author

date

title

mountaineerbr

January 2025

CHATGPT.SH(1) v0.91 \| General Commands Manual

NAME

chatgpt.sh -- Wrapper for ChatGPT / DALL-E / Whisper / TTS

SYNOPSIS

DESCRIPTION

This script acts as a wrapper for ChatGPT, DALL-E, Whisper, and TTS endpoints from OpenAI. Additional service providers such as LocalAI, Ollama, Anthropic, Mistral AI, GoogleAI, Groq AI, GitHub Models, Novita, and xAI APIs are also available.

With no options set, complete INPUT in single-turn mode of the native chat completion.

Handles single-turn and multi-turn modes, pure text and native chat completions, image generation and editing, speech-to-text, and text-to-speech.

Positional arguments are read as a single PROMPT, but some functions such as Whisper and TTS may handle optional positional parameters before the prompt itself.

Chat Completion Modes

Set option -c to start a multi-turn chat mode via text completions with command line history support. This option works with instruct models, defaults to gpt-3.5-turbo-instruct if none set.

Set options -cc to start the chat mode via native chat completions. This mode defaults to the gpt-4o model, which is optimised to follow instructions. Try chatgpt-4o-latest for a model optimised for chatting.

In chat mode, some options are automatically set to un-lobotomise the bot.

While using other providers, mind that options -c and -cc set different endpoints! This setting must be set according to the model capabilities!

Set option -C to resume (continue from) last history session, and set option -E to exit on the first response (even in multi turn mode).

Text Completion Modes

Option -d starts a single-turn session in plain text completions, no history support. This does not set further options automatically, such as instruction or temperature.

To run the script in text completion in multi-turn mode and history support, set command line options -dd.

Set text completion models such as gpt-3.5-turbo-instruct.

Insert Modes (Fill-In-the-Middle)

Set option -q for insert mode in single-turn and option -qq for multiturn. The flag “[insert]” must be present in the middle of the input prompt. Insert mode works completing between the end of the text preceding the flag, and ends completion with the succeeding text after the flag.

Insert mode works with `instruct’ and Mistral `code’ models.

Instruction Prompts

The SYSTEM INSTRUCTION prompt may be set with option -S or via envars $INSTRUCTION and $INSTRUCTION_CHAT.

Option -S sets an INSTRUCTION prompt (the initial prompt) for text cmpls, and chat cmpls. A text file path may be supplied as the single argument. Also see CUSTOM / AWESOME PROMPTS section below.

In multi-turn interactions, prompts prefixed with a single colon “:” are appended to the current request buffer as user messages without making a new API call. Conversely, prompts starting with double colons “::” are appended as instruction / system messages. For text cmpls only, triple colons append the text immediately to the previous prompt without a restart sequence.

To create and reuse a custom prompt, set the prompt name as a command line option, such as “-S .[_prompt_name_]” or “-S ,[_prompt_name_]”.

When the operator is a comma “,”, single-shot editing will be available after loading the prompt text. Use “,,” to actually edit the template file itself!

Note that loading a custom prompt will also change to its respectively-named history file.

Alternatively, set the first positional argument with the operator and the prompt name after any command line options, such as “.[_prompt_name_]”. This loads the prompt file unless instruction was set with command line options.

To insert the current date and time to the instruction prompt, set command line option --time.

Commands

If the first positional argument of the script starts with the command operator forward slash “/” and a history file name, the command ``/session [HIST_NAME]’’ is assumed. This will change to or create a new history file (with options -ccCdPP).

If a plain text or PDF file path is set as the last positional parameter or as an argument to option -S (set instruction prompt), the file is loaded as text PROMPT.

With vision models, append the image file paths and possibly URLs at the end of the prompt.

In chat mode, a PDF, DOC, TXT or URL filepath at the very end of the user prompt is appended as text dump to the current user input.

Make sure file paths containing spaces are backslash-escaped!

Model and Capacity

Set model with ``-m [MODEL]’’, with MODEL as its name, or set it as “.” to pick from the model list. List available models with option -l.

Set maximum response tokens with option “-NUM” or “-M NUM”. This defaults to 1024 tokens and 25000 for reasoning models.

If a second NUM is given to this option, maximum model capacity will also be set. The option syntax takes the form of “-NUM/NUM”, and “-M NUM-NUM”.

Model capacity (maximum model tokens) can be set more intuitively with option “-N NUM”, otherwise model capacity is set automatically for known models or to 2048 tokens as fallback.

List models with option -l or run /models in chat mode.

Option -y sets python tiktoken instead of the default script hack to preview token count. This option makes token count preview accurate and fast (we fork tiktoken as a coprocess for fast token queries). Useful for rebuilding history context independently from the original model used to generate responses.

Image Generations and Edits (Dall-E)

Option -i generates images according to text PROMPT. If the first positional argument is an IMAGE file, then generate variations of it. If the first positional argument is an IMAGE file and the second a MASK file (with alpha channel and transparency), and a text PROMPT (required), then edit the IMAGE according to MASK and PROMPT. If MASK is not provided, IMAGE must have transparency.

The size of output images may be set as the first positional parameter in the command line: “256x256” (S), “512x512” (M), “1024x1024” (L), “1792x1024” (X), and “1024x1792” (P).

The parameter “hd” may also be set for image quality (Dall-E-3), such as “Xhd” or “1792x1024hd”. Defaults=1024x1024.

For Dalle-3, optionally set the generation style as either “natural” or “vivid” as a positional parameter in the command line invocation.

See IMAGES section below for more information on inpaint and outpaint.

Speech-To-Text (Whisper)

Option -w transcribes audio speech from mp3, mp4, mpeg, mpga, m4a, wav, webm, flac and ogg files. First positional argument must be an AUDIO file. Optionally, set a TWO-LETTER input language (ISO-639-1) as the second argument. A PROMPT may also be set to guide the model’s style, or continue a previous audio segment. The text prompt should match the speech language.

Note that option -w can also be set to translate speech input to any text language to the target language.

Option -W translates speech stream to English text. A PROMPT in English may be set to guide the model as the second positional argument.

Set these options twice to have phrasal-level timestamps, options -ww and -WW. Set thrice for word-level timestamps.

Combine options -wW with options -cc to start chat with voice input (Whisper) support. Additionally, set option -z to enable text-to-speech (TTS) models and voice out.

TTS (Text-To-Voice)

Option -z synthesises voice from text (TTS models). Set a voice as the first positional parameter (“alloy”, “echo”, “fable”, “onyx”, “nova”, or “shimmer”). Set the second positional parameter as the voice speed (0.25 - 4.0), and, finally the output file name or the format, such as “./new_audio.mp3” (“mp3”, “opus”, “aac”, and “flac”), or “-” for stdout. Set options -vz to not play received output.

Audio Models

Audio models, such as `gpt-4o-audio’, deal with audio input and output directly.

To activate the microphone recording function of the script, set command line option -w.

Otherwise, the audio model accepts any compatible audio file (such as mp3, wav, and opus). These files can be added to be loaded at the very end of the user prompt or added with chat command ``/audio path/to/file.mp3’’.

To activate the audio synthesis output mode of an audio model, make sure to set command line option -z!

General Considerations

User configuration is kept at “~/.chatgpt.conf”. Script cache is kept at “~/.cache/chatgptsh/”.

The moderation endpoint can be accessed by setting the model name to omni-moderation-latest (or text-moderation-latest).

Stdin text is appended to PROMPT, to set a single PROMPT.

Press <CTRL-X CTRL-E> to edit command line in text editor from readline.

Press <CTRL-J> or <CTRL-V CTRL-J> for newline in readline.

Press <CTRL-L> to redraw readline buffer (user input) on screen.

During cURL requests, press <CTRL-C> once to interrupt the call.

Press <CTRL-\> to exit from the script (send QUIT signal), or “Q” in user confirmation prompts.

A personal OpenAI API is required, set it with option --api-key. See also ENVIRONMENT section.

For complete model and settings information, refer to OpenAI API docs at https://platform.openai.com/docs/.

See the online man page and chatgpt.sh usage examples at: https://gitlab.com/fenixdragao/shellchatgpt.

TEXT / CHAT COMPLETIONS

1. Text Completion Modes

Given a prompt, the model will return one or more predicted completions. For example, given a partial input, the language model will try completing it until probable “<|endoftext|>”,

or other stop sequences (stops may be set with `-s [stop]’).

Restart and start sequences may be optionally set. Restart and start sequences are not set automatically if the chat mode of text completions is not activated with option -c.

Readline is set to work with multiline input and pasting from the clipboard. Alternatively, set option -u to enable pressing <CTRL-D> to flush input! Or set option -U to set cat command as input prompter.

Bash bracketed paste is enabled, meaning multiline input may be pasted or typed, even without setting options -uU (v25.2+).

Language model SKILLS can be activated with specific prompts, see https://platform.openai.com/examples.

2. Chat Modes

2.1 Text Completions Chat

Set option -c to start chat mode of text completions. It keeps a history file, and keeps new questions in context. This works with a variety of models. Set option -E to exit on response.

2.2 Native Chat Completions

Set the double option -cc to start chat completions mode. More recent models are also the best option for many non-chat use cases.

2.3 Q & A Format

The defaults chat format is “Q & A”. The restart sequence “\nQ: ” and the start text “\nA:” are injected for the chat bot to work well with text cmpls.

In native chat completions, setting a prompt with “:” as the initial character sets the prompt as a SYSTEM message. In text completions, however, typing a colon “:” at the start of the prompt causes the text following it to be appended immediately to the last (response) prompt text.

2.4 Voice input (Whisper), and voice output (TTS)

The options -ccwz may be combined to have voice recording input and synthesised voice output, specially nice with chat modes. When setting flag -w or flag -z, the first positional parameters are read as Whisper or TTS arguments. When setting both flags -wz, add a double hyphen to set first Whisper, and then TTS arguments.

Set chat mode, plus Whisper language and prompt, and the TTS voice option argument:

chatgpt.sh -ccwz  en 'whisper prompt'  --  nova

2.5 Vision and Multimodal Models

To send an image or url to vision models, either set the image with the ``!img’’ command with one or more filepaths / urls.

chatgpt.sh -cc -m gpt-4-vision-preview '!img path/to/image.jpg'

Alternatively, set the image paths / urls at the end of the text prompt interactively:

chatgpt.sh -cc -m gpt-4-vision-preview

[...]
Q: In this first user prompt, what can you see?  https://i.imgur.com/wpXKyRo.jpeg

2.6 Text, PDF, Doc, and URL Dumps

The user may add a filepath or URL to the end of the prompt. The file is then read and the text content appended to the user prompt. This is a basic text feature that works with any model.

chatgpt.sh -cc

[...]
Q: What is this page: https://example.com

Q: Help me study this paper. ~/Downloads/Prigogine\ Perspective\ on\ Nature.pdf

In the second example, the PDF will be dumped as text.

For PDF text dump support, poppler/abiword is required. For doc and odt files, LibreOffice is required. See the Optional Packages section.

Also note that file paths containing white spaces must be blackslash-escaped, or the file path must be preceded by a pipe `|’ character.

Multiple images and audio files may be appended the the prompt in this way!

2.7 Command List

While in chat mode, the following commands can be invoked in the new prompt to execute a task or set parameters. The command operator may be either “!” or “/”.

Misc	Commands
`-S`	`:`, `::` [PROMPT]	Append user or system prompt to request buffer.
`-S.`	`-.` [NAME]	Load and edit custom prompt.
`-S/`	`!awesome` [NAME]	Load and edit awesome prompt (english).
`-S%`	`!awesome-zh` [NAME]	Load and edit awesome prompt (chinese).
`-Z`	`!last`	Print last response JSON.
`!#`	`!save` [PROMPT]	Save current prompt to shell history. ‡
`!`	`!r`, `!regen`	Regenerate last response.
`!!`	`!rr`	Regenerate response, edit prompt first.
`!g:`	`!!g:` [PROMPT]	Ground user prompt with search results. ‡
`!i`	`!info`	Information on model and session settings.
`!!i`	`!!info`	Monthly usage stats (OpenAI).
`!j`	`!jump`	Jump to request, append start seq primer (text cmpls).
`!!j`	`!!jump`	Jump to request, no response priming.
`!cat`	-	Cat prompter as one-shot, <CTRL-D> flush.
`!cat`	`!cat:` [TXT\|URL\|PDF]	Cat text, PDF file, or dump URL.
`!dialog`	-	Toggle the “dialog” interface.
`!img`	`!media` [FILE\|URL]	Append image, media, or URL to prompt.
`!md`	`!markdown` [SOFTW]	Toggle markdown rendering in response.
`!!md`	`!!markdown` [SOFTW]	Render last response in markdown.
`!rep`	`!replay`	Replay last TTS audio response.
`!res`	`!resubmit`	Resubmit last STT recorded audio in cache.
`!p`	`!pick` [PROPMT]	File picker, appends filepath to user prompt. ‡
`!pdf`	`!pdf:` [FILE]	Convert PDF and dump text.
`!photo`	`!!photo` [INDEX]	Take a photo, optionally set camera index (Termux). ‡
`!sh`	`!shell` [CMD]	Run shell or command, and edit output. ‡
`!sh:`	`!shell:` [CMD]	Same as `!sh` but apppend output as user.
`!!sh`	`!!shell` [CMD]	Run interactive shell (with command) and exit.
`!url`	`!url:` [URL]	Dump URL text or YouTube transcript text.

