[earakely-scale]

Why

The Claude Code agent pins CLI 2.1.280, which predates Claude Fable 5.2. Behind a gateway it runs claude-fable-5-2 on fallback defaults: a 32K output limit, a 200K window for the bare id (long-tool-list envs then exit with "Autocompact is thrashing"), and an unknown price, so the result's costUSD is a guess. Claude Code 2.1.321 is the first release that knows the model.

What changes

• CLI version. CLAUDE_CODE_VERSION and TOOLS_CLI_VERSION move to 2.1.322, the newest release. ClaudeCodeTool gains Monitor, the one tool the new CLI adds.
• Tool search stays off. 2.1.322 turns tool search on by default. That holds back every MCP tool, with an env of 5 tools as with one of 300, and about half of its own tools (WebFetch, WebSearch, the Cron and Task tools) until the model calls ToolSearch. cli_env sets ENABLE_TOOL_SEARCH=false unless the agent's environment sets it. The CLI then sends the env's tools and its own up front, as 2.1.280 did. The README says so.
• 1M list. claude-fable-5-2 joins _NATIVE_1M_MODELS, so the bridge passes it as claude-fable-5-2[1m] like the other 1M models.

Measured

Both CLIs ran under the bridge's flags against tst/util/fake_anthropic, on a Linux host:

	2.1.280	2.1.322
bare claude-fable-5-2	32K output, 200K window, price unknown	128K output, 1M window, list price
bare claude-opus-5-5	128K output, 200K window	128K output, 1M window
claude-sonnet-5-5, claude-haiku-5-5	32K output, 200K window, price unknown	128K output, 1M window, list price
built-in tools, tool search off	24 (20 for most models)	2.1.280's plus Monitor
MCP tools, default settings	all in tools	none in tools (deferred)
MCP tools, ENABLE_TOOL_SEARCH=false	all in tools	all in tools

The transcript directory naming the bridge relies on (project_dir) is unchanged.

Tests

• tst/integration/cli/run_bundle_claude_code_test.py builds the image with 2.1.322 and runs the real CLI: 9 passed, including the four builtin_tools selections.
• New unit test for the ENABLE_TOOL_SEARCH default and its override; claude-fable-5-2 added to the 1M model cases.
• make unit-test on Linux: 14,405 passed. The 3 failures come from the test copy: one needs a git checkout, two run a generated script with a system python3 that lacks httpx. All 3 pass on macOS.

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge; no new actionable issue was found.

Summary

Updates Claude Code to 2.1.322, adds Monitor, and gives claude-fable-5-2 the [1m] suffix.

• Claude Code sends task tools before the model searches.
• Claude Code sends Fable 5.2 with its 1M window.

_{Reviews (2) · Last reviewed commit: "test(agents): fail the retried-install t..." · Reviewed by Greptile}