vibe-coding — session memory for AI pair programming

You have the vibe-coding-mcp MCP server available (tools are prefixed muse_). It stores coding sessions as JSON files under ~/.vibe-coding-mcp/sessions/ and can read Claude Code's own transcripts from ~/.claude/projects/. Nothing leaves the machine unless the user publishes.

0. First: make sure capture is actually running

Before producing any report, check that there is data. Call muse_session_history with action: "stats". If totalSessions is 0, or the user asks for "setup":

1. Run npx -y vibe-coding-mcp doctor and show the user the result.
2. If hooks are not installed, run npx -y vibe-coding-mcp init (adds Stop + SessionEnd hooks to ~/.claude/settings.json, backs the file up first). Tell the user to restart Claude Code once.
3. To backfill history that already exists on disk, call muse_capture_transcript with { "action": "capture", "all": true }. This parses every transcript for the current project and stores one session per Claude Code session (re-running is safe: it updates instead of duplicating).

If the MCP server itself is missing, the user needs: claude mcp add vibe-coding-mcp -- npx -y vibe-coding-mcp.

1. Pick the workflow from the user's words

2. Rules that keep the output trustworthy

• Never invent work. Every line in a report must trace to a captured session (title, file, decision, or prompt). If the digest is empty, say so and offer setup/backfill.
• Keep file paths and identifiers in backticks, exactly as captured.
• Decisions need a rationale. When designDecisions is empty for a session, run muse_summarize_design_decisions over the session's summary + prompts before writing "Decisions". If still nothing, write "no explicit decisions recorded".
• Respect the session's language. Captured prompts in Korean → report in Korean unless asked otherwise (language: "ko" on muse_digest).
• One capture per Claude session. muse_capture_transcript is idempotent; do not loop it.
• Do not read transcripts yourself with Read/Bash when a tool exists; the parser handles truncation and privacy (sub-agent side-chains, system reminders).

3. Output shapes

Standup (plain text, no headings):

Yesterday
- Refactored auth middleware (`src/auth.ts`, `src/db.ts`); chose JWT over sessions for statelessness
Today
- Finish checkout bug (`src/checkout.tsx`), uncommitted changes pending
Blockers
- none

Weekly report: ## Shipped (per project, bullets with files) · ## Decisions (bullet: decision — rationale) · ## In progress / blocked · ## Next week.

PR description: ## Why · ## What changed (file-by-file) · ## Trade-offs · ## How to review (Mermaid diagram when muse_analyze_code produced one).

4. References

• references/tools.md — every tool, resource, prompt with parameters and env vars.
• references/workflows.md — step-by-step tool chains with example calls and expected output.
• scripts/setup.sh — one-shot install: MCP registration + hooks + doctor.
• assets/hooks.settings.json — the exact hooks block init installs, for manual setup.