v6.30.0

Added

• GPT-6.1 Terra (gpt-6.1-terra) with low, medium, high, xhigh, max, and ultra efforts. It has its own model family and prompt cache (and no auto-fallback chain), uses responses-lite, and ships in both config templates with a low default. The catalog now covers 12 bases and 65 variants. (#299, thanks @WarGloom)
• Codex Fast tier. A host serviceTier, the new per-model or global serviceTier option (fast/priority, flex, default), or a -fast model id now sends service_tier: "priority" to models whose Codex catalog entry offers it. Before this, gpt-5.5-fast was only an alias. gpt-6-sol and gpt-6-luna default to Fast as in stock Codex, and serviceTier: "default" opts out. The tier is re-resolved on model fallback. (#300)
• Codex 0.166 request headers: x-codex-routing-hint, session-id/thread-id (sent alongside the legacy session_id/conversation_id), version under the Codex CLI identity, and x-codex-turn-state sticky routing. The turn-state token is replayed within a turn and on same-account retries. It is tracked per session and account, and dropped at the next user message. (#300)

Changed

• ultra now resolves the way Codex does: to the catalog's multi-agent effort, which is xhigh on gpt-6-astra, gpt-6.1-sol, and gpt-6.1-terra, and to max elsewhere. Native mode now applies the same effort clamps as legacy mode, so ultra never reaches the wire. (#299, #300)
• Request shape follows Codex 0.166. No request carries a top-level instructions field; base instructions are the first developer message in input. Responses-lite groups function and custom tools into the functions namespace and skips additional_tools when a turn has no tools. (#300)
• text.verbosity falls back to the Codex catalog default (low; high on Daybreak Red) in both modes. An explicit reasoningSummary: "none"/"off" now omits the field instead of becoming auto. (#300)
• Shipped templates advertise the Codex working context window: 272000, or 500000 for gpt-6.1-terra, instead of 1050000/872000. OpenCode now compacts when Codex would. (#300)

Fixed

• OpenCode V2 requests now carry prompt_cache_key, plus the matching session_id/conversation_id, derived from the session-affinity headers. The host's AI SDK bridge drops promptCacheKey, so tool-result continuations used to miss the prompt cache. Provider-configured affinity headers are honored too. (#296, #297)
• A plain-string input is no longer dropped when instructions or additional_tools are prepended. (#300)

Full Changelog: v6.29.0...v6.30.0