Script	Settings and UX
`!fold`	`!wrap`	Toggle response wrapping.
`-g`	`!stream`	Toggle response streaming.
`-h`	`!help` [REGEX]	Print help or grep help for regex.
`-l`	`!models` [NAME]	List language models or show model details.
`-o`	`!clip`	Copy responses to clipboard.
`-u`	`!multi`	Toggle multiline prompter. <CTRL-D> flush.
`-uu`	`!!multi`	Multiline, one-shot. <CTRL-D> flush.
`-U`	`-UU`	Toggle cat prompter or set one-shot. <CTRL-D> flush.
`-V`	`!debug`	Dump raw request block and confirm.
`-v`	`!ver`	Toggle verbose modes.
`-x`	`!ed`	Toggle text editor interface.
`-xx`	`!!ed`	Single-shot text editor.
`-y`	`!tik`	Toggle python tiktoken use.
`!q`	`!quit`	Exit. Bye.

Model	Settings
`!Nill`	`-Nill`	Unset max response tkns (chat cmpls).
`!NUM`	`-M` [NUM]	Maximum response tokens.
`!!NUM`	`-N` [NUM]	Model token capacity.
`-a`	`!pre` [VAL]	Presence penalty.
`-A`	`!freq` [VAL]	Frequency penalty.
`-b`	`!best` [NUM]	Best-of n results.
`-j`	`!seed` [NUM]	Seed number (integer).
`-K`	`!topk` [NUM]	Top_k.
`-m`	`!mod` [MOD]	Model by name, empty to pick from list.
`-n`	`!results` [NUM]	Number of results.
`-p`	`!topp` [VAL]	Top_p.
`-r`	`!restart` [SEQ]	Restart sequence.
`-R`	`!start` [SEQ]	Start sequence.
`-s`	`!stop` [SEQ]	One stop sequence.
`-t`	`!temp` [VAL]	Temperature.
`-w`	`!rec` [ARGS]	Toggle Whisper. Optionally, set arguments.
`-z`	`!tts` [ARGS]	Toggle TTS chat mode (speech out).
`!ka`	`!keep-alive` [NUM]	Set duration of model load in memory (Ollama).
`!blk`	`!block` [ARGS]	Set and add custom options to JSON request.
`!vision`	`!audio`, `!multimodal`	Toggle multimodality type.

Session	Management
`-H`	`!hist`	Edit history in editor.
`-P`	`-HH`, `!print`	Print session history.
`-L`	`!log` [FILEPATH]	Save to log file.
`!br`	`!break`, `!new`	Start new session (session break).
`!ls`	`!list` [GLOB]	List History files with glob in name. Intruction prompts: “pr”. Awesome: “awe”. All: “.”.
`!grep`	`!sub` [REGEX]	Search sessions (for regex) and copy session to hist tail.
`!c`	`!copy` [SRC_HIST] [DEST_HIST]	Copy session from source to destination.
`!f`	`!fork` [DEST_HIST]	Fork current session to destination.
`!k`	`!kill` [NUM]	Comment out n last entries in history file.
`!!k`	`!!kill` [[0]NUM]	Dry-run of command `!kill`.
`!s`	`!session` [HIST_FILE]	Change to, search for, or create history file.
`!!s`	`!!session` [HIST_FILE]	Same as `!session`, break session.

: Commands with a colon have their output appended to the prompt.

‡ Commands with double dagger may be invoked at the very end of the input prompt (preceded by space).

E.g.: “/temp 0.7”, “!modgpt-4”, “-p 0.2”, “/session HIST_NAME”, “[PROMPT] /pick”, and “[PROMPT] /sh”.

The “/pick” command opens a file picker (usually a command-line file manager). The selected file’s path will be appended to the current prompt in editing mode.

The “/pick” and “/sh” commands may be run when typed at the end of the current prompt, such as “[PROMPT] /sh”, which opens a new shell instance to execute commands interactively. The output of these commands is appended to the current prompt.

When the “/pick” command is run at the end of the prompt, the selected file path is appended instead.

Command “!block [ARGS]” may be run to set raw model options in JSON syntax according to each API. Alternatively, set envar $BLOCK_USR.

2.8 Session Management

The script uses a TSV file to record entries, which is kept at the script cache directory. A new history file can be created or an existing one changed to with command “/session [HIST_FILE]”, in which HIST_FILE is the file name of (with or without the .tsv extension), or path to, a history file.

When the first positional argument to the script is the command operator forward slash followed by a history file name, the command “/session” is assumed.

A history file can contain many sessions. The last one (the tail session) is always loaded if the resume option -C is set.

Copying and resuming older sessions

To continue from an old session type in a dot “.” as the first positional argument from the command line. This is a shortcut for the “/sub” command, which really copies an old session to the tail the history file so we can resume it.

If you want to change to a custom history file, type in as the first positional argument at the command line “/session [NAME]” or simply “/[NAME]”.

Once the history file is set, the user can resume from an older session with “/grep [REGEX]”, “/copy current” or simply “/sub”.

