModel();
-  const labels = { select:"智能选片", crop:"生成构图方案", grade:"生成调色方案",
-    retouch:"开始修图", finish:"生成所选作品", caption:"生成 3 版文案" };
-  return `<section class="workflow-control"><h3>本步由谁处理</h3>
-    <div class="workflow-mode" role="group" aria-label="AI 执行来源">
-      <button class="${agent ? "" : "on"}" aria-pressed="${!agent}" data-ai-mode="gateway">模型网关</button>
-      <button class="${agent ? "on" : ""}" aria-pressed="${agent}" data-ai-mode="agent">Agent 自带模型</button>
-    </div>
-    <p class="note" id="wf-mode-note">${agent ? "使用已绑定 Agent 的模型、工具与会话上下文。" : "使用配置的模型网关；AI 分析会发送照片预览，不进入 Agent 会话。"}</p>
-    ${S.step === "finish" ? `<label class="note" for="finish-mode">本次制作</label>
-      <select id="finish-mode" style="width:100%;padding:7px;border:1px solid var(--line-2);border-radius:6px">
-        <option value="art" ${S.finishMode !== "local" ? "selected" : ""}>AI 艺术作品 · 使用所选模型来源</option>
-        <option value="local" ${S.finishMode === "local" ? "selected" : ""}>本地画板／九宫格 · 不调用模型</option>
-      </select><p class="note">在下方选择风格或版式，再点击「生成所选作品」。此步不打包。</p>` : ""}
-    ${["crop","grade"].includes(S.step) ? `<label class="note" for="wf-scope">处理范围</label>
-      <select id="wf-scope" style="width:100%;padding:7px;border:1px solid var(--line-2);border-radius:6px">
-        <option value="current" ${S.aiScope !== "all" ? "selected" : ""}>当前第 ${S.sel} 格</option>
+  const sel = S.sel != null ? S.sel : 1;
+  return `<div class="deck-ctl">
+    <div class="deck-row"><span class="deck-k">谁来做</span>
+      <div class="workflow-mode" role="group" aria-label="AI 执行来源">
+        <button class="${agent ? "" : "on"}" aria-pressed="${!agent}" data-ai-mode="gateway">模型网关</button>
+        <button class="${agent ? "on" : ""}" aria-pressed="${agent}" data-ai-mode="agent">Agent 自带模型</button>
+      </div></div>
+    ${["crop","grade"].includes(S.step) ? `<div class="deck-row"><span class="deck-k">处理范围</span>
+      <select id="wf-scope" aria-label="处理范围">
+        <option value="current" ${S.aiScope !== "all" ? "selected" : ""}>当前第 ${sel} 格</option>
         <option value="all" ${S.aiScope === "all" ? "selected" : ""}>全部已选照片</option>
-      </select>` : ""}
-    <div class="row"><button class="btn btn-sm" id="wf-review" onclick="runCurrentStep('review')">检查本步</button>
-      <button class="btn btn-sm btn-ai" id="wf-work" onclick="runCurrentStep('work')">${labels[S.step] || "处理本步"}</button></div>
-    <p class="note" id="wf-connection" ${agent ? "" : 'style="display:none"'}>${agent ? "连接状态正在同步…" : ""}</p>
-    ${agent ? `<button class="btn btn-sm ghost" onclick="openAgentPanel()">管理 Agent 连接</button>` : ""}
-  </section>`;
+      </select></div>` : ""}
+    ${S.step === "finish" ? `<div class="deck-row"><span class="deck-k">本次制作</span>
+      <select id="finish-mode" aria-label="本次制作方式">
+        <option value="art" ${S.finishMode !== "local" ? "selected" : ""}>AI 艺术作品 · 用所选模型来源</option>
+        <option value="local" ${S.finishMode === "local" ? "selected" : ""}>本地画板／九宫格 · 不调模型</option>
+      </select></div>` : ""}
+    <p class="note" id="wf-mode-note">${agent
+      ? "用已绑定 Agent 的模型、工具与会话上下文；检查、处理、追问都在本卡完成。"
+      : "用配置的模型网关；AI 分析会发送照片预览，不进入 Agent 会话。"}</p>
+  </div>`;
+}
+/* One input box serves both 「本步要求」 and 「继续追问」, so the session buttons
+   live next to the trigger instead of under a second textarea. */
+function deckSessionButtons() {
+  const active = !!AP.pending || !!(AP.data && AP.data.active);
+  if (active) {
+    return `<button class="btn btn-sm" id="ap-steer" onclick="controlAgent('steer')">补充到当前执行</button>
+      <button class="btn btn-sm ghost" id="ap-stop" onclick="controlAgent('stop')">停止本轮</button>`;
+  }
+  return `<button class="btn btn-sm" id="ap-send" onclick="sendAgentReply()">继续对话</button>`;
 }
-function agentWorkflowHtml() {
+function agentStreamHtml() {
   if (!useAgentModel()) return "";
-  return `<section class="agent-workflow" aria-label="当前步骤的 Agent 协作">
-    <header>本步回复与反馈 · 同一会话</header>
+  return `<div class="deck-agent">
+    <div class="deck-out-h"><span>Agent 会话</span><span class="sp"></span>
+      <span class="deck-out-s">同一会话 · 回复不会自动跳步</span></div>
     <div class="ap-stream" id="ap-stream">
       <div class="ap-status" id="ap-run-status">正在读取绑定会话…</div>
       <div id="ap-messages"></div><div id="ap-live" class="ap-live" style="display:none" role="status"></div>
       <div id="ap-approval"></div>
       <details class="ap-tools"><summary>本轮真实工具事件</summary><div id="ap-events"></div></details>
     </div>
-    <div class="ap-compose">
-      <textarea id="ap-input" placeholder="继续追问、补充要求，或确认采用刚才的方案。仍然进入同一个 Agent 会话。" aria-label="给绑定 Agent 的回复">${H(AP.draft)}</textarea>
-      <div class="row"><button class="btn btn-sm btn-primary" id="ap-send" onclick="sendAgentReply()">继续对话</button>
-        <button class="btn btn-sm" id="ap-steer" onclick="controlAgent('steer')">补充到当前执行</button>
-        <button class="btn btn-sm ghost" id="ap-stop" onclick="controlAgent('stop')">停止本轮</button></div>
-      <button class="btn btn-sm ghost" style="margin-top:6px" onclick="syncAgentHistory()">同步会话回复</button>
-      <p class="note" style="margin-top:6px">收到回复不会自动跳步。确认完成后，再使用本步底部的「进入下一步」。</p>
-    </div></section>`;
+    <div class="row" style="margin-top:7px">
+      <button class="btn btn-sm ghost" onclick="syncAgentHistory()">同步会话回复</button>
+      <button class="btn btn-sm ghost" onclick="openAgentPanel()">管理 Agent 连接</button>
+    </div>
+  </div>`;
+}
+function deckInput() { return $("deck-intent"); }
+function deckInputText() {
+  const el = deckInput();
+  return el ? el.value.trim() : (S.copilotIntent || "").trim();
 }
 function startAgentPolling() {
   if (!AP.timer) AP.timer = setInterval(() => {
     if (AP.open || useAgentModel() || AP.data?.active) pollAgent();
   }, 850);
 }
-function wireWorkflow() {
+function wireDeckControls() {
   const finishMode = $("finish-mode");
   if (finishMode) finishMode.addEventListener("change", () => {
     S.finishMode = finishMode.value;
@@ -132,8 +178,6 @@ function wireWorkflow() {
   document.querySelectorAll("[data-ai-mode]").forEach(el => el.addEventListener("click", () => setAIMode(el.dataset.aiMode)));
   const scope = $("wf-scope");
   if (scope) scope.addEventListener("change", () => { S.aiScope = scope.value; });
-  const input = $("ap-input");
-  if (input) input.addEventListener("input", () => { AP.draft = input.value; });
   if (useAgentModel()) {
     startAgentPolling();
     AP.seq = -1; AP.approvalId = null;
@@ -187,7 +231,7 @@ async function runCurrentStep(kind) {
   if (useAgentModel()) {
     if (kind === "review") {
       if (!await requireAgentConnection()) return;
-      return sendAgentMessage("review_step", $("cp-intent")?.value || "", S.step);
+      return sendAgentMessage("review_step", deckInputText(), S.step);
     }
     if (S.step === "retouch") return aiEdit();
     const operation = { select:"select9", crop:"compose", grade:S.aiScope === "all" ? "grade_all" : "grade_plan", retouch:"edit", finish:"finish", caption:"caption" }[S.step];
@@ -355,25 +399,31 @@ async function sendAgentMessage(action, message, step = S.step, options = {}) {
   try {
     const data = await post("/api/agent/send", { ...options, action, message, step, slot: S.sel });
     if (!data.ok) { toast(data.error, "err", 6500); return false; }
-    AP.draft = ""; if ($("ap-input")) $("ap-input").value = ""; pollJobs(); pollAgent();
+    AP.draft = "";
+    const el = deckInput();
+    if (el) el.value = "";
+    S.copilotIntent = "";
+    (S.copilotDrafts = S.copilotDrafts || {})[copilotDraftKey()] = "";
+    S.deckBrief = "";
+    pollJobs(); pollAgent();
     return true;
   } finally { AP.pending = false; }
 }
 function sendAgentStep(action) {
-  return sendAgentMessage(action, ($("ap-input")?.value || AP.draft).trim());
+  return sendAgentMessage(action, deckInputText());
 }
 function sendAgentReply() {
-  const text = ($("ap-input")?.value || AP.draft).trim();
+  const text = deckInputText();
   if (!text) { toast("请写下本轮回复或修改要求"); return; }
   return sendAgentMessage("", text);
 }
 async function controlAgent(action, answers) {
-  const text = ($("ap-input")?.value || AP.draft).trim();
+  const text = deckInputText();
   if (action === "steer" && !text) { toast("请先写下要补充的要求"); return; }
   const data = await post("/api/agent/control", { action, text, answers,
     run_id: AP.data?.run_id, request_id: AP.data?.approval?.request_id });
   if (!data.ok) { toast(data.error, "err"); return; }
-  if (action === "steer") { AP.draft = ""; if ($("ap-input")) $("ap-input").value = ""; }
+  if (action === "steer") { AP.draft = ""; const el = deckInput(); if (el) el.value = ""; }
   pollAgent();
 }
 function answerAgent() {

