Oh My Pi 17.3.5: xAI onboarded to Responses API, GLM-5.3 support, Grok-4.5 default, broad provider reliability overhaul

@oh-my-pi/pi-agent-core

• Added automatic retry for transient provider failures during one-shot completions (compaction, handoff, branch summarization, manual /compact)

@oh-my-pi/pi-ai

• retryTransientCompletion support for non-agent LLM calls — retries Anthropic overload/rate-limit errors, HTTP 429/500/502/503/529
• Fixed xAI availability detection: paid-key-only setups now default to xai/grok-4.5 instead of the free SuperGrok catalog
• Fixed xAI Requests sending unsupported parameters (reasoning summary, presence/frequency penalties)
• Fixed Umans usage reporting: corrected raw-request-count → weighted-usage, added soft-cap warning and hard-exhaustion limit
• Fixed omp usage invalidate not fully clearing stale data
• Fixed Cursor HTTP/2 connection errors being treated as session-ending instead of transient
• Fixed OpenAI-compatible streams (DeepSeek) cut off mid-generation being silently accepted
• Fixed DeepSeek resource-exhaustion interruptions not retrying
• Fixed tool-call IDs lost during same-model replay
• Fixed Kimi Code multi-account routing to prefer higher-quota accounts
• Fixed Anthropic custom signing-proxy conversations losing tool-search results and thinking during replay
• Fixed runaway response loops failing gracefully instead of repeating indefinitely
• Fixed xAI rejecting turns due to certain MCP tool schema shapes
• Fixed Alibaba DashScope/Bailian per-minute rate limits misclassified as quota exhaustion
• Fixed Anthropic-compatible streams dropping thinking content
• Updated Alibaba Coding Plan China login flow URL

@oh-my-pi/pi-catalog

• Added GLM-5.3 on z.AI: unified reasoning-effort ladder, mandatory thinking, 1M context, default-model status
• Paid xAI (XAI_API_KEY) switched from Chat Completions to OpenAI Responses API — matches SuperGrok for prompt-cache affinity, reasoning-effort handling, encrypted-reasoning replay
• Default model for xai → grok-4.5; default model for xai-oauth → grok-4.5
• xAI models now request and replay encrypted reasoning content across multi-turn Responses API calls
• Fixed Codex Daybreak Blue/Red showing zero token prices (incorrectly labeled free)
• Fixed Baseten Kimi-K3 catalog metadata for thinking levels
• Fixed opencode-go/deepseek-v4-flash Responses sending forced tool_choice selectors rejected during thinking

@oh-my-pi/pi-coding-agent

• Added Extensions tab group to settings schema
• Routed paid xAI models through Responses API (aligned with SuperGrok), including encrypted reasoning replay
• Default model for XAI_API_KEY → grok-4.5
• Stopped sending presence/frequency penalties and stop sequences to xAI reasoning models
• Fixed hub job/wait lists hiding stale running subagent registrations
• Fixed external thinking scratchpads conflicting with native xAI Grok 4 reasoning
• Fixed llama.cpp model discovery missing /v1 prefix for non-Qwen models (404 errors)
• Fixed prompt caching on open-weight providers (DeepSeek, Qwen, GLM) across directory changes and midnight rollovers
• Fixed omp --fork omitting artifact directory
• Fixed long ask option labels being hard-truncated at terminal width
• Fixed toggling display.showTokenUsage leaving stale rows
• Fixed mid-run auto-compaction blocking live loop while waiting on extension handlers