agent-kit

Agent kit for Claude Code and Codex CLI: one statusline for both, Codex patches, an agent broker and two-way session sync.

• Statusline for Claude Code and Codex (Codex: native footer or a log panel): context, quotas, git across umbrella repositories, runtime activity, Codex/Kimi executor widgets. Dependency-free Node, TUI editor, English and Russian.
• Codex patches: a native footer driven by the same renderer, full diffs, command timing, capacity retries, Shift+Tab Plan/Auto, restricted subagents via a readonly-/plan- task name prefix. You build Codex yourself from the patched sources; no binaries are shipped.
• Faster Codex voice: a realtime prompt for quicker speech (configuration only).
• Agent broker: Codex, Kimi ACP and Claude background workers, interactive decisions and event delivery to orchestrators; finished Codex threads are archived.
• Session sync (experimental): a Claude Code session and a Codex session in the same project mirror each other both ways, with native history, titles and tool calls; continue in Codex when Claude tokens run out and back.

Русская версия: README.RU.md

Fable 5.1 high   ctx ▰▰▱▱▱ 42% 84k/200k   ⚙  Edit render.js   📋 3/7   agents 2
5h ▰▰▰▱▱ 63% ↻ 1h00m   7d ▰▰▱▱▱ 41% ↻ 3d0h   Fable 18%
#128 ✓ approved
▸ api  ⎇ feat/login 1 to push +120 −14 ?1
  web  ⎇ main       2 to pull +8 −2
  Σ +128 −16

Prerequisites

• Claude Code statusline: Node.js 20+. Nothing else.
• Codex patches (Windows x64 for now, macOS/Linux TODO):
  • Git;
  • rustup (winget install Rustlang.Rustup); the Rust version comes from Codex's rust-toolchain.toml;
  • MSVC C++ Build Tools: winget install Microsoft.VisualStudio.2022.BuildTools --override "--wait --passive --add Microsoft.VisualStudio.Workload.VCTools --includeRecommended";
  • the official standalone Codex matching manifest installed (its package is copied as the base);
  • a short CARGO_HOME (a manual junction works) or Windows LongPathsEnabled; the build prints remediation commands if the rusty_v8 checkout would exceed MAX_PATH;
  • tens of GB of disk and 20+ minutes for the first release build.

The installer only checks these and prints the commands; it never installs tools for you.

Quick start

git clone https://github.com/FrostBy/agent-kit && cd agent-kit
node install.js          # or install.cmd / sh install.sh

Pick items with arrows, Space and Enter: Claude statusline, each Codex patch separately, faster voice. Dependencies are added automatically. Non-interactive:

node install.js --list
node install.js --only claude
node install.js --only statusline,full-diff,interaction-modes,voice
node install.js --yes --skip voice

Settings and config.toml are backed up before changes; CLAUDE_CONFIG_DIR and CODEX_HOME are respected. The voice template is based on Codex's realtime backend prompt (Apache-2.0).

Full diff, capacity retries and faster voice are based on booxter's patches, see Credits.

Codex

Native footer (patched CLI)

Build on Windows x64 (see Prerequisites); macOS/Linux stop with an explicit error for now. Only source patches are shipped.

npm run codex:build -- --dev     # omit --dev for release; --features a,b selects patches, default all
npm run codex:install            # activates <version>-agent-kit next to the official release
npm run codex:sync-footer        # after JS changes: update the packaged renderer
npm run codex:restore            # back to the official release

Install copies the official standalone package into <version>-agent-kit, keeps rollback state and never edits config or auth. To upgrade Codex: codex:restore, codex update to the manifest version (from PowerShell or cmd, not Git Bash), build, codex:install.

The footer uses the Claude layout with data from native CLI JSON and ends with an auto/plan/read-only mode line; maxLines in --codex --configure limits its content rows. node scripts/native-codex.js [project] runs a dev build in an isolated home with a loopback provider, without touching your Codex.

Details in patches/README.md: what each patch adds, build/install internals, paths and limitations.

Standalone log inspector (legacy mode)

A separate panel for a terminal pane, reading local Codex session logs (no network, no Claude credentials):

node statusline.js --codex --watch      # q / Ctrl+C to exit; omit --watch for one snapshot
node statusline.js --codex --preview    # sample umbrella with three repositories

--root="Z:/work" sets an umbrella directory, --session=UUID or --session="C:/path/session.jsonl" pins a session; otherwise CODEX_THREAD_ID or the newest session under the root is used. ctx is measured against the session's effective model_context_window; quotas show usage consumed. The log format is internal and may change. Configure with node statusline.js --codex --configure (~/.codex/statusline.json).

