[Bug] Tool calls intermittently emitted as plain text in large context sessions

Bug Description
[Bug] Tool calls intermittently emitted as plain text → "Your tool call was
malformed and could not be parsed. Please retry."

Environment: Claude Code, model Opus 4.8 (1M context).

Symptom:
In a long-running session, the model intermittently emits tool invocations as
visible plain text (the literal function-call wrapper leaks into the assistant
message) instead of a properly structured tool call. The harness rejects it with
"Your tool call was malformed and could not be parsed. Please retry." No tool
actually executes.

Reproduction pattern:

• Happens more frequently as conversation context grows large.
• Affects multiple tool types (file edits and shell commands alike).
• Re-issuing the exact same call in correct form succeeds immediately, so the

failure is in output formatting, not in the tool arguments themselves.

• Occurred repeatedly across many turns in a single long session.

Impact:

• Each occurrence wastes a turn and requires a manual retry.
• Because hooks fire only after a tool call is successfully parsed (Pre/PostToolUse),

there is no way to intercept or auto-retry this at the hook layer — it must be
handled by the model/parser itself.

Request:

• Investigate why structured tool calls degrade to plain text under large context.
• Consider an automatic retry/repair path at the parser layer for malformed tool

calls, since hooks cannot reach this stage.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.179
• Feedback ID: 275a51fa-d9bd-4ded-b17b-076c8921998f

Errors

[]