omninode_dev_setup
Announce at start: "I'm using the omninode_dev_setup skill."
One script, plugins/onex/skills/_bin/lab-onboarding.sh, does the work. It runs
the same way from Claude Code or from a plain terminal, and works on a physical
Mac or inside a macOS VM. This skill starts it where the developer can answer
its prompts, then reports each phase to the developer the moment it finishes.
Everything runs on the developer's own Mac: native onex, and optionally the stack
in Docker, on the developer's own models: any combination of their Gemini, OpenRouter
and OpenAI keys and Ollama running on the Mac. Delegation chooses among them for
each task; the developer never picks one per run. Nothing connects to the lab.
Phases
| # |
Phase |
Changes the machine? |
| 0 |
Preflight: macOS version, CPU, RAM, disk, VM or physical, admin rights, ports; what an earlier setup left (the onex the developer's shell runs and its version, published onex installs, old variables and profile lines, a local-model routes file, the workspace clones' state); then every question the run has (set those aside or stop, models and keys, Docker and its terms, the Mac password), before anything installs |
only what the developer agreed to set aside |
| 1 |
Base tools: Xcode command-line tools, Homebrew, gh, jq, python@3.13, uv |
yes |
| 2 |
Workspace: the canonical clones, and OMNIBASE_PATH and PATH in the shell profile |
yes |
| 3 |
onex, the local identity, then each chosen model set up with onex models add (Ollama installed with one model downloaded) and tested with one delegation pinned to it; one result line per model; fails only if none passed |
yes |
| 4 |
Docker (optional): only if the developer says yes to the one question; Docker Desktop installed, or started if stopped, then the local stack on the same keys |
yes |
| 5 |
Claude Code plugin: the customer plugin onex@omninode-tools (one skill, /onex:delegate), from the public marketplace. With --developer: the full onex tree (onex@omninode-tools-dev, from the omniclaude clone), and omni and onex-overlays when this GitHub login can read the private plugin repository and an SSH key is loaded |
yes |
| 6 |
Verify |
no |
Minimum requirements are printed by phase 0. A Mac below them gets nothing
installed and a recommended VM size instead (exit code 3).
Execution
1. Preflight first, in this session
Always run preflight before anything else and show its table to the developer:
bash "${CLAUDE_PLUGIN_ROOT}/skills/_bin/lab-onboarding.sh" --preflight-only <args>
- Exit 3: the Mac is below the minimum. Relay the failed requirement and the
recommended VM verbatim. Stop. Do not offer to install anyway.
- Exit 1: a fixable condition (not an admin account, a Rosetta shell, not macOS).
Relay the
Next: line and stop.
- Exit 0: relay the "Will set up" line. Native onex is set up on every run.
If the line says Docker can be added, ask
the developer the one question the run would ask: whether to also run the
stack locally in Docker (about 10 GB of memory while it runs, and 10-20
minutes the first time). Pass their answer as
--containers or
--no-containers, so the Terminal run does not ask it again. If Docker
Desktop is installed but stopped, say the run will start it; if it is
missing, say the run will install it. If the line says Docker is not
offered, relay why: native onex alone covers delegations.
- The models (the run asks them before the Docker question). Unless
--provider
was given, ask which ones; any combination, at least one:
- Gemini (a Google AI Studio key), OpenRouter or OpenAI (their key;
OpenAI needs credits on the account). Each key is typed only in the Terminal
window, at a hidden prompt naming its provider, right after preflight and
before anything installs. If they lack a key they chose, they press Enter at
its prompt to skip it, and can add it later with
onex models add <provider>.
An OpenRouter key runs the default model the onex overlay declares
(onex models default-model openrouter reads it), with no model question; it
may be a paid model, so the key may need credit on openrouter.ai. With no
overlay reachable, OpenRouter ends with that reason and nothing stored while
the other models carry on. Another model is one command later:
onex models add openrouter --model <model id>.
- Ollama: no key. It runs a model on the Mac, so it downloads one
(sized to the Mac's memory, as omnimarket's model config declares) and is slower,
especially on Intel or in a VM.
Pass the answer as
--provider with a comma-separated list, e.g.
--provider gemini,ollama.
If --preflight-only was the argument, stop here.
