JiuwenSwarm is an intelligent AI Agent built on openJiuwen. It extends the powerful capabilities of large language models directly to your fingertips through various communication apps you use daily.
Slash Commands Reference
This document categorizes commands by parsing location: TUI Local Parsing and Gateway / Agent Parsing.
For quick reference of current behavior; final implementation follows the code.
Overview: By Parsing Side
TUI Local Parsing (CLI Built-in)
Executed locally in the terminal UI, not through Gateway control pipeline.
| Command | Description |
|---|---|
/clear |
Clear screen |
/color |
Adjust TUI color scheme |
/copy |
Copy last message |
/exit |
Exit |
/help |
Show available commands |
/keybindings |
View, edit, or reset TUI keyboard shortcuts (alias /keybind) |
/theme |
Switch theme |
/config |
Modify configuration (currently local, planned to unify with Gateway) |
/context |
Show context window usage and token breakdown (see below) |
/workspace |
Manage trusted directories (see below) |
/teamskills |
TeamSkills Hub publish/delete (publish/delete) |
/export |
Export current conversation to file or clipboard (see below) |
/status |
Show jiuwenswarm status overview, usage, config (see below) |
/statusline |
Configure the TUI footer status bar with a custom command (see below) |
/permissions |
Manage tool permissions (allow/ask/deny) |
/evolve |
Trigger skill self-evolution for one skill (see below) |
/evolve_list |
Show one skill's evolution records (see below) |
/evolve_simplify |
Simplify and consolidate one skill's evolution records (see below) |
/evolve_rebuild |
Rebuild SKILL.md from archives and evolution records (see below) |
/hooks |
Browse configured hooks (read-only, see below) |
/simplify |
Code simplify review: checks reuse, quality, efficiency and auto-fixes (code.* only, see below) |
/sandbox |
Set sandbox mode (see below) |
/agents |
Manage Agent configs (list, get, create, update, enable, disable, delete, see below) |
/auto-harness |
Auto-Harness task management (run/schedule/issue, see below) |
/btw |
Ask a quick side question without interrupting the main conversation (see below) |
Note:
/modecontrolled switching logic is primarily on Gateway side, see "/modeand/switch" below. The TUI local command additionally supports/mode planand/mode team.normal; see the TUI guide for details.
Gateway / Agent Parsing (Controlled Channel)
Identified by Gateway and forwarded to AgentServer and other backend capabilities.
| Command | Description |
|---|---|
/plan |
Switch to planning sub-mode |
/resume |
Resume historical session (see below) |
/new_session |
Create new session (IM only) |
/mode |
Mode switching (supports first-level entry and direct syntax) |
/switch |
Switch second-level mode within current mode family |
/skills |
Skills management (list, install, uninstall, marketplace, ClawHub, SkillNet) (see below) |
/model |
Model view, add, edit, delete, switch (see below) |
/mcp |
MCP server management (see below) |
/diff |
Interactive change review: per-turn diffs + uncommitted working tree changes (see below) |
/compact |
Compress current context (see below) |
/init |
Project initialization (see below) |
/branch |
Create a branch session from current conversation point (see below) |
/rewind |
Rewind conversation to before a specific turn (see below) |
/memory |
Memory management (see below) |
/cron |
Scheduled task (cron job) management (see below) |
/review |
Code review a pull request (see below) |
/security-review |
Security review of pending changes on the current branch (see below) |
Key Command Details
/workspace (TUI Trusted Directory Management)
Manages directories AI can access for file read, edit, and execute operations.
Subcommands
| Command | Description |
|---|---|
/workspace or /workspace get |
Show system default workspace and current trusted directories list |
/workspace add [path] |
Add trusted directory (defaults to cwd; error if path doesn't exist) |
/workspace set <path> |
Reset trusted dirs to single path (confirmation required if dirs exist) |
/workspace remove <path> |
Remove specified trusted directory |
/workspace clear |
Clear all trusted directories (use default workspace only) |
Concepts
- System default workspace: Fixed path
~/.jiuwenswarm/agent/jiuwenclaw_workspace, always available - Trusted directories (
trusted_dirs): User-authorized accessible directories, managed by TUI, passed to backend Agent
Control Logic
Startup confirmation: TUI prompts user whether to trust current directory
- "Trust" → add current directory as trusted
- "Don't trust" → use default workspace only
Session-level management: Trusted directories are effective for current CLI session, not persisted to file
Backend passing: TUI passes
trusted_dirsvia request params; Agent restricts file operations accordinglyPath restriction: Agent limits file operations within trusted directories; operations outside require user confirmation
Path validation:
addandsetvalidate path existence; error shown if invalid
Aliases
/workspace_dir, /workspace-dir
/mode and /switch (Controlled Channel)
- First-level entry mapping:
/mode agent->agent.plan/mode code->code.normal/mode team->team
- Direct syntax:
/mode agent.plan->agent.plan/mode agent.fast->agent.fast/mode code.normal->code.normal/mode code.team->code.team
- Second-level switching:
- agent family:
/switch plan<->agent.plan,/switch fast<->agent.fast - code family:
/switch normal<->code.normal,/switch team<->code.team
- agent family:
- Invalid combinations (e.g.,
/switch fastundercode.*) return:Invalid command. - Controlled channels do not accept
/mode planor/mode team.normal. - Note: Standalone
/teamcommand removed, use/mode teaminstead.
/resume
/resume list: List historical sessions./resume <conversation_id>: Resume specified session.
Interactive picker (TUI)
Entering /resume or /continue with no arguments opens an interactive session picker (instead of a plain session.list dump).
| Key | Action |
|---|---|
↑ / ↓ |
Move focus between sessions |
Enter |
Resume the focused session |
| type chars | Live search (filter by session ID / title / project dir) |
Backspace |
Delete a search character |
Space |
Preview the focused session info card (title, ID, project dir, branch, message count, last active / created). In preview: Enter resumes, Space/Esc goes back |
Ctrl+R |
Rename the focused session. In edit mode: Enter saves, Esc cancels, empty value clears the title |
Ctrl+A |
Toggle scope between "all projects" and "current project only" |
Ctrl+B |
Toggle git branch filter (only show sessions whose git_branch strictly equals the current project's branch) |
Esc |
Clear search if any; otherwise close the picker |
Space/Ctrl+R/Ctrl+A/Ctrl+B/Escin the list can be rebound under theResumeListcontext via/keybindings(preview/rename sub-states and search text entry stay hardcoded).
Behavior:
- Defaults to listing all projects (press
Ctrl+Ato narrow to the current project). When the current project has no sessions, an (empty) picker still opens so you can pressCtrl+A. - Branch recording & filtering (
Ctrl+B): a session's git branch is recorded (per itsproject_dir) on the first message (HEADfor non-git/detached). When the filter is on, sessions are matched by branch name strictly; legacy sessions without a recorded branch andHEADsessions are filtered out. Note the match is by name only and not repo-aware — with "all projects + branch filter" enabled, same-named branches in different directories are shown together. - Restore scope: resume only restores the conversation context (history, session ID, accent color, workflow snapshot, window title); it does not switch the workspace / current working directory.
/model (View / Add / Edit / Delete / Switch Model)
Manages model configs defined under models.defaults in config.yaml. Supports both text subcommands and an interactive list; delete/edit are done primarily through the interactive list.
Text usage:
/modelor/model list: Open the interactive model list (with current model marker); switch / add / edit / delete can all be done inside the list;/model <name>: Switch directly to the model by name;/model add <name> key=value ...: Add a model config via text form (e.g.,model_name=...,provider=...,api_base=...,api_key=...,reasoning_level=...).
Interactive list hotkeys (after opening the list via
/modelor/model list):Key Action ↑/↓Select model up/down EnterSwitch to the selected model aOpen the form to add a model eOpen the form to edit the selected model dEnter delete confirmation for the selected model EscClose list / cancel current action Delete flow (
d→ confirm): After entering theDelete model: <name>confirmation page—Enter: confirm deletion; the backend locates the entry by its original index (index, i.e. the position inmodels.defaultsbefore filtering out multimodal keys like video/audio/vision), removes it, writes back toconfig.yaml, then triggers an Agent config reload (AGENT_RELOAD_CONFIG). Success response:{type: "model_deleted", name, current}.Esc: cancel, return to the list.
Edit flow (
e): reuses the add form, but leavingapi_keyempty means keep the original key unchanged; only changed fields are submitted.Limitations & validation:
video/audio/visionare multimodal-only keys and cannot be set as the default chat model via/model <name>; use/config editor/config setinstead;- Deleting the last remaining model is rejected (
Cannot delete the last model); an out-of-range index returnsmodel index not found;
Config write behavior:
- Add / edit / delete mutate
models.defaultsinconfig.yaml(compatible with the old structure) and trigger an Agent config reload; - Switching model validates config and environment variable placeholders, updates
MODEL_NAME/MODEL_PROVIDER/API_BASE/API_KEY, writes back to.env.
- Add / edit / delete mutate
Secure display: sensitive fields like
api_key,tokenare masked in the list; only when models share the same name AND identical provider+api_base (genuinely indistinguishable) does the list append the key's last 4 characters[…xxxx].
