fix: keep one-shot skills loaded until NewHEADmaster

5 files changed, 78 insertions, 14 deletions

diff --git a/CHANGELOG.md b/CHANGELOG.md
index bb8fb42..c4f6473 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -76,10 +76,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Skills. Instruction files read from `~/.agents/skills/*/SKILL.md` can be
   loaded into the model's context on demand: the model calls a `load_skill`
   tool, or you type `/skill-name` or mention a skill's name. Loaded skills
-  stay for the whole session in chat mode, drop after the reply in one-shot
-  mode, and show as removable chips. The backend tool loop now dispatches
-  between `web_search` and `load_skill`; `max_searches` caps total tool
-  rounds per turn.
+  show as removable chips and stay loaded for the whole session in chat
+  mode; one-shot mode keeps them in memory until New or the chip is clicked,
+  never in history. The backend tool loop now dispatches between `web_search`
+  and `load_skill`; `max_searches` caps total tool rounds per turn.
 
 ### Fixed
 
diff --git a/README.md b/README.md
index 349c21d..3e05f00 100644
--- a/README.md
+++ b/README.md
@@ -46,7 +46,8 @@ persistent process and toggles like a scratchpad from a Hyprland keybind.
   loaded into the model's context on demand. The model can call a
   `load_skill` tool when a task matches one; you can type `/skill-name` or
   mention a skill's name. Loaded skills show as removable chips, stay for
-  the whole session in chat mode, and drop after the reply in one-shot mode.
+  the whole session in chat mode, and in one-shot mode stay in memory until
+  New or the chip is clicked (never written to history).
 - **Tray icon** for show/hide/quit, hosted by waybar's tray module.
 
 Not in this version: RAG or embedding search over history, multi-user
@@ -491,7 +492,9 @@ A skill is loaded into the model's context on demand, three ways:
 - You mention a skill's name anywhere in the message.
 
 Loaded skills show as removable chips above the input. They stay loaded for
-the whole session in chat mode and drop after the reply in one-shot mode.
+the whole session in chat mode. One-shot mode keeps them in memory (never in
+history) until New is pressed or the chip is clicked, so a skill loaded in a
+scratch window stays active for the rest of it.
 `max_searches` caps the total tool rounds per turn here too, so skills load
 and searches share the same per-turn budget.
 
diff --git a/docs/superpowers/specs/2026-08-25-skills-design.md b/docs/superpowers/specs/2026-08-25-skills-design.md
index 20a3c0f..99baee6 100644
--- a/docs/superpowers/specs/2026-08-25-skills-design.md
+++ b/docs/superpowers/specs/2026-08-25-skills-design.md
@@ -13,8 +13,9 @@ with `name` and `description`, then the instructions.
 This adds a skills subsystem: llamachat reads that directory, lets the model
 load a skill on demand through a tool call, and lets the user load one by
 typing `/name` or mentioning the name naturally. A loaded skill's text stays
-in context for the whole conversation in chat mode, and for the single reply
-in one-shot mode.
+in context for the whole conversation in chat mode. One-shot mode keeps it
+in memory (never in history) until New or the chip is clicked, so a skill
+loaded in a scratch window stays active for the rest of it.
 
 ## Skills directory and format
 
diff --git a/llamachat/ui.py b/llamachat/ui.py
index 58c0156..db705c3 100644
--- a/llamachat/ui.py
+++ b/llamachat/ui.py
@@ -1301,8 +1301,8 @@ class ChatWindow(QMainWindow):
         """Record a skill as loaded; True when it was not already present.
 
         Persisted to the session in chat mode so reopening the conversation
-        restores it. One-shot mode keeps the set in memory only; it is
-        cleared when the reply finishes.
+        restores it. One-shot mode keeps the set in memory only: the chip
+        stays until New is pressed or the chip is clicked.
         """
         if not self.cfg.skills_enabled or name in self.loaded_skills:
             return False
@@ -1590,10 +1590,6 @@ class ChatWindow(QMainWindow):
 
     @Slot()
     def _on_stream_finished(self) -> None:
-        # One-shot mode: a skill loaded mid-reply is gone when the reply ends.
-        if self.mode == MODE_ONESHOT:
-            self.loaded_skills = []
-            self._refresh_skill_chips()
         if self.assistant_message_id is not None:
             self.history.update_message(
                 self.assistant_message_id,
diff --git a/test_llamachat.py b/test_llamachat.py
index 71a2c35..8fb7a0e 100755
--- a/test_llamachat.py
+++ b/test_llamachat.py
@@ -3773,6 +3773,69 @@ def test_skill_parts():
     print("ok  skill parts")
 
 
+def test_oneshot_skill_persists_until_new():
+    """A loaded skill survives a one-shot reply and is cleared by New.
+
+    One-shot used to drop skills when the reply finished, which made the
+    loaded-skill chip flash and vanish. The chip is the indicator that a
+    skill is active, so it must stay until the user clears it or starts a
+    new conversation.
+    """
+    os.environ.setdefault("QT_QPA_PLATFORM", "offscreen")
+    from PySide6.QtWidgets import QApplication
+
+    from llamachat import backend as _backend
+    from llamachat import config as _config
+    from llamachat import models as _models
+    from llamachat.ui import ChatWindow
+
+    app = QApplication.instance() or QApplication([])
+
+    with tempfile.TemporaryDirectory() as tmp:
+        tmp = Path(tmp)
+        skill_dir = tmp / "skills"
+        (skill_dir / "handoff" / "SKILL.md").parent.mkdir(parents=True)
+        (skill_dir / "handoff" / "SKILL.md").write_text(
+            "---\nname: handoff\ndescription: prepare a session handoff.\n"
+            "---\n\n# handoff\nWrite the handoff.\n"
+        )
+
+        cfg = _config.load(tmp / "config.toml")
+        cfg.prompts_dir = tmp / "prompts"
+        cfg.db_path = tmp / "t.db"
+        cfg.skills_dir = skill_dir
+        cfg.skills_enabled = True
+        history = db.History(cfg.db_path)
+        client = _backend.MultiClient(cfg.providers, cfg.request_timeout)
+        presets = _config.parse_presets(cfg.presets_path)
+        store = _models.ModelStore(cfg.models_path)
+        window = ChatWindow(cfg, history, client, presets, store)
+
+        window.session_id = history.create_session("oneshot", "m")
+        text = window._load_skills_from_text("prepare a handoff please")
+        assert text == "prepare a handoff please"
+        assert "handoff" in window.loaded_skills
+        assert window.skill_chips, "a loaded skill must show a chip"
+
+        # A finished one-shot reply must keep the skill loaded until New.
+        window.assistant_message_id = history.add_message(
+            window.session_id, "assistant", "here is the handoff"
+        )
+        window.assistant_buffer = "here is the handoff"
+        window._on_stream_finished()
+        assert "handoff" in window.loaded_skills, "one-shot must keep the skill"
+        assert window.skill_chips, "the chip must stay after a one-shot reply"
+
+        # New drops the skill and the chip.
+        window.new_session()
+        assert window.loaded_skills == []
+        assert not window.skill_chips
+
+        window.close()
+        history.close()
+    print("ok  oneshot skill persists until new")
+
+
 if __name__ == "__main__":
     test_presets()
     test_real_presets()
@@ -3848,4 +3911,5 @@ if __name__ == "__main__":
     test_skills_db()
     test_skills_column_migration()
     test_skill_parts()
+    test_oneshot_skill_persists_until_new()
     print("\nall checks passed")