onboard - finish the Blueprint overlay setup

Context reuse: Reuse any required file already loaded in project instructions or the current session. Read it again only if absent, changed, or exact current bytes or line references are needed.

First action: Before project inspection, preflight, or any other tool call, publish running to blueprint/.state/run.json using the dashboard activity contract in AGENTS.md.

Where this sits in the workflow:

scaffold app  ->  overlay Blueprint  ->  [onboard]  ->  project-plan + build-plan  ->  /overview
(user/tool)       (copied files)          (tune setup)   (user-owned inputs)       (generated context)

/onboard is the fresh-project on-ramp. It assumes the app was scaffolded first and the Blueprint files were overlaid after. Run it before filling in plans or running /overview. Its job is to make the Blueprint fit the real project before planning starts: commands, project title, conventions, ignore rules, and tool adapters. It also asks whether the Blueprint workflow files should be committed with the repo or kept local-only through .gitignore.

Use /adopt instead when the app is brownfield: real routes, shipped features, and project behavior already exist and need to be reflected into the plans.

Input

No argument is required. If the user provides context about the stack, hosting, database, auth, or preferred tool, use it as a hint and verify against files.

Step 0 - confirm Git and make an unborn repository usable

Before reading application code or changing setup files, confirm both Git states:

git rev-parse --is-inside-work-tree
git rev-parse --verify HEAD

If this is not a Git repository, stop and ask the user to initialize one, then rerun /onboard.

An existing first commit may contain only the scaffold or may already contain Blueprint. Both are valid. Do not ask the user to rewrite either history shape.

If Git reports an unborn HEAD, handle it here instead of sending the user away to run Git commands:

1. Inspect status and build a safe scaffold-only candidate from paths outside AGENTS.md, CLAUDE.md, .agents/, .claude/, and blueprint/. Exclude secrets, dependencies, caches, build output, generated state, and anything else that should not enter source control. Include the existing .gitignore when it is safe.
2. Resolve the intended default branch from a remote default when available, then an existing main or master, then Git's configured initial branch, and otherwise main. Preserve the current unborn branch name as the setup branch when it is not the intended default. If the intended name is genuinely ambiguous, ask only that one question.
3. Show the exact candidate and the branch result, then ask once: Create the initial scaffold commit and continue Onboard? (Recommended) State that this creates one local commit and never pushes.
4. On approval, stage only the reviewed candidate, verify the staged diff, and commit it as chore: scaffold application. If needed, rename the unborn branch before committing so the root commit establishes the intended default branch. Then create or return to the named setup branch at that same commit and continue Onboard in the same run.

If there is no safe scaffold candidate, stop with the exact blocker rather than creating an empty or mixed root commit. Never create the commit without explicit approval.

Then confirm this is onboarding, not adoption.

Inspect the repository and the two planning docs:

• If blueprint/project-plan.md and blueprint/build-plan.md are mostly empty or worksheet-like, proceed.
• If the app already has substantial shipped features, stop and recommend /adopt instead.
• If the plans already contain real user-owned content, do not overwrite them. Continue only with setup files such as AGENTS.md, coding-standards.md, .gitignore, and optional notes.

Never run a framework scaffolder. The Blueprint is already overlaid.

Step 1 - survey the project facts

Read only enough to identify the setup:

• package manager and lockfile (pnpm-lock.yaml, package-lock.json, yarn.lock, bun.lockb, etc.)
• manifest scripts (package.json, pyproject.toml, go.mod, Cargo.toml, and similar)
• framework and runtime config (astro.config.*, next.config.*, vite.config.*, tailwind.config.*, database config, test config)
• source layout, route layout, and app/package directories
• existing .gitignore
• blueprint/.state/manifest.json, when present, and its selected adapter list
• which selected tools need .agents/, .claude/, or both
• whether Blueprint workflow paths are already tracked by git
• existing verification commands and .github/workflows/
• blueprint/config.json, when present, and whether it parses cleanly
• project name, from package.json, the folder name, existing docs, or the user

Do not infer more than the files support. Mark uncertain items as > TODO in the summary rather than inventing a convention.

Step 2 - update project entry files

If the root README.md is a copied Blueprint workflow document, replace that obsolete overlay content in the product README slot:

• Detect it conservatively: the first heading is # AI Coding Blueprint, or the opening section clearly describes the Blueprint workflow rather than this app.
• Create a small root README.md stub for the actual project using the detected project name, one-line purpose when known, and the Commands from AGENTS.md. Keep it minimal if the project plan is not filled yet.
• Do not move or copy the workflow document into blueprint/. Agents use the local skills, plans, and context files directly.
• Remove any AGENTS.md claim that a project README explains the Blueprint workflow.

If the root README.md already looks like a real project README, leave it alone. Never replace a project README with Blueprint documentation.

