10 commits landed across OpenClaw in the last 3 hours. This is a correctness-heavy batch — no flashy features, but a tight set of fixes that keep agent-to-agent messages intact and keep media and video requests obeying the same policy rules as everything else.
What Changed
Native quote replies survive agent sends (#105347). Jesse Merhi landed fix(signal): preserve native quote replies from agent sends. When an agent sends a message that’s meant to quote a prior one, the native quote reference is now carried through instead of being dropped — so threading in chat transports stays coherent instead of losing the “replying to” context.
Request-policy patches for media models (MiniMax #102072, DashScope #102094). Alix-007 applied the request policy to two more surfaces: fix(minimax): apply request policy to media generation requests and fix(dashscope): apply request policy to video requests. Both media and video generation now route through the same policy controls as text, closing a gap where those request types could bypass configured limits.
Offline, deterministic test harness (QA #105493, #105488). Peter Steinberger modeled Codex compaction in the parity mock (test(qa): model Codex compaction in parity mock) and kept mock memory embeddings offline (fix(qa): keep mock memory embeddings offline). The QA mocks now behave like the real paths without needing a network round-trip — faster, flakier-free CI.
CI and refactor cleanup (#105472, #105497, #105486, #105482). Steinberger also retained Telegram candidate diagnostics in CI so failures are debuggable, while Vincent Koc removed benchmark and Brave testing aliases and localized the Tencent preview model reference — less dead surface area, more consistent config.
Teams streaming edge case (#102357). Galin Iliev fixed Microsoft Teams streaming so it keeps the chunk after a whitespace collapse, instead of silently dropping the trailing piece of streamed output.
Why It Matters
This batch is the glue that makes OpenClaw feel reliable when you’re not watching. Preserved quote replies keep multi-turn agent conversations readable; the MiniMax and DashScope policy patches mean your configured guardrails actually apply to every request type, not just text. And the offline QA work means the test suite catches regressions without flaking on network.
OpenClaw’s session-fleet work is the longer story here — our OpenClaw Claude session fleet coverage lays out the trajectory these messaging and policy fixes are feeding.
Before you buy any coding agent subscription — compare prices and find exclusive deals at aiFiesta.