/diff (Interactive Change Review)
- Usage:
/diff(no subcommands). - Applicable modes: All modes.
- Data source: TUI sends a
command.diffrequest to the AgentServer (60s timeout). The handler resolves the currentsession_idandproject_dirfrom request metadata, then fetches two data sets in parallel via worker threads:turns— per-turn change sets derived from.agent_historyfile operation logs;gitDiff— uncommitted working tree changes fromgit diff HEAD.
- Response payload:
{ type: "list", turns: [...], gitDiff?: {...} }. On error:{ ok: false, error: "..." }. - Display mode: Opens a full-screen interactive Diff viewer:
- List view: Shows all changed files (working tree
workingand per-turnTurn N) with relative paths, source label, and added/removed line counts; - Detail view: Press
Enteron a selected file to view its full hunk-by-hunk diff with scrolling support.
- List view: Shows all changed files (working tree
- List view keybindings:
↑/↓— Move selection (auto-scrolls);Enter— View full diff for the selected file;Home/g— Jump to top;End/Shift+g— Jump to bottom;Esc/Ctrl+C— Close.
- Detail view keybindings:
↑/↓— Scroll line by line;PgUp/PgDn— Page up / down;Home/g— Go to file top;End/Shift+g— Go to file bottom;←/Esc— Return to list view.
- Fallback: When the TUI does not provide the
enterDiffViewercapability, falls back to inline display (file names, source, and line stats only).
Turn-based diff data source
Per-turn diffs are computed from .agent_history/file_ops_jiuwenswarm*.json logs, not from git. The service reads and merges file operation logs from multiple locations:
- Agent workspace (
~/.jiuwenswarm/agent/jiuwenclaw_workspace/.agent_history/) - User workspace
.agent_history/ - Project directory
.agent_history/(session-specific and global files)
Entries are deduplicated by path normalization and timestamp proximity (±1 second). Turn boundaries are defined by user messages in session history: a turn spans from one user message timestamp to the next. Only turns with file changes are returned; empty turns are filtered out. Turn indices are preserved (aligned with the actual user message count in history) to allow /rewind to map correctly.
Git diff data source
Working tree changes are obtained via git diff HEAD (tracked files only). The git repo root is resolved from project_dir (which may be a subdirectory). Staged renames (including brace shorthand form a/{b => c}/d.txt) are handled to align numstat keys with hunk keys.
Effect boundaries and limits
| Boundary | Value | Behavior |
|---|---|---|
| Max files in detail | 50 | Only the first 50 tracked files get hunks; stats still cover all changed files |
| Max lines per file | 400 | Hunks are truncated beyond 400 lines per file; isTruncated flag is set |
| Max diff size per file | 1 MB | Files whose diff exceeds 1 MB are skipped in hunk parsing; isLargeFile flag is set, stats still counted |
| Max files for details | 500 | If more than 500 files changed, only aggregate stats are returned (no per-file hunks) |
| Git command timeout | 10s | Git commands that exceed 10 seconds return None |
| Git root resolution timeout | 5s | git rev-parse --show-toplevel that exceeds 5 seconds returns None |
What is NOT covered
- Untracked files:
git diff HEADonly covers tracked file modifications. Untracked files are excluded from the git diff section. (Untracked files edited by the agent may still appear in per-turn diffs via file_ops logs.) - Manual/bash edits in turn diffs: Per-turn diffs are derived from the agent's
.agent_historyfile operation logs. Files edited manually or via bash commands are not tracked in turn diffs. - Committed changes: Only uncommitted working tree changes are shown. Committed history is not covered.
- Transient git states: During merge, rebase, cherry-pick, or revert, git diff returns
Noneto avoid showing misleading incoming changes. - Binary files: Binary file changes are counted in stats but no hunks are shown (
isBinaryflag set). - Not a git repo: If
project_diris not in a git repository,gitDiffisNone; only per-turn diffs are returned. - No project_dir: If
project_dircannot be resolved,gitDiffisNone; per-turn diffs may still work using session metadata.
/diffis not a replacement forgit difffrom a full version control perspective. It combines agent-tracked per-turn changes with a snapshot of uncommitted tracked-file changes for quick review within a coding session.
/compact (Context Compression)
- Usage:
/compact(no parameters). - Function: Trigger context compression,清理对话 history but keep summary in context.
- Data source: TUI requests Agent compression service via
command.compact. - Results:
busy: Compression in progress, retry later;compressed: Success, shows before/after token counts and savings ratio;noop: No compression needed, context already optimal.
/context (Context Window Usage)
- Usage:
/context(no parameters, no subcommands). - Function: View the current session's context window occupancy and token usage details.
- Data source: TUI requests Agent context statistics service via
command.context, carrying the currentmode. - Display contents:
- Overview panel: Context window occupancy percentage + progress bar;
context_window(used / limit tokens),occupancy(rate),messages(count); - Token breakdown panel: Shows token usage by
system_prompt,messages,tools, andtotal; - DeepAgent occupancy details (if available): Key-value list of
context_occupancyfields; - DeepAgent usage details (if available): Key-value list of
deepagent_usagefields.
- Overview panel: Context window occupancy percentage + progress bar;
- Threshold warning: When occupancy >= 90%, the overview title shows
Context window 90% full — consider /compact. - Error handling: On request failure, displays
context failed: <error message>.
/init (Project Initialization)
- Usage:
/init(no parameters). - Function: Initialize project AI collaboration config, generates
JIUWENSWARM.mdand optionallyJIUWENSWARM.local.md. - Scope: Only runs in
codemode. - Flow:
- Select scope:
Team-shared(JIUWENSWARM.md),Personal(JIUWENSWARM.local.md), orBoth. - Detect existing configs: Auto-detect
CLAUDE.md,.cursorrules,copilot-instructions.mdetc. - Generate configs: Create project config files based on selection.
- Select scope:
- Auto mode switch: Code initialization runs in
code.normalfor write permission.
/mcp (MCP Server Management)
- Usage:
/mcp list: List all MCP servers (name, transport, enabled status);/mcp show [name]: Show MCP config; withoutnameshows enabled items, withnameshows one server detail;/mcp add --name <name> --transport <stdio|sse|http> ...: Add an MCP server;/mcp update --name <name> ...: Update MCP server config (transport / params / enabled status);/mcp enable <name>: Enable a specific MCP server;/mcp disable <name>: Disable a specific MCP server;/mcp remove <name>: Remove a specific MCP server.
- Transport parameters:
stdio: requires--command; optional--args,--cwd,--env;sse/http(alias ofstreamable-http): requires--url; optional--headers,--timeout_s.
- Examples:
/mcp list/mcp show/mcp show playwright/mcp add --name playwright --transport stdio --command python --args "server.py --transport stdio"/mcp update --name playwright --transport sse --url http://127.0.0.1:9000/sse --headers "Authorization=Bearer xxx"/mcp add --name local-sse --transport sse --url http://127.0.0.1:9000/sse/mcp disable playwright/mcp remove local-sse
- Config and effect:
- Changes are written to
config.yamlundermcp.servers; - After write, Agent config reload is triggered, and runtime MCP server bindings are synced accordingly.
- Changes are written to
/evolve* (Skill Self-Evolution)
These commands are registered and parsed by the TUI, then forwarded as slash text through the normal chat channel. The actual evolution logic runs on the Agent / Team backend:
- Agent mode: handled by
SkillEvolutionRail; onlyagent.planis supported. - Team mode: handled by
TeamSkillEvolutionRailfor team skill evolution. - Code mode and
agent.fastdo not support these commands.
Subcommands
| Command | Description |
|---|---|
/evolve <skill_name> [user_query] |
Trigger evolution for one skill. agent.plan scans the current conversation for tool failures and user corrections; Team mode requires user_query. |
/evolve_list <skill_name> [--sort score] |
Show one skill's evolution records with count, average score, usage/feedback stats, section, and content preview. |
/evolve_simplify <skill_name> [user_intent] |
Generate an approval-gated cleanup plan to merge duplicates, split long records, or remove low-value records. Trailing text is passed to the backend as intent. |
/evolve_rebuild <skill_name> [user_intent] |
Same as control-plane skills.evolution.rebuild: adopt live evolution records and generate a new SKILL.md version (prepare → rewrite → finalize). |
Approval Flow
/evolveand/evolve_simplifydo not silently write changes; the backend pushes a confirmation question and the TUI waits for approval.- Accepting persists/solidifies the generated records; rejecting discards this generation.
- Accepted Team skill evolution syncs the team skill directory.
- While evolution or approval is pending, supplemental user input is queued and sent after evolution completes.
Examples
/evolve pptx improve export error handling
/evolve_list pptx --sort score
/evolve_simplify pptx merge duplicate export-failure records
/evolve_rebuild pptx strengthen Troubleshooting and Examples
/branch (Branch Session)
- Usage:
/branch [name]. - Alias:
/fork. - Function: Create a branch session from the current conversation state, copying the current conversation history.
- Constraints:
- Rejected when the session is busy (
session is busy); - Rejected when the current session has no conversation records.
