Command reference

Every shell command, option, and in-chat slash command.

Use shell commands to start or manage Harness UI; use slash commands inside an interactive conversation. For a workflow rather than a syntax lookup, start with setup or terminal use.

Shell invocation

a13n-harness-ui [GLOBAL OPTIONS] [COMMAND] [COMMAND OPTIONS]

Omitting COMMAND opens the interactive terminal. Put global options before the command; use -h or --help at each level. --version is an eager information option. Help and version do not initialize the App, database, Model, or Environment.

These tables show command-line defaults; Configuration explains effective defaults. --resume cannot be combined with new-session Agent, Environment, or title overrides.

Global options

ParameterType / choicesParser defaultMeaning
--configpathnot setExplicit Harness UI configuration YAML (default: ~/.a13n-harness-ui/a13n-harness-ui.yaml).
--data-rootpathnot setOverride the local Harness UI data root.
--resumetextnot setResume a saved session by ID.
--agenttextnot setSelect a configured Agent for this session.
--environment-modefull-control, sandboxnot setOverride the built-in Environment mode for this session.
--environment-profiletextnot setOverride the custom Environment profile for this session.
--displayconcise, detailednot setPresentation only; overrides display.mode (default concise). /mode switches live.
--no-update-checkbooleanfalseSkip startup update detection for this invocation.
--versionbooleanfalseShow the version and exit.

Commands

Place global options before these subcommands. Authentication and plugin commands change local stores; config inspection also opens application state.

webui

Run one foreground WebUI server with bundled browser assets.

ParameterType / choicesParser defaultMeaning
--hosttext"127.0.0.1"Listener IPv4 or IPv6 address.
--portinteger 1..655358765
--apikey, --api-keytextnot setListener API key; overrides A13N_HARNESS_UI_API_KEY (visible in shell arguments).
--dangerous-skip-permissions, --dangerously-bypass-permissionbooleanfalseDisable Web authentication only; does not change Agent permissions.
--share-computer / --no-share-computerbooleantrueShare native Host Files as the server OS account, independent of Agent permissions.

update

Update this uv-tool installation now, without starting chat or setup.

No command-specific parameters.

setup

Configure a model, context budget, and execution permissions interactively.

ParameterType / choicesParser defaultMeaning
--advancedbooleanfalseAlso choose context, reasoning, tool review, subagents, and instructions.

add agent

Create another agent without changing existing agents or defaults.

ParameterType / choicesParser defaultMeaning
--advancedbooleanfalseAlso customize reasoning, tool review, and instructions.

add model

Create a reusable model without creating or changing an agent.

ParameterType / choicesParser defaultMeaning
--advancedbooleanfalseAlso customize subscription context and reasoning settings.

run

Execute one prompt without an interactive terminal.

ParameterType / choicesParser defaultMeaning
PROMPTtextrequiredRequired positional argument
--resumetextnot setContinue a saved session.
--agenttextnot setAgent used for a new session.
--environment-modefull-control, sandboxnot setBuilt-in execution mode used for a new session.
--environment-profiletextnot setCustom Environment profile used for a new session.
--titletextnot setTitle used for a new session.
--formattext, json"text"

config path

Show the selected configuration and data paths.

ParameterType / choicesParser defaultMeaning
--formattext, json"text"

config validate

Validate the selected source tree.

ParameterType / choicesParser defaultMeaning
--formattext, json"text"

config show

Show the accepted configuration.

ParameterType / choicesParser defaultMeaning
--formattext, json"text"

config subagents

List package-owned subagents and their current inclusion.

ParameterType / choicesParser defaultMeaning
--formattext, json"text"

import subagents

Preview or apply external subagent imports.

ParameterType / choicesParser defaultMeaning
--productclaude-code, cursor, codexrequired
--scopeuser, projectrequired
--project-rootpathnot set
--user-homepathnot set
--applybooleanfalseApply every ready candidate; omission is a dry-run preview.
--formattext, json"text"

plugin install

Install one Content Plugin from a Git repository.

ParameterType / choicesParser defaultMeaning
REPOSITORYtextrequiredRequired positional argument
--plugintextnot setPlugin ID when the repository contains several.
--reftextnot setGit branch, tag, or commit to install.
--formattext, json"text"

plugin list

List installed Content Plugins and their directories.

ParameterType / choicesParser defaultMeaning
--formattext, json"text"

plugin uninstall

Permanently delete the installed Content Plugin directory, including local edits, without another confirmation. Back up edits first; reinstall does not restore them.

ParameterType / choicesParser defaultMeaning
PLUGIN_IDtextrequiredRequired positional argument
--formattext, json"text"

environment list

List Environment modes and profiles.

ParameterType / choicesParser defaultMeaning
--formattext, json"text"

doctor

Inspect App and extension health.

ParameterType / choicesParser defaultMeaning
--formattext, json"text"

auth status

Show Model authentication status.

ParameterType / choicesParser defaultMeaning
PROVIDERcodex, grok, copilotnot setOptional positional argument
--formattext, json"text"

auth key list

List saved key references without returning their key bytes.

ParameterType / choicesParser defaultMeaning
--formattext, json"text"

auth key set

Add or replace a key using a hidden prompt, never a command-line key.

