Accept Bash and text editor calls from Anthropic CodeExecution_20250825 - #10552

Type

• Refactor
• Feature
• Bug Fix
• Optimization
• Documentation Update

Description

With Toolkit.make(AnthropicTool.CodeExecution_20250825()), generateText and streamText fail when Claude runs a Bash command or edits a file, the operations the docs list for this tool version:

LanguageModel.streamText: Invalid output: Missing key
  at ["code"]

The tool's parameter schema only accepts { code }, while the bash_code_execution and text_editor_code_execution calls in the documented response carry command, path and file_text. generateText fails one step earlier: it does not map those two names to the toolkit name, so the call matches no tool.

Changes:

• CodeExecution_20250825 accepts the existing Bash and text editor operation schemas next to the code-only input, so its params type becomes a union, as it already is for CodeExecution_20250522. The exported CodeExecution_20250825_Parameters is unchanged.
• The non-streaming path maps both names to the toolkit name, as the streaming path already does.
• Bash and text editor results are marked providerExecuted, like code_execution_tool_result, so they replay in the assistant message.
• Replay recognizes the view, create and str_replace result types and no longer sends the internal type discriminator in the call input, so the next request carries the server blocks as received.

The new tests run Bash, create, view, str_replace and both error results through generateText and streamText, then compare the follow-up request with the original blocks. They fail on main with InvalidOutputError. The recorded code_execution_20250825 responses in vercel/ai's Anthropic fixtures (anthropic-code-execution-20250825.1.json and .1.chunks.txt) also decode through both APIs and replay as recorded.

pnpm test --run packages/ai/anthropic/test, pnpm lint, pnpm check and pnpm build pass. One full workspace run on Node 24 passes; the latest one hits pre-existing intermittent failures in the cluster Sharding release-warning test and the Cloudflare Readiness initialization test, which fail the same way on the merge base.

Related

Follow-up to #10297 and #10373; the latter described this as a separate problem with the tool's parameter schema.

Closes EFF-3634