- Rejected when the session is busy (
- Behavior:
- Generate a new
session_idand sendsession.forkRPC to the backend (carryingsource_session_id,target_session_id, and optional title). - TUI automatically switches to the new branch session, clears the current transcript, and restores the branch history.
- Prompts the user that they are now in the new branch, and informs them they can use
/resume <original_session_id>to return to the original session.
- Generate a new
- Examples:
/branch— Create an untitled branch/branch fix-login-bug— Create a branch namedfix-login-bug
/rewind (Rewind Conversation)
- Usage:
/rewind [turn_number]. - Alias:
/checkpoint. - Function: Rewind the current session to before a specified turn, supporting conversation-only, code-only, or both.
- Constraints:
- Rejected when the session is busy (
session is busy); - Rejected when there are no conversation turns.
- Rejected when the session is busy (
- Interactive flow:
- Without parameters, displays a list of all conversation turns (with timestamps and file change statistics) for the user to select the target turn.
- After selecting, displays restore options:
- Restore conversation and code — Truncate conversation and restore files to their prior state;
- Restore conversation only — Only truncate conversation, files remain unchanged;
- Restore code only — Only restore files, conversation remains unchanged (shown only when the target turn has file changes);
- Cancel — Abort the operation.
- Calls the corresponding backend RPC based on selection:
both→session.rewind_and_restoreconversation→session.rewindcode→session.restore_files
- After rewind: TUI clears the transcript and reloads history; if the rewound content contains user input, it is automatically filled into the input box.
- Limitation: Rewinding does not affect files edited manually or via bash commands.
- Examples:
/rewind— Interactive turn selection and restore mode confirmation/rewind 2— Directly rewind to before turn 2
/memory (Memory Management)
- Alias:
/mem. - Function: View and manage memory system status, memory files, toggle settings, and directory paths via a tabbed console.
- Subcommands:
| Command | Description |
|---|---|
/memory or /memory edit |
Open the tabbed console and select the edit tab |
/memory edit <path> |
Directly edit the specified memory file (via $EDITOR) |
/memory status |
Open the tabbed console and select the status tab |
/memory toggle |
Open the tabbed console and select the toggle tab |
/memory toggle <key> |
Directly toggle the specified memory system switch |
/memory open |
Open the tabbed console and select the open tab |
- Tabbed console: The console has 4 tabs — edit / status / toggle / open. Interaction keys:
←/→— Switch tabs;↑/↓— Navigate within the current tab;Enter— Execute the selected item;Ctrl+O— Toggle full path display (edit / open tabs); resets to default (relative path) when switching tabs;Esc— Close the console.
editdisplay contents:- Lists memory files in priority order: Project memory (Checked in at
<path>), Local memory (Saved in<path>), User memory (Saved in<path>), plus any rule files; Enteropens the selected file with$EDITOR.
- Lists memory files in priority order: Project memory (Checked in at
statusdisplay contents:- Engine (format
builtin (local), combining the storage engine and storage mode); - Switch row — mode-adaptive (mirrors the current mode's toggle set), shown as
✓ on/✗ off:- agent mode:
memory_enabled(Memory),memory_proactive(Proactive memory),memory_forbidden_enabled(Forbidden filter); - code mode:
memory_enabled(Memory),auto_coding_memory(Auto coding memory),memory_forbidden_enabled(Forbidden filter).
- agent mode:
- Runtime memory statistics — mode-adaptive: agent mode shows "Auto Memory" (files / chars / dir); code mode shows "Coding Memory" (files / chars / dir);
- Project Memory statistics (files / chars / project dir);
- External Memory (provider + enabled status), shown only if configured.
- Engine (format
toggledisplay contents:- Each row shows the toggle key (padded to equal width), a
✓ on/✗ offstatus marker, and a Chinese description. Only the key is shown (no English label); columns are separated by spaces, not·. Example (code mode):→ memory_enabled ✓ on 记忆功能总开关 auto_coding_memory ✓ on 每轮对话后自动提取记忆(需总开关开启) memory_forbidden_enabled ✗ off 过滤敏感信息 - Available keys (mode-adaptive):
- agent mode:
memory_enabled(Master memory switch),memory_proactive(Proactive memory),memory_forbidden_enabled(Forbidden filter); - code mode:
memory_enabled(Master memory switch),auto_coding_memory(Auto coding memory),memory_forbidden_enabled(Forbidden filter).
- agent mode:
- After toggling, a prompt is shown if a session restart is required for the change to take effect.
- Each row shows the toggle key (padded to equal width), a
opendisplay contents:- Lists memory directory paths — mode-adaptive:
- agent mode: Memory Dir / Project Dir / User Project Dir;
- code mode: Coding Memory Dir / Project Dir / User Project Dir.
Enteropens the selected directory in the system file manager directly.
- Lists memory directory paths — mode-adaptive:
- Tab completion:
/memory— Suggests subcommands (edit,status,toggle,open);/memory edit— Suggests memory file paths (viagetDisplayPath, relative/tilde-shortened, deduplicated, only existing rule files);/memory toggle— Suggests toggle keys (mode-adaptive, mirrors the current mode's toggle set);- All completions support prefix filtering.
- Examples:
/memory— Open the tabbed console (edit tab)/memory edit memory/MEMORY.md— Edit a specific memory file/memory status— View detailed status/memory toggle memory_enabled— Toggle the master memory switch/memory open— View memory directory paths
/cron (Scheduled Task Management)
Manage cron jobs via RPC calls to the backend CronController, sharing the same backend logic and data store with the Web UI.
- Alias:
/crontab - Subcommands:
| Command | Description |
|---|---|
/cron or /cron list |
List all cron jobs |
/cron show <job_id> |
Show detailed info for a specific job |
/cron add name=<name> cron_expr=<expression> description=<desc> [other params] |
Create a new cron job |
/cron update <job_id> key=value ... |
Update specific fields of a job |
/cron delete <job_id> |
Delete a job |
/cron toggle <job_id> <on or off> |
Enable or disable a job |
/cron run <job_id> |
Run a job immediately |
/cron preview <job_id> |
Preview upcoming execution times for a job |
addparameters:
| Parameter | Required | Description |
|---|---|---|
name |
Yes | Job name |
cron_expr |
Yes | Cron expression, supports two formats: 5-field (min hour day month dow) or 7-field Quartz (sec min hour day month dow year). 5-field is auto-converted to 7-field (second=0, year=*). Examples: daily 9am = 0 9 * * * (5-field) or 0 0 9 * * ? * (7-field) |
description |
Yes | Job description — the input prompt the Agent receives when executing |
targets |
No | Push channel, default tui; options: tui, web, feishu, whatsapp, wecom, xiaoyi, wechat, dingtalk, or feishu_enterprise:<app_id>. With targets=tui, results broadcast to all connected TUI windows; see Scheduled tasks — Push to TUI |
timezone |
No | IANA timezone, default Asia/Shanghai |
mode |
No | Execution mode, default agent.fast. Options: agent, agent.fast, agent.plan, plan, team, team.plan, code.team. Team modes use streaming multi-agent execution; see Scheduled tasks — Team mode |
timeout_seconds |
No | Per-run timeout in seconds (60–259200). Default 600 for normal modes, 1200 for team modes |
wake_offset_seconds |
No | Wake-up offset in seconds, default 0 |
delete_after_run |
No | Auto-delete after one run, default false |
addexamples:/cron add name=minute-test cron_expr="0 * * * *" description="Tell me the current time" targets=tui/cron add name=morning-brief cron_expr="0 9 * * *" description="Generate today's morning briefing" targets=tui mode=agent.plan/cron add name=model-weekly cron_expr="0 9 * * 1" description="Compare GLM vs DeepSeek and output a report" targets=tui mode=team/cron add name=reminder cron_expr="0 30 17 29 4 ? 2026" description="Don't forget the meeting" targets=tui delete_after_run=true/cron add name=weekly-report cron_expr="0 9 * * 1" description="Generate weekly report" targets=web
updateusage: Only pass the fields you want to change, e.g.,/cron update <id> name=new-name enabled=falseshowdisplay: full job details in key-value format (id, name, status, cron_expr, timezone, description, targets, mode, timeout_seconds, wake_offset_seconds, delete_after_run)listdisplay: sequence number, full job ID, name, cron expression, enabled status, description snippetpreviewdisplay: wake_at and push_at timestamps for each upcoming execution
/skills (Skills Management)
Manage skills lifecycle: listing, installing, uninstalling, marketplace source management, ClawHub and SkillNet online skill registries.
