project-onboard（船上项目）

project-onboard

Reads the current project's file system and determines which of 6 onboarding cases applies, then recommends the exact command sequence.

Triggers: /project-onboard, onboard project, diagnose project state, what do I run first, project setup help

Process

I run a 5-check waterfall in strict priority order. I read real files — I never ask the user questions. The first failing check determines the primary case assignment. Check 4 is non-blocking: a project can simultaneously be Case 6 (healthy) and have local skill issues.

Check 1 — CLAUDE.md exists

Read .claude/CLAUDE.md (also accept CLAUDE.md at project root for global-config repos).

If absent → Case 1: Brand-new project

## Diagnosis

Project state: Case 1 — Brand-New Project

Detected:
- .claude/CLAUDE.md: NOT FOUND
- No Claude configuration present in this project

Warnings:
- None

## Recommended Command Sequence

1. /project-setup     — creates .claude/CLAUDE.md, ai-context/ skeleton, persists context to engram
2. /memory-init       — generates ai-context/ files from real project content
3. /project-audit     — produces audit-report.md with score and findings
4. /project-fix       — applies all corrections from the audit report

## Notes
After /project-fix, re-run /project-audit to verify score ≥ 75 and SDD Readiness = FULL or PARTIAL.
See ai-context/scenarios.md → Case 1 for failure modes and recovery steps.

Stop here if Case 1.

Check 2 — Engram MCP reachable

Check if Engram MCP is reachable (call mem_context).

• If Engram is reachable → SDD infrastructure lives in Engram. Treat the project as healthy for this check and continue to Check 3.
• If Engram is NOT reachable → Case 2 (below).

If Engram not reachable → Case 2: CLAUDE.md present but no SDD persistence

## Diagnosis

Project state: Case 2 — CLAUDE.md Without SDD Persistence

Detected:
- .claude/CLAUDE.md: FOUND
- Engram MCP: NOT REACHABLE
- SDD cannot persist artifacts without Engram

Warnings:
- [list any ai-context/ files found or note if ai-context/ is absent]

## Recommended Command Sequence

1. Ensure Engram MCP server is running and configured
2. /project-audit     — diagnose the full scope of what is missing
3. /project-fix       — adds SDD section to CLAUDE.md
4. /memory-init       — if ai-context/ is empty or absent
5. /project-audit     — verify score improved

## Notes
project-fix will ask before every change — review each proposed action carefully.
See ai-context/scenarios.md → Case 2 for failure modes and recovery steps.

Stop here if Case 2.

Check 3 — ai-context/ has ≥ 3 populated files

Read ai-context/ directory. Count files that exist AND have more than 10 lines. Expected files: stack.md, architecture.md, conventions.md, known-issues.md, changelog-ai.md.

If fewer than 3 populated files → Case 3: Partial SDD, sparse memory layer

## Diagnosis

Project state: Case 3 — Partial SDD (ai-context/ is sparse)

Detected:
- .claude/CLAUDE.md: FOUND
- Engram MCP: REACHABLE
- ai-context/ populated files: [N] of 5 (minimum needed: 3)
- Missing or empty: [list each absent/stub file]

Warnings:
- [list any other issues found, e.g. stale onboarding.md]

## Recommended Command Sequence

1. /memory-init       — regenerates all ai-context/ files from real project state
2. /project-audit     — verify D2 score improved
3. /project-fix       — address any remaining findings

## Notes
/memory-init does not overwrite files that already have substantial content.
See ai-context/scenarios.md → Case 3 for failure modes and recovery steps.

Stop here if Case 3.

Check 4 — Local skills review (non-blocking)

Read .claude/skills/ directory. If it exists and has any subdirectories, record a warning for the diagnosis output. Do not stop — continue to Check 5.

Local skills warning (append to any case diagnosis):

Warnings:
- .claude/skills/ found with [N] local skill(s): [list names]
  Run /project-audit to see Dimension 9 findings (duplicate detection, structural completeness, language compliance).
  Run /project-fix Phase 5 to apply corrections after auditing.

