feat(context)!: deterministic minimum-first layer allocator - #87
Open
LukasParke wants to merge 16 commits into
Open
feat(context)!: deterministic minimum-first layer allocator#87LukasParke wants to merge 16 commits into
LukasParke wants to merge 16 commits into
Conversation
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
|
@LukasParke is attempting to deploy a commit to the Matt Apperson's projects Team on Vercel. A member of the Team first needs to authorize it. |
Collaborator
Author
|
Agent: Code CI/structural/DCO are green where configured; remaining external red checks are |
Replaces the 60/40 layer/history split and its infinite-headroom
special cases: minimums are guaranteed first (proportionally scaled
when overcommitted), the remainder distributes proportional to
declared caps, and 'auto'/omitted budgets get a fixed 2000-token cap
instead of an unbounded share of the window.
Deterministic block sizes are the precondition for byte-stable
rendered layer blocks — which is what makes the prompt-cache anchoring
in this repo actually hold across turns when layer budgets are in
play. Same layers + same policy now always yields the same allocation
regardless of window size.
BREAKING CHANGE: layers with 'auto' or omitted budgets previously
received up to 60% of the post-reserve window (~74k tokens on the
128k default); they now receive a fixed 2000-token cap. Declare an
explicit budget (number or {min,max}) for layers that need more. The
allocator's historyBudget return field is removed (it had no readers).
Part of the ce6d44e9 port from OpenRouter's fork.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
The ported allocator (faithfully to its fork source) scaled declared
minimums against the 25% layer pool rather than the full available
budget, so {min: 10000} at a 32k window received 7000. Minimums are
now satisfied first out of the full post-reserve window (proportional
scale-down only when they overcommit it), and the LAYER_POOL_SHARE
ration applies to the discretionary remainder only. The documented
"minimums first" guarantee now holds numerically.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
…behavior Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
LukasParke
force-pushed
the
lukeparke/openrouter-deterministic-allocator
branch
from
August 13, 2026 07:46
1ebd469 to
6fa44da
Compare
Expose the pure retry-delay calculation for deterministic assertions instead of proxying global setTimeout, which captured sibling test timers under CI concurrency. Signed-off-by: Luke Parke <5702154+LukasParke@users.noreply.github.com>
Collaborator
Author
|
Agent: Rewrote the two unsigned commits with identical tree content and valid SSH/DCO signatures, then replaced global timer interception with deterministic retry-delay unit assertions. DCO and Structural gate pass on the new history; CI rerun is in progress. |
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.
What
autolayer budgets at 2,000 tokenshistoryBudgetresult fieldWhy
The prior 60/40 split could violate declared minimums and treated
autoas infinite headroom, producing unstable allocations. Deterministic block sizes also keep prompt-cache bands byte-stable.This ports item 17 from
fork/port/openrouter-fixescommitsd8fb4f34and59497c48, plus a review fix that conserves fractional residual budgets.Test plan
Breaking changes
autobudgets now cap at 2,000 tokensallocateBudgets()no longer returnshistoryBudgetDependency
Semantically follows compaction items #80 and runtime folding #85. The branch currently includes those prerequisites and must be rebased through main after they merge.