Subcommands
| Command | Description |
|---|---|
/skills or /skills list |
List skills (grouped: Installed / Available to install) |
/skills install <skill> or /skills install <slug@clawhub> or /skills install <name@skillnet> or /skills install <skill@marketplace> or /skills install <path_or_url> |
Install a skill: builtin accepts bare name, ClawHub uses <slug>@clawhub, SkillNet uses <name>@skillnet (auto-searches for URL), marketplace uses <name>@<marketplace>, local paths and URLs auto-detected |
/skills uninstall <name> |
Uninstall a skill by name |
/skills marketplace or /skills marketplace list |
List marketplace sources (name, URL, enabled status, last updated) |
/skills marketplace add <name> <url> |
Add a new marketplace source |
/skills marketplace remove <name> |
Remove a marketplace source (also clears its cache) |
/skills marketplace toggle <name> <on or off> |
Enable or disable a marketplace source (on/true/1 = enable, otherwise disable) |
/skills marketplace clawhub |
View ClawHub token status (configured/not configured) |
/skills marketplace clawhub token <value> |
Set ClawHub CLI token |
/skills marketplace clawhub token |
View ClawHub token status |
/skills skillnet or /skills skillnet search <query> |
Search SkillNet skill registry (shows name, description, author, stars, category, URL) |
/skills skillnet install <skill_url> |
Install a skill from SkillNet by URL (async download, auto-polls progress) |
/skills use <skill_name>, <query> |
Execute a query using a specific skill |
Concepts
- Skill: An extension capability that can be installed from marketplace sources, ClawHub, SkillNet, builtin directory, or local paths, providing additional functionality to the agent.
- Builtin skill: A preset skill shipped with the software. Install using bare skill name (e.g.,
/skills install advanced-daily-report); no marketplace source needed. - ClawHub: An online skill registry (clawhub.ai) hosting community-published skills. Install using
<slug>@clawhubformat, where slug is the skill's unique identifier (not its display name). Requires a ClawHub CLI token to be configured first. - SkillNet: An academic skill registry. Two install methods:
<name>@skillnet(auto-searches to find URL then installs) and/skills skillnet install <url>(direct URL install). - Marketplace source: A remote Git repository that hosts available skills. Each source has a name, URL, and enabled/disabled state.
- Spec: The install identifier format supporting:
<skill>@builtin(builtin),<slug>@clawhub(ClawHub),<skill>@<marketplace>(Git marketplace); bare names without@are auto-detected as builtin if applicable. - Local install: Use
/skills install <path>to install from a local directory (must containSKILL.md) or remote archive URL; paths/URLs are auto-detected and routed to the local import flow. - Install location: The directory where a skill is stored after installation (
~/.jiuwenswarm/agent/jiuwenclaw_workspace/skills/). - Source tag: Each skill in the list is tagged with its source:
[builtin]= builtin,[local]= imported,[clawhub]= ClawHub,[skillnet]= SkillNet,[project]or marketplace name = other.
Grouped List Display
/skills list returns skills in two groups:
- Installed: Skills already in the user's skills directory, ready to use.
- Available to install: Builtin skills not yet installed, plus marketplace skills available for installation. Use
/skills installfirst.
IM vs TUI Differences
Both ultimately request skills.list, but trigger methods and display differ.
| Side | Trigger Method | Behavior |
|---|---|---|
| IM (Feishu etc. controlled channel) | Exact match /skills list (whitespace normalized first) |
Gateway intercepts control message and requests skills.list, results shown as IM notification/card; standalone /skills doesn't go through this control path. |
| TUI (CLI built-in) | Input /skills |
Locally executes built-in command and calls skills.list, displays as grouped list view in session (titles Installed Skills and Available Skills); shows No installed skills when empty. |
For other subcommands (/skills install, /skills uninstall, /skills marketplace add/remove/toggle, /skills use), Gateway does not intercept them — on the IM side they are treated as regular chat messages. These subcommands are only functional on the TUI (CLI built-in) and Web UI paths, where they send RPC requests directly to AgentServer.
Notes
- Timeout:
install,uninstall, andmarketplace togglerequests have a 120-second timeout on the TUI side; other subcommands have no explicit timeout. - Builtin auto-detection: When installing with
/skills install <skill>(no@), the system checks if it matches a builtin skill and redirects to the builtin install flow; if not, a format hint is returned. - Path/URL auto-detection: When installing with
/skills install <path_or_url>(local path like/path/to/skillorC:\skill, or remote URLhttps://...), the system automatically routes to the local import flow (skills.import_local). All URLs go through import_local; SkillNet is not auto-routed from URLs. @skillnetsearch-install: When using/skills install <name>@skillnet, the frontend first callsskills.skillnet.search. Only auto-installs if an exact match by skill_name is found; with no exact match, it only displays search results (with URLs and names) without auto-installing the first result — the user must choose one and install via/skills skillnet install <url>or/skills install <exact_name>@skillnet. This is because SkillNet search is semantic: searching "code" may return "taskflow", "coding-agent" etc. whose names don't contain "code".- ClawHub token required: A ClawHub CLI token must be configured before installing from ClawHub (via
/skills marketplace clawhub token <value>). Without a token,@clawhubinstalls will fail with a message explaining how to set the token. Obtain your token at clawhub.ai. - ClawHub slug vs. display name: ClawHub skills are identified by their unique slug (e.g.,
code-review-security), not their display name (e.g., "Code Review Assistant"). When a direct slug install fails, the system automatically searches ClawHub and displays matching results (with real slugs and summaries) to help you find the correct skill. - ClawHub overwrite confirmation: When the target slug already exists (same name from any source counts as installed), TUI presents an interactive prompt: "Skill xxx is already installed. Do you want to force overwrite?". Choosing "Yes" re-installs with
force: true, replacing the old skill; choosing "No" or exiting keeps the existing skill unchanged. The Web UI bypasses confirmation and usesforce: truedirectly. - SkillNet async install: SkillNet installation is asynchronous — it initiates a download task and returns an
install_id, then TUI automatically pollsinstall_statusevery 800ms until completion or failure (max wait: 15 minutes). Progress is shown asDownloading... (install_id: xxx). - SkillNet overwrite confirmation: Same as ClawHub — TUI prompts the user interactively when a skill already exists. Web UI uses
force: truedirectly. - SkillNet accessible in China: SkillNet search API is hosted at
http://api-skillnet.openkg.cn(OpenKG platform) and is directly accessible in China without VPN. However, the skill content itself is hosted on GitHub, which may require VPN. - Same-name skills cannot coexist: Skills are stored as directories at
skills/{name}/, and the filesystem does not allow two directories with the same name. Installing a skill with the same name from a different source will overwrite the previous one (with user confirmation)./skills useonly uses the skill name and cannot distinguish between sources. - ClawHub network access: ClawHub API is hosted at
https://clawhub.ai. VPN may be required in regions with restricted access to this domain. - Cache cleanup:
marketplace removesends{ name, remove_cache: true }to also clear the local cache for that source. - Auto-refresh:
marketplace add,marketplace remove, andmarketplace toggleautomatically re-list marketplace sources after a successful operation. - Offline handling:
/skills usechecks connection status; if offline, showsoffline: waiting for reconnect before sending /skills use request.
Examples
/skills— List skills (grouped: Installed / Available)/skills list— List skills (explicit subcommand)/skills install advanced-daily-report— Install a builtin skill (bare name auto-detect)/skills install advanced-daily-report@builtin— Install a builtin skill (explicit format)/skills install code-review@clawhub— Install a skill from ClawHub (using slug)/skills install code-review@skillnet— Install from SkillNet (auto-searches for URL)/skills skillnet search code-review— Search SkillNet skill registry/skills skillnet install https://github.com/user/skill-repo— Install via SkillNet subcommand (direct URL)/skills install my-skill@marketplace— Install a skill from Git marketplace/skills install /path/to/my-skill— Install a skill from local directory/skills install https://example.com/skill.zip— Install from remote URL (local import)/skills uninstall my-skill— Uninstall a skill/skills marketplace list— List marketplace sources/skills marketplace add community https://github.com/user/skills-repo— Add a marketplace source named "community"/skills marketplace remove community— Remove the "community" marketplace source/skills marketplace toggle community on— Enable the "community" marketplace source/skills marketplace toggle community off— Disable the "community" marketplace source/skills marketplace clawhub— View ClawHub token status/skills marketplace clawhub token abc123xyz— Set ClawHub CLI token/skills use my-skill, Code and execute a Hello World program.— Use a skill to execute a query
/export (Export Conversation)
Export the current conversation to a file or clipboard.
Usage
/export— Copy conversation to clipboard (no filename argument)/export <filename>— Save conversation to a.txtfile in workspace directory
Subcommands
| Command | Description |
|---|---|
/export |
Copy entire conversation to clipboard; if clipboard unavailable, prompt to specify a filename |
/export <filename> |
Write conversation to filename.txt in workspace directory; if filename lacks .txt extension, it is automatically appended |
Output Format
The exported text renders each conversation entry with a timestamp and role prefix:
[User] <timestamp>— User input[Assistant] <timestamp>— Assistant response[Thinking] <timestamp>— Internal reasoning trace[Tools] <timestamp>— Tool calls with name, summary, and truncated result (max 500 chars)[System] / [Error] / [Info] <timestamp>— System messages[Diff] <timestamp>— Per-turn file change summary
Tab Completion
When typing /export and pressing Tab, auto-generated filename suggestions appear:
<timestamp>-<sanitized-first-prompt>.txt— Based on the first user message (truncated to 50 chars, sanitized)conversation-<timestamp>.txt— Generic timestamped name
Timestamp format: YYYY-MM-DD-HHmmss.
Behavior Details
- Clipboard fallback: If no filename is given and clipboard is unavailable, an error message prompts the user to specify a filename instead.
