release aidev BREAKING SIG 3/5

Oh My Pi v17.3.0: breaking config changes, new OpenAI models, per-agent advisors, Nix support

Breaking: the global advisor.subagents setting is removed (per-agent advisor config replaces it), withGeminiThinkingLoopGuard is renamed to withThinkingLoopGuard, and OpenAICompat.enableGeminiThinkingLoopGuard is removed. New: first-party OpenAI Daybreak Blue, Daybreak Red, and GPT-5.6 Cyber models; per-agent advisors via frontmatter or settings; redesigned /agents fullscreen hub; Nix support with reproducible builds, NixOS/Home Manager modules; Astral ty as a Python LSP fallback. Multiple fixes across Anthropic cache pricing, Ollama chat adapter, xAI usage reporting, and Grok thinking-loop guard. Legacy cost-estimation machinery for OpenCode Go is removed.

PUBLISHED2026-08-13
OBSERVED2026-08-21
AGE13d
SOURCES1

Breaking changes

  • Removed global advisor.subagents setting; subagent advisors are now per-agent via frontmatter or task.agentAdvisor (auto-migration from advisor.subagents: true)
  • Renamed withGeminiThinkingLoopGuard to withThinkingLoopGuard (guards Gemini, DeepSeek, and Grok model-id families)
  • Removed OpenAICompat.enableGeminiThinkingLoopGuard; thinking-loop eligibility is derived solely from model ID family

Added

  • First-party OpenAI Daybreak Blue, Daybreak Red, and GPT-5.6 Cyber models with full API pricing (including long-context rates above 272K input), token limits, tools, and reasoning effort controls (off/low/medium/high/xhigh/max)
  • Per-agent advisors configured via frontmatter advisor field or task.agentAdvisor setting
  • Redesigned /agents fullscreen hub with scope sidebar, type-to-filter search, pinned detail pane, mouse support, and interactive property chips
  • Nix support: reproducible source builds for Linux and macOS, pinned development shell, NixOS and Home Manager modules, offline Bun dependency support
  • Astral ty as a built-in fallback Python LSP server (ordered behind pyright, basedpyright, pylsp)
  • calculateUncachedInputCost() for prompt pricing against active context-length tiers

Fixes

  • Anthropic cache-write pricing correctly handles mixed 5-minute and 1-hour TTL usage
  • Ollama chat adapter forwards sampling parameters (temperature, topP) to the provider
  • OpenAI agent turns no longer end prematurely after a web search with no visible answer
  • Completed model streams release provider concurrency permits promptly
  • xAI usage reporting no longer falls back to stale cache at weekly cycle boundaries
  • Together AI login validates via authenticated models list instead of hardcoded model
  • Grok runaway thinking streams caught by the thinking-loop guard
  • DeepSeek V4 Flash and V4 Pro apply correct effort contracts (low/high/max) across all transports
  • Image input support for qwen3.8-max and newer models in DashScope compatible-mode
  • Perplexity email-OTP login preserves session cookies
  • Various credential-health, login, and config resolution fixes

Removed

  • Legacy local request-cost estimation machinery and database schemas for OpenCode Go
breakingfeaturefixmodel-supportcapability

COMMUNITY

No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.