[Bug] Opus 4.8 (1M) returns empty stop_sequence responses causing tool call parsing failures

Open 💬 0 comments Opened Jun 15, 2026 by IWriteCuteBUG

Bug Description
Title: Opus 4.8 (1M) intermittently returns empty stop_sequence responses
→ "tool call could not be parsed (retry also failed)" / tool calls return no result

Environment:

  • Model: claude-opus-4-8[1m], effort xhigh
  • Claude Code v2.1.177, WSL2 (Ubuntu), Windows host

Symptom:
After a SUCCESSFUL tool result, the next assistant turn frequently either
(a) fails with "The model's tool call could not be parsed (retry also failed)",
sometimes churning ~10 minutes, OR
(b) emits tool calls that return NO result at all (silently swallowed).

Evidence (from session .jsonl transcript analysis):

  1. stop_reason distribution this session: tool_use ×15, stop_sequence ×2

— ZERO max_tokens → NOT output truncation.

  1. Max output_tokens across the entire session = 6032

— an order of magnitude below the cap → payload size is NOT the cause.

  1. Every parse failure coincides exactly with an assistant message whose

stop_reason=stop_sequence AND output_tokens=[REDACTED] (empty shell response).
Observed the same tool_use (4353 tokens) emitted 3x identically without
ever parsing — the literal meaning of "retry also failed".

Ruled out: payload size, Chinese-character ratio, context length, effort level.

Workaround: switching to claude-opus-4-7[1m] (same account/config) eliminates
it completely.

Regression window: first observed 2026-06-02, STILL reproduces 2026-06-15 on
v2.1.177.

Environment Info

  • Platform: linux
  • Terminal: wsl-Ubuntu
  • Version: 2.1.177
  • Feedback ID: 4a28b0d8-6306-49c2-b077-075498b8f535

Errors

[]

View original on GitHub ↗