/ration — Emergency Token Rationing

Haiku can handle: simple edits, running commands, file reads, formatting, test runs. Sonnet can handle: everything except complex architecture and security review.

Kill unnecessary context

• If conversation is long: "Run /clear and I'll re-read only what's needed"
• If there are MCP servers not being used: suggest disabling them temporarily

Step 3: Triage Remaining Work

List the user's remaining tasks and categorize:

MUST DO (worth spending tokens on):
  1. [task] — ~XXX tokens, best model: [model]
  2. [task] — ~XXX tokens, best model: [model]

CAN EXPORT (do on a cheaper platform):
  3. [task] — export to: [platform], prompt below
  4. [task] — export to: [platform], prompt below

SKIP (not worth the tokens right now):
  5. [task] — do this tomorrow when budget resets

Step 4: Export Conversation for Continuation

When tasks should move to another platform, generate a portable context bundle that the user can paste into another AI chat.

Check Available Platforms

Before showing platform-specific instructions, determine what the user has access to:

1. Check if ~/.stingy/config.json exists and has a platforms key — if so, use that list.
2. If no config exists, ask the user:

Which AI platforms do you have access to? (ChatGPT, Gemini, Grok, or just Claude?)

If the user only has Claude, skip the export and focus on model downgrade (Step 2) and session efficiency (Step 5) instead. Tell them:

No export needed — let's maximize what we can do here. Switch to Haiku for simple tasks and save the complex ones for tomorrow's fresh budget.

If the user has at least one other platform, proceed with the export below — but only show instructions for the platforms they confirmed.

Export Format

Write a file to /tmp/stingy-export-{timestamp}.md:

# Context Transfer — [Project Name]

## What I'm Working On
[1-2 sentence summary of the project and current task]

## Key Files
[List only the files relevant to the remaining task, with brief descriptions]

## Current State
[What's been done, what's left, any decisions already made]

## Remaining Task
[Exact description of what needs to be done]

## Key Constraints
[Any rules, patterns, or requirements the AI needs to know]

## File Contents
[Include ONLY the specific file contents needed — not entire files.
 Use the smallest excerpt that gives enough context.]

Then tell the user (showing ONLY the platforms they confirmed access to):

Export saved to: /tmp/stingy-export-XXXXX.md

Include only the relevant instructions from:

• ChatGPT: Open chat.openai.com → Paste the export → Say "Continue from where I left off"
• Gemini: Open aistudio.google.com → Paste the export (Gemini handles huge context well)
• Grok: Open grok.x.ai → Paste the export

Copied to clipboard:

EXPORT_FILE=$(ls -t /tmp/stingy-export-*.md 2>/dev/null | head -1)
if [ -n "$EXPORT_FILE" ]; then
  if command -v pbcopy >/dev/null 2>&1; then
    cat "$EXPORT_FILE" | pbcopy
  elif command -v xclip >/dev/null 2>&1; then
    cat "$EXPORT_FILE" | xclip -selection clipboard
  elif command -v xsel >/dev/null 2>&1; then
    cat "$EXPORT_FILE" | xsel --clipboard --input
  else
    echo "Clipboard tool not found. Copy manually from: $EXPORT_FILE"
  fi
  echo "Export copied to clipboard — paste into any AI chat"
fi

Step 5: Ultra-Efficient Mode Rules

For the rest of this session, enforce these rules:

1. No agent spawns. Do everything inline. Agents duplicate the full context.
2. No full file reads. Use Grep to find the line, then Read with offset/limit.
3. No exploratory searches. Only search if you know what you're looking for.
4. Minimal output. Don't explain what you're doing. Just do it.
5. Batch operations. Combine multiple edits into one tool call where possible.
6. Skip formatting. No tables, no headers, no decoration. Raw information only.
7. No web searches. Use existing knowledge. Web searches are expensive.

Tell the user:

Rationing mode active. I'll be extremely terse and efficient. Estimated remaining capacity: ~[X] tool calls at current model. Priority tasks: [list the MUST DO items]

Step 6: Save State for Tomorrow

If the user can't finish everything today, help them save state.

Ask before touching git

Do not run any git commands without explicit permission. Use AskUserQuestion first:

Would you like me to save your work-in-progress? I can either:

1. Git stash — temporarily shelves your changes (undo with git stash pop)
2. WIP commit — creates a commit you can amend later
3. Skip — leave everything as-is, just write pickup notes

This is optional. Your files on disk won't be affected if you skip.

Only proceed with the option the user picks:

If stash:

# Save uncommitted work
git stash -m "stingy-ration: WIP $(date +%Y-%m-%d)"
echo "Work stashed. Tomorrow: git stash pop"

If WIP commit (stage only tracked files — never use git add -A):

git add -u && git commit -m "WIP: [description of what's in progress]"

If skip: Go straight to pickup notes below.

Pickup notes

Write a brief note for tomorrow's session regardless of the git choice:

cat > /tmp/stingy-pickup-notes.md << 'EOF'
# Pickup Notes — [date]

## What was in progress
[brief description]

## What's done
[list]

## What's left
[list with priority order]

## Key context
[anything the next session needs to know]
EOF
echo "Pickup notes saved to /tmp/stingy-pickup-notes.md"

Rules

• Be TERSE. Every word costs tokens.
• Prioritize ruthlessly. Not everything needs to get done today.
• The export should be copy-paste ready — no "fill in the blanks"
• Include actual file contents in exports, not just file names
• Clipboard copy is mandatory — don't make the user open the file manually
• If the user is on a subscription (not API), the "cost" is daily rate limits, not dollars. Frame advice around "X tool calls remaining" not "$X remaining"