onboarding git:20261020.7f0098bA

Immutable. This exact content never changes and is served at /api/v1/blob/7314686d3ee9861d.

---
name: onboarding
description: "Guides a first-time omo user through setup and the core workflow. Use when a new user asks to get started with omo."
---

# onboarding - the first conversation with omo

## Purpose

This skill introduces OmO through six lanes, in order: the feature tour, migration help, session archaeology, value mapping,
memory recording (which runs through the whole flow, not at the end), and the first-session
init-deep proposal. Three of the lanes are opt-in. When the user declines one, move on without
argument and without repeating the offer.

Detect the user's language from their message and respond in that language for the rest of the
conversation. The skill is written in English; your output is not. Match them exactly, including
tone.

Use Senpi-native tools only: `read`, `bash`, `edit`, `write`, the `memory` tools, the
`onboarding_session_history`, `onboarding_understand` and `onboarding_memory_apply` tools, and skill invocations.

Your context carries the `Onboarding scan` block: the engine already ran the one scan the desktop
app and `omo setup` use, with one row per item it found on this computer (sign-ins, providers, MCP
servers, skills, model choices) and a `Value summary` line. Recent sessions are not in it: they
arrive on a later turn as a short background summary, and `onboarding_session_history` has more. When you
show the user what omo found, print those rows, one per item, and quote the value summary as
computed; never recount by hand. Never assume a tool from another agent product exists here.

Be concrete, never generic. "omo caches your context" is a failure of this skill. "Your last week
of Claude Code sessions read 4.7M tokens from cache at a 78% hit rate; here is what that would
have cost cold" is the bar.

## Entry modes

- **Context on the first prompt:** answer the user's actual request fully first. Then greet them
  briefly in their language and offer the tour in one line. Wait for acceptance before starting
  the lanes below; weave useful onboarding into their task without replacing it.
- **Explicit tour:** `--onboard`, a request for the tour or getting started, or acceptance of the
  offer starts the lanes immediately. In TUI, `--onboard` starts the guided conversation at
  launch; in RPC, it applies to the first real user prompt.

## 1. Feature tour

Open by introducing yourself and giving a short tour of what omo adds on top of a plain coding
agent. The catalog below is baked in at authoring time because the user's machine has no omo or
senpi source tree to explore. Present it conversationally, three to five highlights at a time, and
let the user ask for depth on any item. Do not dump the whole list as a wall of text.

The baked catalog:

- **Eleven-agent roster**: primary workers, plan specialists, architecture consultation, codebase
  explorers, research librarians, and focused review agents are routed by the work rather than
  forced through one general-purpose persona.
- **The Senpi component layer**: startup config and migration, native status, onboarding,
  init-deep advising, anonymous telemetry, ultrawork arming, ulw-execute continuation, ulw-loop
  continuation, todo fan-out reminders, fallback architecture, comment checking, ast-grep, LSP,
  task delegation, memory, and live config watching cooperate as independent components.
- **Senpi-native skills**: `init-deep`, `ultrawork`, `ulw-loop`, `ulw-plan`, `ulw-research`,
  `hyperplan`, `coding-agent-sessions`, and `give-me-tips` provide reusable workflows that the
  agent reads and follows only when relevant.
- **Three MCP tiers**: built-in servers, user or project `.mcp.json` servers, and skill-embedded
  servers give projects a layered tool surface without forcing every integration into core.
- **Team mode**: cooperating agent sessions can share work, messages, and task state when the
  selected omo harness exposes that surface.
- **Goal and boulder state**: durable objective and work-plan state let long work resume from
  recorded progress instead of relying on conversation context alone.
- **Ultrawork and the `ulw` keyword**: don't try to memorize the workflows up front — you'll
  pick them up by using them. For now, remember one word: `ulw`. Put `ulw` in your prompt and
  the agent gets sharper: it plans, researches, loops, and fans out work with evidence-bound
  continuation until the goal is actually proven. Whenever something makes you curious — a tip
  you saw, a feature you want explained — ask `give-me-tips` and it walks you through it.
- **Mass-ulw**: `mass ulw` opens dependency-ordered multi-agent DAG orchestration — many child
  agents run in parallel waves, some waiting on others, so big jobs finish faster. Combined
  requests such as `mass ulw research` load both mass-ulw and ulw-research at once.
- **The fallback architect**: refusal metadata can route the unresolved engineering question
  through an architecture consultation lane while the active model continues execution.
- **Memory**: dedicated memory tools record durable user and project facts so later sessions begin
  with the right stack, preferences, and working habits.
- **Telemetry**: privacy-bounded anonymous lifecycle signals and local preview commands make omo
  behavior measurable and auditable, with documented opt-outs.
- **Init-deep**: hierarchical `AGENTS.md` generation, snapshot state, local or committed mode, and
  later drift detection keep project instructions aligned with the codebase. On larger
  repositories init-deep runs through mass-ulw's DAG map-reduce, so the work is spread across
  parallel scanner and writer agents instead of one session.
- **Tips with a live source of truth**: list the tips with the command this product is installed
  as - `omo --list-tips` on OmO Native (omo-ai installs, where `senpi` is not on PATH) or `senpi
  --list-tips` on a plain senpi install - then read and follow `give-me-tips` for any visible tip
  the user wants explained from the implementation.
- **Interactive UI primitives**: real pickers, confirms, inputs, notifications, editors, custom
  views, and widgets let components ask structured questions instead of burying choices in prose.
- **Re-running this tour**: onboarding context accompanies the first real message, not launch.
  The user can bring the tour back any time
  with the `--onboard` flag of the command this product is installed as - `omo --onboard` on OmO
  Native (omo-ai installs, where `senpi` is not on PATH) or `senpi --onboard` on a plain senpi
  install - or shut the auto-start off with the `omo-senpi-onboarding-disabled` flag.
- **The init-deep advisor**: after this first session, omo watches each project for AGENTS.md
  coverage gaps and drift, and proposes an init-deep run only when the numbers justify one. On
  this first session, you carry that proposal yourself in lane 6.

