Merged and never uploaded: npm refused the token the release publishes with (#662). What this section lists first reached npm in 2.7.11.
Fixed#
planoffered the Batch API to calls that cannot take it. Two kinds, both measured. Twenty calls that had already run on the Batch API, a $0.07 bill at the batch rate, were offered "$0.0350 projected" for batching them: half of a price that was already the half. And a converted Claude Code session was told to "route and batch" every turn of an agent loop, $9.13 projected against the example projects' $12.70 bill, on the assumption that "these calls can wait for a batch window". The assumption was printed, so the plan did not lie; it asked a question the log could have answered, and counted the answer into the total before it came. A record can now say"interactive": true, meaning something waited on the answer, andfrom-claude-code,from-codexandfrom-gemini-clisay it on every line. The batch lever leaves out those calls and the ones already billed at the batch rate, in its own figure and in the combined route-and-batch one, each at its own model's rate. The same projects now read "Route", $7.34 projected, and the plan ends by saying how much input and output it left out and why (batchExcludedUsdin the document). A route that combines nothing no longer says "combined with batching". A log that marks nothing is read exactly as before. The profile's breakdown gainsbatchExcludedInputTokens,batchExcludedOutputTokensandbatchExcludedUsd, andfieldCoveragecountsinteractive; a document from before the fields existed reads them as zero. Thirteen faults planted in the lever, the parser, the roll-up, the plan and its renderers were each caught.- The recovery for a release that lost its tag was one step short.
docs/releasing.mdsaid that a person pushes the tag and the run it starts creates the release. That run verifies the tagged tree first, and a tree that is days old can fail a test that reads the calendar: 2.7.2's has since 2026-11-16, on a promotional price in its last 45 days. The page now says the failed run is expected when the recovery comes late, and names the second step,backfill-releaseswith the version and the commit, which creates the release on the tag once it exists. It also says why the tag itself stays a person's step.