[Bug]: 使用opencli chatgpt image 报错send-failed

[pengshengsongcode]

Description

使用 opencli chatgpt image 报错send-failed

Steps to Reproduce

1、opencli版本
opencli --version
1.7.8

2、chrome opencli extension 版本

3、doctor

opencli doctor
opencli v1.7.8 doctor (node v25.2.1)

[OK] Daemon: running on port 19825 (v1.7.8)
[OK] Extension: connected (v1.0.2)
[OK] Connectivity: connected in 0.1s

Everything looks good!

4、实际命令

opencli chatgpt image -v --op ~/.openclaw/media/outbound --format json "生成一张小猫的图片"

[
{
"status": "⚠️ send-failed",
"file": "📁 -",
"link": "🔗 https://chatgpt.com/new"
}
]

5、其他
node -v
v25.2.1

Expected Behavior

可以生成图片

OpenCLI Version

1.7.8

Node.js Version

Other

Operating System

macOS

Logs / Screenshots

[pengshengsongcode]

尝试使用gemini 方案，遇到了另一个问题，看 gemini 后台图片已经生成成功了，但是状态异常，以及没有下载成功

opencli gemini image -v --sd true "白色小猫咪和黑色小猫咪"
status: ⚠️ no-images
file: 📁 -
link: 🔗 https://gemini.google.com/app/24978c5cd4d9ff57

[Benjamin-eecs]

Fixed in #1218.

Three issues found:

1. ChatGPT redirects away from the conversation after sending the prompt. The image poll loop was checking for images on the home page instead of the conversation page. Fixed by capturing the /c/ URL after send and navigating back during polling.

2. The composer selector [aria-label="Chat with ChatGPT"] is not consistent across ChatGPT UI versions. Added fallback selectors.

3. The default output path ~/Pictures/chatgpt was a literal string (Node.js does not expand ~), so images were saved to a directory named ~ in the working directory.

Tested on macOS with v1.7.8. Image generation and download both work.

[pengshengsongcode]

尝试使用修复 PR，还是有相同问题

opencli doctor
opencli v1.7.8 doctor (node v25.2.1)

[OK] Daemon: running on port 19825 (v1.7.8)
[OK] Extension: connected (v1.0.2)
[OK] Connectivity: connected in 0.1s

Everything looks good!

opencli chatgpt image -v --op ~/.openclaw/media/outbound --format json "生成一张小猫的图片"
[
{
"status": "⚠️ send-failed",
"file": "📁 -",
"link": "🔗 https://chatgpt.com/new"
}
]

~/De/z/Script/OpenCLI111 | on fix/chatgpt-image !1 ok | took 7s | at 16:54:13

[Benjamin-eecs]

You're running the published v1.7.8 binary, not the PR code.

The on fix/chatgpt-image in your shell prompt is just the git branch of your cwd. It doesn't change which binary opencli resolves to. The command still resolves to /opt/homebrew/bin/opencli (or the equivalent on your system), which loads the old clis/chatgpt/ bundled inside the npm package.

To actually run the PR code, invoke it directly from the repo root via node:

node dist/src/main.js chatgpt image -v --op ~/.openclaw/media/outbound --format json "生成一张小猫的图片"

Or use npm link to point the global binary at your local checkout:

cd /path/to/OpenCLI
npm run build
npm link
opencli chatgpt image -v --op ~/.openclaw/media/outbound --format json "生成一张小猫的图片"

I tested the PR code on my own account with the same prompt and it generates and saves the image successfully. The old code only had one composer selector ([aria-label="Chat with ChatGPT"]), which doesn't hit on some ChatGPT UI variants, and that's why you saw send-failed. The PR adds three fallback selectors plus a few other fixes that should cover your case. Please retry.

[pengshengsongcode]

也是一样的问题

npm run build
npm link

@jackwener/opencli@1.7.8 build
npm run clean-dist && tsc && npm run copy-yaml && npm run build-manifest

@jackwener/opencli@1.7.8 clean-dist
node scripts/clean-dist.cjs

@jackwener/opencli@1.7.8 copy-yaml
node scripts/copy-yaml.cjs

@jackwener/opencli@1.7.8 build-manifest
node dist/src/build-manifest.js

✅ Manifest compiled: 632 entries → /Users/pengshengsong/Desktop/zhihu/Script/OpenCLI111/cli-manifest.json
✅ Restored executable permission: dist/src/main.js

up to date, audited 3 packages in 3s

found 0 vulnerabilities

opencli chatgpt image -v --op ~/.openclaw/media/outbound --format json "生成一张小猫的图片"
[
{
"status": "⚠️ send-failed",
"file": "📁 -",
"link": "🔗 https://chatgpt.com/new"
}
]

~/De/z/Script/OpenCLI111 | on fix/chatgpt-image !1

[pengshengsongcode]

也尝试使用gemini 方案，遇到了另一个问题，看 gemini 后台图片已经生成成功了，但是状态异常，以及没有下载成功

opencli gemini image -v --sd true "白色小猫咪和黑色小猫咪"
status: ⚠️ no-images
file: 📁 - 文件：📁 -
link: 🔗 https://gemini.google.com/app/24978c5cd4d9ff57

[Benjamin-eecs]

✅ Manifest compiled: 632 entries → /Users/pengshengsong/Desktop/zhihu/Script/OpenCLI111/cli-manifest.json
...
⚠️ send-failed

Thanks for the build output, that confirms opencli now resolves to the PR code and this is no longer a binary mismatch.

For context: I tested the PR on a free plan ChatGPT account and it generates and saves images successfully with the exact same Chinese prompt, so account tier is not the cause.

Since none of the three composer selectors in the PR ([aria-label="Chat with ChatGPT"], [placeholder="Ask anything"], #prompt-textarea) hit on your page, your ChatGPT must be rendering a different DOM than mine. Could you share what the automation window actually shows when this fails?

opencli browser open https://chatgpt.com/new
opencli browser state | head -80
opencli browser screenshot /tmp/chatgpt.png

Then paste the state output and attach the screenshot. That will tell me whether you're hitting a login wall, a modal, a Cloudflare challenge, or a UI variant whose selectors differ from mine, so I can extend the selector list accordingly.