
claude-work-mode
Session work modes for Claude Code: /work-mode auto | out | wrap, with optional gates
Install with your AI
Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.
Install and set up claude-work-mode (claude-plugin project) into my current project. Found on https://claudeers.com/claude-work-mode Repo: https://github.com/danyuchn/claude-work-mode Homepage/docs: — Detected install method: claude-plugin → /plugin install claude-work-mode@danyuchn/claude-work-mode Category: devtools. Platforms: cli, api. Read the repo's README for exact setup and env vars, then install it and wire it into my project. Claudeers Health Verdict: unknown; community-verified: false. Confirm the source before running anything.
⚠ Unverified / not recently updated — review before pasting a run-this config.
/plugin marketplace add danyuchn/claude-work-mode /plugin install claude-work-mode@danyuchn/claude-work-mode
git clone https://github.com/danyuchn/claude-work-mode
// compatibility
| Platforms | cli, api |
|---|---|
| Operating systems | — |
| AI compatibility | claude |
| License | MIT |
| Pricing | open-source |
| Language | TypeScript |
claude-work-mode
Session work modes for Claude Code, switched with one slash command:
| Mode | What the model is told |
|---|---|
auto | Do not stop to ask about internal, reversible steps (local files, local commits, tests). Hard stops (outgoing messages, installs, secrets, unrecoverable deletes, payments, permission changes, production changes) go to a pending list reported once at the end. Calls that will raise a confirmation prompt are run last, back to back. |
out | Delegate batch work a checker or spot check can verify (extraction, classification, summaries, reviews) to a background subagent or an external CLI; the main thread verifies. A one-time nudge fires when the main thread has read a lot of raw output in one turn without delegating. |
wrap | Converge: open no new branches, finish what is open, list only new or changed decisions, each with a label and a recommendation. |
off | Back to normal. |
Modes are per session, combine freely (/work-mode auto wrap), and wrap wins over out when both are typed. A bare /work-mode opens a multi-select picker. The mode text reaches the model once (with the next tool result or prompt) and again only after a compaction or resume, so it does not bloat every turn.
Demo (60s Walkthrough)
[!TIP] Click the preview above to play the 60-second walkthrough video demonstrating behavioral changes across all three modes.
Install
/plugin install work-mode --marketplace danyuchn/claude-work-mode
Answer y to add the marketplace, then pick a scope. Optional gates:
/plugin install work-mode-gates --marketplace danyuchn/claude-work-mode
Requirements: Claude Code with function-hook plugins (2.1.314 or later). The gates need bash, jq, git and python3.
The optional gates (work-mode-gates)
The modes are written to ~/.claude/state/tempo/<session_id> so shell hooks can read them. hooks/lib/tempo-auto.sh gives a gate two helpers: tempo_auto_on "$INPUT" and tempo_deny "<reason>". Under auto, a gate that would ask denies instead, with a reason telling the model what to do, so work continues without a prompt. The plugin ships three gates:
| Gate | Event | Normally | Under auto |
|---|---|---|---|
scratch-doc-guard.sh | PreToolUse Write | asks before creating a new SUMMARY*/REPORT*/NOTES*/FINDINGS*/ANALYSIS*.md | denies: update the existing doc, use a scratch dir, or answer in the reply |
git-safe-commit-gate.sh | PreToolUse Bash | asks on git commit with paths (with or without --) and on git add <file>... when a file already had unstaged edits | denies: use git add -p, then a plain git commit |
decision-label-guard.py | Stop | blocks a reply whose decisions section holds an unlabeled item, or a tradeoff item that does not state what each side gains | same |
Use tempo-auto.sh in your own gates for anything you want converted. Keep real hard stops (secrets, force pushes, deploys) out of it: they should ask in every mode.
Configuration (optional)
Everything works with no config. Language follows, in order: WORK_MODE_LOCALE, the Claude Code language setting (anything mentioning Chinese picks Traditional Chinese), then LC_ALL / LANG; English otherwise.
To personalise, create ~/.config/work-mode/config.json (or point WORK_MODE_CONFIG at a file). See examples/config.example.json. Every field is optional:
| Field | Default | Meaning |
|---|---|---|
locale | detected | en or zh-TW |
userName / pronouns | the user / they | how the mode texts refer to you (they, he, she) |
autonomous.extraRules | [] | extra lines for the auto text |
dispatch.pickerHint | subagents / external CLI | shown in the picker and the switch reply |
dispatch.instructions | built in | replaces the whole out text, one line per entry |
dispatch.commandPattern | codex exec, gemini, claude -p | a Bash command matching this regex counts as delegating |
dispatch.nudgeThresholdChars | 20000 | raw output read before the nudge fires (env DISPATCH_NUDGE_CHARS overrides) |
dispatch.nudgeTemplate | built in | nudge text; {chars} is filled in |
dispatch.quotaFile / poolRouting | none | optional quota table (see below) |
converge.sectionHeading | Decisions Needed / 待你決定 | heading for decision items |
converge.decisionLabels / tradeoffLabel | outbound payment delete permission tradeoff | labels each decision must start with |
converge.autoCloseCommand | null | a command the model runs on its own when nothing is open, e.g. a session-closing skill |
The decision-label gate reads the same file for its heading and labels; without one it accepts both the English and the Chinese label sets.
Quota-aware routing
If you track usage of external model pools, point dispatch.quotaFile at a JSON array like
[{"pool":"agy","gemini_5h_used":40,"gemini_7d_used":12,"route":"open"},{"pool":"codex","5h_used":97,"7d_used":15,"route":"floor_5h","route_used":97,"route_reset_ts":1767225600,"route_reset_in":"0h45m"}].
With out on, the model then gets one line saying which pool is past its floor, worded by poolRouting, and it is repeated only when the table changes.
Embedding in another mod
tempo.ts exports setTempoStatusSink(label => line). A mod that joins several parts into one status line can register these hooks itself and pass a function that takes the work-mode label and returns the whole line to show.
Development
./check.sh
type-checks the plugin, validates the marketplace and both plugins, and runs the plugin tests, the decision-label self-test and the gate tests. It needs tsc and the Claude Code API types (written when the plugin-authoring skill loads; or set CLAUDE_CODE_TYPES to a claude-code.d.ts).
License
MIT
// faq
What is claude-work-mode?
Session work modes for Claude Code: /work-mode auto | out | wrap, with optional gates. It is open-source on GitHub.
Is claude-work-mode free to use?
claude-work-mode is open-source under the MIT license, so it is free to use.
What category does claude-work-mode belong to?
claude-work-mode is listed under devtools in the Claudeers registry of Claude-compatible tools.
// embed badge
[](https://claudeers.com/claude-work-mode)
// retro hit counter
[](https://claudeers.com/claude-work-mode)
// reviews
// guestbook
// related in Developer Tools
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Curs…
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
AI coding assistant skill (Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and more). Turn any folder of code, SQL schemas, R scripts, shell scripts, docs,…
Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA
