Summary

MAX_OVERFLOW_COMPACTION_ATTEMPTS = 3 is a hardcoded const at pi-embedded-Bcz04p2i.js:1955 (verified in 2026.5.7 tarball). When a long-lived agent hits context-overflow, the runtime tries up to 3 LLM-summarizer compactions before bubbling the error. Each attempt is a billable model round-trip. For agents on per-seat ChatGPT Team/Pro/Enterprise quotas, a sustained overflow loop burns the seat very fast.

Real-world impact

A long-lived agent:<bot>:main session running openai-codex/gpt-5.5:

• 7 context-overflow events in 17 hours.
• Each triggered up to 3 compaction calls.
• ~21 billable compaction calls × ~34K tokens each, on top of normal traffic.
• Seat hit 66-hour cooldown.

The 2026.5.4 post-compaction loop guard helps when the agent is repeating the same (tool, args, result) triple, but it does not protect against compaction itself running 3× per overflow on a session where the summary is structurally too large.

Proposed fix

Expose as agents.defaults.compaction.maxOverflowAttempts (default 3, lower to 1 to fail-fast on overflow). Sibling keys (compaction.mode, compaction.reserveTokens, compaction.recentTurnsPreserve) suggest this slot is the natural home.

Related

• [Bug]: Repeated hard resets on same session key despite high reserveTokensFloor; retry loop re-injects bootstrap context #63216 (repeated hard resets on same session key) — adjacent failure mode
• Generalize post-timeout compaction completion reconciliation across session state #45505 (post-timeout compaction reconciliation)

(Re-filed from badlogic/pi#4 — filed in wrong repo originally.)