6 fixes — 22 files changed, 2050 tests passing, 184 models, 40 providers
#501 — Gemini thought_signature missing in function calls:
Gemini 2.5+ thinking models return thoughtSignature on functionCall parts that must be echoed back. Added provider_metadata field on ContentBlock::ToolUse to carry opaque provider-specific data through the message round-trip. Gemini driver now captures and echoes thought_signature. 7 new tests.
#493 — Generic "Verify API key" error hides real cause:
Refined AUTH_PATTERNS to not over-match (403 from Chinese providers for quota/region issues was misclassified as auth). Error messages now include sanitized excerpts of the raw error with API key redaction. extract_status_code() now parses "API error (NNN)" format. 7 new tests.
#492 — OpenClaw plugin install rate limit:
Added 5-attempt exponential backoff with jitter to all ClawHub API calls. Respects Retry-After header. New RateLimited error variant with user-friendly message.
#487 — Telegram forum topic thread_id:
Telegram adapter now parses message_thread_id from incoming messages and includes it in outbound API calls. Forum topic replies go back to the correct topic. Persisted in delivery record for cron/workflow use. 4 new tests.
#491 — MiniMax M2.5-highspeed model:
Added MiniMax-M2.5-highspeed (1M context, tools+vision) and abab7-chat to catalog. Aliases: minimax-m2.5-highspeed, m2.5-highspeed, minimax-highspeed. Metering rates added.
#381 — Chutes.ai provider:
Added Chutes.ai as OpenAI-compatible provider with 5 models (DeepSeek-V3, DeepSeek-R1, Llama-4-Maverick, Qwen3-235B, Llama-3.3-70B). Provider defaults, metering costs, and tests included.
All live tested with real Groq LLM calls.