2. Start the run where the developer can answer prompts
Every question is asked in that terminal, all of them before anything installs:
the models and their keys, Docker and its licence terms, and the Mac administrator
password when something needs installing. After that the run asks nothing. A
Claude Code tool call has no terminal, so open one:
/usr/bin/osascript -e "tell application \"Terminal\" to do script \"bash '${CLAUDE_PLUGIN_ROOT}/skills/_bin/lab-onboarding.sh' <args>\"" -e 'tell application "Terminal" to activate'
Tell the developer a Terminal window opened, that it asks its questions first
(models, keys, Docker, their Mac password) and then runs on its own, and that
they should answer there, never in this chat. Never ask the developer to paste a password or API key into this
session.
If there is no desktop session (for example over ssh), run the script directly
with the Bash tool in the background instead. With no terminal it asks nothing:
it uses the models --provider names (or every key an earlier run stored), stops
before installing if a chosen model has no stored key, and fails a step that
needs the administrator password with a Next: line saying so.
3. Report each phase as it finishes
The script appends one line per phase event to
${TMPDIR}/omninode-onboarding/status:
phase=2 name="Workspace (…)" result=PASS elapsed="1m 12s" note="…"
model=gemini result=PASS detail="gemini-3.5-flash-lite"
model=openai result=FAIL detail="provider_error: insufficient_quota …"
model=openrouter result=SKIPPED detail=""
phase=3 name="onex, local identity and your models" result=PASS elapsed="2m 03s" note="1 of 3 models ready: Gemini"
phase=4 name="Docker (…)" result=FAIL elapsed="…" step="…" next="…"
result=COMPLETE
Watch that file (the Monitor tool with an until-loop on a new line, or re-read
it every 30–60 seconds). For every new PASS, SKIPPED or FAIL line, tell
the developer at once, in one sentence: the phase, the result, the elapsed
time, and for a FAIL the step and next fields verbatim. Do not wait for
the end to report. Report the model= lines together when phase 3 passes: one
line per model, pass, failed with its detail, or skipped, and for each one
that failed or was skipped, that onex models add <provider> adds it later.
Stop watching at result=COMPLETE, a FAIL, or
result=BELOW_MINIMUM.
A FAIL means the run stopped. Re-running the same command resumes from the
failed phase; completed phases are verified, not redone.
4. Finish
On result=COMPLETE, tell the developer to open a new terminal (so OMNIBASE_PATH
and PATH apply) and to open Claude Code and sign in with their Anthropic account
(the first time it asks). In a new session the /onex: skills
(and /omni: ones, if phase 5 installed them) are loaded. A running session keeps
the plugins it started with. If phase 0 set anything aside, tell the developer
the file that lists each item and how to put it back.
What this skill does NOT do
- Put any secret in this session. Passwords and keys go into the Terminal
window, and from there over stdin into the tool that stores them.
- Show a macOS dialog. Every question is in the terminal; the desktop only gets
a notification when a phase ends.
- Install anything on a Mac below the minimum requirements.
- Connect to the lab in any way: no tailnet, no lab bus identity, no lab models.
- Linux or Windows.
