xdog / Packages / coding / Overview

Interactive coding-agent CLI, with a session SDK over the shared Agent runtime.

Coding combines xdog.ai, xdog.agent and xdog.tui into a Linux terminal coding assistant. Interactive TUI, one-shot print and explicit ACP intake share the same session controller; none introduces a second tool-calling loop.

Highlights

• Native scrollback, a multiline editor, tool/reasoning details and explicit tool approvals; print mode supports text, JSON and Markdown output.
• JSON-persisted sessions, explicit session/branch navigation, searchable history and context compaction. -r opens a current-directory selector, not automatic most-recent-session execution.
• Host-owned filesystem/bash/current-time tools, a separate MCP registry and explicitly trusted extension hooks, commands, basic dialogs and providers. Extension workers do not register model tools.
• Default ask tool policy, separate readonly scope and explicit allow-all for trusted unattended mutation. Tool selection and skill loading grant no authority.
• Host-owned children with independent task knowledge, system prompts, models and approvals, limited to frozen subsets of actual parent tools and available skills.
• xdog-skills is included by default: native Goal commands/tools are independent of skill-body activation and the optional readonly display worker.
• Installed Eyes/Flow capabilities use their public CLIs. Eyes shares an independent user-local daemon; Flow has none. Discovery creates no observation or workflow.

Try it

uv run xdog-coding
uv run xdog-coding -r
uv run xdog-coding --print --thinking-level off --prompt "Explain this repository"

Choose a real configured model with --model MODEL_ID; --list-models lists available models. Reasoning levels are off, minimal, low, medium, high, xhigh, subject to the actual model's capabilities. See the CLI and permission reference.

Durable work without another Agent

Goal records an explicitly delegated outcome for the exact Coding session/branch. The current main Agent plans from its real conversation, works itself or delegates through authorized host tools, saves progress assessments and judges completion. Task JSON files are authoritative. No mandatory reviewer or Goal-owned scheduler exists, and the Goal skill body is not active by default.

Completed remains current/viewable until clear or a newly delegated Goal replaces that completed binding. User waits require actual user evidence; task waits require the main Agent's evidence check and versioned activation. Pause stops proactive reminders and Goal's own observation until explicit user resume, not accepted tools or child work. See skills and Coding design.

ACP callbacks use an explicit endpoint and session. The lightweight receiver can persist inputs while the TUI/owner is absent; an activated owner automatically consumes pending input under the original tool policy. Intake ACKs are not execution results, and interrupted/unknown effects are never automatically replayed. Coding's former JSON-lines frontend-control RPC is removed; ACP and private MCP/extension transports remain separate.

Full user configuration and interaction details remain in the package README.