add: gemini-web-generate skill（整合 CLI + skill）

2026-05-14 14:45:51 +08:00
parent a23b9a5272
commit 5df8934000
11 changed files with 3498 additions and 0 deletions
@@ -0,0 +1,166 @@
+---
+name: gemini-web-generate
+description: "Generate images via Gemini web interface using a headless browser automation CLI. Supports text-to-image, image-to-image, multi-image reference, multi-turn conversations, and session management. Use when: (1) User wants to generate images with Gemini, (2) User says 'generate with Gemini' or 'Gemini 生图', (3) Image-to-image or style transfer tasks, (4) Continuing an existing Gemini image conversation."
+---
+
+# Gemini Web Image Generation
+
+Generate images through Gemini's web interface using a bundled Puppeteer-based CLI that automates the full workflow: navigation, image upload, prompt submission, download, and cleanup.
+
+## Architecture
+
+| Component | Path |
+|------|------|
+| CLI entry | `scripts/cli.js` |
+| Node binary | `/home/dazhi/.nvm/versions/node/v22.22.0/bin/node` |
+| Browser CDP | `http://127.0.0.1:9223` (managed by `browser` tool) |
+
+## Quick Start
+
+All commands use absolute paths for reliability:
+
+```bash
+NODE="/home/dazhi/.nvm/versions/node/v22.22.0/bin/node"
+CLI="<skill-dir>/scripts/cli.js"
+```
+
+### Text-to-Image (default: single mode)
+
+```bash
+$NODE $CLI generate --prompt "a cute cat" --mode single
+```
+
+### Image-to-Image
+
+```bash
+$NODE $CLI generate --prompt "convert to watercolor style" --image /path/to/ref.png --mode single
+```
+
+### Multi-image Reference
+
+```bash
+$NODE $CLI generate --prompt "blend these images" --images "/path/a.png,/path/b.png" --mode single
+```
+
+### Multi-turn Conversation
+
+```bash
+# First turn (session stays open)
+$NODE $CLI generate --prompt "draw a sunset"
+# Continue
+$NODE $CLI generate --session <id> --prompt "add a boat"
+# Last turn (auto-close)
+$NODE $CLI generate --session <id> --prompt "make it night" --mode single
+```
+
+## Core Workflow
+
+### 1. Ensure Browser is Running
+
+```
+browser action=start
+```
+
+### 2. Execute Generation
+
+Default to `--mode single` (auto-close after download). Only omit `--mode` for multi-turn sessions.
+
+```bash
+$NODE $CLI generate --prompt "..." --mode single
+```
+
+CLI handles: open tab → navigate to Gemini → paste reference images → type prompt → submit → wait for generation → download to `output/originals/` → close tab (single mode).
+
+Default timeouts: generation 300s (5min), download 120s.
+
+### 3. Handle Results
+
+**On success**: Move downloaded image and clean up:
+
+```bash
+# Move latest generated image
+LATEST=$(ls -t <skill-dir>/scripts/output/originals/ | head -1)
+mv "<skill-dir>/scripts/output/originals/$LATEST" ~/.openclaw/workspace/media/generated/
+
+# Clean up moved files
+for f in Gemini_Generated_Image_*.png generated-*.png; do
+  if [ -f "/home/dazhi/.openclaw/workspace/media/generated/$f" ]; then
+    rm "<skill-dir>/scripts/output/originals/$f" 2>/dev/null
+  fi
+done
+```
+
+**On timeout or error**: Check status:
+
+```bash
+$NODE $CLI status --session <id> --wait
+```
+
+If status returns `done`, run download:
+
+```bash
+$NODE $CLI download --session <id>
+```
+
+**Add `--screenshot` for diagnostics**: When errors/timeouts are expected, add `--screenshot` to auto-capture a screenshot on failure.
+
+### 4. Deliver to User
+
+Use `message` tool with `media` parameter to send the final image.
+
+## Secondary Workflows
+
+### Continue from Chat URL
+
+```bash
+$NODE $CLI generate --chatUrl "https://gemini.google.com/app/xxxx" --prompt "new instruction"
+```
+
+### List Active Sessions
+
+```bash
+$NODE $CLI sessions
+```
+
+### Find Lost Session
+
+```bash
+$NODE $CLI find_session --chatUrl "https://gemini.google.com/app/xxxx" --open
+```
+
+### Close a Session
+
+```bash
+$NODE $CLI close --session <id>
+```
+
+## Error Handling
+
+| Status | Action |
+|------|------|
+| CLI exits `success` | Move image from `output/originals/` |
+| CLI exits `timeout` | Run `status --session <id> --wait` to check if still generating |
+| Status returns `done` | Run `download --session <id>` |
+| Status returns `error` | Report error, suggest retry with modified prompt |
+| Status returns `generating` | Continue waiting with `--wait` |
+| Session lost | Use `find_session --chatUrl <url> --open` to recover |
+| Browser not started | Run `browser action=start` first |
+| Downloaded image not found | Check `output/originals/` directory, verify filename and size |
+
+## Critical Rules
+
+1. **Browser management**: Use `browser` tool for lifecycle; CLI only handles Gemini interaction
+2. **Always use absolute paths** for node and CLI paths
+3. **Default to `--mode single`**: Auto-close after download unless in multi-turn conversation
+4. **Multi-turn continuation**: Do NOT add `--mode` when continuing; close with `--mode single` on last turn or `close` command
+5. **Reference image path handling**:
+   - ❌ Never use absolute paths with spaces — shell splits them
+   - ✅ `cd` to reference image directory first, then use relative paths
+   - ✅ Use `ls | grep` to dynamically get filenames
+6. **Avoid numbered reference images** (e.g., `01-xxx.jpg`) — Gemini may rate-limit repeated use
+7. **Stop after 3 consecutive failures** — investigate root cause instead of retrying
+8. **Always move downloaded images** to `~/media/generated/` and clean up originals
+9. **Verify downloads**: Check filename and file size in originals before moving
+10. **Use `--screenshot`** for diagnostic capture on errors
+11. **No `--json` needed**: Read CLI text output directly
+12. **Don't use screenshot previews**: CLI downloads original images directly