Hermes 9 月 3 日合并 PR：Muse Spark 1.3 上架多平台

Luke The Dev @iamlukethedev原推发布于 2026.09.04

在 X 上打开原推

Hermes merged only 47 PRs on September 3 The ones worth knowing: • Muse Spark 1.3 family selectable on OpenRouter and OpenCode Zen/Go/Free • Muse Spark gets its real 1M context window (was stuck at 256K) • Agents save task-learned knowledge and corrections into the task's skill • Local models on by default on Windows and macOS • Packaged Desktop no longer crashes on launch (undefined.push in rememberLog) • Subagent fan-out: shared scheduler, shared httpx pools, one pyright per worktree • Finished subagents free parent heap; their transcripts skip FTS • GLM-5.3 on Nous/OpenRouter no longer 400s when thinking is disabled • Sibling Nous 401 recovery adopts a peer refresh instead of rotating the grant again • Bot Mode @mention handoffs resolve on every active gateway • execute_code kernels no longer outlive a killed host • OpenCode requests carry x-opencode-session for backend affinity Here is the full list: DESKTOP IMPROVEMENTS (9 PRs) • fix(desktop): packaged app no longer crashes on launch with undefined.push in rememberLog • fix(desktop): clicking an already-open bot no longer shows a stale transcript • fix(desktop): model-switch confirm toast has a real button, not a hairline text link • fix(desktop): selecting a bot no longer leaves the previous chat on screen • fix(desktop): Bot Mode @mention handoffs resolve on every active gateway • fix(desktop): warm session switch no longer pegs the renderer main thread • fix(desktop): cap concurrent local backend spawns + pool sizing as a live preference • fix(desktop): bound the orphan-reap sweep so slow probes cannot stall boot • fix(computer-use): native Wayland opt-in now reaches gateway cua-driver processes DEVELOPER TOOLS & PLUGINS (21 PRs) • Agent saves task-learned knowledge (incl. user preferences and corrections) to the task's skill; memory is the every-session exception • Finished subagents release their transcripts from the parent heap; subagent transcripts skip the trigram FTS index (schema v30) • All agents share one httpx transport pool per provider identity; per-agent client lifecycle unchanged • Per-child heartbeat, turn-lease and liveness timers run on one shared scheduler thread instead of 2-3 threads per subagent • One pyright serves every git worktree; new roots attach as workspaceFolders instead of spawning a server each • bench: fan-out resource harness for subagent scaling work (threads/RSS/fds/pyright/kernels/transports per N children) • fix(prompt): Muse Spark no longer ends the turn in prose with 0 tool calls on default guidance settings • 2x fmt(js): `npm run fix` auto-fix • fix(context): Muse Spark resolves its 1M window on OpenCode Zen/Go/Free, meta-ai and CommandCode instead of 256K • feat(models): Muse Spark 1.3 family selectable across OpenRouter, OpenCode Zen/Go/Free • fix(vision): providers that reject images in tool results no longer get the multimodal envelope: meta-ai/Muse Spark 400 • fix(cron): preserve active turns across gateway restarts • execute_code kernels no longer outlive a killed host on Windows or POSIX • Remote execute_code kernels are idle-reaped and capped like local ones, never mid-cell • perf(agent): buffer streamed tool-call argument fragments and join once at materialisation • feat(opencode): every OpenCode request now carries x-opencode-session for backend affinity • execute_code parallel cells no longer leak kernel processes; LSP spawns one pyright per repo, not per package • enable local models by default on win32 and darwin • fix(gateway): keep the model picker open responsive on a cold pricing cache • Local/meta ai live catalog SECURITY & SECRETS (1 PRs) • fix: sibling Nous 401 recovery adopts a peer's refresh instead of rotating the grant again INTEGRATIONS & CHAT (2 PRs) • fix(gateway): slash commands report an unreadable transcript instead of replying nothing • fix(relay): upgrade to 0.8.3 STABILITY & FIXES (14 PRs)

这条是引用着另一条写的：x.com/iamlukethedev/status/2095409919742861496