sn-ppt-creative

name	sn-ppt-creative
description	Creative-mode PPT pipeline. One full-page 16:9 PNG per slide. LLM / VLM calls go through sn-ppt-standard/lib/model_client.py (shared thin client). Text-to-image (the actual png rendering) goes through sn-image-base/scripts/sn_agent_runner.py. Falls back to web image search when T2I generation fails. Expects task_pack.json + info_pack.json already written by sn-ppt-entry.
metadata	{"project":"SenseNova-Skills","tier":1,"category":"scene","user_visible":false}
triggers	["sn-ppt-creative"]

sn-ppt-creative

⚠️ This skill must be invoked through /skill sn-ppt-entry. Never start here directly — the entry skill collects parameters and writes task_pack.json + info_pack.json that this skill requires. If you arrived here without those files, stop and tell the user to enter via /skill sn-ppt-entry or "生成 PPT".

Call-routing policy

Kind	Backend
LLM (text)	$PPT_STANDARD_DIR/lib/model_client.py → llm(sys, user)
VLM (image understanding)	$PPT_STANDARD_DIR/lib/model_client.py → vlm(sys, user, images)
T2I (image generation)	$SN_IMAGE_BASE/scripts/sn_agent_runner.py sn-image-generate

Never mix — LLM / VLM through sn-image-base, or T2I through model_client — both violate policy.

Visual asset priority

• Creative mode renders each slide as a generated full-page PNG, so image generation is the first-priority visual path.
• If image generation fails for a page, use web search (sn-search-image) as a fallback to find a real image that fits the page's topic. Each search result includes the image URL, source page, title, and domain for traceability.
• Do not create placeholders. If generation and search both fail, record the page failure and continue; never write fake PNGs, grey boxes, broken-image icons, or "image pending" text.
• Do not mention the search provider name in prompts, visible slide text, progress, or summaries.

Preconditions

• <deck_dir>/task_pack.json exists and ppt_mode == "creative"
• <deck_dir>/info_pack.json exists
• <deck_dir>/pages/ exists
• $SN_IMAGE_BASE env var (OpenClaw-injected) points at the sn-image-base skill root
• $PPT_STANDARD_DIR env var points at the sn-ppt-standard skill root (so we can import model_client)

Any missing → stop and tell user to enter via /skill sn-ppt-entry.

Resume

python3 $SKILL_DIR/scripts/resume_scan.py --deck-dir <deck_dir>
# => {"style_spec_done": bool, "outline_done": bool, "pptx_done": bool,
#     "pages": [{"page_no": 1, "action": "skip|render_only|full"}, ...]}

Dispatch:

Manifest	Do
style_spec_done == false	Run Stage 2
outline_done == false	Run Stage 3
per-page action == "full"	Run Stage 4.1 + 4.2
per-page action == "render_only"	Run Stage 4.2 only (prompt.txt already on disk)
per-page action == "skip"	Skip
pptx_done == false (all pages done or failed)	Run Stage 5

Stage 2 — style_spec.md (LLM or VLM via model_client)

One independent exec tool_call. Two branches based on reference images.

Branch A (no ref images, or all missing on disk) — use model_client.llm:

python3 -c "
import sys, pathlib, json
sys.path.insert(0, '$PPT_STANDARD_DIR/lib')
from model_client import llm

deck = pathlib.Path('<deck_dir>')
tp = json.loads((deck / 'task_pack.json').read_text())
ip = json.loads((deck / 'info_pack.json').read_text())

sys_prompt = open('$SKILL_DIR/prompts/style_from_query.md').read()
user_prompt = json.dumps({
    'params': tp['params'],
    'query': ip.get('user_query'),
    'digest': ip.get('document_digest'),
}, ensure_ascii=False)

md = llm(sys_prompt, user_prompt)
(deck / 'style_spec.md').write_text(md, encoding='utf-8')
print('style_spec.md ok')
"

