Switchyard v0.3.0 — multi-provider harness

The Switchyard rebrand brings a shared terminal interface to DeepSeek, GLM,
Anthropic Claude, and OpenAI GPT, plus native Codex and Claude Code connections.

Added

• Provider and model selection during startup, live model catalogs, and local
  /commands, /model, /provider, /usage, and /session commands.
• API model pricing per million tokens, with cheapest-example sorting and
  explicit unknown prices. Reference rates checked September 21, 2026.
• Native Codex and Claude Code login, streaming, permissions, and session resume.
• A terminal renderer with styled messages, a composer, resize handling,
  scrolling, queued input, and expandable reasoning and tool results.
• Provider-aware transports, automatic caching, balance handling, update checks,
  and Docker tooling.

Changed

• switchyard is the main command and Windows executable. Existing d, dsw,
  dsd, and dswait commands remain available.
• Context compaction replaces the older prefix, retains 15 complete tool-safe
  turns, and uses a deterministic fallback when summarization fails.
• Provider-specific compaction requests and session metadata support API model
  changes while preserving tool-call/result pairs.
• Installer and desktop window branding now use Switchyard.

Compatibility and limits

• The legacy npm package identifier and configuration/session directories remain
  unchanged so existing installs and saved sessions keep working.
• Native Codex/Claude Code sessions resume within their own engine. Their tools
  and compaction are managed by that engine; cross-engine session migration and
  native-worker integration with Switchyard coordination are not yet supported.
• Price examples are references, not spending caps. Model availability and
  provider billing modifiers still apply.
• Windows standalone binaries provide CLI functionality. Electron UI requires
  the source installation and its dependencies.

Validation: syntax checks and 11 focused test suites passed; all five Windows executables passed --help smoke tests.

Download the Windows x64 ZIP below and verify it with SHA256SUMS.txt. Native account connections require Codex CLI or Claude Code installed separately.

v0.1.5 - Electron UI, TUI polish, and web search

What's new

• Added the Electron desktop UI launcher and local UI application.
• Added saved session history, independently scrollable chat history, and UI screenshots.
• Improved TUI feedback with live stream status, clearer tool-call streaming, clickable workspace file links, and context-token counts.
• Added the doctor command and clearer vision/setup status.
• Added provider-aware web search tools and related configuration guidance.
• Added safer guidance for fragile file-editing workflows and shell helper scripts.

Comparison

• 12 commits since v0.1.4.
• Release target: 51751c6.
• Full changelog: v0.1.4...v0.1.5

This release contains only changes committed to GitHub. The newer local multi-agent coordination work is not included.

v0.1.4 - Interrupt, search, and vision fixes

What's new

• Fixed interactive stream interruption: Esc and Ctrl+C can abort an active streamed response, then return to a usable prompt without leaving the terminal raw/frozen.
• Interrupted responses are saved as [interrupted by user], and partial tool calls from interrupted responses are not executed.
• Fixed search_code so path can target either a directory or a specific file; missing paths now report a clear tool error.
• Added OpenAI-backed image analysis via analyze_image_openai for real vision/transcription when OPENAI_API_KEY is configured.
• Clarified view_image as metadata/data-url only, not visual understanding.
• Added PBC upload guidance to the default system prompt so agents use pbc tab upload/pbc upload instead of the Windows file picker.

Verification

• npm run check
• npm run smoke
• npm run build:exe
• Refreshed global d, dsw, dsd, and dswait executables.

v0.1.3 - Agent ops, skills, and image support

What's new

• Added Agent Ops tools for long-running sessions: persistent goals, plans, checkpoints, session health, session summaries, and bounded handoff helpers.
• Added local skill support with --skill, --skills, --skill-root, --list-skills, plus list_skills and read_skill tools.
• Skills now persist across resumed sessions; resuming with --skill pbc refreshes the saved system prompt and stores the skill list in session JSON.
• Added view_image for workspace screenshots, diagrams, photos, and UI images, returning metadata, dimensions, and data URLs when small enough.
• Added web search and fetch tooling so agents can search current information and read page text.
• Added broader workspace/repo tools: batch file reads, search, glob, tree, stats, git status/diff/log/blame, cache, and safer shell aliasing.
• Improved TUI/session reliability, including repaired tool-call history on resume and sequential execution for shell-style tools.
• Added dummy smoke fixtures and project instructions for testing wrapper tools.

Verification

• npm run check
• npm run smoke
• npm run build:exe
• Verified global d --print-system --skill pbc loads the PBC skill.

v0.1.2

What's new

Quiet outfile mode

dsw now supports silent Markdown output for detached subagent workflows:

dsw -p "inspect this repo" --permission full --no-output --outfile result.md

Default outfile content includes:

• Final assistant response
• Files touched by edit tools

Add --full-chat to write the whole transcript, including tool calls, tool results, and reasoning content.

Also included

• --outfile alias for --output
• --no-output validation that requires an output file
• Resume flow now renders recent session history before prompting

v0.1.1 — standalone exe build

What's new

Standalone executables

The installer now builds real .exe files using Node.js SEA (Single Executable Applications) and esbuild:

• dsw.exe / d.exe — interactive agent
• dsd.exe — detached runner
• dswait.exe — output file poller

Exes are copied to %LOCALAPPDATA%\Programs\dsw\ and that directory is added to your user PATH automatically — no npm, no Node.js required to run after install.

How it works

1. esbuild bundles each ESM entry point → single CJS file (local imports inlined, node: builtins kept external)
2. node --experimental-sea-config compiles the bundle into a .blob
3. postject injects the blob into a copy of node.exe
4. d.exe is a copy of dsw.exe for the short alias

Build it yourself

npm install          # installs esbuild + postject
npm run build:exe    # produces dist/exe/*.exe

Install

irm https://raw.githubusercontent.com/gaston1799/deepseek-detached-agent/main/install.ps1 | iex

v0.1.0 — initial release

dsw v0.1.0

First release of the DeepSeek Watch interactive terminal agent.

Install

irm https://raw.githubusercontent.com/gaston1799/deepseek-detached-agent/main/install.ps1 | iex

Or clone and run npm install -g .

What's included

• dsw / d — interactive streaming agent with session memory
• dsd — fire-and-forget detached runner with Claude fallback
• dswait — polls for a detached output file

Tools available to DeepSeek

• read_text_file — with line-range support (start_line / end_line)
• write_text_file — create or overwrite files
• patch_text_file — exact search-and-replace in files
• list_workspace_files, get_runtime_context
• run_cmd, run_powershell — gated by permission level

Permission levels

review (read-only) · ask (confirm before writes/shell) · full (auto-run)