𝙊𝙥𝙚𝙣𝘾𝙤𝙙𝙚 v1.17.9 released. TL;DR:
Agent step limits now behave correctly, GLM-5.2 thinking variants land, Bedrock/Copilot provider handling improves, and the TUI gets cleaner subagent/error behavior.
𝗔𝗴𝗲𝗻𝘁
• Fixed configured agent step limits in the v2 runner: the final step disables tools and forces a text-only summary instead of failing the run.
• Changed agents without a configured step limit to continue past the old hard 25-step local tool cap.
• Changed steered follow-up prompts to stay as normal user messages, preserving message shape and improving prompt-cache stability.
𝗣𝗿𝗼𝘃𝗶𝗱𝗲𝗿𝘀
• Added High/Max thinking variants for GLM-5.2 across OpenAI-compatible and Anthropic-compatible providers.
• Added OpenRouter GLM-5.2 High/XHigh reasoning variants using OpenRouter’s native effort mapping.
• Fixed GitHub Copilot model discovery so configured provider headers are passed through.
• Fixed Devstral model detection to use locale-stable lowercase matching.
𝗟𝗟𝗠
• Fixed Amazon Bedrock Converse topK forwarding by sending it through model-specific request fields as top_k.
𝗧𝗨𝗜
Fixed
• Fixed subagent task rows merging visually with neighboring inline tools.
• Fixed inline spacing around assistant summaries, assistant errors, block tools, and multi-line rows.
• Fixed the Console org switcher so org-load failures render inline instead of tearing down the dialog.
• Fixed background-subagent shortcut visibility so it only appears when the server advertises support.
𝗦𝗲𝗿𝘃𝗲𝗿
• Added /experimental/capabilities so clients can discover enabled experimental runtime features.
𝗦𝗗𝗞
• Updated the generated JS SDK and OpenAPI types for experimental capability discovery.
𝗖𝗼𝗻𝘀𝗼𝗹𝗲
• Added unique-user charts to OpenCode Data home and model pages.
• Fixed model peer rankings to use the same ranking window as the displayed model rank.
𝗭𝗲𝗻
• Improved provider budget telemetry so usage is recorded when spend is tracked and includes per-model budget usage.
No noticeable bundle change
Compare: github.com/anomalyco/open…
English