Branch B (≥1 reference image on disk) — use model_client.vlm:

python3 -c "
import sys, pathlib, json
sys.path.insert(0, '$PPT_STANDARD_DIR/lib')
from model_client import vlm

deck = pathlib.Path('<deck_dir>')
ip = json.loads((deck / 'info_pack.json').read_text())
tp = json.loads((deck / 'task_pack.json').read_text())

refs = [p for p in (ip.get('user_assets') or {}).get('reference_images', []) if pathlib.Path(p).exists()]

sys_prompt = open('$SKILL_DIR/prompts/style_from_image.md').read()
user_prompt = f'PPT 主题/参数: {json.dumps(tp[\"params\"], ensure_ascii=False)}\nuser_query: {ip.get(\"user_query\") or \"\"}'

md = vlm(sys_prompt, user_prompt, images=refs)
(deck / 'style_spec.md').write_text(md, encoding='utf-8')
print(f'style_spec.md ok (from {len(refs)} ref images)')
"

If user_assets.reference_images is non-empty but all paths missing on disk: fall through to Branch A and prepend a line reference_images_missing: <original paths> at the top of style_spec.md.

Stage 3 — outline.json (LLM via model_client)

python3 -c "
import sys, pathlib, json
sys.path.insert(0, '$PPT_STANDARD_DIR/lib')
from model_client import llm

deck = pathlib.Path('<deck_dir>')
tp = json.loads((deck / 'task_pack.json').read_text())
ip = json.loads((deck / 'info_pack.json').read_text())
style = (deck / 'style_spec.md').read_text()

sys_prompt = open('$SKILL_DIR/prompts/outline.md').read()
user_prompt = json.dumps({
    'style_spec_markdown': style,
    'params': tp['params'],
    'query': ip.get('user_query'),
    'digest': ip.get('document_digest'),
}, ensure_ascii=False)

raw = llm(sys_prompt, user_prompt).strip()
if raw.startswith('\`\`\`'):
    raw = raw.split('\n', 1)[1].rsplit('\`\`\`', 1)[0]
data = json.loads(raw)
assert len(data['pages']) == tp['params']['page_count'], 'page_count mismatch'
(deck / 'outline.json').write_text(json.dumps(data, ensure_ascii=False, indent=2))
print(f'outline ok, {len(data[\"pages\"])} pages')
"

On failure (non-JSON / length mismatch): abort.

Stage 4 — per-page: one independent exec per page

4.1 Compose prompt (LLM via model_client) — skip if action == "render_only"

python3 -c "
import sys, pathlib, json
sys.path.insert(0, '$PPT_STANDARD_DIR/lib')
from model_client import llm

deck = pathlib.Path('<deck_dir>')
N = <NNN>
style = (deck / 'style_spec.md').read_text()
outline = json.loads((deck / 'outline.json').read_text())
page = next(p for p in outline['pages'] if int(p['page_no']) == N)

sys_prompt = open('$SKILL_DIR/prompts/page_prompt.md').read()
user_prompt = json.dumps({'style_spec_markdown': style, 'page': page}, ensure_ascii=False)

txt = llm(sys_prompt, user_prompt)
(deck / 'pages' / f'page_{N:03d}.prompt.txt').write_text(txt, encoding='utf-8')
print(f'prompt page {N} ok')
"

# sanitize the written prompt in-place: strip hex/rgb/hsl/CSS/px/em/rem etc
# to prevent T2I server-side prompt-enhance from baking them into the image.
# Silent: no chat-facing notification; removals go to stderr only.
python3 $SKILL_DIR/scripts/sanitize_prompt.py --path <deck_dir>/pages/page_<NNN>.prompt.txt

4.2 Generate image (T2I via sn-image-base)

--negative-prompt 是针对可能带自身 prompt-enhance 的 T2I 后端的最后一道防线： 即使前面的 sanitize 没拦住、或后端重写时引入了新的样式元数据，也通过反向约束压制模型把它们画出来。这段字符串在所有页上都一致。

python $SN_IMAGE_BASE/scripts/sn_agent_runner.py sn-image-generate \
  --prompt "$(cat <deck_dir>/pages/page_<NNN>.prompt.txt)" \
  --negative-prompt "hex color code, #RRGGBB, rgb(), rgba(), hsl(), hsla(), css, json, yaml, code snippet, pixel values, px, em, rem, pt, color palette text, typography label, design spec, style guide, font stack, hex code, layout annotation, dimensional callout, figma-style spec sheet, wireframe annotation, swatch with numbers" \
  --aspect-ratio 16:9 \
  --image-size 2k \
  --save-path <deck_dir>/pages/page_<NNN>.png \
  --output-format json

4.3 Failure handling

• 4.1 failure (model timeout / empty / malformed): record page_no into failed_pages, echo failure line, continue.
• 4.2 failure: same — record, echo, continue.
• No retries. No placeholder PNG. Don't write 1x1 transparent PNGs to fake success.
• .prompt.txt may remain on disk for a later manual re-run of 4.2 only.