It is also possible to copy sessions of a history file to another file when a second argument is given to the command with the history file name, such as “/copy [SRC_HIST_FILE] [DEST_HIST_FILE]”. A dot as file name means the current history file.

Changing session

To load an older session from a history file that is different from the defaults, there are some options.

Change to it with command “!session [NAME]”, and then “!fork” the older session to the active session.

Or, “!copy [ORIGN] [DEST]” the session from a history file to the current or other history file.

In these cases, a pickup interface should open to let the user choose the correct session from the history file.

History file editing

To change the chat context at run time, the history file may be edited with the “/hist” command (also for context injection). Delete history entries or comment them out with “#”.

2.9 Completion Regeneration

To regenerate response, type in the command “!regen” or a single exclamation mark or forward slash in the new empty prompt. In order to edit the prompt before the request, try “!!” (or “//”).

3. Prompt Engineering and Design

Minimal INSTRUCTION to behave like a chatbot is given with chat options -cc, unless otherwise explicitly set by the user.

On chat mode, if no INSTRUCTION is set, minimal instruction is given, and some options auto set, such as increasing temp and presence penalty, in order to un-lobotomise the bot. With cheap and fast models of text cmpls, such as Curie, the `best_of’ option may be worth setting (to 2 or 3).

Prompt engineering is an art on itself. Study carefully how to craft the best prompts to get the most out of text, code and chat cmpls models.

Certain prompts may return empty responses. Maybe the model has nothing to further complete input or it expects more text. Try trimming spaces, appending a full stop/ellipsis, resetting temperature, or adding more text.

Prompts ending with a space character may result in lower quality output. This is because the API already incorporates trailing spaces in its dictionary of tokens.

Note that the model’s steering and capabilities require prompt engineering to even know that it should answer the questions.

It is also worth trying to sample 3 - 5 times (increasing the number of responses with `option -n 3’, for example) in order to obtain a good response.

For more on prompt design, see:

https://platform.openai.com/docs/guides/completion/prompt-design
https://github.com/openai/openai-cookbook/blob/main/techniques_to_improve_reliability.md

See detailed info on settings for each endpoint at:

https://platform.openai.com/docs/

ESCAPING NEW LINES AND TABS

Input sequences “\n” and “\t” are only treated specially in restart, start and stop sequences!

CUSTOM / AWESOME PROMPTS

When the argument to option -S starts with a full stop, such as “-S .my_prompt”, load, search for, or create my_prompt prompt file. If two full stops are prepended to the prompt name, load it silently. If a comma is used instead, such as “-S ,my_prompt”, edit the prompt file, and then load it.

When the argument to option -S starts with a backslash or a percent sign, such as “-S /linux_terminal”, search for an awesome-chatgpt-prompt(-zh) (by Fatih KA and PlexPt). Set “//” or “%%” to refresh local cache. Use with davinci and gpt-3.5+ models.

These options also set corresponding history files automatically.

Please note and make sure to backup your important custom prompts! They are located at “~/.cache/chatgptsh/” with the extension “.pr”.

IMAGES / DALL-E

1. Image Generations

An image can be created given a text prompt. A text PROMPT of the desired image(s) is required. The maximum length is 1000 characters.

2. Image Variations

Variations of a given IMAGE can be generated. The IMAGE to use as the basis for the variations must be a valid PNG file, less than 4MB and square.

3. Image Edits

To edit an IMAGE, a MASK file may be optionally provided. If MASK is not provided, IMAGE must have transparency, which will be used as the mask. A text prompt is required.

3.1 ImageMagick

If ImageMagick is available, input IMAGE and MASK will be checked and processed to fit dimensions and other requirements.

3.2 Transparent Colour and Fuzz

A transparent colour must be set with “-@[COLOUR]” to create the mask. Defaults=black.

By defaults, the COLOUR must be exact. Use the `fuzz option’ to match colours that are close to the target colour. This can be set with “-@[VALUE%]” as a percentage of the maximum possible intensity, for example “-@10%black”.

3.3 Mask File / Alpha Channel

An alpha channel is generated with ImageMagick from any image with the set transparent colour (defaults to black). In this way, it is easy to make a mask with any black and white image as a template.

3.4 In-Paint and Out-Paint

In-painting is achieved setting an image with a MASK and a prompt.

Out-painting can also be achieved manually with the aid of this script. Paint a portion of the outer area of an image with alpha, or a defined transparent colour which will be used as the mask, and set the same colour in the script with `option -@’. Choose the best result amongst many results to continue the out-painting process step-wise.

