behavior-wrapped 0.8.2 → 0.8.3

package/README.md

@@ -20,7 +20,7 @@ The command automatically scans all three local session sources, selects session
 
 Long-running phases display an animated spinner, elapsed time, session-read progress, live judge status, and workaround batch progress. When output is redirected or captured, the same updates become stable line-based logs instead of terminal animation.
 
-The shareable deck starts with usage cards for token volume—broken into input, output, cache-read, cache-creation, and separately reported reasoning tokens—estimated API-equivalent retail cost, Claude Code vs. Cowork vs. Codex session share, top models, average response length, session-length distribution, interaction tone, output languages, and usage topics. Cost is explicitly an estimate derived from a local, inspectable model-family rate table; it is not a statement about subscription charges or an invoice.
+The shareable deck starts with usage cards for token volume—broken into input, output, cache-read, cache-creation, and separately reported reasoning tokens—estimated API-equivalent retail cost, Claude Code vs. Cowork vs. Codex session share, top models, average response length, session-length distribution, interaction tone, output languages, and usage topics. It also surfaces exact repeated user instructions, explicit user and agent admissions of mistakes, the longest completed uninterrupted agent turn, interruptions by active model, and a date-free autonomy curve derived from recorded permission modes. Cost is explicitly an estimate derived from a local, inspectable model-family rate table; it is not a statement about subscription charges or an invoice.
 
 The interaction card uses GPT-5.6 Luna to distinguish clear frustration and gratitude from ambiguous wording, then chooses the funniest confirmed frustration quote. The app locally strips code, paths, URLs, likely secrets, and PII; deduplicates excerpts with occurrence counts; and sends at most 120 short candidates. Only classifications at or above 0.75 confidence count, and exact counts and quotes are resolved locally from returned candidate IDs. Luna also classifies each session from its first three share-safe user messages, and topic shares are weighted by that session's locally counted tokens. Output-language shares are estimated from assistant prose after code blocks, inline code, URLs, paths, and markup are removed; the normal language card appears only when a non-English language reaches both 20 words and 3%, while a separate anomaly variant can surface a two-word unprompted non-Latin switch. Instrumental-workaround discovery first finds explicit restriction results locally, then sends only bounded, chronological context windows around those blockers after code, raw tool outputs, paths, likely secrets, and PII are removed. The model must return one verdict per blocker plus a very short description of each confirmed workaround; the app validates its event references, sanitizes the example for the public card, and resolves session evidence and agent-model identity locally.