v0.64.1

Fixed

• The chat bot gets what 0.64 said a person waiting gets.

  • What was wrong: 0.64.0 turned the limits into warnings, and that reached chimera chat and the desktop app only. The Discord bot (serve --discord, and the same bot started from the app's Messaging switch) kept the six-step wall, ended a long thread on a context overflow, and could not show a warning, because the gateway called send, which takes no callbacks. A reply cut short by a limit read like a finished one.
  • Now: the bot runs with the same five settings as the terminal (one function, attended, instead of five keywords copied to four places and missed in three). Its replies carry the turn's warnings and the reason a reply was cut short, one ⚠ line each, under the answer. The HTTP /chat reply stays the answer alone, and the HTTP gateway keeps the step wall, because the webhook jobs run through it with nobody waiting.

• The terminal compacts at the context its model was measured to read.

  • What was wrong: run without --model, the context budget was sized for an empty model name. The default model, measured to read 255k, compacted at about 51k, and /model kept the old model's budget.
  • Now: the budget is sized for the model the run calls (the default when none is given) and follows a switch.

• Warnings reach every terminal command, and a finished background job reaches every chat.

  • --max-usd wrapped the agent in a turn whose signature had no on_notice, so every warning of a conversation with a ceiling was dropped. chimera tui and chimera agent never asked for them.
  • A shell command that outlives its timeout becomes a job on every surface, and only the Code screen was told when one ended. chat, assist, tui and the bot now hear it on the next turn, as the Code screen does.

• A slow Intel dmg no longer holds back the update of every platform (#732, in this tag; listed under Unreleased in the CHANGELOG).

  • What was wrong: on the macos-15-intel runner, Tauri's bundle_dmg.sh went from 1.5 minutes (v0.61.2) to failing after 10 (v0.64.0, twice), with the same 272 MB dmg. The failed job skipped latest.json, so v0.64.0 was never offered by the in-app updater nor marked Latest.
  • Now: the Intel job bundles the app and its updater only, and the Intel dmg is built by hdiutil in its own step (25 s in the dry run), which may fail without taking the release down.