package/.agents/skills/trip-moments/scripts/clawbridge.py

@@ -591,13 +591,22 @@ def _studio_instruction_impl(action: str, ctx: dict, state: dict, base: str = ""
                 "   （也可逐格调用 POST /api/framing；ratio 为 null 表示保留原图）\n"
                 "4. 只回几行：列出你把哪几格定成了什么画幅、保留了哪几格原图。")
 
-    if action == "caption":
-        return (head + "\n请做：为这组九宫格写朋友圈文案。\n"
-                + _slot_lines(slots) + "\n\n"
-                "写 3 版不同语气（文艺诗意 / 纪实行记 / 松弛日常），每版 1-3 句、可直接发。\n"
-                "**只写画面里真实存在的东西**；括号里是选片时的标注，必要时核图确认。"
-                "如果哪几张的标注和画面不符，在最后用一行提醒我。\n"
-                "不要写文件，直接在这个会话里给出三版。")
+    if action == "caption":
+        return (head + "\n请做：为这组九宫格写朋友圈文案。\n"
+                + _slot_lines(slots) + "\n\n"
+                "写 3 版不同语气（文艺诗意 / 纪实行记 / 松弛日常），每版 1-3 句、可直接发。\n"
+                "**只写画面里真实存在的东西**；括号里是选片时的标注，必要时核图确认。"
+                "如果哪几张的标注和画面不符，在最后用一行提醒我。\n"
+                "写好后**必须登记到工作台**：只说在会话里不算完成，用户在工作台文案面板上"
+                "会一直看到「还没有文案」，打包时也拿不到：\n"
+                f"   POST {base}/api/caption\n"
+                '   {"narrative": "一句话叙事线（可留空）",\n'
+                '    "versions": [{"style": "文艺诗意", "text": "…"},\n'
+                '                 {"style": "纪实行记", "text": "…"},\n'
+                '                 {"style": "松弛日常", "text": "…"}],\n'
+                '    "tags": ["可选标签"]}\n'
+                "（versions 至少一版；登记后工作台会立即显示并可逐版复制）\n"
+                "最后只回一两句：三版各是什么语气。不要复述全文，用户在工作台上就能看到。")
 
