Build on abs

The whole project, in four places

There's very little to learn. Everything abs does lives in one of these:

# the program — commands, persona, hooks, usage, all of it
abs.sh

# voice, both directions (local, no cloud)
transcribe.py   # voice note  → text
speak.py        # text        → voice note

# your runtime state, owner-only (never in git)
~/.abs/          # rc.json, access.json, usage.json …

The agent's behaviour isn't in a config file — it's the system prompt abs appends at launch, built in build_prompt().

Point it at a different model

Set a default model per profile, applied on every launch. An explicit abs --model … on the command line always wins for that one session.

abs config model opus        # default for this profile
abs config model --clear       # back to Claude Code's default
abs --model sonnet           # just this session

To change how the default is stored or applied, the logic is in cmd_config() (stores it) and cmd_run() (passes --model at launch).

Rewrite the persona & tone

How abs talks to you — when it messages, how short it keeps things, what it calls itself — is plain English in a file you own, and you don't need to touch the source to change it. abs prompt opens it on the Persona tab; ^R puts the shipped wording back.

abs prompt                   # the page: read it, edit it, reset it
abs persona list             # abs, ceo, cto, friend, and yours
abs persona create mine      # start one from the shipped text
abs --persona cto            # just this session

~/.abs/persona.md            # the default persona, if you'd rather edit by hand
~/.abs/personas/<name>.md    # the named ones

The prompt is assembled in three slots — bridge mechanics, then your persona, then safety. Safety is appended last on purpose, so a persona can change the voice and the priorities but never talk abs past a destructive-command rail. Over 16,000 characters, or containing a <channel tag, is refused at save rather than silently swapped.

See everything it tells Claude

Nothing abs adds to a session is hidden. One command opens a tabbed page over every layer: the locked bridge mechanics, your persona, the wording each remote command injects, this project's CLAUDE.md, your global one, and what Claude has remembered.

abs prompt                   # the page — F1…F7 or ^PgUp/^PgDn to move
abs prompt show persona      # or do it without the page at all
abs prompt diff persona      # what you changed vs what ships
abs prompt reset persona     # back to the shipped wording

^S saves, ^R resets, ^N adds a hook phrase or a memory fact, ^T deletes one, ^Q or Esc leaves — and if anything is unsaved it asks first. The System tab is read-only: it is what makes the bridge work, so breaking it would break the tool rather than personalise it.

Add a remote command of your own

The ABS … phrases you send from Telegram are handled by a session hook, not by the model, which is why they work even when a message tries to talk the model out of them. The wording each one injects lives in ~/.abs/hooks.json, and you can add your own.

# ~/.abs/hooks.json — send "ABS REVIEW" as a whole message
{ "ABS REVIEW": "Stop and review the diff before going further." }

Three phrases can't be reworded — the two that mute and the two that cut the bridge act in the hook and never reach the model at all, so an entry under those names is ignored. Everything else is yours, and the Hooks tab of abs prompt is the editor.

Replace the voice engine

Installing voice is opt-in — it's a few GB of models, so nothing downloads until you ask. Once installed, replies go out as a voice note by default wherever the machine can speak. You don't need a dev checkout any more; an installed abs keeps its engines in ~/.abs/voice:

abs voice setup       # install the speech engines
abs voice status      # what's installed, what's missing, how to fix it
abs voice samples     # one voice note per voice — choose by ear

The speaking engine is Kokoro: 82M parameters, built for the CPU, a note in seconds on a laptop. The alternative wants a GPU and is the only one that can clone a voice from a reference clip, so it's a separate opt-in:

abs voice setup --chatterbox         # adds the GPU engine, for cloning
abs config voice-sample <clip>       # point it at your reference audio

Underneath, it's standalone scripts with a dead-simple contract. Keep the same in/out and you can drop in any engine — a different Whisper size, a cloud STT, another TTS voice.

# in:  an audio file path  →  out: transcript on stdout
transcribe.py <file.oga>

# in:  text  →  out: a voice note sent to Telegram
speak_kokoro.py "the text to speak"     # CPU, the default
speak.py "the text to speak"            # chatterbox, for cloning

abs only ever calls these files, so swapping engines never touches the main script.

Tune how fast it goes quiet

Auto-silent trips after a set number of back-to-back terminal prompts. Want it to hold on longer, or mute sooner? Change one constant.

readonly SILENT_STREAK=3    # prompts at the terminal before it mutes

The counting itself lives in cmd_silent_hook() just below — that's where a Telegram message resets the streak.

Add your own command

Commands are a plain case in main(). Write a cmd_yours() function, add one line to the dispatch, and abs yours works. No registration, no plugin system.

cmd_hello() { ok "hi from $PROFILE"; }

# in main()'s case block:
hello) cmd_hello ;;

Change what the status bar shows

The bottom-bar dot and usage are rendered by cmd_statusline(); the 5h / week glance itself is formatted in usage_glance_str(). Edit either to change the text, order, or what's shown.