Nightfall CLI

A minimal coding agent harness in Python, built to show how the pieces of a coding agent fit together.

This is the Nightfall CLI repository, a small coding-agent harness built from scratch.

Getting started

Install the project with uv:

uv sync

Configure an OpenAI-compatible endpoint and key in ~/.agents/env:

BASE_URL=https://your-endpoint/v1
API_KEY=your-api-key
MODEL=your-model-name

Start the agent from the repository root:

uv run nightfall

Use uv run nightfall --resume to resume the latest saved session or uv run nightfall --debug to display raw model responses. The sandbox backend is selected automatically for the host operating system.

Features

• Interactive terminal chat
• Tools for running shell commands, reading and writing files, and making targeted edits.
• Configurable model and OpenAI-compatible API endpoint.
• Tool permissions and shell sandboxing on macOS, Linux, and Windows
• Skills loaded from project and user .agents/skills directories.
• Prompt autocomplete: / completes commands, /skill: completes skills, and @ completes paths.
• Subagents for exploring a codebase in a separate context window.
• Todo tracking for tasks with multiple steps.
• Saved chat sessions, with /sessions to reopen them and /rewind to go back in the conversation.
• Automatic context compaction, plus /compact to trigger it manually.
• Git branch context and reminders when files change between turns.
• Optional OpenTelemetry traces and metrics (see below).

Telemetry

Tracing is off unless an exporter is configured, so the default install does not pay for it. Install the extra and point it at a collector:

uv sync --extra otel
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318   # base URL, no /v1/traces
uv run nightfall

A session then reports a span per turn, one per model call and one per tool call, plus counters for turns, tool calls, errors and tokens (input, output, reasoning and cached). NIGHTFALL_TELEMETRY=console prints the spans to the terminal instead of shipping them, which is the quickest way to see what is being recorded.

Only names, counts and sizes are ever recorded - never prompts, replies, tool arguments, tool output or file contents. Prompts, file contents and tool output stay in the transcript; telemetry reports how long the session was and what it touched.

The standard OTEL_* variables work as they do everywhere else (OTEL_SERVICE_NAME, OTEL_EXPORTER_OTLP_HEADERS, and the rest). The exporter is given five seconds to hand over a batch before the CLI exits, so an unreachable collector costs a warning rather than a hang.