AUDIO / WHISPER

1. Transcriptions

Transcribes audio file or voice record into the set language. Set a two-letter ISO-639-1 language code (en, es, ja, or zh) as the positional argument following the input audio file. A prompt may also be set as last positional parameter to help guide the model. This prompt should match the audio language.

If the last positional argument is “.” or “last” exactly, it will resubmit the last recorded audio input file from cache.

Note that if the audio language is different from the set language code, output will be on the language code (translation).

2. Translations

Translates audio into English. An optional text to guide the model’s style or continue a previous audio segment is optional as last positional argument. This prompt should be in English.

Setting temperature has an effect, the higher the more random.

Provider Integrations

For LocalAI integration, run the script with option --localai, or set environment $OPENAI_BASE_URL with the server Base URL.

For Mistral AI set environment variable $MISTRAL_API_KEY, and run the script with option --mistral or set $OPENAI_BASE_URL to “https://api.mistral.ai/”. Prefer setting command line option --mistral for complete integration.

For Ollama, set option -O (--ollama), and set $OLLAMA_BASE_URL if the server URL is different from the defaults.

Note that model management (downloading and setting up) must follow the Ollama project guidelines and own methods.

For Google Gemini, set environment variable $GOOGLE_API_KEY, and run the script with the command line option --google.

For Groq, set the environmental variable $GROQ_API_KEY. Run the script with option --groq. Whisper endpoint available.

For Anthropic, set envar $ANTHROPIC_API_KEY and run the script with command line option --anthropic.

For GitHub Models, $GITHUB_TOKEN and invoke the script with option --github.

For Novita AI integration, set the environment variable $NOVITA_API_KEY and use the --novita option. Novita AI offers a range of LLM models, including the highly recommended Llama 3.1 model. For an uncensored model, consider sao10k/l3-70b-euryale-v2.1 or cognitivecomputations/dolphin-mixtral-8x22b.

And for xAI’s Grok, set environment $XAI_API_KEY with your API key.

Run the script with option --xai and also with option -cc (chat completions.).

Some models also work with native text completions. For that, set command-line option -c instead.

ENVIRONMENT

BLOCK_USR

BLOCK_USR_TTS
Extra options for the request JSON block (e.g. “"seed": 33, "dimensions": 1024”).

CACHEDIR
Script cache directory base.

CHATGPTRC
Path to the user configuration file.

Defaults="~/.chatgpt.conf"

FILECHAT
Path to a history / session TSV file (script-formatted).

INSTRUCTION
Initial initial instruction message.

INSTRUCTION_CHAT
Initial initial instruction or system message in chat mode.

LC_ALL

LANG
Default instruction language in chat mode.

MOD_CHAT, MOD_IMAGE, MOD_AUDIO,

MOD_SPEECH, MOD_LOCALAI, MOD_OLLAMA,

MOD_MISTRAL, MOD_GOOGLE, MOD_GROQ,

MOD_AUDIO_GROQ, MOD_ANTHROPIC, MOD_GITHUB,

MOD_NOVITA, MOD_XAI
Set default model for each endpoint / provider.

OPENAI_BASE_URL

OPENAI_URL_PATH
Main Base URL setting. Alternatively, provide a URL_PATH parameter with the full url path to disable endpoint auto selection.

PROVIDER_BASE_URL
Base URLs for each service provider: LOCALAI, OLLAMA, MISTRAL, GOOGLE, ANTHROPIC, GROQ, GITHUB, and NOVITA.

OPENAI_API_KEY

PROVIDER_API_KEY
Keys for OpenAI, GoogleAI, MistralAI, Groq, GitHub and Novita AI APIs.

OUTDIR
Output directory for received image and audio.

RESTART

START
Restart and start sequences. May be set to null.

Restart=“\nQ: ” Start="\nA:" (chat mode)

VISUAL

EDITOR
Text editor for external prompt editing.

Defaults="vim"

CLIP_CMD
Clipboard set command, e.g. “xsel -b”, “pbcopy”.

PLAY_CMD
Audio player command, e.g. “mpv –no-video –vo=null”.

REC_CMD
Audio recorder command, e.g. “sox -d”.

COLOR THEMES

The colour scheme may be customised. A few themes are available in the template configuration file.

A small colour library is available for the user conf file to personalise the theme colours.

The colour palette is composed of $Red, $Green, $Yellow, $Blue, $Purple, $Cyan, $White, $Inv (invert), and $Nc (reset) variables.

Bold variations are defined as $BRed, $BGreen, etc, and background colours can be set with $On_Yellow, $On_Blue, etc.

Alternatively, raw escaped color sequences, such as \u001b[0;35m, and \u001b[1;36m may be set.

Theme colours are named variables from Colour1 to about Colour11, and may be set with colour-named variables or raw escape sequences (these must not change cursor position).

