Skip to content

Size sidecar context budget from the model's real window - #454

Merged
TheGreatAxios merged 2 commits into
mainfrom
cl-7170-size-sidecar-context-budget-from-the-models-real-window
Aug 29, 2026
Merged

TheGreatAxios merged 2 commits into
mainfrom
cl-7170-size-sidecar-context-budget-from-the-models-real-window

Conversation

@TheGreatAxios

Copy link
Copy Markdown
Contributor

https://linear.app/abklabs/issue/CL-7170

Size the sidecar context budget from InferenceSource quirks / advertised window; compact on message.received before hard overflow. This is not a 32k product cap.

Compaction used an 8k-token default (32k characters) whenever quirks
did not pin numCtx, so frontier hosted models overflowed far inside
their advertised windows. Resolve the window from quirks, then the
catalog's advertised native size, and compact on message.received
when over the headroomed budget but still under the hard limit.
The reactor forbids compact+infer in one cycle and does not re-enter
the director after compact. Replacing infer with compact on
message.received left the user message unanswered until the next inbound
event. Compact stays on the non-final tool.done path, which still
re-infers.
@TheGreatAxios
TheGreatAxios force-pushed the cl-7170-size-sidecar-context-budget-from-the-models-real-window branch from a800ff4 to 8d0687d Compare August 29, 2026 04:46
@TheGreatAxios
TheGreatAxios merged commit 08d1738 into main Aug 29, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant