YUIBoard：本地只读的 AI 辅助开发进度看板

English · 中文 · 日本語

A local, read-only progress board for AI-assisted development.

YUIBoard reads the planning docs in your repository (spec tasks.md files, ADRs, plans, TODO lists), your GitHub pull requests, and your local Claude Code / Codex session logs. From these it works out where every task stands: not started, in progress, awaiting review, awaiting merge, or done. It shows the result in your browser as six views: an overview, a Gantt chart, a dependency graph, what to do next, what needs your attention, and diagnostics. It can also show a pane right inside Claude Code that follows the task you're working on (see Watch it inside Claude Code).

Why

When AI agents write most of the code, the hard part is keeping track of what has happened. One task is half-done in a session, another has a draft PR, and a third was merged but never checked off in the plan. YUIBoard puts that picture back together from material you already have:

• No new tracker. The task docs your AI already writes are the source of truth. There are no tickets to file and nothing to keep in sync by hand.
• Evidence over claims. A task can count as done only once it is checked in the doc, and a linked PR that is still open or merged but not yet pulled locally keeps it from being done. A checked task with no valid PR, or whose local sync can't be verified, still counts as done but is flagged "Checked in doc only" or "Local sync not verified". Every mismatch shows up under "Needs attention", and every status links back to the line, PR or session it came from.
• Sees AI work as it happens. A Claude Code or Codex session that is working on a task shows up on that task within seconds.
• Local and read-only. It never writes to your repository, your GitHub account or your session logs. It sends no telemetry and calls no AI APIs.
• Zero dependencies. It runs on Node 24's standard library and a native ES-module frontend, with no build step and no npm install.

The UI is available in English, 中文 and 日本語. It follows your browser language, and you can switch from the top bar.

Quick start

Requirements: Node 24.x and git. To see pull requests you also need the GitHub CLI, signed in with gh auth login; read-only access is enough.

Try it first with a fictional project (made-up docs, pull requests and an AI session). No config or GitHub sign-in needed:

git clone https://github.com/funnysubaru/yuiboard.git
cd yuiboard
npm run demo

The demo runs on the first free port from 4318, writes only to your system's temp directory, and deletes its files when you press Ctrl+C.

To watch your own repository:

cp devboard.config.example.json devboard.config.json   # point it at your repository (see Configuration)
node src/server.mjs                                     # or: npm start

Open the URL printed at startup (default http://127.0.0.1:4317/). To keep the config file elsewhere:

node src/server.mjs --config /path/to/devboard.config.json

• If the port is taken, the server exits with an error. A config mistake affects only the project it belongs to: that project shows an error page naming the exact field, and the other projects keep working.
• Server logs and devboard-ctl output are in English by default. Set DEVBOARD_LANG=zh or ja (or run with a Chinese or Japanese LANG) to switch. When devboard-ctl up creates the LaunchAgent or systemd service, it saves the language of the terminal you ran it from as DEVBOARD_LANG, because services usually start without LANG.
• Caches go to cache/<project id>/ and manual links to data/<project id>/; both are git-ignored. Deleting cache/ only makes the next start slower.
• Write your docs and PR titles the way CONVENTION.md describes. Most existing tasks.md files already work as they are.

Add a project with AI

Rather than editing the config by hand, you can ask your AI agent: "add this project to devboard". The bundled skill skills/devboard-add-project finds the spec tasks.md files, ADRs, plans and TODO docs, confirms the list with you, writes the config through ops/devboard-ctl.mjs, and restarts the service. Saying "remove this project from devboard" removes it.

To install the skill, run this in the YUIBoard directory. It must be a symlink, because the skill uses it to find its way back to this directory.

mkdir -p ~/.claude/skills && ln -s "$PWD/skills/devboard-add-project" ~/.claude/skills/devboard-add-project
mkdir -p ~/.agents/skills && ln -s "$PWD/skills/devboard-add-project" ~/.agents/skills/devboard-add-project

The AI changes devboard.config.json only through devboard-ctl save / remove, which back up the file to devboard.config.json.bak first. Repositories you add are only ever read.

Watch it inside Claude Code (mod)

This repository is also a Claude Code plugin. It opens a YUIBoard pane next to the conversation, in both the terminal and the desktop app. The pane shows the project the session belongs to: the overall numbers, a card with the task this session is working on, the task tree, Needs attention and Diagnostics. It reads from your running YUIBoard service and changes nothing. Tested with Claude Code 2.1.293–2.1.296.

Both screenshots show the fictional demo project. The session asked about "sync-engine task 2.3", so the card at the top shows that task and why it has its status, and the tree opens to it.

