aidentity / index

OpenSenseNova/SenseNova-Skills · MIT

sn-ppt-workbench

Use when the user asks to open, preview, inspect, or visually edit an existing SenseNova HTML presentation in the PPT Workbench, using an explicit deck_dir or the latest deck in the current workspace.

stable

Contents

sn-ppt-workbench

Start or reuse the AI PPT editing WebUI for an existing HTML slide deck. The

helper prefers Hermes/canvas/port-forward URLs when available. Without a

forwarded URL, native hosts bind to localhost by default, while Docker/WSL bind

to 0.0.0.0 so remote users can reach the WebUI through the host network.

Use Cases

  • The user already generated a deck and asks to open the WebUI/editor/workbench.
  • The user provides a deck_dir containing task_pack.json,

slides/slide_NN.html, pages/page_NNN.html, or a root-level HTML deck.

  • The user says to inspect or edit previously generated slides without regenerating them.

Hard Rules

  1. Do not run sn-ppt-entry, sn-ppt-standard, page-html, batch-page-html, or export.
  2. Do not edit slide files directly. The WebUI handles user-visible editing.
  3. If deck_dir is missing, auto-detect only under $(pwd)/ppt_decks.

Other roots are considered only when the caller explicitly passes

--search-root; never scan home folders by default.

  1. If no usable deck is found, ask the user for the absolute deck_dir.
  2. Workbench startup is best-effort. If it skips or fails, report the reason and do not attempt generation as a fallback.
  3. Prefer a forwarded/canvas URL over localhost when the user may be on SSH, IM, or another remote device.
  4. When launched from the unified PPT flow, use the exact absolute

task_pack.deck_dir; opening the Workbench must not create or switch tasks.

  1. Treat task_pack.json as progress/resume state and outline.md as the

user-editable Story. Existing runtime builds may expose only part of these

artifacts; report compatibility gaps instead of regenerating the deck.

Invocation

Use the helper directly:

python3 $SKILL_DIR/scripts/open_workbench.py --deck-dir "<deck_dir>" --agent-session-id "${HERMES_SESSION_KEY:-}"

On native Windows where python3 is unavailable:

python $SKILL_DIR/scripts/open_workbench.py --deck-dir "<deck_dir>" --agent-session-id "%HERMES_SESSION_KEY%"

If the user did not provide a deck path, let the helper auto-detect:

python3 $SKILL_DIR/scripts/open_workbench.py --agent-session-id "${HERMES_SESSION_KEY:-}"

Pass same-session bridge details when available:

python3 $SKILL_DIR/scripts/open_workbench.py \
  --deck-dir "<deck_dir>" \
  --agent-session-id "${HERMES_SESSION_KEY:-}" \
  --agent-managed 1 \
  --agent-provider hermes \
  --agent-transport gateway \
  --webui-base-url "${HERMES_WEBUI_BASE_URL:-}" \
  --gateway-base-url "${HERMES_GATEWAY_BASE_URL:-}"

--agent-transport may be omitted: the helper auto-selects gateway when the

Hermes Gateway endpoint and key resolve (environment or Hermes .env files),

then webui / acp / rest when only those details exist. Explicit flags and

WORKBENCH_AGENT_TRANSPORT always win. With no resolvable bridge the bottom

chat stays read-only by design.

For non-Hermes interactive providers, pass structured bridge details instead of

scraping CLI/TUI output:

python3 $SKILL_DIR/scripts/open_workbench.py \
  --deck-dir "<deck_dir>" \
  --agent-managed 1 \
  --agent-provider openclaw \
  --agent-transport rest \
  --agent-base-url "${WORKBENCH_AGENT_BASE_URL:-}" \
  --agent-api-key "${WORKBENCH_AGENT_API_KEY:-}"

python3 $SKILL_DIR/scripts/open_workbench.py \
  --deck-dir "<deck_dir>" \
  --agent-managed 1 \
  --agent-provider codex \
  --agent-transport acp \
  --acp-command "${WORKBENCH_ACP_COMMAND:-codex-acp}"

python3 $SKILL_DIR/scripts/open_workbench.py \
  --deck-dir "<deck_dir>" \
  --agent-managed 1 \
  --agent-provider box-agent \
  --agent-transport acp \
  --acp-command "${WORKBENCH_ACP_COMMAND:-${BOX_AGENT_ACP_COMMAND:-box-agent-acp}}"

When this helper runs inside Box Agent, the exported BOX_AGENT_PYTHON or

BOX_AGENT_SKILL_TOOLS_ROOT environment automatically selects box-agent,

acp, and the installed box-agent-acp command. Explicit

WORKBENCH_AGENT_* flags or environment variables still take precedence.

For WorkBuddy, pass --agent-provider workbuddy only as provider identity. Tell

the user clearly that WorkBuddy currently has no supported API or ACP interface

for the interactive workbench bridge, so the WebUI can show WorkBuddy status but

cannot send bottom-chat modification turns back to WorkBuddy.

CLI Subcommands (direct launch fallback)

When bypassing open_workbench.py and invoking the workbench launcher directly, use subcommands:

node <launcher> start   --deck-dir "<path>" [--port 0] [--agent-managed]
node <launcher> status  --deck-dir "<path>"
node <launcher> stop    --deck-dir "<path>"

The old argument-only form (no subcommand) is deprecated — it will print a usage error.

Restart pitfall: start silently reuses an existing server instance for the same deck-dir (response contains "reused": true). If you updated the workbench code, you MUST run stop first, then start — otherwise the old process keeps serving stale code:

node <launcher> stop  --deck-dir "<path>"
node <launcher> start --deck-dir "<path>" --port 0

Bind Host

The workbench resolves its bind address from --host CLI flag → WORKBENCH_HOST env var → default 127.0.0.1. When the user needs remote/LAN access, set WORKBENCH_HOST=0.0.0.0:

export WORKBENCH_HOST=0.0.0.0
node <launcher> start --deck-dir "<path>" --port 0

The response url field will then show the LAN IP instead of 127.0.0.1.

Output Handling

Parse the single JSON line from stdout.

When status == "ok", echo the returned editor_url, public_url, or workbench.editorUrl for existing-deck editing requests. The helper may also return generation_url; that URL is the generation progress page at /progress, while the editor is /editor.

PPT 编辑工作台已启动:<editor_url>

If workbench.reused == true, say it reused the existing server.

If bind_host is 0.0.0.0, the URL is a Docker/WSL LAN fallback. Tell the user it must be reachable from their device/network; if not, ask them to provide a Hermes/canvas/port-forward URL and rerun with --public-url, or an explicit bind host with --host.

When status == "skipped" or status == "failed", echo the reason field. If reason == "nodejs_missing", ask whether the user wants NodeJS/dependencies installed; if they decline, do not open the WebUI. Ask for a valid deck_dir only if the reason says no deck was found.

Environment

The helper locates the editor launcher in this order:

  1. SENSENOVA_PPT_WORKBENCH_CLI
  2. PPT_WORKBENCH_CLI
  3. The runtime packaged beside this Skill
  4. The nearest src/ppt-editor source checkout or packaged release in the current worktree
  5. Common ~/Repository/ppt-editor fallbacks

The helper does not build the editor during an end-user launch. Packaged Skills

must include workbench-runtime; source checkouts must build src/ppt-editor

explicitly before using its launcher. The repository-level webui/ directory

is a separate product and is not a launcher fallback for this Skill.

Remote URL discovery order:

  1. Explicit --public-url
  2. WORKBENCH_PUBLIC_URL, HERMES_WORKBENCH_PUBLIC_URL
  3. Hermes/canvas/forwarding env vars such as HERMES_CANVAS_URL, HERMES_PORT_FORWARD_URL, CANVAS_URL, PORT_FORWARD_URL, TUNNEL_URL
  4. Codespaces/Gitpod URL patterns
  5. Localhost on native hosts, or LAN fallback from the machine's non-loopback IPv4 address on Docker/WSL

Gateway Auth Setup

For writable interactive agent bridging (chat panel in workbench), Hermes keeps

using the existing WebUI/Gateway protocol. OpenClaw should be exposed through a

Gateway/REST-compatible HTTP endpoint. Codex, Claude Code, and Box Agent should

use ACP JSON-RPC adapters, configured with --acp-command or

WORKBENCH_ACP_COMMAND. Box Agent defaults to its installed box-agent-acp

entry point.

Never parse provider CLI/TUI output for chat synchronization.

For writable Hermes Gateway bridging, the Gateway API server and workbench must

authenticate with matching API keys.

Key env vars:

VariableSet ByValue
API_SERVER_KEYGateway (.env)Shared secret
WORKBENCH_GATEWAY_API_KEYWorkbench envSame as API_SERVER_KEY; preferred explicit bridge key
HERMES_GATEWAY_BASE_URLWorkbench envhttp://127.0.0.1:8642
WORKBENCH_AGENT_PROVIDERWorkbench envhermes, openclaw, codex, claude-code, box-agent, or workbuddy
WORKBENCH_AGENT_TRANSPORTWorkbench envwebui, gateway, rest, or acp
WORKBENCH_AGENT_BASE_URLWorkbench envOpenClaw/Gateway REST endpoint
WORKBENCH_AGENT_API_KEYWorkbench envREST provider bearer token
WORKBENCH_ACP_COMMANDWorkbench envACP adapter command, e.g. codex-acp
BOX_AGENT_ACP_COMMANDBox Agent / Workbench envOptional override for the default box-agent-acp adapter command

Quick steps:

  • Gateway: set API_SERVER_KEY in .env, restart gateway
  • Workbench: set WORKBENCH_GATEWAY_API_KEY to same value, set HERMES_GATEWAY_BASE_URL=http://127.0.0.1:8642
  • If no explicit bridge key is set, the Node relay falls back to API_SERVER_KEY, HERMES_GATEWAY_API_KEY, then backend-only AI_GATEWAY_API_KEY (ppt-editor-dev by default). Prefer the explicit bridge key for custom/shared gateways.

Progress Echo

Always send a short progress message before and after the helper:

WhenExample
Before helper正在启动 PPT 预览工作台,不会重新生成幻灯片...
SuccessPPT 编辑工作台已启动:http://127.0.0.1:18087
Generation progress URL available生成进度工作台:http://127.0.0.1:18087/progress
Skip/failurePPT 工作台未启动:<reason>

Provenance

Authored by OpenSenseNova in OpenSenseNova/SenseNova-Skills. The content is mirrored here under its upstream license.

Related

lead-magnets When the user wants to create, plan, or optimize a lead magnet for email capture or lead generation. Also use when the user mentions "lead magnet," "gated content," "content upgrade," "downloadable," "ebook," "cheat sheet," "checklist," "template download," "opt-in," "freebie," "PDF download," "resource library," "content offer," "email capture content," "Notion template," "spreadsheet template," "what should I give away for emails," "course as a skill," or "package our content as an agent skill." Use this for planning what to create and how to distribute it. For interactive tools as lead magnets, see free-tools. For writing the actual content, see copywriting. For the email sequence after capture, see emails. coreyhaines31/marketingskills MIT wizard Generate an interactive bash wizard that walks a human through steps only they can perform. Use when provisioning infrastructure, setting up credentials or CI secrets, walking an unfamiliar third-party dashboard, or running a one-off migration or cutover. Don't invoke this for steps the agent can perform itself. mattpocock/skills MIT sn-da-excel-workflow Excel 数据分析多步编排器。覆盖:(1) 读取多 Sheet Excel 文件并统计行数,(2) 大文件检测(≥10k 行自动 Parquet 优化),(3) 数据清洗(缺失值、文本标准化、无效字符),(4) 条件筛选与分类提取,(5) 跨 Sheet 统计聚合,(6) 导出 Excel/CSV 并提供下载链接。覆盖从数据读取到报告生成全流程,按步骤编排 capability 子 skill。**遇到以下任一情况就主动使用本 skill,不要自行写几行 pandas 就回答**:①用户出现触发词:Excel 分析 / 表格分析 / 数据分析 / 数据清洗 / 数据统计 / 数据筛选 / 数据可视化 / 数据导出 / 汇总统计 / 透视表 / 分组统计 / 交叉分析 / 趋势分析 / 对比分析 / 异常值检测 / 去重 / 缺失值处理 / Excel 报告 / 生成报表 / analyze Excel / data analysis / data cleaning / pivot table;②用户上传或指定了 .xlsx / .xls / .csv 文件并要求分析、清洗、统计或可视化;③任务涉及多 Sheet 读取、条件筛选、分类汇总、图表生成中的任意一项;④用户要求导出带格式的 Excel 报告或下载链接。仅不用于:不涉及表格数据的纯文本处理、图片分析(使用 sn-da-image-caption)、单个公式计算的简单问答。 OpenSenseNova/SenseNova-Skills MIT large-file-parquet-analysis-and-highlight 当Excel文件总行数超过1万行时,通过转换为Parquet格式提升读取性能,提取目标指标并计算最大值,最后将结果输出为Excel并对特定行进行高亮标注。 OpenSenseNova/SenseNova-Skills MIT excel-conditional-comparison-and-large-file-processing 对比Excel多表中的特定系数并对异常值进行颜色标记。 OpenSenseNova/SenseNova-Skills MIT