REQUIRED PACKAGES

Bash
cURL, and JQ

OPTIONAL PACKAGES

Optional packages for specific features.

Base64 - Image endpoint, vision models
Python - Modules tiktoken, markdown, bs4
ImageMagick/fbida - Image edits and variations
SoX/Arecord/FFmpeg - Record input (Whisper)
mpv/SoX/Vlc/FFplay/afplay - Play TTS output
xdg-open/open/xsel/xclip/pbcopy - Open images, set clipboard
W3M/Lynx/ELinks/Links - Dump URL text
bat/Pygmentize/Glow/mdcat/mdless - Markdown support
termux-api/termux-tools/play-audio - Termux system
poppler/gs/abiword/ebook-convert/LibreOffice - Dump PDF or Doc as text
dialog/kdialog/zenity/osascript/termux-dialog - File picker

BUGS

Bash “read command” may not correctly display input buffers larger than the TTY screen size during editing. However, input buffers remain unaffected. Use the text editor interface for big prompt editing.

If readline screws up your currrent input buffer, try pressing <CTRL-L> to force it to redisplay and refresh the prompt properly on screen.

File paths containing spaces may not work correctly in the chat interface. Make sure to backslash-escape filepaths with white spaces.

Folding the response at white spaces may not worked correctly if the user has changed his terminal tabstop setting. Reset it with command ``tabs -8’’ or ``reset’’ before starting the script, or set one of these in the script configuration file.

If folding does not work well at all, try exporting envar $COLUMNS before script execution.

Bash truncates input on “\000” (null).

Garbage in, garbage out. An idiot savant.

The script logic resembles a bowl of spaghetti code after a cat fight.

LIMITS

The script objective is to implement most features of OpenAI API version 1 but not all endpoints or options will be covered.

This project doesn’t support “Function Calling” or “Structured Outputs”.

OPTIONS

Service Providers

--anthropic, --ant
Set Anthropic integration (cmpls/chat).

--github, --git
Set GitHub Models integration (chat).

--google, -goo
Set Google Gemini integration (cmpls/chat).

--groq
Set Groq integration (chat).

--localai
Set LocalAI integration (cmpls/chat).

--mistral
Set Mistral AI integration (chat).

--novita
Set Novita AI integration (cmpls/chat).

--openai
Reset service integrations.

-O, --ollama
Set and make requests to Ollama server (cmpls/chat).

--xai
Set xAI integrations (cmpls/chat).

Configuration File

-f, --no-conf
Ignore user configuration file.

-F
Edit configuration file with text editor, if it exists.

$CHATGPTRC="~/.chatgpt.conf".

-FF
Dump template configuration file to stdout.

Sessions and History Files

-H, --hist [/HIST_FILE]
Edit history file with text editor or pipe to stdout.

A history file name can be optionally set as argument.

-P, -PP, --print [/HIST_FILE]
Print out last history session.

Set twice to print commented out history entries, inclusive. Heeds options -ccdrR.

These are aliases to -HH and -HHH, respectively.

Input Modes

-u, --multiline
Toggle multiline prompter, <CTRL-D> flush.

-U, --cat
Set cat prompter, <CTRL-D> flush.

-x, -xx, --editor
Edit prompt in text editor.

Set twice to run the text editor interface a single time for the first user input.

Set options -eex to edit last buffer from cache.

Interface Modes

-c, --chat
Chat mode in text completions (used with options -wzvv).

-cc
Chat mode in chat completions (used with options -wzvv).

-C, --continue, --resume
Continue from (resume) last session (cmpls/chat).

-d, --text
Single-turn session of plain text completions.

-dd
Multi-turn session of plain text completions with history support.

-e, --edit
Edit first input from stdin or file (cmpls/chat).

With options -eex, edit last text editor buffer from cache.

-E, -EE, --exit
Exit on first run (even with options -cc).

-g, --stream (defaults)
Set response streaming.

-G, --no-stream
Unset response streaming.

-i, --image [PROMPT]
Generate images given a prompt. Set option -v to not open response.

-i [PNG]
Create variations of a given image.

-i [PNG] [MASK] [PROMPT]
Edit image with mask and prompt (required).

-q, -qq, --insert
Insert text rather than completing only. May be set twice for multi-turn.

Use “[insert]” to indicate where the language model should insert text (`instruct’ and Mistral `code models’).

-S .[PROMPT_NAME], -.[PROMPT_NAME]

-S ,[PROMPT_NAME], -,[PROMPT_NAME]
Load, search for, or create custom prompt.

Set .[PROMPT] to load prompt silently.

Set ,[PROMPT] to single-shot edit prompt.

Set ,,[PROMPT] to edit the prompt template file.

