trip-moments-skill 1.1.1 → 1.1.2

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/selftest.py

@@ -2930,24 +2930,29 @@ 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 +2963,34 @@ 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 +3027,12 @@ 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 +3049,7 @@ 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 +3066,40 @@ 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 +3129,10 @@ 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

@@ -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:
@@ -2948,6 +2949,41 @@ 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.
 

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

@@ -2242,6 +2242,26 @@ async function handleJobDone(j) {
           （<code>POST /api/grade</code>），或改用「模型网关」直接生成方案。</div>`;
       }
       toast("Agent 没改动工作台，所以还没有三档候选可选", "err", 7000);
+    } else if (akind === "caption") {
+      // 文案 is the one step with no adopt concept: it lands in the workspace and
+      // the user copies a version. Before /api/caption existed an agent could
+      // only say the three versions in the conversation, so this very panel kept
+      // answering 「还没有文案」 right after the user asked for them.
+      const caps = (S.data.captions && S.data.captions.versions) || [];
+      if (caps.length) {
+        renderCaptions();
+        toast(`Agent 已登记 ${caps.length} 版文案，在文案面板逐版复制`, "ai", 6000);
+      } else {
+        const step = (j.meta || {}).step || jobStep(j);
+        const box = step === S.step ? outBox() : null;
+        if (box) {
+          box.innerHTML = `<div class="note" style="margin-top:10px"><b>Agent 没把文案登记到工作台</b><br>
+            三版文案可能只写在会话里了，所以本步仍是空的、打包也拿不到。
+            回答在下方「本步回复与反馈」里；可让 Agent 重新调用 <code>POST /api/caption</code>
+            登记，或改用「模型网关」生成。</div>`;
+        }
+        toast("Agent 没有登记文案，文案面板仍是空的", "err", 7000);
+      }
     }
     return;
   }

package/bin/install.js

@@ -16,7 +16,7 @@ const path = require('path');
 const os = require('os');
 const { execSync, spawnSync } = require('child_process');
 
-const VERSION = '1.1.1';
+const VERSION = '1.1.2';
 const SKILL_NAME = 'trip-moments';
 
 const c = {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "trip-moments-skill",
-  "version": "1.1.1",
+  "version": "1.1.2",
   "description": "AI 朋友圈旅行出图助手 (trip-moments skill) — 工业级全流程旅行摄影智能出图与朋友圈九宫格套件",
   "main": "bin/install.js",
   "bin": {