v3.18.2

Context-window auto-detection, plus the Claude-side companion to the v3.18.1 provider fixes.

Added

• Context window limits are auto-detected each run instead of hardcoded. Resolution order, most authoritative first: (1) your context_limits override in ~/.kodiqa/config.json (per-model or per-provider); (2) the official value from the provider's own /models API — Groq (context_window), Mistral (max_context_length), and OpenRouter (context_length) report it, and local Ollama models read it from ollama show; (3) a maintained fallback table for OpenAI / DeepSeek / Qwen / Claude, which don't expose it via API. Everything reads warm caches (no per-turn network calls) and is offline-safe.

Fixed

• DeepSeek context window corrected from a stale 64K to 1,000,000 (deepseek-v4-pro/flash). The auto-compact warning was firing at ~45K and would start summarizing the conversation ~16× sooner than necessary. You can still cap it lower via context_limits to control cost.
• Claude path: the v3.18.1 mixed-history / 400-no-failover fixes now apply to the Claude API too. _build_claude_messages enforces Anthropic's tool_use/tool_result invariant defensively — it drops the "review"/"lint" orphan tool-result ids that would 400 every Claude edit/lint turn, converts OpenAI-format turns left by cross-provider failover instead of silently dropping them, and backfills unanswered tool_use blocks. A Claude 400/422 also no longer cascades through failover.