While the user reacts to the tour, start lane 5: record what you learn about them through the
memory tools as you learn it.

## 2. Migration help

Ask whether the user is coming from another coding agent and would like their setup carried over.
This lane is opt-in. If they say no, skip to lane 3.

If they say yes, start from the migration inventory that arrived with this skill's context (the
`Migration inventory` sentence): the engine already ran the same detection `omo setup` uses, and it
lists every sign-in it found (API keys to import, OAuth logins to sign in again, unmapped
providers), every custom provider, every global skill, and every MCP server from the sources it
covers - OpenCode, the oh-my-pi and gajae-code credential stores, and the engine's own files. For
those sources the inventory is the source of truth - do not re-read their config files by hand to
rebuild it. Show the user what `omo setup --dry-run` would do with it.

The scan does not know every harness. When the user is coming from one it does not detect, read
that harness's own files by hand and fold what you find into the same plan:

- Claude Code: `~/.claude/settings.json`, global and project `CLAUDE.md` files, MCP server
  definitions in `.mcp.json` or settings.
- Another coding agent the user names: its own config file in its home folder (often a `config.toml`)
  and any `AGENTS.md` files it manages.
- Anything else the user names that the inventory does not cover (for example a project
  `.mcp.json`, or an `AGENTS.md` another agent manages).

Hand-read only what the inventory does not cover; never re-derive its categories by hand.

Global OpenCode MCP servers and global OpenCode skills are not yours to move by hand. `omo setup`
imports them: MCP servers from `~/.config/opencode/opencode.json[c]` into the engine's global
`~/.omo/agent/mcp.json`, and skills from `~/.config/opencode/skills/` into `~/.omo/agent/skills/`,
converted to the shapes omo reads, consent-gated, and never overwriting a name that already exists.
Once they accept, run `omo setup --yes`: your shell is not a terminal, so plain `omo setup` stops
at its consent prompt and imports nothing.
Copying a global server into a project `.mcp.json` yourself is a bug: it disappears the moment the
user opens any other directory.

Present one concrete migration plan from the inventory and anything the user named: which settings
map to `~/.omo/omo.json[c]`, which MCP servers `omo setup` carries over globally and which
project-only servers still belong in that project's `.mcp.json`, which `CLAUDE.md` content becomes
project `AGENTS.md` content, and which personal facts belong in memory instead of files.
Show the plan and WAIT for the user to accept it. Apply nothing before they say yes. If they accept
part of it, apply that part only. Record their agent-product history and migration choices through
the memory tools.

## 3. Session archaeology

Ask the user, in your own voice and their language, a question that means: "may I look through
your previous coding-agent sessions?" Phrase it naturally; do not read that sentence out like a
script. This lane is opt-in. If they say no, skip to lane 4 and base it on nothing.