- Filename normalization: Any extension is replaced with
.txt; e.g.,/export my-chat.jsonbecomesmy-chat.txt. - Write location: Files are saved to
ctx.getWorkspaceDir()(orprocess.cwd()as fallback).
Examples
/export— Copy conversation to clipboard/export my-chat— Save tomy-chat.txtin workspace/export 2026-05-09-debug-session.txt— Save with explicit timestamp name
/simplify (Code Simplify Review)
Parsed locally by the TUI, this command sends a dedicated RPC command.simplify to get a server-generated three-phase review prompt, then injects it as an Agent message (logAsUser: false). The Agent automatically reviews changed code for reuse, quality, and efficiency, and directly fixes issues found.
- Scope: reuse / quality / efficiency only. Security vulnerabilities (injection, XSS, hard-coded secrets, auth flaws, etc.) are out of scope — do not fix or report them here. Use
/security-reviewfor a read-only security report. - Alias: None.
- Applicable modes:
code.*only. Non-code mode shows an error promptingRun /mode code first. - Parsing location: TUI local (not a Gateway controlled channel); unavailable in IM.
Usage
| Command | Description |
|---|---|
/simplify |
Review current git changes (or recently edited files) and auto-fix issues |
/simplify <target> |
Add focus: file path, module name, or specific review dimension |
Execution Flow
- TUI validates: Confirms current mode starts with
code.; errors otherwise. - RPC request: Calls
command.simplify(with optionaltarget), 30-second timeout. - Server generates prompt: Builds a three-phase review instruction from
_SIMPLIFY_PROMPT_TEMPLATE; appends## Additional Focussection iftargetis provided. - Injects into Agent: TUI calls
ctx.sendMessage(prompt, ..., { logAsUser: false })to inject the prompt; the Agent begins execution. - Offline handling: Shows a retry message if offline.
Three-Phase Review
Phase 1 — Identify Changes: Run git diff (or git diff HEAD for staged changes) to find changed files. If no git changes, review recently edited files from the conversation.
Phase 2 — Launch Three Review Agents in Parallel (use sub-agent tools if available; otherwise perform all reviews directly):
| Review Dimension | Focus Areas |
|---|---|
| Code Reuse Review | Existing utilities/helpers that could replace new code; duplicated functionality; hand-rolled logic that could use an existing utility |
| Code Quality Review | Redundant state; parameter sprawl; copy-paste variations; leaky abstractions; stringly-typed code (use existing constants/enums); unnecessary JSX nesting; unnecessary comments (keep only non-obvious WHY) |
| Efficiency Review | Unnecessary work (redundant computations, repeated I/O, N+1 patterns); missed concurrency; hot-path bloat; recurring no-op updates; TOCTOU anti-patterns; memory leaks / missing cleanup; overly broad operations |
Phase 3 — Fix Issues: Aggregate all findings and fix each issue directly. Skip false positives without argument. Briefly summarize what was fixed (or confirm the code was already clean).
Examples
/simplify— Review all changes/simplify src/auth/— Focus on changes undersrc/auth//simplify focus on error handling patterns— Emphasize error handling
/sandbox (Sandbox Mode Management)
Enter / leave jiuwenbox sandbox mode and tune its runtime policy. Calls command.sandbox on the agent server.
Subcommands
| Command | Description |
|---|---|
/sandbox or /sandbox status |
Show current runtime (enabled, landlock, excluded_commands, files.allow_write, files.deny_write) |
/sandbox enable |
Enter sandbox mode (spawns jiuwenbox if needed, rebuilds agent) |
/sandbox disable |
Leave sandbox mode (rebuilds agent; stops jiuwenbox only if jiuwenswarm started it) |
/sandbox exclude add <pattern> |
Add a shell glob whose matches run locally instead of in the sandbox |
/sandbox exclude remove <pattern> |
Remove a pattern |
/sandbox exclude list |
List current excluded_commands |
/sandbox files allow <path> |
Allow write access to <path> inside the sandbox (shown as rw) |
/sandbox files deny <path> |
Deny write access to <path> inside the sandbox (read still allowed, shown as ro) |
/sandbox files remove <path> |
Remove <path> from the user-configured allow & deny sets |
/sandbox files list |
List effective allow_write / deny_write |
/sandbox help |
Print usage |
Concepts
- Platform support:
/sandboxis Linux-only (jiuwenbox depends on Linux kernel features such as bwrap, Landlock, and Linux namespaces). On a Windows or macOS agent-server, every/sandboxsub-command returns aSANDBOX_BAD_REQUESTerror. If the TUI runs on Windows/macOS but the agent-server is on a Linux host, the command works — what matters is the agent-server's platform. - Write policy semantics:
allow/denycontrol write access (rw/ro) inside the sandbox, not Unix octal modes. Enforcement uses bwrap bind mounts +--remount-ro; Landlock is defense-in-depth (whenlandlock.compatibility=disabled, bwrap is primary). - Nested paths: Supported: parent allow + child deny (e.g. allow
/tmp, deny/tmp/secret). Not supported: child allow + parent deny (parent deny wins); the server rejects such configs. - Effective write policy:
files.allow_write/files.deny_writein the status panel show the merged view of auto-managed and user-configured entries, each labeled(rw)or(ro). - preserve_file_sharing_mode: Controlled by jiuwenswarm config, not by
/sandbox. Onlymountis supported: intrinsic files andproject_dirare bind-mounted into the sandbox andproject_dir/config/config.yamlis explicitly added todeny_write. Writing any other value into config.yaml is rejected by the server. - excluded_commands:
fnmatchper simple-command leaf (full leaf text or command name). All matches → whole command on host; none → whole command in sandbox; mixed leaves → local bash orchestrates and wraps remote leaves withjiuwenbox sandbox exec(CLI required). - Add / remove are strict:
exclude addrejects a pattern that is already in the list;exclude removerejects a pattern that is not in the list.files allow|denyrejects a path that is already in the same bucket, and rejects a path that exists in the opposite bucket (allow vs deny conflict) — runfiles removefirst if you want to flip it.files removerejects paths that have no matching user-configured entry. - enable / disable: Triggers an agent rebuild. The response lists
rebuilt_modes(typicallyagent.*/code.*) and the jiuwenbox endpoint.
Examples
/sandbox enable— turn on sandbox mode/sandbox status— see runtime + effective files/sandbox files allow ./tmp/— allow sandbox write access to./tmp/(rw)/sandbox files deny ./tmp/secret/— deny write under an allowed parent (ro)/sandbox exclude add "git *"— letgitrun on the host instead of inside the sandbox
/keybindings (Keyboard Shortcuts)
View, edit, or reset TUI keyboard shortcuts. Config file: ~/.jiuwenswarm-tui/keybindings.json.
Usage
| Command | Action |
|---|---|
/keybindings |
Same as /keybindings edit |
/keybindings edit |
Create or open the config file; reload after the external editor closes |
/keybindings list |
List effective shortcuts grouped by context |
/keybindings reset |
Delete user config and restore built-in defaults |
Alias: /keybind.
Notes
- Built-in defaults are merged with user JSON per context; set a key to
nullto unbind a default. - Key ids must match pi-tui
matchesKeyformat (ctrl/shift/alt+ main key); chords are not supported. - Unsupported keys:
win/cmd/super/meta: These modifiers require the Kitty keyboard protocol, which is not available in standard terminals (Windows CMD, VS Code integrated terminal, etc.). Bindings using these will never fire.ctrl+shift+letter: On legacy terminals like Windows CMD,ctrl+shift+landctrl+lproduce the same byte — the terminal cannot distinguish them. This combination is not recommended. Use Windows Terminal, WezTerm, or another VT-mode capable terminal if you need these combos.- Chords (space-separated multi-key sequences, e.g.
"ctrl+x ctrl+k"): Not supported in the current version.
- Non-rebindable:
ctrl+c,ctrl+d,ctrl+m(reserved keys). - Select-list navigation, config editor text input, Resume preview/rename sub-states, etc. remain hardcoded.
See the Chinese TUI User Guide · Keyboard shortcuts for the full context/action reference.
/hooks (Browse Hooks Configuration)
View a summary of all hooks configured in config.yaml (read-only).
Usage
/hooks(no parameters, no subcommands)
Data Source
The TUI requests the Gateway via the hooks.list RPC, which loads the hooks section from config.yaml and returns a summary.
Display Contents
/hooks displays hooks configuration in three levels:
- Event List (Level 1): All events sorted by hook count in descending order. Each row shows event name and hook count; the description column shows hook count distribution per matcher.
- Status Panel:
Source— Configuration source (config.yaml)Global Status— Global toggle state (enabled/DISABLED)Total Hooks— Total hook count across all eventsActive Events— Events with at least 1 hook / total events (out of 17)
- Hook Detail Cards (Level 2): Grouped by
Event > Matcher, each hook shows:Type—command(shell command) orprompt(LLM review)Command/Prompt— The hook contentTimeout— Timeout in secondsShell— Execution shell (command hooks only)Status— Status message
When No Hooks Are Configured
If config.yaml has no hooks configured, displays No hooks configured. with a hint to use /config edit to configure them.