1---2name: omninode-dev-setup3description: Take a Mac from bare to a working local onex in one run — tools, workspace, onex on the developer's models (any combination of their own Gemini, OpenRouter and OpenAI keys and Ollama on the Mac, at least one), and optionally the local container stack — reporting each phase as it finishes. Never connects to the lab4mode: full5version: 1.0.06level: basic7debug: false8category: onboarding9tags:10 - onboarding11 - setup12 - macos13 - local14author: OmniClaude Team15skill_kind: methodology16args:17 - name: --preflight-only18 description: "Check this Mac against the minimum requirements and stop; changes nothing"19 required: false20 - name: --containers21 description: "Answer the Docker question yes in advance (also run the stack locally in Docker)"22 required: false23 - name: --no-containers24 description: "Answer the Docker question no in advance (native onex only)"25 required: false26 - name: --provider27 description: "The models, comma-separated: any of gemini, openrouter, openai (the developer's own keys) and ollama (on this Mac, no key). Default: asked"28 required: false29 - name: --ollama-model30 description: "With --provider ollama: the model to download (default: chosen from the Mac's memory)"31 required: false32 - name: --workspace33 description: "Workspace directory (default: $OMNIBASE_PATH, else ~/Code/omnibase; OMNI_HOME is never read). A directory holding clones this setup did not create is refused"34 required: false35 - name: --set-aside36 description: "Answer yes in advance to setting aside what an earlier setup left (an onex alias or function, old variables in the shell profile, a published onex, a local-model routes file, clones holding work): moved or commented out, never deleted. Without a terminal, a run that finds any stops unless this is passed"37 required: false38 - name: --restart39 description: "Run every phase again instead of resuming"40 required: false41metadata:42 source: https://skillmd.com/skills/omninode-ai/omninode-dev-setup43---4445# omninode_dev_setup4647**Announce at start:** "I'm using the omninode_dev_setup skill."4849One script, `plugins/onex/skills/_bin/lab-onboarding.sh`, does the work. It runs50the same way from Claude Code or from a plain terminal, and works on a physical51Mac or inside a macOS VM. This skill starts it where the developer can answer52its prompts, then reports each phase to the developer the moment it finishes.5354Everything runs on the developer's own Mac: native onex, and optionally the stack55in Docker, on the developer's own models: any combination of their Gemini, OpenRouter56and OpenAI keys and Ollama running on the Mac. Delegation chooses among them for57each task; the developer never picks one per run. Nothing connects to the lab.5859## Phases6061| # | Phase | Changes the machine? |62|---|---|---|63| 0 | Preflight: macOS version, CPU, RAM, disk, VM or physical, admin rights, ports; what an earlier setup left (the onex the developer's shell runs and its version, published onex installs, old variables and profile lines, a local-model routes file, the workspace clones' state); then every question the run has (set those aside or stop, models and keys, Docker and its terms, the Mac password), before anything installs | only what the developer agreed to set aside |64| 1 | Base tools: Xcode command-line tools, Homebrew, gh, jq, python@3.13, uv | yes |65| 2 | Workspace: the canonical clones, and OMNIBASE_PATH and PATH in the shell profile | yes |66| 3 | onex, the local identity, then each chosen model set up with `onex models add` (Ollama installed with one model downloaded) and tested with one delegation pinned to it; one result line per model; fails only if none passed | yes |67| 4 | Docker (optional): only if the developer says yes to the one question; Docker Desktop installed, or started if stopped, then the local stack on the same keys | yes |68| 5 | Claude Code plugin: the customer plugin `onex@omninode-tools` (one skill, `/onex:delegate`), from the public marketplace. With `--developer`: the full onex tree (`onex@omninode-tools-dev`, from the omniclaude clone), and `omni` and `onex-overlays` when this GitHub login can read the private plugin repository and an SSH key is loaded | yes |69| 6 | Verify | no |7071Minimum requirements are printed by phase 0. A Mac below them gets nothing72installed and a recommended VM size instead (exit code 3).7374## Execution7576### 1. Preflight first, in this session7778Always run preflight before anything else and show its table to the developer:7980```bash81bash "${CLAUDE_PLUGIN_ROOT}/skills/_bin/lab-onboarding.sh" --preflight-only <args>82```8384- Exit 3: the Mac is below the minimum. Relay the failed requirement and the85 recommended VM verbatim. **Stop.