Coding agents

wsx runs Claude Code in each workspace by default. You can pick a different coding agent for all new workspaces or for one workspace.

wsx config set coding_agent codex            # default for new workspaces
wsx workspace create backend --agent pi      # this workspace only

A workspace created from inside another workspace inherits its agent kind. See wsx workspace create.

wsx runs the binary of the same name from PATH. Override the path with WSX_<AGENT>_BIN, and the model with the agent's model variables. See Environment variables.

Capabilities

• "wsx instructions" is the process doctrine, the auto-rename prompt, and the repo's custom instructions.
• Agents without automatic status still report through wsx status set. See Status and recap.
• Session resume is covered in Multi-agent workspaces.

claude

• Instructions go in --append-system-prompt. Related repos go in --add-dir.
• wsx adds status hooks through --settings.
• Yolo passes --dangerously-skip-permissions.
• Remote control is on by default. Turn it off with the remote_control setting.

pi

• Instructions go in --append-system-prompt.
• wsx loads a small extension (pi -e) that reports each session start, so each pi agent resumes its own session.
• WSX_PI_MODEL and WSX_PI_PROVIDER pick the model for new sessions.
• pi reads skills from ~/.claude/skills.

hermes

• Hermes has no flag for extra instructions, so wsx writes them into a <!-- BEGIN wsx-managed --> block at the end of the worktree's AGENTS.md on each start.
  • When wsx creates AGENTS.md, it copies the repo's CLAUDE.md into it.
  • An untracked AGENTS.md is added to .git/info/exclude.
  • A tracked AGENTS.md shows as modified while the block is present.
• On resume, wsx picks the newest Hermes session started since the workspace's first Hermes spawn. If two worktrees start Hermes within a few seconds, it can pick the wrong one.
• Added Hermes agents always start fresh.
• WSX_HERMES_MODEL overrides the model in ~/.hermes/config.yaml. WSX_HERMES_PROVIDER is passed as --provider, which Hermes's default REPL mode may ignore; set the provider in config.yaml to be sure.
• The detail bar shows chat and tool counts, but not edited files.

codex

• Requires Codex 0.146.0 or newer.
• Instructions go in -c developer_instructions=…. Nothing is written to the worktree. Codex reads CLAUDE.md when the repo has no AGENTS.md.
• Instructions are set only on a fresh start. A resumed session keeps the ones it started with, so edits to custom instructions or related repos don't reach it.
• Status comes from Codex's notify after each turn.
• Your ~/.claude/commands/*.md are mirrored into a Codex plugin on each start, so commands such as /pull-request work in Codex.
• wsx turns off Codex's startup update prompt, which would block input. Update Codex yourself, for example with npm install -g @openai/codex.
• Yolo passes --dangerously-bypass-approvals-and-sandbox.
• The detail bar doesn't show edited files.

omp

omp and pi are separate harnesses with different binaries. You can install both.

• Instructions go in --append-system-prompt. Related repos go in --add-dir.
• Yolo passes --approval-mode yolo. Otherwise omp uses your tools.approvalMode.
• WSX_OMP_MODEL sets the model. It accepts provider/id.
• On each start wsx passes --config <state dir>/omp-config.yml. This overlay turns on omp's ~/.claude skills, commands and Claude plugins, which omp 18 turns off by default.
  • It overrides your own settings for those keys in wsx sessions, and replaces your enabledProviders list.
  • It enables every skill and command under ~/.claude, not only wsx's.
  • Your ~/.omp/agent/config.yml is never edited.
• Pinned chips that name a bundled skill are sent as /skill:<name>. See Agent skills.