     if action == "title":
         return (head + f"\n请做：给第 {slot_no} 格拟标题。\n"

package/.agents/skills/trip-moments/scripts/journey_ui.py

@@ -366,33 +366,26 @@ async function renderJourneyMap() {
   }
 }
 
-function copilotHtml() {
+/* Quick direction chips for the deck. The input they used to own is now THE
+deck input (#deck-intent) — one box per step, right above the trigger button,
+ so the old note telling the user to click the button 「上方」 is gone. */
+function deckChipsHtml() {
   const s = curSlot();
-  // Retouch already has a complete preset + instruction + submit workflow.
-  // A second quick-action card duplicated the input and submitted prematurely.
   if (!s?.present || !["crop", "grade"].includes(S.step)) return "";
   const buttons = S.step === "crop"
     ? [["compose","保护主体","保留主体完整与原有空间感，不要机械套用竖裁"],["compose","更有张力","寻找更有叙事张力的画幅，保留重要建筑与人物"]]
     : S.step === "grade"
     ? [["grade_plan","自然通透","自然通透，克制饱和度，保持真实肤色"],["grade_plan","电影暖调","电影感暖色调，但不要压死暗部，肤色保持自然"],["grade_plan","清冷氛围","清冷克制的旅行画册影调，不要过度偏青"]]
     : [];
-  return `<section class="copilot"><h3>AI 要求（可选）</h3>
-    <p>选方向只填写要求，不会发起任务；确认后点击上方「生成${S.step === "crop" ? "构图" : "调色"}方案」。</p>
-    <div class="cp-actions">${buttons.map(([action,label,intent]) =>
-      `<button data-cp-action="${action}" data-cp-intent="${H(intent)}">${label}</button>`).join("")}</div>
-    <input id="cp-intent" aria-label="本步 AI 要求" placeholder="可选：选择或填写本步要求" value="${H(S.copilotIntent || "")}" maxlength="800">
-    </section>`;
+  return `<div class="cp-actions">${buttons.map(([action,label,intent]) =>
+      `<button data-cp-action="${action}" data-cp-intent="${H(intent)}">${label}</button>`).join("")}</div>`;
 }
 
 function wireCopilot() {
-  const input = $("cp-intent");
-  if (input) input.addEventListener("input", () => {
-    S.copilotIntent = input.value;
-    (S.copilotDrafts = S.copilotDrafts || {})[copilotDraftKey()] = input.value;
-  });
   document.querySelectorAll("[data-cp-action]").forEach(el => el.addEventListener("click", () => {
     S.copilotIntent = el.dataset.cpIntent;
     (S.copilotDrafts = S.copilotDrafts || {})[copilotDraftKey()] = S.copilotIntent;
+    const input = $("deck-intent");
     if (input) input.value = S.copilotIntent;
     document.querySelectorAll("[data-cp-action]").forEach(button =>
       button.setAttribute("aria-pressed", String(button === el)));