ParameterType / choicesParser defaultMeaning
REFERENCEtextrequiredRequired positional argument

auth key delete

Delete a saved key after confirmation. Future Model resolution using it can fail.

ParameterType / choicesParser defaultMeaning
REFERENCEtextrequiredRequired positional argument
--yesbooleanfalseConfirm the action without prompting.

auth sources

List supported saved account sources without exposing credentials. Source selection is currently supported for Copilot only.

ParameterType / choicesParser defaultMeaning
PROVIDERcodex, grok, copilotrequiredRequired positional argument
--formattext, json"text"Output format

auth select

Explicitly replace this Host's account and credential-source binding. Use a login listed by auth sources copilot; source paths are resolved by the Host, not supplied on the command line.

ParameterType / choicesParser defaultMeaning
PROVIDERcodex, grok, copilotrequiredRequired positional argument; currently only Copilot supports selection
--sourcenative, copilot_cli_filerequiredCredential source kind
--accounttextrequiredSelected GitHub login
--formattext, json"text"Output format

auth logout

Remove the selected locally stored Model credentials. For a shared Copilot CLI file source, this also deletes the selected account's supported token fields from that file and affects Copilot CLI. Other accounts are preserved; the Host does not fall back to another source.

ParameterType / choicesParser defaultMeaning
PROVIDERcodex, grok, copilotrequiredRequired positional argument
--formattext, json"text"

login

Authenticate a compatible Model provider. Copilot supports device authorization only and defaults to the public Copilot App identity; subscribers do not need to register an OAuth App. See models and authentication for permissions and compatibility limits.

ParameterType / choicesParser defaultMeaning
PROVIDERcodex, grok, copilotrequiredRequired positional argument
--allow-account-switchbooleanfalse
--device-code, --browserbooleantrueDevice authorization (default) or a local browser callback.
--formattext, json"text"

In-chat commands

Type these in the terminal composer, not your shell. Tab completes supported syntax; /help and /? show native help. There are no /setup, /login, /approve, /deny, or /result commands. Decisions use their typed selector.

“While busy” means the command parser allows it during work; it is not permission to bypass a pending interaction or operate on an unavailable resource. /steer requires a currently steerable root operation. Attachment commands change the draft; active Enter steering sends its text and attachments together. See Use the terminal.

Command grammarAliasesWhile busyPurpose
/help [command]/?YesShow command help and keyboard shortcuts.
/mode [concise|detailed]—YesSwitch output detail without changing execution.
/theme [auto|dark|light]—YesChoose a terminal theme.
/mouse [on|off]—YesToggle wheel capture; off preserves native selection/copy.
/attach path—YesAttach a file or image to the current draft.
/paste-image—YesRead clipboard images explicitly.
/recover—NoRestore an unsent prompt.
/status—YesShow model, context, environment, and subscription usage.
/ps—YesInspect observed background processes and their last reported status.
/subagents [execution-id|next]—YesInspect this conversation's child executions and retained output.
/usage [details|subscription|reset]—YesShow recorded Thread usage; subscription/reset inspect Codex limits.
/goal task description—NoStart a Goal with bounded self-checks against the authored objective.
/steer message—YesAdd guidance while the agent is working.
/import—NoPreview and optionally enable external subagents with parent inheritance.
/agent [agent-id]—NoSwitch agent, including its model, instructions, and tools.
/model [model-id|default|defaults]—NoSelect a Project Model; default clears its preference; defaults configures global media understanding with Save/Cancel.
/fast [on|off|reset]—NoToggle supported Fast processing without saving configuration.
/pro [on|off|reset]—NoSelect Pro or Standard reasoning mode; reset inherits the Model configuration.
/thinking [level]—NoShow or change reasoning effort for subsequent turns.
/environment [mode]—NoShow or select execution permissions for subsequent turns.
/new—NoStart a fresh session; keep all saved history.
/resume [session-id]—NoSearch, preview, and name saved sessions, or resume one by ID.
/history—NoBrowse retained messages (Ctrl+T).
/notes—NoShow saved notes with full contents within display budgets.
/config—NoFind your configuration files.
/review request-id—NoInspect a pending request against the selected continuation.
/cancel—YesStop the current task.
/quit/exitYesEnd this session.

/thinking opens the selected Model's supported choices; command completion uses the same list. default inherits the Model's configured settings. Depending on the model and adapter, explicit choices can include off, minimal, low, medium, high, xhigh, or max. Unsupported choices are rejected, and unknown models offer only default with an explanation. /environment selects full-control or sandbox where supported. Native /steer retains its whole trailing message rather than splitting it into shell words.

Consequential actions and limitations

  • auth key delete REFERENCE --yes skips the delete confirmation explicitly; it does not make deletion reversible.
  • plugin uninstall ID removes installed files and edits immediately. It is not merely disabling a source.
  • update requests installation without a second confirmation. A startup update check still requires confirmation before installation.
  • import subagents previews unless --apply is supplied; imported files are not automatically enrolled in every Agent.
  • One-shot run does not open interactive approval selectors; suspended/failed operations exit nonzero. Resume interactively for decisions.
  • webui --dangerous-skip-permissions changes HTTP authentication only, not Agent execution permissions. Read Browser server before exposing a listener.

Task-oriented guides

On this page