Install it in a Claude Code session:

/plugin marketplace add funnysubaru/yuiboard
/plugin install yuiboard@yuiboard

Or from a shell in one step (Claude Code 2.1.275 or later):

claude plugin install yuiboard --marketplace funnysubaru/yuiboard

The plugin and your YUIBoard service must be the same version. When you update the service (git pull, then restart it), update the plugin too: claude plugin update yuiboard@yuiboard.

To try changes from a checkout instead, load the plugin straight from it (the repository root is the plugin): claude --plugin-dir /path/to/yuiboard for one terminal session, or add "CLAUDE_CODE_PLUGIN_DIRS": "/path/to/yuiboard" to env in ~/.claude/settings.json and start a new session. Loaded this way, the plugin's id is yuiboard@inline instead of yuiboard@yuiboard.

Options:

echo '{"url":"http://127.0.0.1:4317"}' | claude plugin configure yuiboard@yuiboard --values-stdin

• The pane opens by itself once per session, as soon as the session can be matched to a project: the session is linked to one of its tasks, or its working directory is inside a watched workspace. After you close it, it stays closed. Type /yuiboard to open it again.
• The card shows the tasks the session is linked to. Name the task the way you would in a PR title, such as "<doc> task 3.2"; a bare number isn't enough. When several tasks are active, the one with the latest activity comes first. Subagents count separately, so each dispatched task stays "running" while its own subagent works.
• In the terminal, use the keyboard: press Ctrl+X then Tab to enter the pane, Tab or the arrow keys to move, Enter to press, and Esc to go back to the prompt. In fullscreen mode with at least 110 columns, the pane docks on the right and takes mouse clicks. Task numbers and "Open in browser" open the web page. Terminals without OSC 8 links print the address after the label.
• Option changes reach running sessions. Changes to the plugin's files don't; start a new session.
• The plugin also brings the skill skills/devboard-add-project. An installed plugin runs from a copy, so this copy of the skill asks where your YUIBoard checkout is before it writes anything. If you also symlinked the skill as described above, Claude Code lists it twice; either one works. The ~/.agents/skills link for Codex is not affected.

Run as a service (macOS / Linux)

Run devboard-ctl to install the service or restart it. You don't need to edit any plist or unit file by hand.

node ops/devboard-ctl.mjs up                # macOS: restart the LaunchAgent for this directory, or create one from ops/devboard.plist.template
                                            # Linux: same, with a systemd user service (ops/devboard.service.template)
node ops/devboard-ctl.mjs up --no-service   # just start a background process (PID in cache/devboard.pid)
node ops/devboard-ctl.mjs check             # validate the config only
node ops/devboard-ctl.mjs find <repo path>  # find the project registered for a repository (read-only)
node ops/devboard-ctl.mjs save --new < project.json        # add a project; refused if the id or repository is already registered
node ops/devboard-ctl.mjs save --update <id> < patch.json  # patch only name / docSources / repo.github
node ops/devboard-ctl.mjs remove <id>                      # remove a project (cache/<id> and data/<id> are kept)
node ops/devboard-ctl.mjs revert <revision>                # undo the save / remove that returned this revision

The service is reachable at http://devboard.localhost:4317/ as well as 127.0.0.1. It still listens only on the loopback address. Browsers that support .localhost resolve the name to your machine, so you don't need to edit hosts files.

Details: up, save / remove, launchd, systemd, logs

• up first checks whether the configured port is served by a devboard this directory manages. If it is, up restarts it. If another program holds the port, up moves to the next free port and writes it back to server.port. If the devboard on the port was started by hand, up reports its PID and leaves it alone. up prints one line of JSON; on ok:false it exits with code 1 and gives a machine-readable reason (unmanaged, bad-port, not-ready, …).

• save / remove take a lock (devboard.config.json.lock; they wait up to 10 s, then report config-busy, or stale-lock if the lock was left by a process that has exited). They copy the config to devboard.config.json.bak, write atomically, and run check. If the global section or the touched project fails validation, they restore the original file (reason: "check-failed"). Errors that other projects already had don't block the save. On success they return a revision; to undo, use revert <revision> rather than copying .bak back. They don't restart the service, so run up afterwards.

• macOS: the LaunchAgent label is com.devboard.agent (with a hash suffix if another directory already uses it). The plist goes to ~/Library/LaunchAgents/, and the service runs on the Node binary that ran up. It starts at login and launchd restarts it if it exits. Logs: ~/Library/Logs/devboard.log and devboard.error.log. To manage it by hand:

  launchctl print gui/$(id -u)/<label>        # status
  launchctl kickstart -k gui/$(id -u)/<label> # restart after changing the backend
  launchctl bootout gui/$(id -u)/<label>      # stop; delete ~/Library/LaunchAgents/<label>.plist to disable autostart

  Logs are not rotated automatically. To rotate them with macOS's newsyslog (each log rotates at 10 MiB and 3 archives are kept), run this from the YUIBoard directory:

  sed "s|/Users/YOU|$HOME|" ops/devboard.newsyslog.conf > cache/devboard.newsyslog.conf
  launchctl bootout gui/$(id -u)/<label> &&
    /usr/sbin/newsyslog -r -f "$PWD/cache/devboard.newsyslog.conf"
  launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/<label>.plist   # run this even if rotation failed

  Don't just rename a log while the service is running, because the process keeps writing to the old file handle.