Update the Commands section of AGENTS.md to match real scripts and commands. Remove the shipped <!-- blueprint:onboarding-required --> marker and the For a standard Next.js project instruction when replacing the placeholder commands. Status uses the dedicated marker, with the old sentence retained only as a migration fallback, to distinguish a fresh overlay from a tuned project. Include only commands that exist or are intentionally available:

• dev server
• build
• preview or start
• lint, format, typecheck, and test, if configured
• verify, when a real combined verification command already exists
• useful app-specific commands, if obvious

If no test command exists, say so explicitly. Do not claim tests are a gate until a real test command is configured.

Follow and preserve the proportional-engineering contract in AGENTS.md when adapting commands, standards, and context to the detected project.

If CLAUDE.md exists and still has the placeholder # Project Name, replace it with the detected project name. Keep @AGENTS.md. Remove direct imports of project-overview.md, current-feature.md, coding-standards.md, and ai-interaction.md; workflow skills read those files only when relevant. Preserve unrelated user imports. Do not move detailed app context into CLAUDE.md; that belongs in AGENTS.md and the generated project overview.

Step 3 - tune coding standards

Update blueprint/context/coding-standards.md so it matches the detected stack. Keep stable, tool-agnostic sections such as writing style, comments, scope, and testing philosophy. Replace stack-specific defaults that do not apply.

Cover the practical conventions the build loop needs:

• framework and rendering model
• package manager
• project structure
• styling approach
• data access and API boundaries, if known
• validation and error handling expectations
• test gate status
• build and verification commands, via AGENTS.md

If the project is too new to reveal a convention, leave a concise > TODO rather than pretending a pattern exists.

Step 4 - check project configuration and AI interaction rules

Read blueprint/config.json. A missing file means built-in defaults and is not an error. If the file exists but is invalid, stop and show the exact invalid key or value before changing other setup files.

Keep project configuration deterministic. Ask before changing preferences and edit only values the user actually chose, such as branch prefixes, UI evidence, logic-test strictness, regular or Continuous quality gates, review execution, or Continuous Mode limits. Independent review defaults to when-sensitive for regular and Continuous work, and its execution defaults to automatic. Audit, and check default to manual. Try guide defaults to when-testable, which adds a compact Try It Now demonstration only when a safe meaningful observation exists. Preserve these defaults unless the user chooses different policies. A manual independent-review gate disables automatic selection for that workflow without disabling explicit independent audits. Never put commands, product requirements, communication prose, secrets, or permission for commits, merges, pushes, deployments, publication, destructive actions, failed checks, or finding waivers into config.

Unless the user already chose these values, ask one short Implementation style question using the current tool's selectable prompt when available:

1. Efficient (Recommended) - one feature-level review packet, a final code walkthrough option, and no step checkpoint prompts. Write workflow.stepReview: "feature" and workflow.checkpointCommits: "disabled".
2. Guided - pause for approval after every step and offer optional checkpoint commits, followed by the same final code walkthrough option. Write workflow.stepReview: "every" and workflow.checkpointCommits: "enabled".
3. Custom - ask separately when review should happen and whether checkpoint commits should be offered, then write the selected low-level values.

These are onboarding presets, not a third configuration field. Never write an implementationStyle key. Show the current two values before asking, preserve them if the user chooses not to change them, and explain that either value can be edited later. A later /implement run reads the current configuration. The final code walkthrough is not a configuration setting and remains available with every implementation style.

Read blueprint/context/ai-interaction.md and update only obvious mismatches. Usually the default review loop should stay intact. Flag preferences for the user instead of guessing, such as:

• whether review should happen once per feature (the lower-context default) or after every step for teaching, close pairing, or high-risk work
• whether optional step checkpoint commits should be enabled. Explain that the previous workflow requires per-step review and enabled checkpoints together; changing only stepReview restores the approval pauses, not checkpoint prompts
• whether branches should use a different naming pattern
• whether /check should require browser evidence for UI work
• whether audit, independent review, or check should stay manual, run only for their documented conditional case, or run for every regular or Continuous work item
• whether try guides should use the recommended when-testable policy, remain manual, use the narrower when-user-facing policy, or run for every item
• whether a selected independent review should use the default automatic isolated reviewer or review.independentExecution: "manual" for a fresh-session handoff. Explain that automatic execution uses a fresh isolated reviewer only when the active adapter can expose its exact identity and model, and otherwise stops with the manual handoff

If no changes are needed, say so.

Step 5 - point to optional CI setup

Do not create or change Verify commands or GitHub workflows during onboarding. Report any verification command or CI already present. When equivalent automatic pull-request checks are absent, mention the optional standalone setup:

Run /ci or $ci when you want automatic GitHub checks.

Explain that CI is not required to continue with planning or the Blueprint build loop. The /ci skill owns project-specific Verify and GitHub workflow setup.