Agent broker and session sync (experimental)

Broker. Run Codex (shared app-server), Kimi (kimi acp) or Claude (persistent stream-json) as background workers. Questions, permissions, cancellation and session resurrection use the same broker commands. Codex and server-owned Kimi web orchestrators can receive events directly; Claude supports the event waiter and optional wake hook. Kimi/Claude queue busy messages for the next turn (steer=false). Point --state at the scratchpad the statusline reads. send/wait print and ACK the result; send revives an exited worker from its saved session. See the short broker README for commands, executable overrides, the legacy --kimi-print flag and isolated live tests.

node scripts/agents/broker.js --agent codex --name reviewer --state /tmp/scratchpad --prompt /tmp/prompt.md --cwd /path/to/project

Brokers exit after 15 idle minutes (--idle-min overrides); send revives a dead broker using saved settings and its Codex/Kimi session, without replaying the initial prompt. Mesh hooks ignore normal/idle exits; the synchronous SessionEnd hook closes session brokers, and a detected owner PID also binds their lifetime.

Interactive mesh. Use broker.js answer|approve|deny --name X --state DIR --request ID for pending requests. interrupt/exit work during a turn. Codex's --approval-policy defaults to never; Claude never bypasses permissions. The optional Claude mesh-hook.js uses asyncRewake on SessionStart/Stop; install it explicitly with node scripts/agents/mesh-hook.js install (remove with uninstall). See setup and the live validation matrix. Autonomous Claude idle wake and native TUI rendering still need acceptance; loopback CLI protocol tests are covered separately.

Session sync is off by default, including with --yes:

node install.js --only sync
node scripts/agents/sync.js uninstall

It adds global hooks to CLAUDE_CONFIG_DIR/settings.json and CODEX_HOME/config.toml and activates per project only after both clients have been used in that directory. Each source session gets a mirror in the other client, titled <source title> · Claude/Codex: resume it there and continue. The mirror carries the latest compacted context (about 50k tokens, SESSION_SYNC_CONTEXT_TOKENS) with native history and tool calls; later turns arrive as hook context. Broker agents, subagents and noninteractive sessions are excluded. Nothing is executed and no credentials are read, but transferred text can contain private data.

Internals, limits, state layout and manual checks: docs/sync.md.

Install

Install it globally from GitHub, then point Claude Code at it:

npm install -g github:FrostBy/agent-kit
agent-kit --install

--install writes statusLine and subagentStatusLine into ~/.claude/settings.json as node "<install path>/statusline.js", so each render runs the installed copy and fetches nothing. The previous values are saved and --uninstall puts them back. The settings file is left untouched if it does not parse.

To update, run the npm install -g line again.

Configure

agent-kit --configure

The editor shows a live preview on sample data. Keys: ↑↓ select, a add a widget, d delete, K/J move, | separator, Enter widget options, n new line, g global settings, s save, q quit. In mintty or Git Bash run it through winpty.

The Claude config lives in ~/.claude/statusline.json (override with AGENT_STATUSLINE_CONFIG or CLAUDE_STATUSLINE_CONFIG). Missing values come from the defaults, and --init writes them out. The package name and command are agent-kit; agent-statusline and claude-statusline-ultimate remain aliases. Codex's internal codex-resources/agent-statusline directory and agent-statusline-state.json filename remain unchanged for compatibility.

{
  "lines": [
    ["model", "effort", "fast", "|", "context", "|", "activity", "|", "todos", "|", "subagents"],
    ["limit-5h", "|", "limit-7d", "|", "limit-model"],
    ["cost", "|", "cost-today", "|", "cache", "|", "pr"],
    ["git", "|", "worktree"]
  ],
  "widgets": { "context": { "bar": 8 }, "git": { "max": 6, "branchMax": 30 } },
  "colors": { "model": "#c792ea", "branch": "magenta" }
}

Key	Default	Meaning
lines	see above	Widgets per line. "|" starts a new group. An item can be an object with options: { "id": "text", "value": "🔥" }
separator	three spaces	Between groups
gapAfter	true	Empty line after the status line, above Claude Code's own mode line
money	auto	auto: money only on API billing or extra usage; always: always
lang	auto	en, ru, or auto from the system locale
icons	unicode	unicode, ascii for plain terminals, nerd for Nerd Fonts
colors		ANSI names (cyan, brightRed, gray…) or #rrggbb. NO_COLOR turns colours off
thresholds	50 / 80	Percent where values turn yellow and red
pricing		Override model prices, $ per 1M tokens: { "claude-opus-5-5": { "in": 4, "out": 20, "read": 0.2 } }
usageTtl, todayTtl	120, 300	How often the usage API and today's spend refresh, seconds

Other commands: --preview renders sample data in the current folder, --list [--md] lists every widget with its options.

Widgets

In Claude mode, add codex and kimi to lines manually. exclude lists paths separated by ;, hiding the widget in those folders and descendants (case-insensitive on Windows). idleAfter hides sessions after inactivity in minutes (default 10); maxAge remains an additional age limit. Kimi logs above 16 MiB are hidden. Reads are local, without credentials. The widgets first use agent-broker state files in input.scratchpad_dir, matching Codex by threadId and Kimi by sessionId: up to three active agents, otherwise the most recently finished agent if still fresh. Freshness and turn status come from the session log; done/error state files are ignored. Agent names appear in gray, with Codex account quotas once at the end. Without a matching fresh broker agent, selection falls back to cwd: only codex exec sessions for Codex, and local Kimi sessions for kimi -p. Kimi shows limit 5h for 5 hours after a 5-hour quota error; limit month (monthly) and limit (unknown window) remain for 30 days after the event.