• Linux: the unit is devboard.service (with a hash suffix if taken). It starts at login and restarts 10 s after it exits, and it appends logs to cache/devboard.log. To run it without logging in, use loginctl enable-linger. Where no systemd user instance is available (containers, some WSL setups), up falls back to a background process and says so in warnings.

  systemctl --user status devboard.service
  systemctl --user restart devboard.service
  systemctl --user disable --now devboard.service   # stop and disable autostart

• The service runs the code currently in this directory. After you switch branches or change src/, restart it. If you change only web/, a hard refresh in the browser is enough.

• To update to a new release: run git pull in this directory, restart the service with node ops/devboard-ctl.mjs up (or launchctl kickstart -k / systemctl --user restart as above), then, if you use the Claude Code pane, run claude plugin update yuiboard@yuiboard so the plugin matches the service.

Configuration

Copy devboard.config.example.json to devboard.config.json. Relative paths are resolved against the config file's directory, and ~ expands to your home directory.

{
  "server": { "host": "127.0.0.1", "port": 4317 },
  "defaults": { "hoursPerDay": 8, "defaultDurationDays": 1, "sessionRunningMinutes": 10, "sessionStaleDays": 3,
                "githubRefreshMinutes": 5, "recentClosedPrDays": 14, "recentChangesDays": 7 },
  "projects": [{
    "id": "my-app", "name": "My App",
    "repo": { "path": "~/Projects/my-app", "github": "acme/my-app" },
    "docSources": [
      { "type": "spec-dir", "label": "Claude Code spec", "path": ".claude/specs" },
      { "type": "adr-dir", "label": "ADR", "path": "docs/adr" },
      { "type": "task-list", "label": "Task list", "files": [] }
    ],
    "archived": [], "excluded": [],
    "branchLines": [
      { "id": "main", "name": "Main", "base": "main", "docFilter": "^(?!v2-)" },
      { "id": "v2", "name": "v2", "base": "release/v2",
        "titlePrefix": "[V2]", "labels": ["v2"], "docFilter": "^v2-" }
    ],
    "github": {},
    "sessions": [{ "source": "claude-code" }, { "source": "codex" }]
  }]
}

Docs, sessions and manual links refresh every 5 seconds; GitHub every githubRefreshMinutes. Pages update live over SSE.

Read-only guarantees

• Repositories: YUIBoard runs only allow-listed read-only git commands (rev-parse, merge-base --is-ancestor, for-each-ref, log, worktree list) with GIT_OPTIONAL_LOCKS=0, ignoring any inherited GIT_* variables such as GIT_DIR. It never fetches or checks out, never writes .git or the index, never competes with you or your AI for git locks, and never changes any file in the repository.
• GitHub: YUIBoard only reads, through gh. Your token stays with gh and never appears in the config, caches, logs or pages. If the token has more than read access, the page tells you.
• AI session logs: files are opened read-only and read up to the last complete line, and nothing is created in session directories. Caches and pages keep only metadata and short excerpts, with anything that looks like a key or token masked. Full conversations are never stored or uploaded.
• The server itself writes only to cache/ and data/ in its own directory. Pages accept only same-origin requests, and the write endpoints (manual links, refresh now) check Origin and Content-Type. There is no telemetry.
• test/readonly.test.mjs runs a full refresh against a temporary repository and session directory. It checks that git status, .git and the session files (including their mtimes) are unchanged afterwards.

Platforms

• Tested: macOS 15 with Node 24. The full test suite also runs in CI on Ubuntu and macOS.
• Linux: the systemd user service is tested only against a stand-in, not yet on a real machine. Directories whose file names are NFD-normalized may fail to open from "View source".
• Windows: handled by design but untested. Known risks are the encoding of Claude Code session directory names, finding gh.exe (set github.ghPath if needed), shared reads of session files that are being written, and rename failing with EPERM while antivirus or indexing holds the cache file.
• Everywhere: paths are normalized to NFC and shown with /, CRLF and LF are both accepted, and git / gh are called through execFile, never a shell.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for the dev setup, tests and the hard constraints (zero dependencies, read-only).

License

Apache License 2.0. Copyright 2026 JimmyFu.