bullswarm 0.38.8 → 0.38.9

package/CHANGELOG.md

@@ -2,6 +2,13 @@
 
 ## Unreleased
 
+## 0.38.9 — automatic pricing catches up when a rate card arrives; codex long-context pricing
+
+- pricing: the automatic pricing pass now prices attempts whose token counts were measured but whose model had no rate card yet. It only revisited unknown or byte-estimated attempts, so a model that gained its price in a release (Sonnet 5.5 in 0.38.8, the gpt-6 rows in 0.37.3) stayed `unpriced` until someone ran `reprice --apply`. Such attempts are now priced from their own recorded counts at today's card (no transcript lookup, no ledger tries), and a new package version or a changed connector override re-reads every run once. On the live home this prices 92 attempts (about $63).
+- pricing: codex long context. OpenAI charges more for a request with over 272K input tokens; the codex rows for gpt-6-astra, gpt-6.1-sol, gpt-6-luna and gpt-5.6-sol now carry those rates (`longContext`, read 2026-10-09; the page lists none for gpt-6-sol, gpt-5.6-terra or gpt-5.6-luna). Codex's transcript reader records each request from its `last_token_usage` and tags it short or long, so a long request is priced at the long rates — but only when the requests add up exactly to the window's total; otherwise the one combined record (short rates) is kept. Checked on a real gpt-5.6-sol session: 4 of 111 requests were long, adding exactly the hand-computed $0.50 ($11.84 → $12.34). An attempt measured only from codex's stream total stays at the short rates.
+- pricing: Claude Sonnet 5.5 cache reads are $0.10 per million (0.05x input, the pricing page's footnote 2, read 2026-10-09); 0.38.8 carried $0.20. Claude Code's own reported cost matched $0.20 on 41 of 45 attempts, so its bundled table differs from the published page; the dashboard's API price follows the page.
+- pricing: command-code prices gpt-6-luna ($0.10 / $0.01 / $0.125 / $0.50) and claude-sonnet-5-5 ($2 / $0.20 / $2.50 / $10) at commandcode.ai's published rates (input / cache read / cache write / output); 20 command-code attempts on those models had no price.
+- tests: the published-price guard covers codex (short and long context), grok and command-code, not only Claude.
 ## 0.38.8 — dashboard money says only what its source supports; proof counts on v3 runs
 
 - dashboard: money says only as much as its source supports, in fewer words. An API price built from the provider's own token counts — reported, or summed from its transcript — prints plainly (`$4.01`, was `≈$4.01`): exact arithmetic at the dated rate card, checked on 2026-10-02 against all 251 priced codex attempts and 163 of 167 Claude attempts' own reported cost. `~` stays on prices estimated from text size. A total that leaves attempts out reads `≥$9.52` everywhere (was `at least $9.52 api · 6 unmeasured` or `≈`); the Runs day header and the Run page's spend rule (`25 of 26 attempts priced`) keep the counts. Every plan amount is a share of an account-wide meter, so it always carries `≈`, and a meter that did not move shows `—` (was `≈ $0.000`).