session-start

Session Start — Full Context Loading (MANDATORY)

This skill is REQUIRED at the start of every session. It must run automatically — do not wait for user confirmation. Do not skip any step. Read ALL memory files without exception.

Load the complete project context at the beginning of a new conversation to ensure continuity across sessions. This restores accumulated knowledge, reviews recent work, and presents an orientation resume.

Context comes from two sources:

• Shared team knowledge — bundled with this plugin in references/ (architecture blueprints, known issues, coding standards). Available to every teammate.
• Personal memory — each user's local auto-memory directory (user profile, feedback preferences). Private per developer.

Workflow

Execute all steps. Parallelize independent operations for speed.

Step 1: Load Shared Team Knowledge (REQUIRED — read ALL files)

Read references/MEMORY.md (bundled with this plugin at ${CLAUDE_PLUGIN_ROOT}/references/MEMORY.md) to get the index of shared knowledge files.

Then read EVERY reference_*.md and project_*.md file in the same ${CLAUDE_PLUGIN_ROOT}/references/ directory. These contain:

• References — architecture blueprints, coding standards, schemas, deployment guides, security patterns
• Project — known issues, ongoing work context, decisions

Load ALL of them in parallel. Do not skip any. Do not summarize without reading. Every file must be opened and read.

Step 2: Load Personal Memory (REQUIRED — read ALL files)

Read the user's local auto-memory index at: ~/.claude/projects/<sanitized-cwd>/memory/MEMORY.md

Then read EVERY file referenced in MEMORY.md — not just some, ALL of them:

• User — profile, role, workflow preferences
• Feedback — corrections and confirmed approaches that must be applied to ALL work in this session
• Reference — all reference memories
• Project — all project memories

Feedback memories are CRITICAL and NON-NEGOTIABLE — they encode how this specific user wants to work (e.g., commit style, code review preferences). These apply session-wide and must never be ignored.

If the local memory directory doesn't exist, that's fine — skip and continue with shared knowledge only. But if it exists, read EVERY file in it.

Step 3: Read Project Guidelines

Read AGENTS.md from the project root (if it exists). It contains supplementary rules about:

• Default workflows
• Tool usage preferences
• Framework-specific guidelines

These rules supplement CLAUDE.md and apply to all work in the session.

Step 4: Fetch Git & GitHub State

Run these commands in parallel:

git log --oneline -20                    # Recent commit history
git branch --show-current                # Current branch
git status --short                       # Working tree state
gh pr list --limit 10 --state all        # Recent PRs (open + closed + merged)

Step 5: Present the Resume

Compile everything into a concise resume:

## Session Resume

**Branch:** <current-branch>
**Working tree:** <clean / dirty — list changed files if dirty>

### Recent Commits
<Last ~10 commits grouped by feature area, as bullet points>

### Pull Requests
<Open PRs first, then recently closed/merged — or "None open">

### Known Issues
<Remaining items from project_known_issues.md — bulleted list>

### Session Rules (from personal feedback memories)
<One-line reminders of each active feedback rule, or "No personal memory loaded">

Keep it scannable — the full context is already loaded, so the resume just needs to orient.

Step 6: Show Capabilities & Ready for Work

After presenting the resume, display a short capabilities banner so the user knows what's available:

### Ready — here's what I can do

| Category | Capabilities |
|----------|-------------|
| **Build** | Brainstorm → spec → plan → implement (full superpowers workflow) |
| **Test** | TDD, run tests, coverage analysis |
| **Debug** | Systematic debugging with root cause analysis |
| **Review** | Code review PRs, review specs/plans before you see them |
| **Git** | `/commit`, `/commit-push-pr`, branch management, PR creation |
| **Frontend** | Next.js pages, components, i18n, auth flows, browser testing |
| **Backend** | NestJS services, gRPC, guards, Firestore, Prisma, Redis |
| **CI/CD** | Workflows, deploy configs, gateway OpenAPI, Docker |
| **Research** | Web search, library docs (context7), codebase exploration |
| **Memory** | I remember your preferences and project context across sessions |

Type what you need — or just describe the problem and I'll pick the right workflow.

Then wait for the user's first task. Do not start any work unprompted.

Important Notes

• This skill is MANDATORY — it must run at the start of every session without user confirmation. Never bypass it.
• Read ALL memory files — do not skip, summarize, or selectively load. Every reference, feedback, user, and project memory must be opened and read.
• Memory files are point-in-time snapshots. Before acting on claims about code (file paths, function names, flags), verify against the current codebase.
• If gh CLI is not authenticated or fails, skip the PR list and note it — don't block the session.
• If any memory file is missing, note it but continue loading the rest.
• The resume should feel like a handoff briefing — quick to scan, enough to jump straight into work.

Updating Shared Knowledge

To update the shared team knowledge, edit the files in this plugin's references/ directory and push to the repo. All teammates will get the updates on their next plugin sync.