package/.agents/skills/trip-moments/scripts/selftest.py

@@ -2917,10 +2917,35 @@ def t_claw_grade_candidates_and_output_visibility():
     html = studio_ui.STUDIO_HTML
     assert 'id="ai-out"' not in html, \
         "还有面板在用重复的 id=\"ai-out\"（getElementById 只会拿到第一个，结果会画进隐藏面板）"
-    assert html.count('class="ai-out"') >= 6, \
-        "六个 step 应各有一个输出区，实际 %d 个" % html.count('class="ai-out"')
-    assert '[data-pane="${step}"] .ai-out' in html, \
-        "输出区没有按当前 step 解析"
+    assert 'id="ai-out-${H(step)}"' in html, \
+        "操作台没有给每一步自己的输出区 id（结果可能画到别的步）"
+    assert '[data-pane="${step}"] .ai-out' in html, "输出区没有按当前 step 解析"
+    assert html.count("actionDeckHtml(") >= 7, \
+        "六个 step 应各接一个操作台（含函数定义共 7 处），实际 %d 处" % html.count("actionDeckHtml(")
+    assert html.count('pane("') >= 6, "六个 step 面板数量不对"
+    assert "getElementById(`ai-out-${step}`)" in html, \
+        "outBox() 没有优先按 step 精确取输出区"
+
+    # 操作台的存在意义：写要求 → 点按钮 → 看结果 在同一张卡里，顺序不能倒。
+    # 旧布局把触发按钮放在 rail 顶部、输入框放在中间卡片、结果放在最底部，
+    # 于是五处正文直接叫用户「点击上方…」——用户抱怨“上下跑”。这里把那个
+    # 设计约束固化成断言，而不是只写进注释。
+    deck = html[html.index("function actionDeckHtml("):]
+    deck = deck[:deck.index("async function loadDeckBrief")]
+    assert deck.index('id="deck-intent"') < deck.index('id="wf-work"'), \
+        "操作台里触发按钮又跑到输入框上面了（写→发 的顺序被倒过来）"
+    assert deck.index('id="wf-work"') < deck.index('class="ai-out"'), \
+        "操作台里结果区不在触发按钮下面（点完还要再找结果）"
+    for bad in ("点击上方", "上方选择", "上方唯一的", "本页顶部"):
+        assert bad not in html, f"还在叫用户往上看：「{bad}」——操作台就是为了去掉这种跳转"
+    # 一个输入框：旧布局有「本步要求」和「继续追问」两个框，用户得猜往哪个里写。
+    for gone in ('id="cp-intent"', 'id="ap-input"', 'id="inst"'):
+        assert gone not in html, f"旧输入框 {gone} 又回来了（应只剩 #deck-intent 一个）"
+    assert "function deckInputText" in html and "deckInputText()" in html, \
+        "发送路径没有统一读操作台的输入框"
+    assert html.count('id="deck-intent"') == 1, \
+        "#deck-intent 应只在操作台模板里出现一次，实际 %d 次" % html.count('id="deck-intent"')
+
     assert "p.export_zip || p.export_dir" not in html, \
         "导出目录的 Windows 路径又进了 href（浏览器里点了没反应）"
     assert "/api/reveal" in html, "出品台没有「在资源管理器里打开」的入口"
@@ -2930,24 +2955,29 @@ def t_claw_grade_candidates_and_output_visibility():
 
 
 def t_agent_model_grade_candidates():
-    """Agent 自带模型轨道也必须给出三档候选 + 预览/采用入口（2026-10-04 用户实测）。
-
-    用户挂着 Agent 会话点「生成调色方案」，页面既没有 克制/标准/浓郁 三行，也没有每行
-    的 预览/采用 按钮。缺口是三处协同造成的：发给 agent 的简报把那句「工作台会自动生成
-    三套候选」改写掉了、agent_turn 的 work() 不像 clawhive 轨道（studio.py 里 grade
-    分支）那样从落地的配方推导候选、前端又在 useAgentModel() 下把候选面板整个隐藏。
-
-    这里用真实 HTTP 跑 /api/agent/send：会话的 run() 换成「通过工作台自己的 /api/grade
-    把配方落地」——真实 agent 就是这么做的——然后断言三档候选真的随任务结果回到页面；
-    再跑一轮什么都不改的 agent，断言不会拿旧配方冒充本轮的方案。
+    """Agent 自带模型轨道：每一步的结果都要落到工作台上，不能只在会话里说。
+
+    2026-10-04 用户实测两处空白：
+      a) 挂着 Agent 会话点「生成调色方案」，页面既没有 克制/标准/浓郁 三行，也没有每行的
+         预览/采用 按钮。三处协同缺口：简报把「工作台会自动生成三套候选」改写掉了、
+         agent_turn 的 work() 不像 clawhive 轨道那样从落地的配方推导候选、前端又在
+         useAgentModel() 下把候选面板整个隐藏。
+      b) 点「生成 3 版文案」后，文案面板仍然回「还没有文案」——因为 Agent 根本没有能
+         登记文案的接口（选片/裁切/调色/修图/出片都有，只有文案没有）。
+
+    这里用真实 HTTP 跑 /api/agent/send，会话的 run() 换成「通过工作台自己的接口落地结果」
+    ——真实 agent 就是这么做的——然后断言：三档候选真的随任务结果回到页面；Agent 只说不
+    做时不拿旧配方冒充本轮方案；文案登记后 state 真的带得回来、空数组会被拒。
     """
     import tempfile as _tf
     import threading
     import time as _time
+    import urllib.error
     import urllib.request
     from pathlib import Path as _P
 
     import agent_session
+    import clawbridge
     import studio
     import studio_ui
 