Hooks Concept Overview
Hooks are extension logic that executes automatically when specific events fire. 17 events are supported:
| Event | Execution Layer | Trigger |
|---|---|---|
PreToolUse |
Agent Rail | Before a tool call |
PostToolUse |
Agent Rail | After a successful tool call |
PostToolUseFailure |
Agent Rail | After a failed tool call |
Stop |
Agent Rail | After agent response completes |
PermissionRequest |
Agent Rail | On permission request |
PermissionDenied |
Agent Rail | On permission denied |
SubagentStart |
Agent Rail | When a sub-agent starts |
SubagentStop |
Agent Rail | When a sub-agent stops |
BeforeModelCall |
Agent Rail | Before a model call |
AfterModelCall |
Agent Rail | After a model call |
UserPromptSubmit |
Gateway | User submits a message |
SessionStart |
Gateway | Session starts |
SessionEnd |
Gateway | Session ends |
Notification |
Gateway | Notification is sent |
ConfigChange |
Gateway | Configuration changes |
InstructionsLoaded |
Gateway | Instructions are loaded |
Setup |
Gateway | Initialization |
Two hook types are supported:
| Type | Description | Key Parameters |
|---|---|---|
command |
Executes a shell command (subprocess). Receives JSON context via $ARGUMENTS env var. Exit code 0 = success, 2 = block. |
command, timeout (default 30s), shell (default bash) |
prompt |
Invokes LLM review. $ARGUMENTS in the template is replaced with JSON context, $TOOL_NAME with the tool name. LLM response JSON with decision: "block" blocks the operation. |
prompt, timeout (default 15s), model |
- Blocking: Exit code 2 (command) or
decision: "block"(prompt) blocks the current operation (e.g., skip tool call) and feeds the reason back to the model. - Input Modification: PreToolUse hooks can modify tool input parameters via
modifiedInputin stdout JSON. - Additional Context: Extra information can be injected into tool results or model context via
additionalContextin stdout JSON. - Global Toggle:
hooks.disable_all_hooks: trueinconfig.yamldisables all hooks.
Configuration Example
hooks:
PreToolUse:
- matcher: "write_file"
hooks:
- type: command
command: "echo 'write_file about to execute' >> /tmp/hooks.log"
timeout: 10
- matcher: "bash|run_command"
hooks:
- type: prompt
prompt: "Review if this command is safe: $ARGUMENTS"
timeout: 20
SessionStart:
- matcher: "*"
hooks:
- type: command
command: "echo 'Session started: $ARGUMENTS' >> /tmp/hooks.log"
Example
/hooks— Browse all current hooks configuration
/agents (Agent Management)
Manage custom agents (subagents) throughout their full lifecycle: view, create, update, enable/disable, and delete. Agent definitions are stored as Markdown files and support four-level source priority merging.
- Parsing location: TUI local parsing, calling backend
agents.*endpoints via RPC. - Applicable modes: All.
- Note: This command is registered as hidden (
hidden: true) and does not appear in/helplistings but can be used directly.
Subcommands
| Command | Description |
|---|---|
/agents or /agents list |
List all agents (name, source, enabled status, description summary) |
/agents get <name> |
View full details of a specific agent (including System Prompt) |
/agents create [--project|--local] <name> <description> |
Create a custom agent; LLM auto-generates the prompt |
/agents update <name> [--generate] <new description> |
Update agent description; add --generate to have LLM rewrite the prompt |
/agents enable <name> |
Enable a custom agent (cannot operate on builtin agents) |
/agents disable <name> |
Disable a custom agent (cannot operate on builtin agents) |
/agents delete <name> |
Delete a custom agent (cannot operate on builtin agents) |
Agent Sources & Storage
| Source | Storage Location | Priority | Manageable |
|---|---|---|---|
builtin |
In-code builtins | Lowest | Cannot enable/disable/delete |
local |
<workspace>/.jiuwenswarm/agents-local/ |
Local | Full lifecycle management |
user |
~/.jiuwenswarm/agents/ |
User | Full lifecycle management (default create location) |
project |
<workspace>/.jiuwenswarm/agents/ |
Highest | Full lifecycle management |
Agents with the same name are resolved by project > user > local > builtin priority; shadowed agents are marked with shadowed_by.
Agent Definition Fields
| Field | Description |
|---|---|
name |
Agent name (unique identifier) |
description |
Brief description |
prompt |
System Prompt text |
source |
Origin (builtin / user / project / local) |
file_path |
Agent definition file path |
model |
Specified model (null = use default) |
tools |
Available tool list |
disallowed_tools |
Disallowed tool list |
color |
Display color |
permission_mode |
Permission mode |
memory_scope |
Memory scope |
when_to_use |
When-to-use description |
max_iterations |
Max iterations (default 200) |
skills |
Associated skill list |
enabled |
Enabled status (true / false / null) |
shadowed_by |
Which source shadows this agent (null = active) |
/agents create Behavior
- Argument parsing:
--project/--localare positional flags and must precede the name (e.g.,/agents create --project my-agent description). - LLM generation: By default, the current model auto-generates
when_to_useandsystem_prompt; falls back to a built-in template on failure. - Auto-enable: After creation, automatically writes
react.subagents.<name>.enabled = truetoconfig.yamland hot-reloads the configuration. - Timeout: 60 seconds.
- Output: Displays LLM generation marker, storage location, and file path.
/agents update Behavior
- No description: When no description is provided, shows current agent details (same as
get) and prints usage. --generate: Explicitly triggers LLM prompt rewriting; without this flag, the request's template values are used.- Auto hot-reload: Configuration is automatically reloaded after update.
/agents enable / disable Constraints
- Builtin agents (
source == "builtin") cannot be enabled/disabled; the backend returns an error. - The operation writes to
config.yaml'sreact.subagents.<name>.enabledand hot-reloads.
/agents delete Constraints
- Builtin agents cannot be deleted.
- After deletion, the entry is automatically removed from
config.yaml'sreact.subagentsand hot-reloaded.
/agents get Display
Displays all agent definition fields as key-value pairs, with the full System Prompt text appended at the end.
Tab Completion
The get, update, enable, disable, and delete subcommands support Tab completion on agent names (fetched via the agents.list RPC).
Examples
/agents # List all agents
/agents list # Same as above
/agents get Explore # View Explore agent details
/agents create bug-hunter Root cause analysis expert # Create user-level agent
/agents create --project proj-agent Project-level # Create project-level agent
/agents create --local local-agent Local use only # Create local agent
/agents update bug-hunter --generate Better description # Update with LLM prompt rewrite
/agents enable bug-hunter # Enable agent
/agents disable bug-hunter # Disable agent
/agents delete my-agent # Delete agent
/status (Show Status)
Display jiuwenswarm runtime status: overview, usage statistics, or config editor.
Usage
/statusor/status overview— Show core identity, model/API info, MCP servers, and config sources/status usage— Show session token usage statistics/status config— Enter interactive config editor
Subcommands
| Command | Description |
|---|---|
/status |
Show full status overview (version, session, model, connection, MCP servers, config) |
/status overview |
Same as /status — explicit overview subcommand |
/status usage |
Show session token usage (input, output, total, per-model breakdown) |
/status config |
Enter interactive config editor (same as /config edit) |
Overview Display Sections
When /status is run, four key-value panels are displayed:
- Core identity: version, session ID, session name (or prompt to
/rename), cwd, current mode - Model & API: model name, provider, API base URL, connection status
- MCP servers: each server's name, transport type, and enabled/disabled state
- Config sources: config file path and all settings source paths
Usage Display
/status usage shows token consumption for the current session:
- Total input tokens, output tokens, and total tokens
- Per-model breakdown: model name, token count, input/output split
Interactive Mode
If the TUI provides an interactive StatusView (ctx.enterStatusView), /status opens the full status UI with tabs. The subcommand argument selects the initial tab:
/status→ opens on overview tab/status usage→ opens on usage tab/status config→ opens on config tab
If StatusView is unavailable, the command falls back to inline key-value display.
Data Sources
- Overview data:
command.statusRPC request to AgentServer - Usage data:
ctx.getUsageSummary()from local session tracking - Config data:
config.getRPC request to AgentServer
Examples
/status— Show full overview/status overview— Show overview (explicit)/status usage— Show token usage/status config— Open config editor
/statusline (TUI Footer Status Bar)
Configure the TUI footer status bar with a custom shell command that dynamically displays session info (mode, model, cwd, etc.), modeled after Claude Code's /statusline implementation.
Subcommands
| Command | Description |
|---|---|
/statusline or /statusline get |
View current status line configuration |
/statusline set <shell-command> |
Set the status line command (its output will appear in the TUI footer) |
/statusline clear |
Remove the status line configuration (footer bar will hide) |
/statusline help |
Show usage guide (writing patterns, practical examples, field list) |
/statusline json |
Show the actual current JSON data values (useful for debugging jq expressions) |
Concepts
- StatusLine: A text area at the bottom of the TUI that displays user-defined dynamic information, supporting multi-line output. When a custom statusline is configured, the built-in status line is automatically hidden to avoid redundant information.