Step 6 - check ignore files, visibility, and adapters

Update .gitignore for common generated files from the detected stack while preserving existing entries. Typical examples include dependencies, build output, framework caches, logs, environment files, test output, temporary files, and OS or editor files.

Ask how Blueprint workflow files should be handled in git, unless the user already gave a preference:

Blueprint visibility?

1. Commit Blueprint workflow files
   Portable. Best for teams and working across machines.

2. Keep Blueprint workflow files local
   Adds .agents/, .claude/, blueprint/, and CLAUDE.md to .gitignore.
   Keeps AGENTS.md public as the lightweight project agent guide.

Recommend option 1 by default. If the user chooses option 2:

• Add this block to .gitignore, preserving existing entries:

  # AI Blueprint local workflow files
  .agents/
  .claude/
  blueprint/
  CLAUDE.md
  

• Keep AGENTS.md tracked. It remains the lightweight public project guide for commands and conventions.

• Make AGENTS.md public-safe: keep project description, commands, testing gate, and coding conventions, but remove or avoid Blueprint workflow explanations, hidden adapter paths, workflow-document pointers, and core skill lists that would expose the local-only workflow.

• Explain that local-only mode hides the workflow contents from the repo, but the .gitignore names still reveal the ignored paths.

• Explain that Blueprint state, specs, findings, and history will not travel with the repo; another machine needs the Blueprint reinstalled or restored locally.

• If any of .agents/, .claude/, blueprint/, or CLAUDE.md are already tracked, say .gitignore will not hide tracked files. Ask before running git rm --cached -r .agents .claude blueprint CLAUDE.md, and only run it if the user explicitly approves. Never delete the local files.

Then report which selected tools and adapter folders are needed:

• When a valid blueprint/.state/manifest.json exists, its adapters list is the authoritative installer selection. The presence of .agents/ means its files are compatible with Codex and GitHub Copilot. Google Antigravity and OpenCode can use them too; it does not mean all four tools were selected.
• Do not ask the user to select adapters again when that valid manifest exists. Keep and report the exact selection. If a required adapter tree is missing, report the mismatch and point to /doctor instead of guessing or deleting another tree.
• Without a valid manifest, explain that folder detection cannot distinguish Codex or GitHub Copilot from Google Antigravity or OpenCode. Ask which tools the user actually uses instead of assuming all of them are selected.
• Codex only: keep AGENTS.md, .agents/, and blueprint/; CLAUDE.md and .claude/ can be deleted.
• Claude Code only: keep AGENTS.md, CLAUDE.md, .claude/, and blueprint/; .agents/ can be deleted.
• GitHub Copilot only: keep AGENTS.md, .agents/, and blueprint/.
• Google Antigravity only: keep AGENTS.md, .agents/, and blueprint/.
• OpenCode only: keep AGENTS.md, .agents/, and blueprint/.
• OpenCode with Claude Code: OpenCode can reuse .claude/; no separate .opencode/skills/ copy is needed.
• Mixed tools: keep only the compatible adapter trees required by the selected tools. Never create .opencode/skills/ or an Antigravity-specific skill tree. Both tools use the supported shared trees.

Do not delete adapters unless the user explicitly asks.

Step 7 - hand off to planning

Stop with a concise onboarding report:

• stack and package manager detected
• project name used for entry files
• README handling, especially if the copied Blueprint README was moved
• Blueprint visibility choice
• tracked-file warning if local-only mode was chosen after files were already tracked
• files changed
• project configuration state and any user-selected overrides
• commands now available
• testing gate status
• verification command and GitHub checks status
• adapter recommendation
• TODOs or uncertainties
• exact next files for the user to fill in:
  • blueprint/project-plan.md
  • blueprint/build-plan.md

Make the direct path clear: the user can write or develop those files through any conversation, then run /overview. Also mention /discovery or $discovery as an optional deep planning conversation for users who want guided help. Do not start it, make it a prerequisite, or imply that directly written plans are less complete.

End with the next command:

/overview

For Codex, also mention:

$overview

Rules

• Setup files are fair game; planning docs are user-owned.
• /discovery is optional and never runs as part of onboarding. The direct plan-writing path must remain fully supported.
• Never overwrite real project-plan.md or build-plan.md content.
• Never run scaffolders or install dependencies unless the user explicitly asks.
• Reflect the stack that exists, not the stack the default Blueprint mentions.
• Be honest about tests. No test command means no required test gate yet.
• Keep AGENTS.md public in local-only mode unless the user explicitly asks for a more advanced setup.
• Do not untrack Blueprint files with git rm --cached without a separate explicit approval.
• Keep changes small and explain what changed.

Formatting

Format the output to match the project's conventions in blueprint/context/ai-interaction.md: concise, scannable markdown, with lists for enumerations and tables for matrices rather than dense paragraphs.