5. Why not summaries?

Summarization-based compaction solves the immediate problem — the window is full — and creates two others.

1. The paraphrase you must trust

A summary drops the exact flag, the line number, the error text — and leaves the model no handle to go get them back. What survives in context is a rewrite it must believe. When the task turns on a detail, there is no “reload the original”: it was discarded, and the durable record is invisible to the agent. This isn’t a tuning problem; it is the mechanism.

dsh-chapters inverts it: the archive is verbatim text written from the session log (a file copy), and the index is a Table of Contents whose every line is a path back to that text. The guarantee is not “the chapter contains the whole history” — it is “every byte remains retrievable”, with retrieval a first-class agent action.

2. The summarization tax

Producing a summary means prefilling the whole history into one auxiliary call — and repeating that every time the window fills:

History	30 tok/s	100 tok/s	300 tok/s
32K	18 min	5 min	2 min
100K	55 min	17 min	6 min

On local hardware that is not “slow”, it is stop-working slow. The index this plugin carries instead cost, measured, ~102 tokens where the equivalent transcript ran ~2,476 — and produced with zero inference. The field number to anchor on: a real 10-session production run archived 680K tokens across 21 compactions at 0 model tokens; the stock equivalent pays that 680K as prefill, repeatedly, plus the generation.

And the tax you don’t want to hear either

Honesty is a feature here: compaction is not cheaper for prompt caching. Replacing a span mid-conversation invalidates everything after it — every compaction scheme’s cold tail (measured on llama.cpp, it re-prefills). Continuation pays one cold prefill (header + TOC + note) like any post-compaction request does. What continuation buys instead: verbatim reload, an untouched cacheable parent, reversibility, and branchable history. If a vendor pitches compaction as a caching win, they are selling you the one axis this design concedes by name.

Next: 6. The four invariants — the rules that make the above trustworthy rather than merely convenient.