菜单
Fleet operations — use when a task involves other machines (discovering, reaching, or reasoning about them), installing/updating/removing any user-level tool, converging a machine, or working with paseo or the agent-fleet repo.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | fleet |
| description | Fleet operations — use when a task involves other machines (discovering, reaching, or reasoning about them), installing/updating/removing any user-level tool, converging a machine, or working with paseo or the agent-fleet repo. |
The fleet is defined by the agent-fleet repo (~/agent-fleet on every
managed machine). One exact-pinned pixi-global manifest (base + per-machine
overlay) declares each machine's tools; a machine converges to it — the
manifest is the complete state, and converging installs what's missing AND
deletes what's unlisted.
| Question | Source of truth | Command |
|---|---|---|
| What exists / is online? | Tailscale (self-updating) | tailscale status |
| Specs, GPUs, roles? | Rackpeek (curated, can lag) | ssh pablo-dl-server docker exec rackpeek rpk summary (also rpk servers describe <name>, rpk systems list) |
| Fleet-managed? | agent-fleet repo | overlay at manifests/machines/<lowercased short hostname>.toml (the LOCAL hostname, which may differ from the Tailscale name — e.g. the M5 laptop is pablos-m5-pro on the tailnet but macbook-pro-3.toml here, and the DGX is dgx-spark vs spark-9232.toml); ~/agent-fleet checked out |
Stable facts: pablo-dl-server is the conductor (always-on, hosts rackpeek).
The DGX Spark (dgx-spark on the tailnet, local hostname spark-9232)
is linux-aarch64 — account for it in platform availability.
Macs are osx-arm64.
SSH rides the tailnet: any hostname from tailscale status is an SSH target
(ssh pablo-ubuntu, ssh dgx-spark; ~/.ssh/config may hold shorter
aliases). Patterns that matter:
ssh -o ConnectTimeout=5 -o BatchMode=yes <host> '…'
— BatchMode fails fast instead of hanging on a password prompt.export PATH="$HOME/.pixi/bin:$PATH".pkill -f <pattern> over SSH can match and kill your own SSH session's
command line — prefer pgrep -a first, then kill exact PIDs.Moving files — scripted transfers always use SSH; Taildrop is for humans:
scp file <host>:path.COPYFILE_DISABLE=1 tar czf - . | ssh <host> 'tar xzf - -C <dest>'
(COPYFILE_DISABLE strips macOS xattr noise).tailscale file cp) — its
niche is devices without SSH. No destination control: macOS receives to
~/Downloads; headless Linux needs a second step (tailscale file get).
Alpha; requires the tailnet "Send Files" setting.Tailscale itself is OS-layer (never pixi-managed); the tailnet is also the auth for Tailscale SSH on some machines (e.g. spark).
manifests/base.toml (fleet-wide) or manifests/machines/<host>.toml
(one machine). Exact == pins; base packages must solve on linux-64,
linux-aarch64, AND osx-arm64. Channels: conda-forge → own channels
(prefix.dev/ai-demos, prefix.dev/my-skill-forge) →
prefix.dev/github-releases as gap-filler.pixi run check, commit, push.~/agent-fleet/scripts/sync.sh.Done when every target machine prints ✓ <host> converged to <rev> at the
pushed revision — not before.
JS CLIs get rattler-build recipes in the ai-demos repo (npm-tarball pattern,
see recipes/ruler there) and arrive via the channel.
Skills are agent-skill-<name> conda packages from
https://prefix.dev/my-skill-forge. Three steps, each explicit:
noarch: generic,
agentskills validate in tests — copy an existing recipe). Merging
publishes to the channel.agent-skill-<name> = "==<version>" under
[envs.agent-skill-forge.dependencies] in agent-fleet's
manifests/base.toml, push.link_skills in sync.sh symlinks it into
~/.claude/skills and ~/.codex/skills (work profile shares via
~/.claude-work/skills).Publishing alone does NOT propagate: the pin is the membership act (weekly
pin-bump only bumps versions of skills already pinned — it never adds new
ones). Done when the skill dir appears under ~/.claude/skills/ on every
converged machine.
ONE declared MCP set in git lands on claude (personal), claude-work, AND
codex on every machine: [mcp_servers.<name>] blocks in agent-fleet's
.ruler/ruler.toml → pixi run render (committed .mcp.json) → converge
(scripts/mcp_apply.py applies at user scope through each CLI's own
mcp add/remove — it never writes their state files directly).
claude mcp add / codex mcp add for a fleet server —
hand edits are drift and revert on the next converge. Hand-added servers
under OTHER names are ignored (the applier only touches declared names).mcp: skip <name> … note and apply automatically once the
pin catches up (e.g. rerun viewer-mcp needs >= 0.34).tmux = "3.3.*"): the range IS the hold — pin-bump ignores it.# pin-bump: hold on an == pin line: frozen exact pin.Mon 05:00 UTC: ai-demos autobump PRs recipe bumps (auto-merges on green CI, publishes). Mon 06:00: agent-fleet pin-bump PRs base.toml updates — human-gated; merging converges the fleet on next sync. Rollback = revert the pin PR.
Fleet-managed on every machine (2026-07-06): CLI daemon from the pixi
paseo pin, run by a service unit — systemctl --user status|restart paseo
on Linux (linger required for boot start), launchctl print|kickstart -k gui/$UID/sh.paseo.daemon on macOS (starts at LOGIN). NEVER paseo daemon start by hand — the unit owns the daemon. paseo daemon status for health.
Config: ONE file in git, agent-fleet/config/paseo/config.json — edit,
push, converge (scripts/paseo_apply.py seds in the machine's tailnet IP
and restarts only on fleet-initiated change via the .fleet-config-applied
shadow copy; UI edits are drift that reverts on the next fleet change).
State lives in ~/.paseo (machine-local, survives binary swaps). There is
NO daemon password — auth is the daemon keypair + relay pairing. One web UI:
https://pablo-dl-server.ilish-ruler.ts.net:8767 (PWA); pair a daemon into
it with paseo daemon pair → Add connection.
Read references/internals.md — sync.sh's
pipeline and safety gates, capture/overlay onboarding, and known failure
signatures (dead OAuth masquerading as "model unavailable", stale stamps,
app configs broken by version bumps).
Model routing for delegated work — use when spawning subagents or workflows, choosing a model for a task, or handing work to the Codex runtime. (Claude Code side; other agents rarely delegate.)
Prove what the Rerun viewer rendered — pixel evidence over logs. Use when a .rrd, blueprint, or Rerun rendering must be visually verified, when a timeline sweep or video of a recording is wanted, when an .rrd must be embedded in an HTML page, or when a gradio/WebViewer surface needs browser validation.
Python conventions for every repo — typing, dataclasses, return types, docstrings. Use whenever touching Python code.
Create self-contained single-file HTML artifacts when an explanation, plan, report, review, comparison, diagram, deck, prototype, or lightweight editor would be more readable, useful, or reusable in a browser than in markdown. Use for making dense information scannable and beautiful with semantic HTML, inline CSS, inline JavaScript, responsive layout, typography, charts, tables, annotations, tabs, collapsibles, timelines, diagrams, and copy/export affordances. Especially use when the user wants an artifact they are likely to open, read, share, or iterate on directly as a .html file.
Create a new single-purpose CV node in the monorepo with API layer (class-based node with verbose Rerun logging, result logging helper, config + result dataclasses), Gradio UI with embedded Rerun viewer (streaming binary stream), CLI entry point (tyro + RerunTyroConfig), and bundled example data. Use when adding a new model, predictor, or pipeline step as a standalone reusable app.
Performs conda-forge operations. Fixes failing builds by analyzing CI logs, creates new packages via staged-recipes, adds cross-compilation and ARM support to feedstocks, and migrates recipes from v0 to v1 format. Use when working with conda-forge feedstocks, staged-recipes, build failures, recipe migrations, or when the user mentions conda-forge.