pi-prompt-reviewer0.3.0

pi-prompt-reviewer

Clear prompts lead to better coding results, but typos, awkward wording, and small misunderstandings can waste precious context, especially for people writing in a second language. This extension puts a fast, lightweight reviewer in front of your main pi session so weak prompts get cleaned up before they create extra churn, making prompt review easy enough to become a habit and helping you avoid the frustration of watching a long task fail because the original request was unclear.

pi-prompt-reviewer is a pi extension that reviews your prompt before it is sent to the main session.

Features

• intercepts normal prompts before they are sent
• rewrites prompts for clarity while preserving intent
• can include recent conversation context when useful
• lets you choose the target language, reviewer model, and thinking level
• remembers the target language, reviewer model, and thinking level across sessions
• loads the reviewed prompt back into the editor automatically
• makes Enter insert a new line by default to prevent accidental submissions
• keeps reviewed submission available through Ctrl+Shift+Enter regardless of the Enter behavior
• reviews follow-up messages submitted with Alt+Shift+Enter
• provides configurable shortcuts for reviewed and unreviewed submission
• lets you submit without review via Ctrl+Enter by default
• lets you queue a follow-up without review via Alt+Enter by default
• lets you restore the original prompt with a command or shortcut
• displays token usage and cost for the review step

Install

pi install npm:pi-prompt-reviewer

After installing or editing the extension, reload pi:

/reload

How it works

With the default configuration:

1. Type a normal prompt.
2. Press Ctrl+Shift+Enter.
3. The extension reviews it with a lightweight model.
4. The reviewed prompt is loaded back into the editor.
5. A review widget appears above the editor.
6. Press Ctrl+Shift+Enter to send the reviewed prompt, restore the original first, or press Ctrl+Enter to submit the current editor contents without review.

Enter and Shift+Enter both insert a new line by default. The configured submitForReview shortcut always starts review or sends the reviewed prompt, regardless of the Enter behavior.

Configuration

The extension reads persistent settings from:

<agent-dir>/extensions/prompt-reviewer.json

The agent directory defaults to ~/.pi/agent. Copy examples/prompt-reviewer.json to that path or add the new fields to your existing file:

{
  "targetLanguage": "match input",
  "reviewerModel": "auto",
  "reviewerThinking": "off",
  "preventAccidentalSubmission": true,
  "keybindings": {
    "submitForReview": "ctrl+shift+enter",
    "submitWithoutReview": "ctrl+enter",
    "submitFollowUpWithoutReview": "alt+enter"
  }
}

preventAccidentalSubmission defaults to true. It controls only the plain Enter key:

• when true, Enter inserts a new line, just like Shift+Enter
• when false, Enter submits the editor contents for review

The configured submitForReview shortcut remains active in both modes. The submission defaults are:

• Ctrl+Shift+Enter: submit for review, or send the reviewed prompt
• Ctrl+Enter: submit without review; while the agent is running, send it as a steering message
• Alt+Enter: queue a follow-up without review
• Alt+Shift+Enter: submit a follow-up for review

Use different shortcuts for the three configurable actions. Alt+Shift+Enter is reserved for reviewed follow-ups, and other keys already reserved by pi should also be avoided. Modified Enter combinations depend on terminal support. Key names use pi syntax, such as ctrl+enter or alt+shift+enter.

After changing the file, run /reload.

Bypasses

These inputs are not reviewed:

• slash commands such as /help
• ! shortcuts
• prompts with image attachments

To skip review once for a normal prompt, prefix it with a backslash:

\send this directly without review

To submit the current editor contents immediately without review, use the configured submitWithoutReview shortcut. The default is:

Ctrl+Enter

To queue the current editor contents as a follow-up without review, use the configured submitFollowUpWithoutReview shortcut. The default is:

Alt+Enter

Usage

Submit follow-up messages

While the agent is running, press Alt+Shift+Enter to submit a follow-up for review. After the reviewed text is loaded into the editor, press Ctrl+Shift+Enter to queue it while preserving its follow-up behavior.

Press Alt+Enter to queue the current editor contents as a follow-up without review.

Enable or disable prompt review

/prompt-review on
/prompt-review off
/prompt-review toggle

Show status or help

/prompt-review status
/prompt-review help

Restore the original prompt after review

/prompt-review revert

Default shortcut:

Ctrl+Alt+R

Configure context mode

/prompt-review context
/prompt-review context off
/prompt-review context always

Context modes:

• off: do not send recent conversation context
• always: always send the previous user prompt and last assistant reply when they exist

Configure target language

/prompt-review language
/prompt-review language match input
/prompt-review language English
/prompt-review language Brazilian Portuguese

The default is match input, which keeps the reviewed prompt in the same language as your input prompt. Set a specific language to have the reviewer translate the reviewed prompt as needed.

Configure reviewer model

/prompt-review model
/prompt-review model auto
/prompt-review model <model-pattern>

Examples:

/prompt-review model openai-codex/gpt-5.4-mini
/prompt-review model haiku

Notes:

• auto prefers a lightweight available model
• the default auto-selected model may not be supported by your subscription
• explicit reviewer model changes are tested before they are saved
• if the test fails, the extension warns you and keeps the previous reviewer model

Configure reviewer thinking

/prompt-review thinking
/prompt-review thinking off
/prompt-review thinking minimal
/prompt-review thinking low
/prompt-review thinking medium
/prompt-review thinking high
/prompt-review thinking xhigh

Recommended default:

• model: auto
• thinking: off

This is usually the best balance of speed, cost, and review quality.

Thinking changes are also tested before they are saved. If the test fails, the extension warns you and keeps the previous reviewer thinking level.

Target language, reviewer model, reviewer thinking, accidental-submission prevention, and all three configurable submission keybindings are loaded from the configuration file. The enabled/disabled state and context mode remain session-specific.

Retry behavior

If the first reviewer run returns no text, the extension retries once using the current session model with thinking set to off.