docs: clarify enhance_prompt browser and OpenCode usage#15
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 58201b9834
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| "mcpServers": { | ||
| "ace-tool": { | ||
| "command": "npx", | ||
| "args": [ |
There was a problem hiding this comment.
I checked OpenCode's MCP docs before reviewing this section: local servers must be declared under mcp.<name> with type: "local" and a single command array. This example uses the Claude/Desktop-style mcpServers object plus separate command/args fields instead, so anyone who pastes it into opencode.json will not get ace-tool registered in OpenCode.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2658492137
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| IMPORTANT AFTER THE TOOL RETURNS: | ||
| - Treat the returned text as a refined version of the user's instructions. | ||
| - You MUST continue fulfilling the user's original request using that enhanced prompt. | ||
| - Do NOT stop after displaying or quoting the enhanced prompt. | ||
| - Unless the user explicitly asked only for prompt rewriting, proceed with implementation or the next task step. |
There was a problem hiding this comment.
The new post-call guidance treats every return value as a usable refined prompt, but EnhancePromptTool::execute() still converts failures into normal text results like Error: ... instead of surfacing an MCP error. If the enhancer API fails, or the browser review is cancelled/times out, an agent that follows this description will continue the task using the error message as its prompt rather than stopping and surfacing the failure to the user.
Useful? React with 👍 / 👎.
| "mcpServers": { | ||
| "ace-tool": { | ||
| "command": "npx", | ||
| "args": [ | ||
| "ace-tool-rs", |
There was a problem hiding this comment.
I checked OpenCode's MCP docs (https://opencode.ai/docs/mcp-servers): local servers must be declared under mcp.<name> with type: "local" and a single command array. This new example still uses the Claude/Desktop-style mcpServers object plus separate command/args, so readers who paste it into opencode.json will not get ace-tool registered in OpenCode.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 528ad108ec
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| - You MUST continue fulfilling the user's original request using that enhanced prompt. | ||
| - Do NOT stop after displaying or quoting the enhanced prompt. | ||
| - Unless the user explicitly asked only for prompt rewriting, proceed with implementation or the next task step. |
There was a problem hiding this comment.
In MCP clients that follow tool descriptions literally, this new post-call rule can override explicit staging instructions from the user. A request like “enhance this prompt, show it to me, and wait for approval before changing files” is not “only prompt rewriting”, so the current wording tells the agent it MUST continue and Do NOT stop, which can skip the approval checkpoint and start implementation earlier than requested.
Useful? React with 👍 / 👎.
1 participant
Motivation
enhance_promptappeared to hang after the MCP tool call because the tool by default opens a local Web UI and waits for a browser confirmation, so documentation should explain the default flow and the non-browser option.Description
OpenCodesection to bothREADME.mdandREADME-zh-CN.mdshowing a recommended--no-webbrowser-enhance-promptconfiguration example and documenting that the default behavior calls the prompt-enhancer API then opens a local Web UI for user review.Testing
git diff --checkwhich returned no problems.Codex Task