You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
[0.7.0] - 2026-11-07
Changed
Under an enabled subagent model-selection allowlist, a mounted child's first attempt is allowlist-admitted: the walk starts on the first dispatchable route of the role chain that is on the allowlist, so a child can start on a non-head route.
A mount is declined when the role chain has no allowlisted dispatchable route — the child keeps pure inheritance instead of being mounted on a chain the host allowlist cannot serve.
A mounted child's walk refuses before any attempt (UNDISPATCHABLE_EFFECTIVE_HEAD, no request sent) when the host subagent policy becomes unprovable or no longer admits the cooled head, instead of dispatching an unfiltered route first.
Fallback now applies only to sessions whose model selection is FallbacksChain / Auto; a real-model selection is served by llm-retry alone, with no plugin fallback.
A walk switches only while the failed attempt has emitted no output; after the first chunk the attempt is committed and the failure is forwarded verbatim, so no second model's output is appended behind a partial one.
Subagents mounted through the dispatch seam inherit a chain; explicit routes (provider + model in the spawn request) and external backends keep their own route unchanged.
The alwaysModeRetryCap and roleAutoMatch config keys were removed with no successor; a stored section that still carries them is reported by the legacy-key notice and startup warn, then ignored.
Upgrade @deepseek-ai/dsh-* peers to ^0.2.2-rc.1.
Added
Add a localized preview image to the README header and the Plugins-page detail view.