Add house styles, watermarks, Plan mode storyboards, agent CLI auto-update and leaner agent sessions

AGENTS.md

@@ -29,6 +29,11 @@ starter scripts and manual rendering launchers have been retired.
   validates and queues each video once. Script/runtime failures return diagnostics to
   the same scoped chat, then retry with the original output settings, up to three
   repairs. Cancellation, provider errors and newer user edits must stop recovery.
+- Plan mode (`storyboard.py`, per-turn `mode`) writes only `storyboard.md` in the
+  project or video workspace: Studio restores any plan-turn change to `script.py` or
+  `narration.txt`, and `video_workflow.py` never validates or queues a plan turn.
+  Approval is the hidden `.storyboard.json` hash of the approved text, so any edit
+  returns it to draft; approving starts a Build turn that reads the file by path.
 - The render queue is server-owned in `studio.py` (one render at a time, up to 20
   waiting, ordered by `_seq`). Agent settings `queue_start` (`auto`/`manual`) and
   `queue_on_error` (`fix`/`skip`/`pause`) drive it; `QUEUE_PAUSED` is `None` to
@@ -52,6 +57,22 @@ starter scripts and manual rendering launchers have been retired.
 - Folder and ZIP imports use `/api/projects/import`, preserve nested assets and
   create a new project transactionally. Reject traversal, duplicate Windows paths,
   ZIP symlinks and packages over 512 MB / 2000 files; never install package requirements.
+- House styles (`styles.py`) live in `DATA_ROOT/.studio/styles.json`; built-ins are
+  read-only. Projects and videos store only a `style` id (absent: parent, then default).
+  Studio writes `studio_style.py` beside `script.py` before renders and agent turns,
+  only when its text changes so render cache keys stay stable, and adds a short style
+  guide to the agent instructions.
+- Watermarks (`watermark.py`): a global default in `.studio/watermark.json`, logos
+  rasterised into `.studio/watermark/logos/`, per-project/video overrides in metadata.
+  The worker overlays the final movie with PyAV (the bundled ffmpeg is audio-only)
+  after narration and the render-cache checkpoint, so the animation cache survives.
+- Agent CLI updates (`agent_updates.py`) check the npm registry at most daily and run
+  only while no turn uses that provider; new turns wait and repairs retry. Only npm and
+  native installs are updated; custom binaries are reported as managed externally.
+  `MANIM_STUDIO_AGENT_UPDATES=0` disables the background check.
+- Keep agent sessions lean: Claude runs with Studio's own tool list, no user MCP servers,
+  hooks, slash commands or CLAUDE.md; Codex gets feature/MCP `-c` overrides. Flags are
+  feature-detected per CLI binary, so unknown flags never break a run.
 - Narration (`narration.py`) uses Kokoro-82M through `kokoro-onnx` and ONNX Runtime
   on the CPU; do not add PyTorch. Model files download on demand into
   `DATA_ROOT/.studio/models/kokoro` from pinned release URLs with SHA256 checks.