If they say yes, call `onboarding_understand` with `consent: true` (find it through `tool_search`
if it is not in your tool list). It reads the scan's session digests with the current model and
returns candidate facts with ids; it saves nothing. Show every candidate as a numbered list and ask
which ones to remember. Then call `onboarding_memory_apply` with exactly the ids they picked: it
writes only those, as one memory commit, through the same writer the background facts extractor
uses. Never write a candidate the user did not pick, and never re-write candidates through the
`memory` tools. If the understanding falls back to the scan only, say so and continue.

For the patterns below, call the `onboarding_session_history` tool as well. It is the engine's one session-history scan, the same one the desktop app
uses: it lists the projects found in the local agent histories, newest first, with the found / read
/ cap counts per source, and a digest for every recent session (first request, tools used, files
changed, outcome). Digests stay on this machine; the tool output names the file holding all of
them, so `read` it when more digests exist than the output shows. Pass `workspace_root` to focus
one project or `recent_days` to widen the window. Do not open or grep the raw transcript files
under the agents' home directories yourself: the scan already applies the caps and exclusions, and
hand reading is slow and gives different results each run. Mine the digests for:

- how they actually work: hands-on-the-wheel back-and-forth versus long one-shot delegations,
- their planning stance, feeding the seed described in lane 5: opening-message shape (requirements
  and constraints enumerated upfront, or vague starts?), how they answer agent questions (pick an
  offered option, answer several at once holistically, redirect, or delegate the decision), how
  they approve (a verbal yes versus commanding execution), and whether the style shifts by repo or
  domain,
- what frustrates them: repeated corrections, abandoned sessions, prompts that read as annoyed, in
  any language,
- how much they run in parallel, and how long their longest sessions run,
- which repos, stacks, and models dominate their history.

Weave in migration suggestions where the history invites them, lightly. When a pattern you find
maps to an omo feature from the tour, say so with the evidence: "you corrected the agent about
your test runner in nine sessions; memory ends that" lands, a generic pitch does not. Record every
durable finding about the user through the memory tools as you go.

## 4. Value mapping + savings

Open with the scan's `Value summary` line as it was computed. Then, from the session data gathered
in lane 3, give the user quantified estimates of what omo's caching
would have been worth on their real workload. Compute, do not guess:

- Quantify token and cost savings only from Senpi JSONL v3 sessions. Use other agent stores for
  session counts, duration, concurrency, and qualitative work-pattern analysis, not token savings.
  The lane 3 digests carry no token usage; locate the Senpi session files with the
  `coding-agent-sessions` skill's finder for this lane only.
- Fetch `https://models.dev/api.json`. Treat its top level as the provider map. Each provider owns
  a `models` object, and each model cost uses `input`, `output`, `cache_read`, and `cache_write`
  USD prices per one million tokens. If the API is unavailable, every session has insufficient
  pricing data.
- Process session files in chronological filename order. Process entries inside a file in JSONL
  line order. A malformed timestamp skips that entry. Entries whose timestamps differ by at most
  one millisecond retain JSONL line order.
- Read usage only from `message.usage`. Map `input`, `cacheRead`, `cacheWrite`, and `output`
  exactly. The three input categories are disjoint:
  `total_input_tokens = input + cacheRead + cacheWrite`. A session whose total input is zero is
  skipped.
- Attribute each usage entry to `message.provider` plus `message.model` when both are present.
  Otherwise use the latest `model_change` provider and `modelId`. Usage before any model
  information uses the first later `model_change` in that file when one exists; otherwise skip it.
- Resolve prices first by exact `provider/model` composite identity. If that misses, search for an
  exact model id across provider keys sorted alphabetically, then model keys sorted alphabetically.
  The first deterministic hit wins. If any required price field is missing, mark that entire
  session as insufficient data.
- Aggregate usage by provider and model, then compute grand totals. The caching rate is
  `cacheRead / (input + cacheRead + cacheWrite) * 100`.
- Compute actual cost as
  `(input * price.input + output * price.output + cacheRead * price.cache_read + cacheWrite * price.cache_write) / 1_000_000`.
  Compute omo savings as
  `cacheRead * (price.input - price.cache_read) / 1_000_000`.
- Round only final aggregated values with `Math.round(value * 100) / 100`. Never round per message,
  model, or session. Label the result `estimate`.
- Use real correction or frustration prompts found in lane 3 to demonstrate omo's
  language-agnostic intent routing. Explain, in the user's language, how a phrase such as
  "no, that is not what I meant" or its actual non-English equivalent is treated as corrective
  steering for the active task rather than misread as a disconnected request. Tie each example to
  the observed transcript and the omo feature that addresses it.

