REPL basics¶

Launch¶

$ hai

To exit, use /quit or Ctrl + D.

Chat¶

Any message that isn't a /command or !tool is a chat message that triggers a chat response from the LLM.

[0] hi

↓↓↓

Hello! How can I help you today?

REPL Prompt¶

On the left-hand side is the index of the message in the conversation. For example, the first message is always [0] or [1] if a system-prompt is set.

On the right-hand side is a set of status information:

The current working directory which is important for tools that use the local filesystem.
The token count of the conversation. See Cost estimation.
The active LLM model.
The current timestamp.

New Chat¶

To start a new conversation:

/new
/n -- shortcut

This removes all messages unless the REPL is in task mode.

Reset Chat¶

Resetting a conversation removes all conversational messages but keeps loaded data (/load, /load-url, /asset-load):

/reset
/r -- shortcut

Load a file¶

To load a file into the chat conversation, use:

/load <path>

To be vigilant about token usage, check the token count on the status line.

Load a URL¶

To load a URL into the chat conversation, use:

/load-url <url>

PNG and JPG responses will be rendered in the terminal and can be used with image-capable LLM models.

HTML responses are automatically converted to markdown for significantly improved token efficiency. If this is undesirable, use:

/load-url(raw) <url>

Execute program and load stdout/stderr¶

To execute a program, use:

/exec <prog>

Both the standard output and standard error will be added to the chat conversation. If the status code is non-zero, it is added to the conversation as well.

The LLM is not prompted for a response until your next message.

The shortcut for execution is double-exclamation:

!!<prog>

Programs are executed using bash or powershell on Windows but can be changed in default shell.

Example¶

[0]: /exec ls ~/Documents

OR

[0]: !!ls ~/Documents

fetch_and_analyze_tweets.py
psql_index_notes.txt
...

[2]: what do my docs say about me?

↓↓↓

You engage in technical, organizational, and analytical work. You handle SQL, PostgreSQL, and security topics.

Change working directory¶

To change your current working directory, use:

/cd <path>

Changing the working directory is important for loading files, executing programs, and using tools that interact with your local filesystem.

Switch LLM model¶

To switch LLM models, use:

/ai <model>

For information on available models, see LLM models.

Forget a message¶

If you make a mistake, you can forget the last interaction with:

/forget [<n>]

By default, one interaction is removed from the conversation, but n can be specified to forget multiple interactions at once.

For the opposite effect—retaining interactions instead of removing them—see /keep.

Tab completion¶

hai supports tab completion for commands and some arguments. When tab completion is unavailable, it falls back to local file path completion.

Command history¶

Command history is stored in ~/.hai/history. To access it in the REPL:

Arrow up/down cycles through recently used commands.
Ctrl + R performs a reverse-search with partial string matching.

Save and resume chats¶

Resume your last chat with:

/chat-resume

When is a chat saved?

Your last chat is saved locally whenever you exit hai or start a new conversation with /new or /reset.

To save a chat for the long term as an asset:

/chat-save

By default, chats are saved as assets named chat/<timestamp>. A descriptive title is automatically generated and stored in the asset metadata for easier discovery. For example:

[0] /ls
chat/2025-04-08-203003 [Public/Private Key Management for Encryption and Signing]

Resume a saved chat:

/chat-resume <name>

Save with a custom name:

/chat-save [<name>]

Cost estimation¶

The REPL's right-hand prompt shows the number of tokens in the current conversation. To understand how much the converation has cost so far and what the input cost is for the next prompt, use:

/cost

This is especially useful when files, urls, images, and assets have been loaded for context.

Tokenizer

Token counts are estimated using the GPT-3.5 tokenizer because of its smaller size and therefore faster loading time. Unscientifically, token counts can differ by as much as 20%.

Tokens for images

Token counts for images are only accurate for OpenAI models because all images are assumed to consume the hard coded number of tokens for a "low detail" image in the OpenAI API.

Pin and prep messages¶

To add a message that doesn't prompt the LLM and is retained on /reset, use:

/pin <message>

To add a message that doesn't prompt the LLM but is cleared on /reset, use:

/prep <message>

Using /prep directly is uncommon. A shortcut for /prep is to add two blank newlines at the end of a message. This is helpful if you want to add more content (such as a copy-and-paste) before requesting a response.

System prompt¶

To set a system prompt, use:

/system-prompt <message>

Copy to clipboard¶

To copy the last message verbatim to your machine's clipboard, use:

/clip

For a version that lets you prompt the LLM to add to your clipboard, see !clip.