Summary

• cap V2 provider turns at 32,000 output tokens, matching the existing runtime behavior
• keep model capability limits intact while giving compaction and the provider an explicit request budget
• cover models whose advertised output limit consumes their entire context window

Root cause

V2 left generation.maxTokens unset. Grok 4.5 advertises a 500,000-token context and 500,000-token output limit, so the provider defaulted the request to 500,000 output tokens. Even a fresh hi session then requested roughly 500,333 total tokens and failed immediately. Compaction also treated the full 500,000 tokens as reserved output.

Source report: https://x.com/dkfiander/status/2075682541026128183

Test plan

• bun test test/session-runner.test.ts --filter 'caps output when the catalog output limit consumes the full context window'
• bun test test/session-compaction.test.ts test/session-runner-model.test.ts
• bun typecheck from packages/core

The repository-wide pre-push typecheck was attempted but the 3.2 GB host exhausted process resources (EAGAIN) and Bun crashed while Turbo ran 36 packages concurrently. The affected package typecheck passed independently.