** Do not offer to install anyway.86- Exit 1: a fixable condition (not an admin account, a Rosetta shell, not macOS).87 Relay the `Next:` line and stop.88- Exit 0: relay the "Will set up" line. Native onex is set up on every run.89 If the line says Docker can be added, ask90 the developer the one question the run would ask: whether to also run the91 stack locally in Docker (about 10 GB of memory while it runs, and 10-2092 minutes the first time). Pass their answer as `--containers` or93 `--no-containers`, so the Terminal run does not ask it again. If Docker94 Desktop is installed but stopped, say the run will start it; if it is95 missing, say the run will install it. If the line says Docker is not96 offered, relay why: native onex alone covers delegations.97- The models (the run asks them before the Docker question). Unless `--provider`98 was given, ask which ones; any combination, at least one:99 - **Gemini** (a Google AI Studio key), **OpenRouter** or **OpenAI** (their key;100 OpenAI needs credits on the account). Each key is typed only in the Terminal101 window, at a hidden prompt naming its provider, right after preflight and102 before anything installs. If they lack a key they chose, they press Enter at103 its prompt to skip it, and can add it later with `onex models add <provider>`.104 An OpenRouter key runs the default model the onex overlay declares105 (`onex models default-model openrouter` reads it), with no model question; it106 may be a paid model, so the key may need credit on openrouter.ai. With no107 overlay reachable, OpenRouter ends with that reason and nothing stored while108 the other models carry on. Another model is one command later:109 `onex models add openrouter --model <model id>`.110 - **Ollama**: no key. It runs a model on the Mac, so it downloads one111 (sized to the Mac's memory, as omnimarket's model config declares) and is slower,112 especially on Intel or in a VM.113 Pass the answer as `--provider` with a comma-separated list, e.g.114 `--provider gemini,ollama`.115116If `--preflight-only` was the argument, stop here.117118### 2. Start the run where the developer can answer prompts119120Every question is asked in that terminal, all of them before anything installs:121the models and their keys, Docker and its licence terms, and the Mac administrator122password when something needs installing. After that the run asks nothing. A123Claude Code tool call has no terminal, so open one:124125```bash126/usr/bin/osascript -e "tell application \"Terminal\" to do script \"bash '${CLAUDE_PLUGIN_ROOT}/skills/_bin/lab-onboarding.sh' <args>\"" -e 'tell application "Terminal" to activate'127```128129Tell the developer a Terminal window opened, that it asks its questions first130(models, keys, Docker, their Mac password) and then runs on its own, and that131they should answer there, never in this chat. **Never ask the developer to paste a password or API key into this132session.**133134If there is no desktop session (for example over ssh), run the script directly135with the Bash tool in the background instead. With no terminal it asks nothing:136it uses the models `--provider` names (or every key an earlier run stored), stops137before installing if a chosen model has no stored key, and fails a step that138needs the administrator password with a `Next:` line saying so.139140### 3. Report each phase as it finishes141142The script appends one line per phase event to143`${TMPDIR}/omninode-onboarding/status`:144145```146phase=2 name="Workspace (…)" result=PASS elapsed="1m 12s" note="…"147model=gemini result=PASS detail="gemini-3.5-flash-lite"148model=openai result=FAIL detail="provider_error: insufficient_quota …"149model=openrouter result=SKIPPED detail=""150phase=3 name="onex, local identity and your models" result=PASS elapsed="2m 03s" note="1 of 3 models ready: Gemini"151phase=4 name="Docker (…)" result=FAIL elapsed="…" step="…" next="…"152result=COMPLETE153```154155Watch that file (the Monitor tool with an until-loop on a new line, or re-read156it every 30–60 seconds). For every new `PASS`, `SKIPPED` or `FAIL` line, tell157the developer at once, in one sentence: the phase, the result, the elapsed158time, and for a FAIL the `step` and `next` fields verbatim. Do not wait for159the end to report. Report the `model=` lines together when phase 3 passes: one160line per model, pass, failed with its `detail`, or skipped, and for each one161that failed or was skipped, that `onex models add <provider>` adds it later.162Stop watching at `result=COMPLETE`, a `FAIL`, or163`result=BELOW_MINIMUM`.164165A FAIL means the run stopped. Re-running the same command resumes from the166failed phase; completed phases are verified, not redone.167168### 4. Finish169170On `result=COMPLETE`, tell the developer to open a new terminal (so `OMNIBASE_PATH`171and `PATH` apply) and to open Claude Code and sign in with their Anthropic account172(the first time it asks). In a new session the `/onex:` skills173(and `/omni:` ones, if phase 5 installed them) are loaded. A running session keeps174the plugins it started with. If phase 0 set anything aside, tell the developer175the file that lists each item and how to put it back.176177## What this skill does NOT do178179- Put any secret in this session. Passwords and keys go into the Terminal180 window, and from there over stdin into the tool that stores them.181- Show a macOS dialog. Every question is in the terminal; the desktop only gets182 a notification when a phase ends.183- Install anything on a Mac below the minimum requirements.184- Connect to the lab in any way: no tailnet, no lab bus identity, no lab models.185- Linux or Windows.