Newer Anthropic models regress on non-Claude-Code tool schemas

Category: Agentic coding
Status: discussion
Sources: Armin Ronacher post, HN discussion
Summary: Armin Ronacher reports that newer Anthropic models (Opus 4.8, Sonnet 5) more often emit malformed tool calls against non-Claude-Code edit-tool schemas, adding invented keys such as requireUnique, oldText2, type, and id around otherwise byte-correct oldText and newText payloads. He attributes the regression to reinforcement learning around Claude Code's forgiving harness, which uses parameter aliases and silently filters unknown keys, training stronger priors for its flat edit schema. He reports about 20% failure with Opus 4.8, halved by stripping thinking blocks, and eliminated by strict tool-invocation mode.
Why it matters: Teams building agent harnesses outside Claude Code may need strict schema validation to keep newer models reliable.