# Changelog One line per released version. Versions before 1.0.2 were development-only and were never published to npm. ## 1.1.0 — 2026-09-13 Rebuilt on a session projection of this plugin's own, with the wording and the scope of every figure made explicit. The browser half no longer reads the conversation at all. - Reworked into a session-level ledger: the shares now cover only the turns that actually reasoned, so switching to a model that does not think no longer makes the numbers slide down. - Added coverage (`4 / 6 turns`, shown only when part of the conversation did not reason) and a line saying how much of the total was worked out from the thinking text rather than reported by the provider. - Removed the per-reply figure: the DSH reply panel already reports that turn's reasoning tokens, so it was a duplicate. The plugin now owns one readout. - Added a turn-by-turn view behind the details: one row per turn with its thinking tokens, the turn's own output tokens, where that number came from, the model that ran it, and the share of its own output; filters for the three sources, and a jump button on every row. - Fixed the model column: the shipped per-turn usage data exists for only a few turns of a real log, so the column now comes from the plugin's own fold, which records the provider and model each settlement itself. - **Rebuilt on a session projection.** The host half folds the session's whole event log once (`assistant/message` settlements, six numbers per turn, model names in a dictionary) and publishes it as the `thinkingStats` projection; the browser half reads that one value. The figures are therefore whole-session by construction — paging and compaction cannot change them — and identical after a reload. - **All loading is gone**, along with the per-turn cache, the `~`/paging controls for it and the browser's local storage. There is nothing to load: the host already folded the log. The client no longer needs the session or conversation service, a conversation window, the node store, or the trajectory fold. - The overview's scope line is now computed from what the fold actually holds rather than asserted: the whole session with its turn count, or the turns that produced no reply, or this session's own turns when it was forked, or totals only when the host serves no projection. Whenever the count is not complete, a `?` opens the same kind of message as the estimate's, explaining that reason — and the per-turn list's own cap carries one too. - Measured on a real 88-turn session: 8 988 events folded in ~5 ms total (0.6 µs per event), 3.0 KB of persisted state (36 B per turn) and a 2.4 KB published view — smaller than DSH's own `turnOutline` record for the same session. - The overview carries no per-model split: the table's model column answers that question per turn, and an aggregate would have needed extra fold state for a duplicate. - The per-turn table is unchanged otherwise (columns, source filters, 100 rows a page, jump button), and an estimated row now marks both its thinking and its share with `~` and colours them as an estimate. - The fold opens a turn on its `turn/start` boundary rather than on its first settlement, so a turn that errored or was interrupted — and therefore never produced a message — is still listed, with no figures, instead of looking like a turn the conversation lost. It also counts how many turns actually settled, which is what the scope line measures its gaps against. - Worked-out figures are now priced **per script** instead of at a flat four characters per token: CJK/Kana/Hangul and fullwidth forms at 1 character per token, Cyrillic, Greek, Arabic, Hebrew, Armenian, the Indic scripts, Thai and Georgian at 2.5, Latin letters, digits and punctuation at 4. Two `replace` passes per reasoning block, measured at 1.7 ms over a session's 2.3M characters of reasoning against a 5 ms fold. The fold's `stateVersion` is bumped to 2 so stored checkpoints are rebuilt rather than mixed. - The estimate's own line states `Estimated ≈ 5.1K tok` rather than a bare `~5.1K tok`, so the figure reads as the admission it is. - A host without the projection falls back to the official whole-session token projection and says so, rather than inventing per-turn figures. - The plugin now has a unit spec for the fold (`test/projection.mjs`) which, given `DSH_TEST_LOG`, re-folds a real recorded session and checks every total against an independent pass over the same events. - The table renders **one page of 100 turns**, with Newer/Older and a page count; only the page in view is laid out. - Fixed the per-turn table's header in English: `Thinking source` was cut to `Thinking sou…` and `Of output` wrapped onto a second line, because the columns were sized for the shorter Chinese labels. The two columns are now as wide as the longest label in either language, and no header cell wraps. - Fixed the `?` on the overview's footnotes: inline after a sentence it needed its own space instead of sitting on the full stop it closes. - Fixed the estimate's explanation card running off the bottom of the window. It is a paragraph of prose, so its height is now measured and the card is clamped inside the viewport, rather than placed as if it were a fixed height that was shorter than the real card; a window too short for it scrolls the card instead of cutting it. The placement also stops recomputing a fresh position on every render. - Follows the interface language, in the two languages DSH ships: only a `zh`-prefixed `` reads Chinese, and every other value — a third language, or the empty attribute a document has before the locale service publishes — reads English, which is DSH's own fallback. Switching language updates the readout without a reload. - Fixed dismissal: an outside click or Escape closed nothing, because a dismissal set the panel's view to a value the render test did not read as closed — only the logo could close it. Both levels now close from anywhere, and the listener runs in the capture phase so a component that stops pointer propagation cannot swallow it. - Every figure derived from the text fold is marked `~`: on the pill, on both shares, on the table's own total and on each estimated row. - The estimate's `?` opens a short explanation as its own message beside the panel, instead of a hover tooltip that never appeared for touch or keyboard. - The scope note is the details panel's last line, below the door to the turn-by-turn view, and states what the figures cover. - Rewrote the README in plain language and replaced the screenshot. ## 1.0.8 - README: documented how to install from npm. ## 1.0.7 - The details panel now follows the DSH interface language (Chinese and English), and each row states the total it is a share of. - New screenshot. ## 1.0.6 - Replaced the hover tooltip with a details panel styled like DSH's own. ## 1.0.5 - Rendered the per-reply figure as a matching icon-and-text pill. ## 1.0.4 - Followed the newer DSH reply-tail styling so the figure lined up with the built-in pills. ## 1.0.3 - Renamed the project and package to `dsh-thinking-token-stat` and updated every repository link. - Documented that models which do not expose reasoning show no figure. ## 1.0.2 - First npm release, under the current package name.