Check 5 — Orphaned SDD changes

Search engram for active SDD changes: mem_search(query: "sdd/", project: "{project}"). Filter for changes that have a state artifact but no archive-report, and check if they have tasks and verify-report artifacts.

If any change is missing tasks OR verify-report artifacts → Case 5: Orphaned or stale changes

## Diagnosis

Project state: Case 5 — Orphaned SDD Changes

Detected:
- .claude/CLAUDE.md: FOUND
- Engram MCP: REACHABLE
- ai-context/: adequate ([N] populated files)
- Orphaned changes:
  - [change-name]: missing [tasks | verify-report]
  - [change-name]: missing [tasks | verify-report]

Warnings:
- [local skills warning if Check 4 triggered]

## Recommended Command Sequence

1. /sdd-status        — see all active changes and their current phase
2. For each orphaned change, one of:
   /sdd-apply <name>  — if tasks.md is missing (change never implemented)
   /sdd-verify <name> — if implemented but never verified
   /sdd-archive <name> — if complete but never archived
3. /project-audit    — verify D3 shows no orphaned changes

## Notes
To discard a dead-end change: run /sdd-archive <name> to close it in engram.
See ai-context/scenarios.md → Case 5 for failure modes and recovery steps.

Stop here if Case 5.

Check 6 — All healthy (default)

All checks passed. The project is fully configured.

## Diagnosis

Project state: Case 6 — Fully Configured

Detected:
- .claude/CLAUDE.md: FOUND
- Engram MCP: REACHABLE
- ai-context/: adequate ([N] populated files)
- Engram: no orphaned SDD changes

Warnings:
- [local skills warning if Check 4 triggered]
- [stale onboarding.md warning if Last verified > 90 days: "Run /project-update to refresh user docs"]

## Recommended Command Sequence

For a well-understood change:
  1. /sdd-propose <change-name>  — create proposal (orchestrator fast-forwards planning automatically)
  2. /sdd-apply <change-name>    — implement the task plan

For a complex or vague change:
  1. /sdd-explore <topic>        — investigate the area before committing to a change
  2. /sdd-propose <change-name>  — create proposal from exploration findings
  3. /sdd-apply <change-name>    — implement the task plan

## Notes
Multi-phase flows (propose → spec+design → tasks in one shot, or full cycles with exploration) are handled
by the orchestrator as meta-commands typed directly in conversation — not invoked as skills.
See ai-context/quick-reference.md for entry-point guidance.

Check 7 — Runtime sync hint (non-blocking, global-config mode only)

If install.sh AND skills/ exist at the project root (i.e., the cwd is the agent-config meta-repo), append the following line to the Warnings: section of the case diagnosis output — regardless of which case was assigned:

- Run /claude-folder-audit to verify ~/.claude/ is in sync with this repo (installation drift check).

This check MUST NOT change the case assignment, alter the Recommended Command Sequence, or interrupt the existing 6-check waterfall. It is a hint only.

Stale docs warning logic

After determining the primary case, check ai-context/onboarding.md, ai-context/scenarios.md, and ai-context/quick-reference.md for the > Last verified: field. If any file's date is more than 90 days from today, append to the Warnings section:

- [filename] is stale ([N] days since last verification). Run /project-update to refresh.

Rules

• Do not ask the user any questions — detection is fully automatic from file-system state
• Detection is derived from real file presence/absence — no hardcoded case IDs or lookup tables
• Priority order is strict: once a case is assigned at check N, do not also assign a lower-priority case (exception: Check 4 local skill flag is always surfaced as a warning, non-blocking)
• Emit structured output only — no raw directory listings, no stack traces, no internal file dumps
• Make no file-system changes — this skill is 100% read-only
• Always include a ## Notes section pointing to the relevant case in ai-context/scenarios.md for failure modes