Conversation Pruner

What it does

The Conversation Pruner turns a long chat history and a token budget into an explicit pruning plan: every message marked keep, summarize, or drop. System messages, pinned turns, the opening request, and the current message are never pruned; recent turns are kept newest-first until the budget fills; the squeezed middle folds into one running summary when the compressed form fits, and drops when it doesn’t. Runs 100% client-side.

How to use it

1. Paste your conversation as a JSON array — one object per message: { "role": "user", "tokens": 120 }. Add "pinned": true to protect a turn.
2. Set the history budget: your context window minus what you reserve for the reply.
3. Read the plan: the narrative line, any warnings, and the per-message verdict list. Copy the plan into your runbook or client code.

Examples

A 1,090-token chat against a 1,000-token budget:

#0 system    KEEP        220   (system — protected)
#1 user      KEEP        100   (opening request — protected)
#2 assistant SUMMARIZE   310 ┐
#3 user      SUMMARIZE   130 ┘ folded into a 90-token summary
#4 assistant KEEP        260   (newest turn that fits)
#5 user      KEEP         88   (current request — protected)

Projected: 880 tokens — fits. Against a 700-token budget the same fold group no longer fits compressed, so turns #2–#3 drop instead and the plan still lands at 670.

Good to know

• The summary cost model is explicit: 60 fixed tokens + 10% of the folded content — a 400-token middle becomes a ~100-token summary. Swap in your own numbers once you know your model’s summary behavior.
• “Fits” counts only what the plan keeps plus an applied summary; a rejected summary costs nothing.
• Token counts per message are yours to supply (prompt + framing). The Token Estimator can produce them from raw text.
• Runs 100% client-side. Your transcript never leaves the browser.
• Related: Token Estimator, Context Window Planner, Rate Limit Planner.