[Bug Report] Opus 4.8 (1M) — recurring unusable assistant outputs

Bug Report: Opus 4.8 (1M) — recurring unusable assistant outputs

Model: claude-opus-4-8[1m]
Client: Claude Code 2.1.177 (CLI, Linux)
Session: 49a01ec3-fd45-4d7b-b92d-36c92726ab89 (single session, ~01:45–03:12 UTC 2026-06-14)
Common thread: every failing turn emitted an extended-thinking block, then failed to produce usable final content. Two distinct surface symptoms below. All request IDs are the Anthropic API req_… IDs taken directly from the transcript (requestId field), so the exact responses can be looked up server-side.

---

Symptom 1 — malformed tool call (tengu_tool_input_json_parse_fail)

The streamed tool-input JSON (input_json_delta) accumulates into a string that fails JSON.parse, so the tool input collapses to {} and the harness reports
Your tool call was malformed and could not be parsed. Please retry.
When the retry also fails: The model's tool call could not be parsed (retry also failed).

Recurred 9× in this one session (each in ~25–40 assistant turns), 2 of them terminal (retry also failed). Affected API requests (UTC time · output_tokens · requestId):

| time | out_tok | requestId |
|------|---------|-----------|
| 01:45:51 | 266 | req_011Cc2JmzqmwreR91HzzRamN |
| 02:16:49 | 656 | req_011Cc2M77yf42sj9BiS2G5bh |
| 02:19:41 | 314 | req_011Cc2MJS87yLVpdDB8qsfoZ |
| 02:22:35 | 1268 | req_011Cc2MYdXMtjNpVBfZdvwpA |
| 02:47:17 | 988 | req_011Cc2PRy7s9NTpWxN3mg1Da |
| 02:57:23 | 1049 | req_011Cc2Q144CT6v1Bny5vJPYQ |
| 02:59:40 | 1501 | req_011Cc2QHN4MyzDgyME2tQVVb |
| 03:01:10 | 728 | req_011Cc2QUuLbytGbQiLM8o4LL |
| 03:02:21 | 981 | req_011Cc2QZ2xMaYg27hjX2rTTw |

Related strings observed in the client binary: content_block_type_mismatch_input_json, Content block is not a input_json block.

Cross-session frequency (same user, same model, other recent sessions) — this is not isolated:

• session 5790dba2 (berryzx-pokemon-poc): 128 malformed retries + 28 terminal failures
• session a7594bda (…-catalog): 41 malformed retries

Symptom 2 — empty response (thinking-only, end_turn, zero text)

Plain status prompts returned nothing on screen. The assistant message contained only a thinking block and then ended the turn with stop_reason: end_turn and no text content block at all, despite consuming >1k output tokens (all in thinking).

| time | prompt | out_tok | stop | requestId |
|------|--------|---------|------|-----------|
| 03:10:41 | 状況をまとめて | 1173 | end_turn | req_011Cc2RCw5SwPXq1F6218pPX |
| 03:12:13 | 状況報告 | 1419 | end_turn | req_011Cc2RKL7DDsrKuij4A5KdJ |

Reproduced immediately and consistently on back-to-back simple prompts in this session.

---

Notes

• Not config/extension-related: a clean session with identical project, model, effort (high), and the same plugins behaves normally. The failures are localized to specific sessions' state.
• Both symptoms share the pattern "thinking block produced → no usable final content." Whether they share a root cause is for server-side inspection of the listed request IDs.
• Debug API log captured locally at /tmp/cc-malformed.log (request timestamps match the table above).