Problem

When a session's accumulated context exceeds the model's window, some providers (notably OpenRouter/OpenAI) return:

{"stopReason": "length", "output": 0, "content": []}

instead of an explicit context overflow error. This causes:

1. Silent empty replies — the user gets no response
2. Session permanently stuck — every subsequent message gets the same empty response
3. No auto-compaction — the existing overflow recovery only triggers on explicit error text

The workaround is manually deleting the session file, which most users won't know to do.

Fixes #14064

Root Cause

In pi-embedded-runner/run.ts, the overflow detection (line ~473) only checks:

• promptError containing context overflow keywords
• lastAssistant.stopReason === "error" with overflow-like error text

When stopReason === "length" with output: 0, neither condition fires — the run is treated as a successful (but empty) completion.

Fix

Add a isSilentOverflow detection before the existing overflow check:

const isSilentOverflow =
  !aborted &&
  !promptError &&
  lastAssistant?.stopReason === "length" &&
  lastAssistant.usage?.output === 0;

When detected, this feeds into the existing contextOverflowError flow, triggering auto-compaction → retry — the same recovery path used for explicit overflow errors.

Testing

All 75 existing pi-embedded-runner tests pass. The change is additive — it only catches a case that previously fell through silently.

Greptile Overview

Greptile Summary

Added detection for silent context overflow when providers return stopReason: "length" with zero output tokens instead of explicit error messages. The fix properly triggers auto-compaction by feeding into the existing contextOverflowError flow, resolving the issue where sessions would become permanently stuck with empty responses.

• prevented silent context overflow by detecting stopReason === "length" with output === 0
• triggers existing auto-compaction recovery flow for silent overflows
• minor style inconsistency with type casting that could be improved

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The change is well-targeted and additive, catching a previously unhandled edge case without modifying existing logic paths. All 75 existing tests pass, and the fix properly integrates with the existing overflow recovery flow. The only issue is a minor style inconsistency in type casting.
• No files require special attention