[SaladDay]

Each Harness now receives the model provider's token limits through its native setting, and Core rejects a limit the Harness cannot run.

• Codex: context_window → model_context_window. Codex has no native output limit, so a set max_output_tokens returns model_provider_token_limits_invalid.
• Claude Code: CLAUDE_CODE_MAX_CONTEXT_TOKENS and CLAUDE_CODE_MAX_OUTPUT_TOKENS. Native uses the first only for a model it does not know and caps the second at the model's limit.
• MiniMax Code: the custom model's limit.context and limit.output, unchanged.
• The declaration's MaxOutputTokens is the only source of this fact. /core/v1/harnesses projects it as model_configuration_support.accepts_max_output_tokens, and Web shows the Max output tokens field only when it is true.
• Docs: the model-execution protocols table gains a native token limits column, and the core-errors row is updated, in English and Chinese.
• Claude workspace launches drop daemon CLAUDE_CODE_MAX_*_TOKENS, so only the provider sets them; the client accepts an output limit without a context window, as Core does.

Release notes: a Codex Session, saved Agent or deployment default that stores max_output_tokens was accepted and ignored; it now fails at use with model_provider_token_limits_invalid. Remove the field.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}