Veritas

A terminal-native chat client (TUI) for OpenAI-compatible APIs, written in Rust.

Streams responses token-by-token over SSE against any /v1/chat/completions endpoint — Ollama, llama.cpp server, OpenAI, OpenRouter, vLLM, etc.

Features

• Token-by-token SSE streaming with Esc-to-cancel, animated bouncing braille bar for pre-token thinking, and shimmer status line paired with live spinner
• Full markdown rendering for agent replies — headings, lists, tables, links, blockquotes, and syntax-highlighted code blocks
• PgUp/PgDn and Ctrl-U/Ctrl-D scrolling with follow-tail
• Shift+Enter for newlines, /clear to reset the conversation
• Sessions: auto-saved JSON transcripts, Ctrl-S picker overlay, markdown export
• Live sidebar telemetry: RAM (sysinfo), AMD GPU VRAM + utilization (kernel DRM sysfs), context-window usage from server-reported tokens when available
• Slash commands with status-line feedback
• Live / command popup: fuzzy-finds the closest slash command above the input with a one-line description; ↑/↓ select, Tab completes, Enter runs (D031/D033)
• Sidebar hide/show via Ctrl-B or /sidebar; auto-hides below 72×20 (D032)
• Single-file theme (src/theme.rs) for reskinning

Configuration

Everything comes from environment variables:

Variable	Default	Purpose
OPENAI_BASE_URL	http://localhost:11434/v1	API base (Ollama default shown)
OPENAI_API_KEY	(empty)	Bearer token; omitted from requests when unset
OPENAI_MODEL	(unset — auto-detected)	Model id; probed from the engine's GET /models when unset (D030)
VERITAS_THEME	default	Theme preset: default, amber, green, or monochrome

Engine-owned parameters are not client config: requests omit temperature unless you override it at runtime with /temp, and the context window is probed from the engine (D009).

Sidebar telemetry

• RAM — always shown (sysinfo).
• VRAM / GPU % — AMD GPUs (amdgpu) via the kernel's DRM sysfs, sampled every 500 ms; widgets are hidden on anything else. Never faked.
• CTX — used tokens come from the server's usage in stream chunks when it sends them; otherwise a chars/4 estimate is shown. The window total is probed from the engine at startup and on /model (Ollama /api/show or llama.cpp server /props); on engines without model-info endpoints the window shows n/a and no ratio gauge. The gauge shifts green → yellow (≥70%) → red (≥90%) as it fills.

Commands

Type in the input box:

Command	Effect
/help	Show help overlay (keybindings and commands)
/copy [n]	Copy latest code block (or block n) to system clipboard
/sessions, /session	Open the session picker (same as Ctrl-S)
/export <file>	Write the transcript as markdown
/model <id>	Switch model (takes effect on next send; re-probes ctx window); bare /model re-detects from the engine (D030)
/temp <0.0-2.0>	Override sampling temperature — bare /temp shows the effective setting
/clear	Reset the conversation (next save starts a new session)
/sidebar	Toggle sidebar visibility (same as Ctrl-B; auto-hides below 72×20)
/quit	Exit the app (same as Ctrl-C / Ctrl-Q)

Sessions

Every exchange is auto-saved as JSON under $XDG_DATA_HOME/veritas/sessions (default ~/.local/share/veritas/sessions), one file per conversation, named sess-<unixsecs>.json. Sessions are titled from the first user message. Press Ctrl-S for the picker — ↑/↓ select, Enter loads, Esc cancels.

Run

export OPENAI_BASE_URL="http://localhost:11434/v1"
cargo run --release

Keys

• Enter — send · Shift+Enter — newline
• Esc — stop streaming / dismiss / popup / close overlays · Ctrl-S — session picker · Ctrl-B — toggle sidebar · Ctrl-C / Ctrl-Q — quit
• Ctrl-Y — copy latest code block to system clipboard (OSC 52 + desktop tool)
• /help — show help overlay; ? is entered as chat text
• / at the start of input — live fuzzy command popup: ↑/↓ select · Tab completes the highlighted signature · Enter runs it · Esc dismisses
• PgUp/PgDn or Ctrl-U/Ctrl-D — scroll