Set .?, or .list to list all prompt files.

-S, –awesome /[AWESOME_PROMPT_NAME]

-S, –awesome-zh %[AWESOME_PROMPT_NAME_ZH]
Set or search for an awesome-chatgpt-prompt(-zh).

Set // or %% instead to refresh cache.

-T, --tiktoken

-TT, -TTT
Count input tokens with python Tiktoken (ignores special tokens).

Set twice to print tokens, thrice to available encodings.

Set the model or encoding with option -m.

It heeds options -ccm.

-w, --transcribe [AUD] [LANG] [PROMPT]
Transcribe audio file speech into text. LANG is optional. A prompt that matches the speech language is optional. Speech will be transcribed or translated to the target LANG.

Set twice to phrase or thrice for word-level timestamps (-www).

With options -vv, stop voice recorder on silence auto detection.

-W, --translate [AUD] [PROMPT-EN]
Translate audio file speech into English text.

Set twice to phrase or thrice for word-level timestamps (-WWW).

-z, --tts [OUTFILE|FORMAT|-] [VOICE] [SPEED] [PROMPT]
Synthesise speech from text prompt. Takes a voice name, speed and text prompt.

Set option -v to not play response automatically.

Model Settings

-@, --alpha [[VAL%]COLOUR]
Set transparent colour of image mask. Def=black.

Fuzz intensity can be set with [VAL%]. Def=0%.

-Nill
Unset model max response tokens (chat cmpls only).

-NUM

-M, --max [NUM[-NUM]]
Set maximum number of response tokens. Def=1024.

A second number in the argument sets model capacity.

-N, --modmax [NUM]
Set model capacity tokens. Def=auto, Fallback=4000.

-a, --presence-penalty [VAL]
Set presence penalty (cmpls/chat, -2.0 - 2.0).

-A, --frequency-penalty [VAL]
Set frequency penalty (cmpls/chat, -2.0 - 2.0).

-b, --best-of [NUM]
Set best of, must be greater than option -n (cmpls). Def=1.

-B, --logprobs [NUM]
Request log probabilities, also see -Z (cmpls, 0 - 5),

-j, --seed [NUM]
Set a seed for deterministic sampling (integer).

-K, --top-k [NUM]
Set Top_k value (local-ai, ollama, google).

--keep-alive, --ka=[NUM]
Set how long the model will stay loaded into memory (ollama).

-m, --model [MODEL]
Set language MODEL name. Def=gpt-3.5-turbo-instruct/gpt-4o.

Set MODEL name as “.” to pick from the list.

--multimodal, --vision, --audio
Set model multimodal model type.

-n, --results [NUM]
Set number of results. Def=1.

-p, --top-p [VAL]
Set Top_p value, nucleus sampling (cmpls/chat, 0.0 - 1.0).

-r, --restart [SEQ]
Set restart sequence string (cmpls).

-R, --start [SEQ]
Set start sequence string (cmpls).

-s, --stop [SEQ]
Set stop sequences, up to 4. Def="<|endoftext|>".

-S, --instruction [INSTRUCTION|FILE]
Set an instruction text prompt. It may be a text file.

--time, --no-time
Insert the current date and time (timestamp) to the instruction prompt.

-t, --temperature [VAL]
Set temperature value (cmpls/chat/whisper), (0.0 - 2.0, whisper 0.0 - 1.0). Def=0.

Miscellaneous Settings

--api-key [KEY]
Set API key to use.

--fold (defaults), --no-fold
Set or unset response folding (wrap at white spaces).

-h, --help
Print the help page.

--info
Print OpenAI usage status (requires envar $OPENAI_ADMIN_KEY).

-k, --no-colour
Disable colour output. Def=auto.

-l, --list-models [MODEL]
List models or print details of MODEL.

-L, --log [FILEPATH]
Set log file. FILEPATH is required.

--md, --markdown, --markdown=[SOFTWARE]
Enable markdown rendering in response. Software is optional: bat, pygmentize, glow, mdcat, or mdless.

--no-md, --no-markdown
Disable markdown rendering.

-o, --clipboard
Copy response to clipboard.

-v, --verbose
Less verbose.

Sleep after response in voice chat (-vvccw).

With options -ccwv, sleep after response. With options -ccwzvv, stop recording voice input on silence detection and play TTS response right away.

May be set multiple times.

-V
Dump raw JSON request block (debug).

--version
Print script version.

-y, --tik
Set tiktoken for token count (cmpls/chat, python).

-Y, --no-tik (defaults)
Unset tiktoken use (cmpls/chat, python).

-Z, -ZZ, -ZZZ, --last
Print JSON data of the last responses.

Files

README.md

Latest commit

History