Model & session

id	description	options
codex	External Codex: model, context, tokens, quotas	exclude="", maxAge=120, idleAfter=10, tokens=false
kimi	External Kimi: model, context, tokens, limit	exclude="", maxAge=120, idleAfter=10, tokens=false
model	Model
effort	Effort level
thinking	Extended thinking enabled
fast	Fast mode
plan	Plan: Pro / Max 5x / Max 20x / API
session-name	Session name	max=24
duration	Session duration
api-time	Share of time waiting for the API
version	Harness version
style	Output style, if not default
vim	Vim mode
agent	Agent name (--agent)
permission-mode	Permission mode, if not default (from transcript)

Context

id	description	options
context	Context: bar, %, used/window; accounts for Claude's autoCompactWindow	label="ctx", bar=5, percent=true, used=true, left=false
context-breakdown	Context by type: fresh input / cache read / cache write

Limits

id	description	options
limits	Codex quota windows, with their actual durations	label="", bar=5, countdown=true, pace=true, paceMargin=10
limit-5h	5-hour limit	label="5h", bar=5, countdown=true, pace=true, paceMargin=10
limit-7d	Weekly limit	label="7d", bar=5, countdown=true, pace=true, paceMargin=10
limit-model	Per-model weekly limits (Fable etc.), usage API	countdown=false, pace=true, paceMargin=10
limit-spend	Gateway spend limit
extra-usage	Purchased credits (extra usage), usage API
week-breakdown	Where the week went: Claude Code / chats / Cowork, usage API	min=1

Money & edits

id	description	options
cost	Session cost
burn-rate	$ per hour, current session
cost-today	Today's spend across all sessions (from transcripts)	sessions=false
session-lines	Lines changed by Claude this session

Prompt cache

id	description	options
cache	Prompt cache: warm/cold, when it expires	under=5
cache-hit	Cache hit ratio and misses

Git

id	description	options
git	Git: branch, commits to push/pull, +/− lines, untracked; in an umbrella folder a row per changed repo and Σ	depth=2, max=4, showClean=false, branchMax=24, rows=true, aheadBehind=true, lines=true, untracked=true, stash=false, budgetMs=400
worktree	Worktree name
repo	Repository owner/name, clickable
pr	PR/MR: number and review state, clickable	state=true

Folders

id	description	options
dir	Current folder	full=false
project-dir	Launch folder, if different from current
added-dirs	Folders added via /add-dir

Live activity

id	description	options
activity	Tool running right now	detail=true
todos	Todo / task progress
subagents	Running subagents

Misc

id	description	options
clock	Clock	seconds=false
text	Custom text	value="", color="muted"
cmd	Output of your own command (cached)	command="", ttl=30, color="muted"

How it works

• Claude Code pipes session JSON to the command and shows what it prints. Only the sources the current layout needs are loaded.
• Git runs in parallel across repos with a 400 ms budget. Past it, the last known state is shown and the scan finishes in a detached process, so the render never waits.
• The usage API and today's spend refresh in detached processes on their TTL.
• Today's spend is priced per model from the usage of each response in today's transcripts, subagents included, with duplicate content blocks counted once. Files are read incrementally, gigabyte-sized ones in chunks.
• Git is read from the folder the session was launched in, so a cd into one repo does not hide the others.
• The cache lives in agent-statusline under the system temp folder.

Windows Terminal

Cascadia Mono, the default font, lacks some symbols. Windows draws them from a fallback font, wider than a cell, and they overlap the next character. The default icons were checked by eye in Windows Terminal: missing glyphs are emoji or get an extra space. Claude Code also trims leading spaces on every line, so the indent of the git rows starts with U+2800, a blank braille cell.

Privacy

Per-model limits, extra usage and money on a subscription need the account usage. For that the status line reads the Claude Code OAuth token from ~/.claude/.credentials.json and calls api.anthropic.com/api/oauth/usage, the endpoint behind /usage. It is undocumented and may change. The token goes nowhere else. Leave limit-model, extra-usage, week-breakdown and the money widgets (cost, burn-rate, cost-today) out of the layout, and no request is made.

Development

npm test
node statusline.js --preview

Credits

• booxter: the original Codex patches for full diffs in compact mode, capacity retries and the stream retry limit, and the faster-voice idea (booxter/nix). The target Codex version is defined in the manifest.
• GordonBeeming/codex-statusline (Apache-2.0): the base of the native footer patch.
• OpenAI Codex (Apache-2.0): patched sources and the voice prompt template. See patches/NOTICE.

Licence

MIT © Vladislav Romanovsky