- Shell command: The configured shell command is automatically executed every 2 seconds; its stdout output is rendered as the status bar text.
- JSON input: Each execution receives current session info as JSON, which can be parsed with
jqor other tools. On POSIX (Linux/macOS), JSON is passed via stdin pipe; on Windows, due to MSYS2 pipe inheritance limitations, the system automatically writes JSON to a temp file and replaces$(cat)in the command with$(cat "filepath")— the user doesn't need to modify their command format. - Prerequisites: Requires
jq(https://stedolan.github.io/jq/) for JSON parsing; Windows users also need to add Git Bash'susr\bindirectory to the system PATH (e.g.,E:\Git\usr\bin).
JSON Input Fields
The command receives the following JSON data on each execution:
| Field | Description |
|---|---|
session_id |
Current session ID |
session_name |
Session title (set via /rename) |
cwd |
Current working directory |
mode |
Current mode (agent.plan / agent.fast / code.normal / code.team / team) |
model |
Current model name |
provider |
Model provider |
version |
jiuwenswarm version |
connection |
Connection status (idle / connecting / connected / reconnecting / auth_failed) |
theme |
Current theme name |
accent_color |
Current accent color name |
transcript_mode |
Transcript display mode (compact / detailed) |
transcript_fold_mode |
Fold mode (none / tools / thinking / all) |
is_processing |
Whether agent is working (true / false) |
is_paused |
Whether paused (true / false) |
is_interrupted |
Whether interrupted (true / false) |
cancellable_work |
Whether there is running work (true / false) |
streaming_state |
Streaming state (idle / streaming / tool_call / tool_result) |
last_error |
Last error message or null |
evolution_status |
Evolution status (idle / running) |
active_subtask_count |
Number of active subtasks |
todo_count |
Number of todo items |
trusted_dirs |
Trusted workspace directories (array of path strings) |
usage.total_input_tokens |
Total input tokens for session |
usage.total_output_tokens |
Total output tokens for session |
usage.total_tokens |
Total tokens for session |
context_window.context_window_size |
Model max context window tokens (e.g. 200000) |
context_window.used_percentage |
Context occupancy percentage (0-100) |
context_window.remaining_percentage |
Context remaining percentage (0-100) |
Command Writing Template
Use the following template to write commands. input=$(cat) reads JSON into a variable, then echo "$input" | jq -r .field extracts each field. // "default" is jq's fallback syntax — when a field is null or empty, the default value is used.
General formula:
/statusline set 'input=$(cat); field1=$(echo "$input" | jq -r '.field1 // "default"'); field2=$(echo "$input" | jq -r '.field2 // "default"'); echo "format string"'
Recommended universal command (shows mode, model, tokens, context %, connection):
/statusline set 'input=$(cat); mode=$(echo "$input" | jq -r '.mode // "?"'); model=$(echo "$input" | jq -r '.model // "?"'); tokens=$(echo "$input" | jq -r '.usage.total_tokens // 0'); pct=$(echo "$input" | jq -r '.context_window.used_percentage // 0'); conn=$(echo "$input" | jq -r '.connection // "?"'); echo "$mode | $model | ctx:${pct}% | tokens:$tokens | $conn"'
Field extraction quick reference:
| Field to display | jq syntax |
|---|---|
| Session name | jq -r '.session_name // ""' |
| Working directory | jq -r '.cwd // "?"' |
| Mode | jq -r '.mode // "?"' |
| Model name | jq -r '.model // "?"' |
| Provider | jq -r '.provider // "?"' |
| Version | jq -r '.version // "?"' |
| Connection | jq -r '.connection // "?"' |
| Is processing | jq -r '.is_processing // false' |
| Is paused | jq -r '.is_paused // false' |
| Streaming state | jq -r '.streaming_state // "idle"' |
| Last error | jq -r '.last_error // ""' |
| Evolution status | jq -r '.evolution_status // "idle"' |
| Subtask count | jq -r '.active_subtask_count // 0' |
| Todo count | jq -r '.todo_count // 0' |
| Trusted dirs | `jq -r '(.trusted_dirs // []) |
| Total input tokens | jq -r '.usage.total_input_tokens // 0' |
| Total output tokens | jq -r '.usage.total_output_tokens // 0' |
| Total tokens | jq -r '.usage.total_tokens // 0' |
| Context window size | jq -r '.context_window.context_window_size // 0' |
| Context used % | jq -r '.context_window.used_percentage // 0' |
| Context remaining % | jq -r '.context_window.remaining_percentage // 0' |
More Examples
/statusline— View current configuration/statusline set 'input=$(cat); model=$(echo "$input" | jq -r .model); echo "$model"'— Show model name only/statusline set 'input=$(cat); proc=$(echo "$input" | jq -r .is_processing); model=$(echo "$input" | jq -r .model); echo "$proc | $model"'— Show processing state and model/statusline set 'input=$(cat); pct=$(echo "$input" | jq -r .context_window.used_percentage); rem=$(echo "$input" | jq -r .context_window.remaining_percentage); cw=$(echo "$input" | jq -r ".context_window.context_window_size / 1000"); echo "ctx:${pct}% used (${rem}% left, ${cw}K window)"'— Show context window occupancy with percentage bar/statusline set 'input=$(cat); pct=$(echo "$input" | jq -r ".context_window.used_percentage // 0"); if [ "$pct" -ge 90 ]; then warn="⚠HIGH"; elif [ "$pct" -ge 70 ]; then warn="~MED"; else warn="OK"; fi; echo "ctx:${pct}% $warn"'— Show context % with threshold warning (≥90% HIGH, ≥70% MED)/statusline set 'input=$(cat); err=$(echo "$input" | jq -r .last_error); if [ "$err" != "null" ] && [ "$err" != "" ]; then echo "error: $err"; else echo "ok"; fi'— Show error when present, otherwise "ok"/statusline set 'input=$(cat); dirs=$(echo "$input" | jq -r '.trusted_dirs // [] | join(" ")'); mode=$(echo "$input" | jq -r '.mode // "?"'); echo "$mode | dirs:$dirs"'— Show mode and trusted workspace directories/statusline clear— Remove status line configuration/statusline help— View usage guide (writing patterns, practical examples, available fields)/statusline json— View actual current JSON data values (useful for debugging jq expressions)
Behavior Details
- Poll frequency: The configured command runs every 2 seconds automatically.
- Timeout protection: Individual executions timeout after 3 seconds; no impact on subsequent polls.
- Output limit: Command output over 10KB is truncated; display width auto-fits the TUI terminal width.
- Failure silence: Command execution failures don't show errors; previous successful output is kept or the bar hides.
- Persistence: Configuration is saved in
~/.jiuwenswarm-tui/config.jsonunder thestatusLinefield; restored on TUI restart. - Alias:
/sl - Windows adaptation: The system automatically replaces
$(cat)with reading from a temp file; the user's command format remains unchanged. Git Bash'susr\binmust be in the system PATH.
Config File Structure
{
"statusLine": {
"type": "command",
"command": "input=$(cat); mode=$(echo \"$input\" | jq -r '.mode // \"?\"'); model=$(echo \"$input\" | jq -r '.model // \"?\"'); pct=$(echo \"$input\" | jq -r '.context_window.used_percentage // 0'); tokens=$(echo \"$input\" | jq -r '.usage.total_tokens // 0'); echo \"$mode | $model | ctx:${pct}% | tokens:$tokens\"",
"padding": 0
}
}
/auto-harness (Auto-Harness Task Management)
Manage Auto-Harness task creation, execution, and monitoring. Auto-Harness generates harness extension packages via automated pipelines, supporting two pipeline types:
- optimize_expert_harness (backend value
extended_evolve_pipeline): Generate a local harness extension package - optimize_meta_harness (backend value
meta_evolve_pipeline): Submit PR (requires git config)
During pipeline execution, extension packages are activated automatically by default — no manual user confirmation is needed. Logs display harness.extension_ready (extension ready, showing directory and component info) and harness.activate_interaction (activation confirmation prompt) events.
Configuration Requirements
Using the optimize_meta_harness pipeline requires the following fields to be configured (via /config edit or /status config):
| Field | Required | Description |
|---|---|---|
git.user_name |
Yes | Git commit username |
git.user_email |
Yes | Git commit email |
git.fork_owner |
Yes | Fork repository owner (e.g., SnapeK) |
gitcode.access_token |
No | GitCode API token (can also be provided via environment variable GITCODE_ACCESS_TOKEN) |
If configuration is incomplete, the task creation will prompt the missing fields.
Subcommands
| Command | Description |
|---|---|
/auto-harness run [--pipeline <pipeline>] <query> |
Execute a one-time Auto-Harness task |
/auto-harness schedule start --interval <hours> [--pipeline <pipeline>] <query> |
Create a scheduled task |
/auto-harness schedule list |
List all tasks |
/auto-harness schedule status <task_id> |
View task details |
/auto-harness schedule logs <task_id> [--history <n>] |
View task execution logs |
/auto-harness schedule cancel <task_id> |
Cancel a task |
/auto-harness schedule delete <task_id> |
Delete a task |
/auto-harness issue fix <issue_numbers> |
Create fix tasks for GitCode issues |
/auto-harness issue scan [--repo <repo>] [--page <n>] [--labels <labels>] [--force-refresh] |
Scan repo GitCode issues |
/auto-harness issue status |
View GitCode issue processing status |
/auto-harness issue delete <issue_numbers> |
Delete issue processing records |
/auto-harness run (One-time Execution)
- Usage:
/auto-harness run [--pipeline <pipeline>] <query> - Flow:
- If pipeline is not specified, interactively select the pipeline type
- If
optimize_meta_harnessis selected, automatically check git config completeness - Create and execute a one-time task
- Automatically enter real-time log streaming mode (similar to
tail -f)
- Examples:
/auto-harness run Optimize database query performance— No pipeline specified, interactive selection/auto-harness run --pipeline optimize_expert_harness Optimize context compression— Specify pipeline
/auto-harness schedule start (Create Scheduled Task)
- Usage:
/auto-harness schedule start --interval <hours> [--pipeline <pipeline>] <query> - Parameters:
--interval/-i(required): Execution interval in hours; options:1,2,4,8,12,24--pipeline/-p(optional): Pipeline type; interactively selected if not specified<query>(required): Optimization target description
- Flow:
- If pipeline is not specified, interactively select
- If
optimize_meta_harnessis selected, check git config - Interactively confirm whether to run immediately
- Create the scheduled task
- Examples:
/auto-harness schedule start --interval 4 Optimize context compression/auto-harness schedule start -i 2 -p optimize_meta_harness Submit database optimization PR
/auto-harness schedule logs (View Execution Logs)
- Usage:
/auto-harness schedule logs <task_id> [--history <n>] - Modes:
- Default: Stream current running logs in real-time (
tail -fmode); Ctrl+C to interrupt --history <n>: View historical execution logs (viewmode,nis the history index, 0 = most recent)
- Default: Stream current running logs in real-time (
/auto-harness issue (GitCode Issue Auto-Fix)
Manage GitCode issue auto-processing: scan issue matrix, create fix tasks, view status, clean up records.
Requires git.user_name, git.user_email and gitcode.access_token (or GITCODE_ACCESS_TOKEN env var) to be configured.
Subcommands
| Command | Description |
|---|---|
/auto-harness issue fix <issue_numbers> |
Create fix tasks for GitCode issues |
/auto-harness issue scan [--repo <repo>] [options] |
Scan repo issues |
/auto-harness issue status |
View issue processing status |
/auto-harness issue delete <issue_numbers> |
Delete issue processing records |
/auto-harness issue fix (Create Fix Task)
- Usage:
/auto-harness issue fix <issue_numbers> - Parameters:
<issue_numbers>: Issue number(s), comma-separated, e.g.1272,1271,1270--repo <repo>: Target repository (jiuwenswarm/agent_core); interactively selected if not specified
- Issues with bound PRs (open or merged) are automatically skipped
- Examples:
/auto-harness issue fix 1286/auto-harness issue fix 1272,1271,1270
/auto-harness issue scan (Scan Issue)
- Usage:
/auto-harness issue scan - Parameters:
--repo <repo>: Target repository; interactively selected if not specified--page <n>: Page number, default 1--labels <labels>: Label filter, comma-separated; defaults to bug type only--force-refresh: Force refresh from GitCode API (uses cache by default)
- Displays: issue number, title, labels, difficulty, last updated
- Examples:
/auto-harness issue scan/auto-harness issue scan --repo jiuwenswarm --page 1/auto-harness issue scan --repo agent_core --force-refresh
/auto-harness issue status (View Status)
- Usage:
/auto-harness issue status(no parameters) - Lists all issue processing records in table format: number, status, stage, progress, details
- Example:
/auto-harness issue status
/auto-harness issue delete (Delete Records)
- Usage:
/auto-harness issue delete <issue_numbers> - Parameters:
<issue_numbers>: Issue number(s) to delete
- Examples:
/auto-harness issue delete 123/auto-harness issue delete 123 456
/btw (By-the-way Side Question)
Parsed locally by the TUI, this command sends a dedicated RPC command.btw to AgentServer to run an isolated, tool-free, single-turn LLM query against the current conversation context. It answers a quick side question without interrupting the main conversation.
- Alias: None.
- Applicable modes: All.
- Constraint: A question must be provided; returns
no_contextwhen no conversation context exists yet.
Usage
| Command | Description |
|---|---|
/btw <question> |
Ask a side question based on current conversation context |
Behavior Details
- Required argument:
/btwmust include a question; otherwise showsUsage: /btw <your question>. - Thinking indicator: Displays
💭 Answering: <question>(dim style) while the request is in flight. - RPC timeout: 120 seconds.
- Server-side handling:
- Backend receives the request via
command.btwRPC and obtains the current Agent instance. - Shares the main Agent's system prompt (project context, skills, CLAUDE.md, etc.) and retrieves recent conversation messages as context.
- Builds a dedicated btw prompt with a
<system-reminder>telling the model: no tools available, single response only, main Agent is not interrupted. - Calls the model directly (no tools, single-turn), without modifying conversation history (read-only).
- Backend receives the request via
- Return statuses:
ok: Displays💡 /btw <question>+ answer content.no_context: ShowsNo conversation context available yet — send a message first.failed: Shows error message orCouldn't answer the side question.
Examples
/btw what does git status do?/btw What is the time complexity of this code?
/review (Code Review a Pull Request)
When entered in the TUI, sends the raw /review text as a chat message to the Gateway. The Gateway recognizes it, injects a review prompt, and the Agent uses gh CLI to review the PR.
In IM controlled channels (Feishu etc.), the Gateway intercepts /review, injects the prompt, and forwards to AgentServer for execution.
- Alias: None.
- Applicable modes: All (Agent, Code, Team).
- Parsing location: Gateway controlled channel (
scope: "gateway"); TUI sends as a chat message.
Usage
| Command | Description |
|---|---|
/review |
Without arguments, the Agent runs gh pr list to show open PRs |
/review <PR number or URL> |
Review a specific PR: Agent runs gh pr view/diff and analyzes |
Behavior Details
- TUI execution: Sends
/review [args]as a user message viactx.sendMessage(); showsoffline: waiting for reconnect before sending review requestif offline. - Gateway interception (IM side):
- Matches exact
/reviewor prefix/review <arg>. - Argument max 2048 bytes; control characters rejected with an error notice.
- Injects the review prompt into
msg.params["query"]and continues forwarding to AgentServer.
- Matches exact
- Agent execution: Upon receiving the review prompt, the Agent uses
ghCLI:- Without arguments: runs
gh pr listto display open PRs. - With arguments: runs
gh pr view <number>for details andgh pr diff <number>for the diff. - Analyzes changes and provides a comprehensive review (correctness, conventions, performance, test coverage, security).
- Without arguments: runs
- No git/gh pre-check: The Gateway does not check whether
gitorghis installed; the Agent handles missing tools on its own.
Examples
/review— List open PRs in the current repo/review 123— Review PR #123
/security-review (Security Review)
When entered in the TUI, sends the raw /security-review text as a chat message to the Gateway. The Gateway recognizes it, injects a security review prompt, and the Agent uses git commands to analyze pending changes on the current branch.
In IM controlled channels (Feishu etc.), the Gateway intercepts /security-review, injects the prompt, and forwards to AgentServer for execution.
- Alias: None.
- Applicable modes: All (Agent, Code, Team).
- Parsing location: Gateway controlled channel (
scope: "gateway"); TUI sends as a chat message.
Usage
| Command | Description |
|---|---|
/security-review |
Review all pending changes on the current branch vs origin/HEAD |
/security-review <additional instructions> |
Add focus instructions or constraints (e.g., "focus on auth module") |
Behavior Details
- TUI execution: Sends
/security-review [args]as a user message viactx.sendMessage(); showsoffline: waiting for reconnect before sending security review requestif offline. - Gateway interception (IM side):
- Matches exact
/security-reviewor prefix/security-review <arg>. - Argument max 2048 bytes; control characters rejected with an error notice.
- Injects the security review prompt into
msg.params["query"]and continues forwarding to AgentServer.
- Matches exact
- Agent execution: Upon receiving the security review prompt, the Agent performs:
- Repository context research:
git status,git diff --name-only origin/HEAD...,git logto understand the change scope. - Comparative analysis:
git diff origin/HEAD...to review diffs file by file. - Vulnerability assessment across these categories:
- Input validation vulnerabilities
- Authentication and authorization issues
- Cryptography and key management
- Injection and code execution
- Data exposure
- Uses subtasks to identify vulnerabilities and parallel subtasks for false-positive filtering; only reports findings with >80% confidence.
- Outputs a structured Markdown report: file, line number, severity, category, description, exploit scenario, fix recommendation.
- Repository context research:
- Hard exclusion list: Does not report DoS, secret storage, rate limiting, race conditions, and similar issue types.
- No git pre-check: The Gateway does not check whether
gitis installed; the Agent handles this on its own.
Examples
/security-review— Review all pending changes on the current branch/security-review focus on authentication module security— With additional focus instructions
Planned Features
(None currently)