@@ -2958,19 +2988,34 @@ def t_agent_model_grade_candidates():
     old_started = getattr(studio.StudioHandler, "started", None)
     box = {}
 
+    def post(path, payload):
+        req = urllib.request.Request(
+            f"http://127.0.0.1:{box['port']}{path}",
+            data=json.dumps(payload).encode("utf-8"),
+            headers={"Content-Type": "application/json"}, method="POST")
+        with urllib.request.urlopen(req, timeout=30) as r:
+            return r.status, json.loads(r.read())
+
     def fake_run(self, message, progress=None, cancelled=None):
-        # A real agent on this track applies its recipe through the workbench's
-        # own endpoint, so the candidates must be derived from APPLIED state.
-        text = "第 5 格我上了电影青橙。"
-        if box.get("apply"):
-            req = urllib.request.Request(
-                f"http://127.0.0.1:{box['port']}/api/grade",
-                data=json.dumps({"slot": 5, "preset": "cine",
-                                 "grade": {"exposure": 0.2, "highlights": -20,
-                                           "shadows": 15, "dehaze": 12}}).encode("utf-8"),
-                headers={"Content-Type": "application/json"}, method="POST")
-            with urllib.request.urlopen(req, timeout=30) as r:
-                assert r.status == 200, f"agent 落地配方失败: HTTP {r.status}"
+        # A real agent on this track lands its result through the workbench's own
+        # endpoint, so everything downstream must be derived from APPLIED state.
+        mode = box.get("mode")
+        text = "这一轮我没动工作台。"
+        if mode == "grade":
+            status, _ = post("/api/grade", {"slot": 5, "preset": "cine",
+                                             "grade": {"exposure": 0.2, "highlights": -20,
+                                                       "shadows": 15, "dehaze": 12}})
+            assert status == 200, f"agent 落地配方失败: HTTP {status}"
+            text = "第 5 格我上了电影青橙。"
+        elif mode == "caption":
+            status, _ = post("/api/caption", {
+                "narrative": "沿着海岸线走了一天",
+                "versions": [{"style": "文艺诗意", "text": "潮声比人声多"},
+                             {"style": "纪实行记", "text": "今天走了两万步"},
+                             {"style": "松弛日常", "text": "风大，头发乱"}],
+                "tags": ["海边"]})
+            assert status == 200, f"agent 登记文案失败: HTTP {status}"
+            text = "三版写好了，已登记到工作台。"
         # Mirror how the real run() ends: the terminal event is what takes the
         # session out of ACTIVE, so the next turn is allowed to start.
         self.live_text = text
@@ -3007,12 +3052,12 @@ def t_agent_model_grade_candidates():
         box["port"] = port
         threading.Thread(target=httpd.serve_forever, daemon=True).start()
         try:
-            def send_turn():
+            def send_turn(step="grade", operation="grade_plan"):
                 req = urllib.request.Request(
                     f"http://127.0.0.1:{port}/api/agent/send",
-                    data=json.dumps({"step": "grade", "action": "work_step",
-                                     "operation": "grade_plan", "slot": 5,
-                                     "message": "请用你自己的模型给我调色"}).encode("utf-8"),
+                    data=json.dumps({"step": step, "action": "work_step",
+                                     "operation": operation, "slot": 5,
+                                     "message": "请用你自己的模型处理这一步"}).encode("utf-8"),
                     headers={"Content-Type": "application/json"}, method="POST")
                 with urllib.request.urlopen(req, timeout=60) as r:
                     assert r.status == 202, f"Agent 回合没进任务队列: HTTP {r.status}"
@@ -3029,7 +3074,7 @@ def t_agent_model_grade_candidates():
                 raise AssertionError(f"agent_turn 没跑完: {jid}")
 
