humanish 0.111.0 → 0.111.2

package/docs/contracts/schemas.md

@@ -978,14 +978,25 @@ Core-owned fields:
   kept it through `previous_response_id`) or `explicit_context` (the server kept
   none). `explicitReason` says why: `configured` (`zeroDataRetention`) or
   `zdr_rejection` (the organization rejected server-side state; `switchedAt` and
-  `switchedAtRequest` say when). In `explicit_context` each request carries the
+  `switchedAtRequest` say when). With `zdr_rejection`, `rejection` names the
+  answer that made the switch: `zero_data_retention` (the error named it),
+  `previous_response` (a 400 that refused or could not find the
+  `previous_response_id`) or `stored_item` (a 404 for an item the server never
+  kept, "Item with id '...' not found"). In `explicit_context` each request carries the
   conversation from the client, with `store: false` and the reasoning returned
   in encrypted form. It carries the opening message, every exchange of a reply
   with the outputs that answered it, and the latest reply. Past an estimated
-  64,000 input tokens the opening screenshot is dropped first, then the oldest
-  exchanges become lines of a note, sent as an assistant message, that keeps
-  their reasoning summaries, messages and actions as text. The two newest
-  exchanges are always carried whole, even when they alone pass the budget. `summarizedTurns` counts the earlier exchanges the latest
+  64,000 input tokens it is cut down to 32,000 in one step: the opening
+  screenshot is dropped first, then the oldest exchanges become lines of a note,
+  sent as an assistant message, that keeps their reasoning summaries, messages
+  and actions as text, with any text that came back with their screenshots.
+  Past 32,000 characters the note keeps its first four lines and its newest
+  ones, and names the turns between them as not listed.
+  The two newest exchanges are always carried whole, even when they alone pass
+  the budget. Between cuts each request starts with the whole previous request,
+  so the provider's prompt cache can serve all of it; the request after a cut is
+  billed as new input. `tokenUsage.turns` has each request's `cachedInput` and
+  `cacheWriteInput`. `summarizedTurns` counts the earlier exchanges the latest
   `explicit_context` request no longer carried whole: lines of the note, or
   only counted once the note reached its cap.
   `requests[]` has one entry per participant request: its `mode` and, in

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "humanish",
-  "version": "0.111.0",
+  "version": "0.111.2",
   "description": "Synthetic user research for apps, CLIs, and agent-facing product flows.",
   "keywords": [
     "cli",