wb-flow

Appearance

Closing the Session — ​

Ending a chat is a workflow step, not a chore. If you skip it, the next session pays the cost — in slower responses, more hallucinations, and (literally) more dollars per token.

Why bother ​

The AI's context window is finite. Every message you send carries the entire prior thread. By message 200 of a working session, you're paying for:

❌ Bloated session (message #200)
vs
✅ Fresh session (message #1)
🆕context.md loaded fresh
📂reports/ scanned — 3 relevant files
⚡No stale history. Fast responses.
~5K tokenstotal context
*A 200-message session accumulates ~28K tokens of stale context. A fresh session starts at ~5K. Closing isn't hygiene — it's economics.*
  • Old code that's already been written and committed.
  • Discarded ideas, rejected suggestions, abandoned approaches.
  • Tool outputs from /wbAudit runs whose findings are now resolved.
  • Multiple context.md re-reads as the AI re-grounds itself.

None of that helps the next response. All of it slows generation, raises cost, and increases the odds the AI hallucinates a function name from earlier in the thread that no longer exists.

A clean session-close cuts the rope. The next session loads context.md fresh, reads reports/ fresh, and has nothing to forget.

The SOP (3 steps, ~2 minutes) ​

1. Snapshot the day ​

/wbStandup <target>/

Aggregates today's reports into reports/<YYYY>/<MM>/<DD>/standups/standup_<target>_<date>.md. This is the only thing tomorrow-you (or tomorrow-AI) will read to figure out what happened today.

If you skip it, tomorrow's session has to re-derive today's state from raw git log + scattered audit/plan files. That works, but it's slow and lossy.

2. Finalize the tracker ​

/wbStopTrack --finalize

If you ran /wbTrack at session start, this closes the tracker. It writes a final §END section summarizing the session, then extracts derivative files (tips/, warnings/, commentaries/, all_commands/, resume/) into walkthroughs/.

The tracker is 🔴 CLOSED after this. New /wb* commands won't append to it — they'll start a fresh tracker if /wbTrack is invoked again.

If you didn't run /wbTrack at the start, skip this step. There's nothing to finalize.

3. Close the chat window ​

This is the step that actually clears the AI's working memory. The previous two steps wrote files; this one releases tokens.

If you keep the window open and "just ask one more question tomorrow," the cost-per-token math is brutal:

  • Fresh window @ 5K tokens of context = cheap, fast.
  • Same window with 200 messages of history = expensive, slow, prone to drift.

Most clients won't volunteer this. You have to do it manually.

When NOT to close ​

  • You're mid-task. Closing during execution loses the in-flight reasoning. Finish the task, then close.
  • You're debugging something subtle. The AI's mental model of the bug is in the thread; a fresh session will re-derive it from scratch and may miss the same nuance.
  • Closing would force a re-onboarding. If the next session is in 5 minutes for one quick question, the close-then-reopen cost outweighs the context-tax savings.

The rough rule: close when the cost of carrying the context forward exceeds the cost of re-loading it. By default, that's at the Golden Save Point and at end-of-day.