Stage 5 — pptx 打包（一次独立 exec）

所有页图生成后（含部分失败的情况），把 pages/page_*.png 平铺打包成 16:9 整册 PPTX，每张图满版一页。由 scripts/build_pptx.py 完成，模型只负责执行脚本。

python3 $SKILL_DIR/scripts/build_pptx.py --deck-dir <deck_dir>
# => {"deck_id": "...", "output": "<deck_dir>/<deck_id>.pptx",
#     "total_slides": N, "included_pages": [...], "missing_pages": [...]}

行为约定：

• 输出路径默认 <deck_dir>/<deck_id>.pptx；可用 --output 覆盖。
• 页序按 outline.json 的 page_no 排；缺失 outline.json 时按 page_001..page_NNN 走。
• 缺失的 PNG 会插入空白页并在 stderr 记录一行，不中止；这样跟 Stage 4 的"失败跳过"语义一致。
• 脚本失败（依赖缺失 / 写盘失败）：echo 失败原因，不中止整个 skill，仍进入 Stage 6 收尾；PNG 已在磁盘上。 如果 python-pptx 缺失导致失败：🚫 不要尝试 pip install python-pptx 或任何替代方案。PNG 页面已经是最终交付物，直接进入 Stage 6。

Stage 6 — closing

Emit:

创意模式已完成。

📁 输出目录：<deck_dir>
📄 结果文件：
  - style_spec.md
  - outline.json
  - pages/page_001.png ~ page_NNN.png（失败 M 页：page_..., page_...）
  - <deck_id>.pptx（整册，缺失页插入空白）

⚠️ 未完成：
  - page_007：生图返回超时，已跳过（pptx 中为空白页）

下一步：
  - 可直接打开 <deck_id>.pptx 查看整册
  - 或在 pages/ 目录查看 PNG

Progress echo — MANDATORY

Stage	Example
After resume_scan	已进入 sn-ppt-creative，共 N 页
After Stage 2	[1] style_spec.md ✓
After Stage 3	[2] outline.json ✓（N 页）
Per page-prompt (4.1)	[prompt 3/10] ✓
Per page-image (4.2)	[图 3/10] page_003.png ✓ or [图 3/10] ✗ 超时
After Stage 5	[pptx] <deck_id>.pptx ✓（N 页，缺失 M 页） or [pptx] ✗ <reason>
Closing	full summary above

• Each echo is a chat reply, not a log write.
• Per-page echo is the heartbeat for Stage 4.
• On failure, echo failure line with reason before moving on.

🚫 Hard rules

1. Do NOT loop inside a single exec. One page = one tool_call.
2. Do NOT fake images. Failed T2I → record failed, move on. No 1x1 placeholder PNGs.
3. Do NOT use model_client.t2i — T2I must go through sn-image-base. model_client handles only LLM / VLM.
4. Do NOT use sn-text-optimize or sn-image-recognize from sn-image-base — those must go through model_client.llm / model_client.vlm.
5. Do NOT retry on first failure. If the same stage fails twice in a row with the same error, treat it as permanent and move on.
6. Do NOT generate editable JSON from PNG (out of scope).
7. Language integrity. All user-visible text MUST match the user's query language. A single English slide in a Chinese deck is a regression.
8. Do NOT use python-pptx, pptxgenjs, or any alternative PPTX builder. scripts/build_pptx.py is the ONLY way to produce a PPTX. Never pip install python-pptx or write Node scripts that import pptxgenjs. If PPTX build fails, the PNG pages are the final deliverable.
9. Do NOT fabricate data. All numbers and factual claims MUST come from the user's documents or web search. Use qualitative descriptions if no data source is available.
10. Wait for responses. If you ask the user a question, do NOT proceed until they reply. Never assume default values.
11. Multi-round edits: regenerate. When the user requests changes, re-run the affected pipeline stages. Do NOT sed/perl/patch files in-place.
12. Validate paths before writing. All output goes under <deck_dir>/ — the absolute path written in task_pack.json. Before writing any file, verify the parent directory exists. Never write to /workspace/, /tmp/, ~/, ./, or any hallucinated path.