[Bug] Frequent "tool call could not be parsed" errors with parallel/non-ASCII tool calls
Bug Description
Bug Description
Frequent "The model's tool call could not be parsed (retry also failed)." errors
during a normal session. It happens repeatedly — often multiple times within a
single user turn — across different tool types (Bash, AskUserQuestion, parallel
Bash calls). The retry shown in the UI also fails, so the turn is interrupted and
I have to re-prompt the model to continue.
Frequency / Impact
- High frequency: occurred ~6+ times in a single short session.
- Not tied to one specific tool — seen on plain Bash, on AskUserQuestion, and on
parallel (multiple-tool-use-in-one-message) Bash calls.
- User-visible: each occurrence dumps a "Your tool call was malformed and could
not be parsed. Please retry." line into the transcript and stalls progress.
Repro (best guess — not yet deterministic)
- Run a normal coding/ops session in Claude Code (Opus 4.8).
- Issue requests that lead the model to emit tool calls, including parallel
tool calls and AskUserQuestion.
- Intermittently, the tool call fails to parse and the retry also fails.
I could not pin down a 100% reproduction; it appears intermittent rather than
input-specific. Capturing the raw model output that failed to parse (and logging
it to the feedback payload automatically) would help diagnose whether it's a
truncation, an encoding issue, or a schema-serialization problem.
Notes
- Model: Claude Opus 4.8 (claude-opus-4-8)
- Possibly correlated with parallel tool calls and/or non-ASCII (Japanese) content
in tool arguments, but unconfirmed.
- The "Errors" array in the feedback form was empty even though the failures were
clearly visible in the UI — the malformed tool call does not seem to be captured
in the structured error log.
---
(日本語補足)
通常セッション中に "tool call could not be parsed" が頻発(1セッションで6回以上)。
Bash 単体・AskUserQuestion・並列 Bash 呼び出しのいずれでも発生し、UI 上の自動リトライも
失敗してターンが中断する。再現は input 依存ではなく断続的。日本語(非ASCII)引数や並列
ツール呼び出しとの相関が疑われるが未確定。パースに失敗した生の model output を feedback に
自動添付してもらえると原因切り分け(truncation / encoding / schema 化の失敗)が進む。
Environment Info
- Platform: darwin
- Terminal: ghostty
- Version: 2.1.156
- Feedback ID: 4aa76c7e-3db0-4fc1-8e44-543cdc931eb5
Errors
[]This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