Opus 4.8 (1M context) — frequent "tool_use malformed" client errors despite tools executing successfully

Environment

• Model: claude-opus-4-8-1m (switched to claude-opus-4-7-1m as workaround mid-session — see below)
• Claude Code version: 2.1.92
• Session type: forked from a prior session (this is a variable I could not isolate solo — see Repro)
• OS: Windows 11 + Git Bash
• Workflow: long CI-debugging session with many tool calls (Bash, Edit, Read, WebFetch, subagents)

Symptom

Roughly 5 out of ~15 tool calls on Opus 4.8 fail with:

Your tool call was malformed and could not be parsed. Please retry.

The assistant turn appears empty to the user, and a UserPromptSubmit-style retry prompt arrives next turn. However, the tool actually executed on the server side — verified independently each time:

• Edit: subsequent git diff confirms the file was modified exactly as intended
• Bash (gh pr create): the PR was actually created on GitHub (visible at the returned URL)
• Subagent (Agent tool): subagent ran to completion and the structured result was retained in the next assistant turn's context

So the tool call reached the server, ran, and produced correct output — only the client-side parse / surfacing of the assistant message that contained the tool call failed.

Repro / variable isolation

After ~5 occurrences across an hour on Opus 4.8, I switched the same session to Opus 4.7 (1M) via /model. 0 occurrences across 6+ subsequent tool calls in the same session, same workflow, same kinds of tools (Bash, Read, Edit, gh CLI). Strong signal that the regression is model-side rather than environmental.

The one variable I could not isolate solo: the session was forked from a prior conversation. I don't know whether a fresh 4.8 session (no forked history) reproduces — would appreciate help confirming. Three possible outcomes:

• Fresh 4.8 also reproduces → model-level regression in 4.8
• Only forked 4.8 reproduces → interaction between 4.8 and replayed conversation history
• Neither → close

Hypothesis

Plausibly XML close-tag drift in long / nested <parameter> content — e.g. the model emitting </invoke> where </parameter> was expected, causing the client-side parser to drop or mis-attribute the tool_use block. There is a known mode where long args with mixed content trip this; 4.8 appears to have raised the failure rate noticeably.

Observed triggering tool calls in this session:

• Edit with multi-line new_string containing mixed CJK comments + YAML code
• Bash with HEREDOC commit messages (multi-line, CJK + ASCII)
• WebFetch with a longer prompt arg
• (one occurrence after a system-reminder injection arrived between the user message and my tool call — could be coincidence, flagging in case)

Workaround

Switch to Opus 4.7 (1M context) via /model.

Impact

Each failure costs one round-trip + a context-polluting "retry" prompt, and in a long session it accumulates. Worse, because the tool did execute, a careless user might double-execute side-effectful operations (e.g. re-create a PR, double-commit) if they don't notice the empty turn and just re-issue the request.

Happy to provide

• Session transcript excerpts (contains code from a private repo; can redact)
• Specific timestamps / turn boundaries on request