v1.28.0

What's Changed

• fix(sdk): keep OpenHands provider by @enyst in #3548
• test(agent-server): add cross lease coverage by @enyst in #3559
• fix(llm): don't emit explicit cache_control markers for Gemini by @VascoSch92 in #3586
• test(sdk): pin shell parser node shapes by @Fieldnote-Echo in #3590
• Harden workflow GitHub context handling by @enyst in #3372
• feat: add client-defined tools (no-op tools via JSON spec) by @xingyaoww in #3306
• fix(sdk): reject client tool name collisions by @enyst in #3614
• feat(models): add claude-fable-5 model support by @juanmichelini in #3592
• [codex] Count tokens with tokenizer chat templates by @neubig in #3593
• [codex] Sanitize malformed tool call recovery by @neubig in #3575
• Expose active LLM profile in settings API by @enyst in #3552
• feat(terminal): reject Python/JSON literals passed as bash commands by @juanmichelini in #3582
• feat(agent-server): add OpenAI chat completions gateway by @enyst in #3545
• fix(sdk): Support ACP model switching for custom/unknown providers via set_config_option by @Hades32 in #3556
• Consume cached View in prepare_llm_messages by @csmith49 in #3328
• Skip PR description check for release PRs by @enyst in #3639
• Extend workspace session cookie lifetime to effectively never expire by @rbren in #3637
• feat(agent-server): harden OpenAI gateway client compatibility by @enyst in #3626
• Wrap condenser LLM errors as NoCondensationAvailableException by @csmith49 in #3647
• fix(llm): recognize claude-fable-5 as a reasoning model by @juanmichelini in #3642
• Release v1.28.0 by @all-hands-bot in #3638

Full Changelog: v1.27.1...v1.28.0