ops-fleet
GitHub提供 Claude 会话集群的只读监控面板，实时展示本地及远程会话状态、CRS 中继健康度、账户利用率及 Token 类型分布，支持 TUI 和快照模式。
Trigger Scenarios
Install
npx skills add Lifecycle-Innovations-Limited/claude-ops --skill ops-fleet -g -y
SKILL.md
Frontmatter
{
"name": "ops-fleet",
"effort": "low",
"maxTurns": 6,
"description": "Read-only fleet dashboard — every Claude session (local + remote) with token type (oauth\/api\/bedrock\/crs), CRS relay state, account utilization, and per-session detail. No actions.",
"allowed-tools": [
"Bash",
"Read"
],
"argument-hint": "[--once] [--tui] [--all] [--no-color] [--no-ec2]"
}
OPS ► FLEET
Unified, read-only dashboard over the whole Claude session fleet — local claude --bg
sessions plus any remote box reached over SSM — built from claude agents --json,
claude daemon status, the account-rotation daemon state, and the CRS (Claude Relay
Service) relay. It renders, it never mutates.
Runtime Context
The dashboard sources everything live; nothing needs to be parsed by this skill:
- Sessions:
claude agents --json(authoritative; nameless sessions fall back to their session id). - CRS relay: health at
http://127.0.0.1:${CRS_PORT:-3005}/health, allowlist at~/.claude/scripts/account-rotation/crs-allowlist.json, recent 429/529 from the relay window. - Account pool: the rotation daemon's keychain/util state (per-account util%, reset countdown).
- Remote (EC2/FRA) sessions: a background-refreshed SSM cache (
~/.claude/state/fleet-ec2.json, TTL ~45s); degrades torefreshing…/ empty when unavailable.
These paths are the standard account-rotator install locations
(scripts/install-account-rotator-linux.sh). If the rotator isn't installed the CRS /
account / remote rows simply degrade to ? — the session table still renders.
Invocation
Run the bin script and relay its output verbatim (strip ANSI for the chat block):
${CLAUDE_PLUGIN_ROOT}/bin/ops-fleet $ARGUMENTS
Behavior:
- In a real terminal it defaults to a live full-screen TUI (alt-screen, adaptive width,
qto quit). - With stdout captured (slash command / cron) it auto-falls back to a one-shot snapshot.
- Flags:
--once/--snapshotforce a single render ·--tui/--watch [secs]force the loop ·--allinclude completed sessions ·--no-colorplain ·--no-ec2skip the remote SSM round-trip. - Override the CRS port with
CRS_PORT=NNNN(default 3005; 3000 is the legacy alias).
Output
ALWAYS render the dashboard to the user first — re-emit the snapshot inside a fenced code
block (keep the box-drawing/bars/layout, strip ANSI). Below it add a one-line summary:
total sessions + token-type breakdown, anything blocked or flapping (high att:), CRS
health, and any account ≥90% util. The dashboard is the deliverable; the summary is the footnote.
End with a single line noting the live full-screen TUI is ${CLAUDE_PLUGIN_ROOT}/bin/ops-fleet
(or --tui) in a real terminal.
Version History
- 64bad13 Current 2026-08-12 09:00
