Save Work

Save Session

Menu-driven session save for non-CLI interfaces (Cowork desktop, Telegram, CC CLI). Same CONTEXT file format as retrospect:save-context / retrospect:load-context — files are interchangeable.

Target: 1200-1500 tokens MAX for CONTEXT file

Phase 1: Detect Current Project (parallel)

Glob: CONTEXT-*-llm.md (current dir, exclude done/)
Read: CLAUDE.md (first 5 lines, for project name from H1 — required)

Identify:

• Project name (from CLAUDE.md H1 heading — all CLAUDE.md files must have # Project Name as first line)
• Existing CONTEXT streams (filenames → stream names)

If no CLAUDE.md found in current dir or parents → warn: "⚠️ No project context detected. Use /switch first to select a project."

Before Phase 2: run quality self-check from reference.md — if session is trivial, confirm with user before proceeding.

⚠️ AskUserQuestion Guard

After EVERY AskUserQuestion call, check if answers are empty/blank. If empty: output "⚠️ Questions didn't display (known bug).", present options as numbered text list, WAIT for user reply.

Phase 2: Ask Save Type + Stream (single menu when possible)

0-1 existing streams → merge type + stream into 1 question:

💾 Save {stream-name} as:

1. 🔄 Checkpoint — I'll continue later
2. ✅ Done — this work is finished
3. 🅿️ Parking — switching to something else

Which one? (1/2/3)

Where {stream-name} = existing stream name, or auto-generated from project + conversation topic if no streams exist.

2+ existing streams → 2 questions. First ask stream:

📝 Save to which session?

1. 📍 {stream-1} (existing)
2. 📍 {stream-2} (existing)
3. ✨ New session

Which one?

Then ask save type (same 3-choice menu as above).

New stream confirmation (when auto-generated or user picks "New"):

📝 Stream name: {suggested-name}
OK? (yes / type a different name)

Mapping:

• 1 → status: building
• 2 → status: done
• 3 → status: parked

Stream naming: pattern ^[a-zA-Z0-9_-]{1,50}$. Reserved: default → CONTEXT-llm.md, baseline.

Phase 3: Drift Check (ref/ vs wip/)

Skip if no ref or wip folder exists in the current project.

Folder aliases: ref/ or reference/ (whichever exists). wip/ or work-in-progress/ (whichever exists). Use the actual folder name found for all subsequent operations.

Quick scan — must complete in < 5 seconds, no deep content analysis.

Step 1: Collect ref keys (parallel Glob + Read)

• Glob {ref|reference}/*.md — get filenames
• For each ref file, Read first 5 lines to extract: filename, version suffix (v1/v2), H1 title

Step 2: Scan wip for stale refs (parallel Grep per wip file)

• Grep each {wip|work-in-progress}/*.md for ref/ or reference/ string matches
• Detect:
  • Stale links: wip references ref/X-v1.md but ref/X-v2.md exists
  • Dead links: wip references ref/Y.md that doesn't exist
  • Version-locked terms: extract key terms from CLAUDE.md ## Identity or ## Decisions section (persona name, positioning label, locked values), grep wip for contradicting old terms
  • Mature wip: wip files significantly newer than corresponding ref, with [Proposition] markers removed, versioned headers, or dated validations — may be ready to promote to ref

Step 3: Report (only if issues found)

⚠️ Drift detected:
  - wip/{file}.md links ref/{old}-v1.md → v2 exists
  - wip/{file}.md says "{old term}" → ref says "{new term}"
  - wip/{file}.md links ref/{gone}.md → file missing
  - 🆙 wip/{file}.md looks mature — may be ready to promote to ref

🔧 Want me to fix these? (yes/no)

If user says yes → invoke /sync-ref-wip skill with the detected file pairs. It handles direction detection, diff display, and user approval.

If user says no or skips → continue to Phase 4. Save drift warnings in CONTEXT file under ## Session > unexpected: for future reference.

Phase 4: Synthesize & Write

From conversation (last 15-20 messages), synthesize:

1. Next — 3 tasks (IMPORTANT: use heading "Next" — Claude Code compaction grep-matches next/todo/pending keywords for survival priority)
2. Session — progression, decisions, thinking, unexpected (780 tokens max)
3. Hot Files — max 10 discussed/edited
4. Focus & Goal — 1-2 sentence focus + goal

Write CONTEXT file using template from reference.md.

Filename: default → CONTEXT-llm.md, otherwise CONTEXT-{stream}-llm.md

Phase 5: Auto-Archive

If status is done or parked:

Bash: mkdir -p done && mv CONTEXT-{stream}-llm.md done/

Exceptions — do NOT move:

• CONTEXT-llm.md (default stream)
• CONTEXT-baseline-llm.md
• If user explicitly says "keep here"

Phase 6: Confirm

💾 Saved: CONTEXT-{stream}-llm.md
📊 Status: {emoji} {status}
📍 Focus: {1-line focus}
📋 Next tasks:
  - {task 1}
  - {task 2}
  - {task 3}

If archived: 📦 Archived to done/ (status: {status})

Rules

1. Never ask for stream name as free text first — always offer existing streams or auto-suggest
2. 3 choices max per menu — don't overwhelm
3. Auto-detect over ask — project name, stream name, status should be inferred when possible
4. Same CONTEXT file format as retrospect:save-context — files must be readable by both /load-work and retrospect:load-context
5. Client-agnostic output — works in terminal, Cowork, Telegram
6. Token budget — CONTEXT file stays under 1500 tokens
7. No shell scripts — pure tool calls only (Glob, Read, Write, Edit). Only Bash for mkdir -p done && mv