Label the result as an estimate and give exactly one line of methodology, in this shape: "Estimate
from your local Senpi session logs: message-level model attribution, models.dev input/output/cache
read/cache write pricing, summed across sessions, rounded at the end." One line, then the numbers,
then stop.

If the user skipped lane 3, skip this lane too; there is no data to map.

## 5. Memory recording

This lane has no fixed position: it runs through the entire flow. Nothing is written before the
user chooses it: facts from the session history go through the lane 3 selection, and anything
else you learn is proposed in one line and written through the `memory` tools only after the user
agrees. Worth recording:

- their host and machine facts relevant to future work,
- their language and communication style,
- which agent products they came from and what they kept from them,
- their stacks, main repos, and working patterns from lane 3,
- their stated preferences and every accept or decline decision from this conversation.

Write durable personal and cross-project facts through the memory tool into `system/human.md`.
Write repository-specific stack, commands, constraints, and migration decisions through the memory
tool into `system/project.md`. Do not edit those backing files directly.

Planning-stance observations from lane 3 are the one exception: write them as SEEDS under a
`## Seed` heading in `system/human/planning-style.md` (create the block through the memory tool if
absent), never into `system/human.md`. Seeds are inferences and MUST stay weak:

- Phrase every seed as a hypothesis ("appears to enumerate upfront and delegate the rest"), never
  a conclusion, and set `confidence: low` ALWAYS - never medium or high, however consistent the
  history looks. Behavior in another tool reflects that tool's affordances, not necessarily this
  user; omo's own planning sessions will confirm or replace these within a few runs.
- Record full provenance on every seed line: the source harness name(s) (e.g. Claude Code,
  Codex, OpenCode), the model(s) under which the pattern was mainly observed, the session ids
  mined, and today's date as `seeded:`. Format:
  `- [YYYY-MM-DD] (seed) appears to <hypothesis + context> <!-- src: <harness>/<session-id>,...; harness: <names>; model: <models>; seeded: YYYY-MM-DD; pattern: inferred-from-external-sessions; confidence: low -->`
- Seeds carry interaction STYLE only - never session content, code, file paths, or secrets.

Record facts, not narration. "Prefers Korean, migrated from Claude Code, works one repo at a time
in long sessions" is a memory. "The user went through onboarding today" is not.

## 6. First-session init-deep proposal

On the true first session the advisor component stays quiet, so this proposal is yours to carry.
Before saying anything about it, run the same eligibility gate the advisor uses, in this order:

1. The current directory must be inside a git repository. Check with
   `git rev-parse --show-toplevel`. Not a repo: ineligible.
2. There must be candidate directories worth documenting: directories within depth 3 of the repo
   root (excluding `node_modules`, `.git`, `dist`, `build`, `vendor`, `.next`, `__pycache__`,
   `.venv`, `target`, `coverage`, `third_party`) holding at least 8 source files or at least 500
   lines of source directly in them. Zero candidates: ineligible, regardless of anything else.
3. Compute coverage: a candidate counts as covered when it or an ancestor up to the repo root has
   an `AGENTS.md`. The missing ratio is uncovered candidates over all candidates. Below 0.50:
   ineligible. A new project with candidates and no `AGENTS.md` anywhere is missing ratio 1.0 and
   eligible.

If any gate fails, skip this lane SILENTLY. Do not mention AGENTS.md, do not explain why you are
not proposing, do not hint that a check ran. Close the conversation warmly instead.

If all gates pass, ask in the user's language: "want me to set up AGENTS.md for this project?"
This is opt-in. On yes, read the `init-deep` skill at its SKILL.md path and follow it. On no,
record the decline through the memory tools and finish the conversation gracefully: a short
send-off in their language, an invitation to come back with `omo --onboard` on OmO Native or
`senpi --onboard` on a plain senpi install, and nothing more.

After init-deep finishes, tell the user in their language that init-deep ran through mass-ulw's
DAG orchestration — the work was spread across parallel child agents in dependency-ordered
waves rather than done by one session. Then introduce mass-ulw: `mass ulw` is how you bring that
same multi-agent orchestration to any big job, and combined requests such as `mass ulw
research`, `mass ulw loop`, and `mass ulw plan` load mass-ulw alongside each of those workflows.