             # ① agent 通过工作台接口落了配方 -> 必须带三档候选
-            box["apply"] = True
+            box["mode"] = "grade"
             job = wait_job(send_turn())
             assert job["status"] == "done", f"agent 回合失败: {job.get('error') or job}"
             res = (job.get("result") or {}).get("result") or {}
@@ -3046,11 +3091,40 @@ def t_agent_model_grade_candidates():
             assert (res.get("text") or ""), "agent 的话也要在同一个结果里"
 
             # ② 一轮什么都不改的 agent -> 不得拿旧配方冒充本轮方案
-            box["apply"] = False
+            box["mode"] = "none"
             job2 = wait_job(send_turn())
             res2 = (job2.get("result") or {}).get("result") or {}
             assert not (res2.get("candidates") or []), \
                 "agent 没动工作台却给了候选（旧配方被冒充成本轮方案）"
+
+            # ③ 文案：agent 登记后必须进 workspace（能刷新回来、算作本步已完成）
+            box["mode"] = "caption"
+            job3 = wait_job(send_turn(step="caption", operation="caption"))
+            assert job3["status"] == "done", f"文案回合失败: {job3.get('error') or job3}"
+            with urllib.request.urlopen(f"http://127.0.0.1:{port}/api/state", timeout=30) as r:
+                state = json.loads(r.read())
+            caps = (state.get("captions") or {}).get("versions") or []
+            assert len(caps) == 3, f"agent 登记的文案没进 workspace: {state.get('captions')}"
+            assert caps[0]["style"] == "文艺诗意" and caps[0]["text"], caps
+            assert (state.get("captions") or {}).get("narrative") == "沿着海岸线走了一天"
+
+            # ④ 空文案必须被拒：不能把「没写」当成「已登记」
+            try:
+                post("/api/caption", {"versions": []})
+                raise AssertionError("空 versions 不应被接受")
+            except urllib.error.HTTPError as exc:
+                assert exc.code == 400, f"空文案应回 400，实际 {exc.code}"
+            try:
+                post("/api/caption", {"versions": ["   "]})
+                raise AssertionError("全空白文案不应被接受")
+            except urllib.error.HTTPError as exc:
+                assert exc.code == 400, f"全空白文案应回 400，实际 {exc.code}"
+
+            # ⑤ 简报必须告诉 agent 登记哪个接口，否则它只会写在会话里
+            brief = clawbridge.studio_instruction(
+                "caption", {"slots": [1]}, state, f"http://127.0.0.1:{port}")
+            assert "/api/caption" in brief, "文案简报没给出登记接口"
+            assert "POST" in brief and "versions" in brief, "文案简报没有可照抄的请求体"
         finally:
             agent_session.AgentSession.run = real_run
             httpd.shutdown()
@@ -3080,6 +3154,10 @@ def t_agent_model_grade_candidates():
         "agent 只说不做时页面又保持沉默（用户分不清是 bug 还是没改动）"
     assert "S.aiCandidates = null" in branch, \
         "agent 没落地配方时没有清掉上一轮的候选（旧方案会被当成本轮产物）"
+    assert "akind === \"caption\"" in branch, \
+        "agent_turn 完成后没有处理文案（面板又会一直回「还没有文案」）"
+    assert "没把文案登记到工作台" in branch, \
+        "agent 只把文案写在会话里时页面不说明"
 
