Repository navigation
feat(models): route families and the cascaded models block through launches and the door - #502
Merged
Merged
Conversation
…unches and the door A launch now resolves the models block through the whole cascade, the global block as its shallowest layer, and uses it for the /model rows, the credential check and the tier settings. When a deeper layer changed the block the launch records it in its session's registry entry and the door resolves that session's names against it; otherwise the door follows the global block live. The door resolves a `family:<key>` name or member to the targets of the family across providers, takes each candidate's capabilities from its family, and reads what a request needs (image, tools, thinking, the size of the context it re-sends) so `adaptive` can rule out a target that cannot serve it. The picker lists one row per family offered by more than one target between the virtual and qualified rows, and a virtual row no longer disappears when its first member names no tier. Tier pointers set the tier model settings of a session that routes by name. The launch credential check expands family members, and `model explain` follows the same expansion. A target whose own model name contains a slash (`hf:zai-org/GLM-5.1`) was read as a qualified ref on the hop's next turn and refused; a name already resolved to its target is now served as it is.
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The families, tier pointers and cascaded definitions the plan describes were
stored and checked but never reached a request: the launcher read only the
global block, nothing applied a tier pointer, the door took no account of
families, and the door's capabilities for every candidate were hand-set to
"everything".
A launch now resolves the models block through the whole cascade (the global
block is its shallowest layer) for its /model rows, its credential check and
its tier settings. A launch whose deeper layers changed the block records the
result in its registry entry, which the door reads by the session id the launch
stamps on its requests, so the door resolves that session against the block it
launched with; one that changed nothing leaves the door following the global
block live, as before.
The door resolves
family:<key>names and members to every target of thefamily, reads each candidate's capabilities from its family, and passes what a
request needs to the selection engine so
adaptiverules out a target thatcannot serve an image, tools or thinking. The picker gains one row per family
offered by more than one target, and a virtual row is no longer dropped when its
first member names no tier. Tier pointers set the tier model settings of a
session that routes by name. The launch credential check expands family
members, and
model explainfollows the same expansion.Also fixes a defect found while testing this: a target whose own model name
contains a slash (Synthetic's
hf:zai-org/GLM-5.1, a row in the picker) was readas a qualified ref on the hop's next turn and refused. A name already resolved
to its target is now served as it is.