[TheGreatAxios]

Summary

• Compaction occupancy is 60% of the context window minus the reserved completion budget, so a prompt that fits the advertised window cannot still overflow once maxTokens is added.
• A user message or tool follow-up whose usage or live estimate plus that reservation meets the window compacts before the next infer.
• An inference error whose text is a context overflow compact-and-retries even when the provider did not tag the category context_overflow. The recovery cap stays at 2.
• A server sentence that names maximum context length of N pins N for that model for the rest of the process, ahead of the advertised window, without writing settings.

A custom endpoint can advertise a larger window than it enforces. The pin is what makes the next threshold use the server's limit after the first rejection.

Verification

• bun test src/agent/compaction.test.ts src/provider/context-window-thresholds.test.ts src/provider/context-window.test.ts — 100 pass
• pre-commit tsc --noEmit and bun build pass
• full bun run test was not run

Fixes CL-10833

[linear-code]

CL-10833

[corbits-triage]

Triage: Pass

No changes needed from you. A maintainer will review this pull request.

[TheGreatAxios]

Superseded by the v0.3.41 branch, which includes this compaction commit alongside the Option-chord work.