 
 CASES = {

package/.agents/skills/trip-moments/scripts/studio.py

@@ -1646,7 +1646,7 @@ class Workspace:
             "slots": slots,
             "pool": pool,
             "pool_meta": {"total": len(pool) + sum(1 for s in slots if s["present"])},
-            "captions": self.captions,
+            "captions": self.captions,
             "gps": _load_json(self.outdir / "gps.json", []),
             "products": products,
             "selection": getattr(self, "selection_extra", {}),
@@ -2146,7 +2146,8 @@ class StudioHandler(BaseHTTPRequestHandler):
                 "/api/reset": self._api_reset,
                 "/api/undo": self._api_undo,
                 "/api/edit-source": self._api_edit_source,
-                "/api/edit-result": self._api_edit_result,
+                "/api/edit-result": self._api_edit_result,
+                "/api/caption": self._api_caption,
                 "/api/edit-scan": self._api_edit_scan,
                 "/api/art": self._api_art,
                 "/api/art-preview": self._api_art_preview,
@@ -2172,8 +2173,8 @@ class StudioHandler(BaseHTTPRequestHandler):
             if fn is None:
                 self.send_json({"ok": False, "error": f"未知接口 {path}"}, 404)
                 return
-            if path in {"/api/framing", "/api/grade", "/api/swap", "/api/replace",
-                        "/api/reset", "/api/undo", "/api/edit-source"}:
+            if path in {"/api/framing", "/api/grade", "/api/swap", "/api/replace",
+                        "/api/reset", "/api/undo", "/api/edit-source", "/api/caption"}:
                 # Snapshot + mutation + render must be atomic even when an
                 # agent and a browser edit the same workspace concurrently.
                 with self.ws._lock:
@@ -2227,6 +2228,13 @@ class StudioHandler(BaseHTTPRequestHandler):
             self.send_json(link.sync_history())
         elif path == "/api/agent/disconnect":
             self.send_json(link.disconnect())
+        elif path == "/api/agent/brief":
+            # Lazy: the brief can be several KB (the art one embeds a compiled
+            # prompt card), so it is fetched when the deck's details is opened
+            # rather than riding along on every /api/state poll.
+            self.send_json({"ok": True,
+                            "brief": str(self.ws.extra.get("agent_brief") or ""),
+                            "at": str(self.ws.extra.get("agent_brief_at") or "")})
         elif path == "/api/agent/control":
             self.send_json(link.control(str(body.get("action") or ""), body))
         elif path == "/api/agent/send":
@@ -2316,6 +2324,12 @@ class StudioHandler(BaseHTTPRequestHandler):
                                 f"修图状态：{'已采用文件缺失' if s.get('edit_missing') else '新修图待采用' if s['pending_edit'] else '已采用修图' if s['edit_adopted'] else '原片处理版'}"
                                 for s in snapshot["slots"] if s["present"])
             outbound = context + "\n当前作品：\n" + summary + "\n\n用户本轮输入：\n" + message
+            # Keep the brief the deck shows the user. Transparency is the point:
+            # 「它有没有照做」 becomes checkable without reading the conversation.
+            with self.ws._lock:
+                self.ws.extra["agent_brief"] = context
+                self.ws.extra["agent_brief_at"] = _now()
+                self.ws.save()
             link.reserve(message)
             def work(job):
                 result = link.run(outbound, lambda text, pct: (setattr(job, "detail", text), setattr(job, "progress", pct)),
@@ -2948,6 +2962,41 @@ class StudioHandler(BaseHTTPRequestHandler):
                         "summary": self.ws.grade_summary(slot),
                         "state": self.ws.state()["slots"][n - 1]})
 
+    def _api_caption(self, body: dict) -> None:
+        """Persist caption versions produced outside the workbench (both agent tracks).
+
+        The local gateway track has always written captions into the workspace
+        (see _api_ai), but an agent had no equivalent endpoint: it could only say
+        the three versions in the conversation, so the 文案 step kept showing
+        「还没有文案」 — on the very screen where the user had just asked for them —
+        and the set never counted as finished. Registration stores exactly what
+        the caller sends; nothing is invented here.
+        """
+        raw = body.get("versions")
+        if not isinstance(raw, list):
+            self.send_json({"ok": False, "error": "versions 必须是数组"}, 400)
+            return
+        versions = []
+        for item in raw:
+            if isinstance(item, str):
+                text, style = item.strip(), ""
+            elif isinstance(item, dict):
+                text = str(item.get("text") or "").strip()
+                style = str(item.get("style") or "").strip()
+            else:
+                continue
+            if text:
+                versions.append({"style": style or f"版本 {len(versions) + 1}",
+                                 "text": text[:4000]})
+        if not versions:
+            self.send_json({"ok": False, "error": "至少需要一版非空文案"}, 400)
+            return
+        tags = [str(t).strip() for t in (body.get("tags") or []) if str(t).strip()][:12]
+        saved = self.ws.save_captions({"narrative": str(body.get("narrative") or "")[:600],
+                                       "versions": versions, "tags": tags})
+        self.send_json({"ok": True, "captions": saved, "versions": saved["versions"],
+                        "revision": self.ws.revision})
+
     def _api_edit_result(self, body: dict) -> None:
         """Register a retouch result produced outside the workbench.