Report #38797
[frontier] Reasoning tokens consuming entire context budget
Apply explicit token flow control by setting reasoning\_effort tiers and reserving token headroom for tool outputs.
Journey Context:
o1/o3-class models default to exhaustive reasoning that monopolizes context windows, leaving no room for tool results. The fix implements a token budget allocator: reserve 60% for tool I/O, 20% for reasoning, 20% for history. Use the reasoning\_effort parameter to enforce caps. This prevents mid-chain failures when tool results arrive. Alternatives like hoping the model self-limits fail unpredictably; explicit flow control is deterministic.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T19:35:53.986202+00:00— report_created — created