What is context collapse?

It is the point where the accumulated conversation plus the reserved response no longer fit in the model's context window. Past that point the provider truncates or rejects the request, so older turns silently drop and the model loses track of earlier context.

How are tokens estimated here?

The detector uses a character-based heuristic of roughly four characters per token for each turn. It is an estimate suited to finding the approximate overflow point, not an exact tokenizer count.

How do I avoid hitting the limit?

Common strategies are summarizing older turns into a compact recap, dropping the least relevant messages, or moving long reference material into a retrieval step instead of the prompt. The tool shows how many tokens you need to reclaim.

Is my conversation uploaded?

No. All counting happens locally in your browser. Nothing you paste is sent to a server, stored, or logged.

Context Collapse Risk Detector

Context collapse risk detector

Long conversations fail quietly: once the running history plus your next reply exceed the model’s window, the oldest turns get truncated and the model starts forgetting. This detector paste-and-counts a conversation turn by turn, shows the running token total, and flags the exact turn where you cross your usable context limit — so you can summarize or trim before quality drops.

How it works

You paste the conversation with one turn per line (or blank-line separated). Each turn’s characters are converted to an estimated token count (about four characters per token), and the tool keeps a running total. From your chosen model context window it subtracts the tokens you reserve for the response to get the usable budget. The first turn whose running total exceeds that budget is marked as the collapse point, and the tool reports how many tokens you must reclaim to fit.

Tips and notes

Reserve realistic output room. If your replies are long, reserve more — forgetting this is a common cause of mid-conversation truncation.
Summarize, do not just drop. Replacing old turns with a compact recap keeps continuity while reclaiming tokens; pair this with the summarization cost planner.
Move reference material out. Long documents belong in a retrieval step, not pasted into every turn.
Estimates only. For exact counts, verify with your model’s tokenizer; this is for spotting the approximate overflow point fast.