[XieX]

Summary

AIC-3585. Spec: launchdarkly/ai-sdks-monorepo#110. The same change for TypeScript is launchdarkly/js-ai-sdk#183; Go is not affected.

claude-agents ran with the Claude Code CLI's own defaults instead of the AI Config's. The config's max_tokens was dropped, because ClaudeAgentOptions has no output-token field. With no thinking setting, the CLI chose one: a 31,999-token budget on claude-haiku-4-5. Python wasn't checked live; the code path is the same as the TypeScript one that was.

_options_parameters (handler.py) is shared by the handler's invoke and stream and by to_claude_agents. It now:

• Sends max_tokens to the CLI. A positive integer max_tokens up to 2^53 − 1 becomes env={"CLAUDE_CODE_MAX_OUTPUT_TOKENS": "<n>"}. That's the same bound as the JS SDK's safe-integer check. The Agent SDK adds that to the inherited os.environ. A whole-number float such as 200.0 counts as 200, as it does in JavaScript. Any other value is dropped. With no max_tokens, env stays {} and the CLI keeps its default, as before.
• Turns thinking off by default. When the config sets neither thinking nor max_thinking_tokens, it passes thinking={"type": "disabled"}. A malformed thinking that gets dropped counts as not set. An explicit thinking, or max_thinking_tokens on its own, is passed as before. The CLI leaves thinking out for models that reject disabled thinking, such as claude-fable-5.

max_tokens is read separately rather than added to _CLAUDE_AGENT_OPTIONS_FORWARDED_KEYS, since it isn't a ClaudeAgentOptions field and the drift test checks that list against the dataclass. env stays in the excluded list: the max_tokens value is the only thing written into it. tests/forwarding_spec.py adds max_tokens to the Claude Agents list.

Breaking change

A config that sets no thinking now runs without thinking. Before, the CLI chose: adaptive thinking on models that support it, or a fixed budget on others. Configs that relied on that must set thinking in their model parameters, for example {"type": "adaptive"}. The README's new "Output limit and thinking" section has the migration note.

Tests

• A new TestOutputLimitAndThinkingDefault, on both invoke and stream, covers:
  • max_tokens: 200 sets the variable and disables thinking, checked against the SDK's own CLI command builder.
  • The environment the SDK's transport gives the CLI still holds os.environ. open_process is stubbed to capture it.
  • 200.0 is read as 200.
  • With no max_tokens, env is empty.
  • 0, -5, 1.5, 2**53, inf, "200", True, None and a dict are each dropped.
  • effort alone gets the default, and thinking with max_thinking_tokens passes both as given.
  • A configured thinking replaces the default.
  • max_thinking_tokens on its own leaves thinking unset.
• TestNativeGraphModelParameters covers four node cases: max_tokens (and checks that the real SDK builds --thinking disabled for it), its own thinking, an invalid max_tokens, and max_thinking_tokens alone (--max-thinking-tokens 2048).
• Seven existing TestThinkingIsRebuilt / TestMalformedObjectValuesAreDropped cases expected a dropped thinking to be None, and now expect the default.
• max_tokens was removed from test_ui_keys_the_sdk_rejects_are_dropped_without_raising, since it's no longer dropped.
• make test passes (3686 passed, 18 skipped). make lint, make format-check and mypy pass.

🤖 Generated with Claude Code

---

Note

Overview
Breaking: configs that omit thinking no longer inherit Claude Code’s implicit thinking defaults; the handler now passes thinking={"type": "disabled"} unless the config sets thinking or max_thinking_tokens (malformed thinking falls back to the same default).

The shared _options_parameters path (handler invoke/stream and to_claude_agents) now honours max_tokens: valid positive integers (including whole floats, capped at 2^53−1) are sent to the CLI as env={"CLAUDE_CODE_MAX_OUTPUT_TOKENS": "<n>"} because ClaudeAgentOptions has no output-token field. Invalid values are ignored; omitting max_tokens leaves CLI defaults unchanged.

README documents output limits, thinking behaviour, and migration. Cross-SDK forwarding_spec adds max_tokens to the Claude Agents forwarded list; new handler, native-graph, and updated thinking-rebuild tests lock in the env/thinking rules and CLI command shape.

^{Reviewed by Cursor Bugbot for commit cd91993. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jeffdupont]

This matches mono#110 and js-ai-sdk#183. The same max_tokens bounds (a positive whole number up to 2^53 - 1, with 200.0 read as 200), the same disabled-thinking default, and the same rule that a dropped malformed thinking counts as not set. I checked env against the SDK: no host or skills option also sets env, so nothing overwrites it. One small test gap, below, is the same one I raised on the JS PR.

[jeffdupont]

cd91993 covers the test gap from my last review: the native-graph test now pins the disabled default for a malformed thinking ({"type": "enabled"} with no budget) and for effort alone, so every output-limit case mono#110 lists runs through invoke, stream and a graph node. Nothing else changed since 837053d, and the behaviour still matches js-ai-sdk#183.