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
Carries the unfinished scope of #59, which closed after slices 1 and 2 shipped. Three targets remain, and two of them are blocked on a design problem, not on effort.
Blocked on design — both are reached on every run
Slice 2 established the mechanism (skill-local reference, bare-relative citation, load site gated on a flag) and proved it at runtime. It does not help these two, because relocating an always-reached block makes the body heavier: shrunken body, plus the whole file, plus load overhead.
ba-review Step 5 resolvers (~308 ln today, was ~290 when [roadmap] Shrink the always-resident prompt surface in review.md and propose.md #59 was filed).## Step 5: Resolution is reached on every run past the NO_CHANGES exit. Restoring conditionality means splitting by branch, which the slice-2 brainstorm's locked design forbids ("moves whole — never split"), and which duplicates the ~51-line post-apply guard reachable from both. Two further findings from slice-2 planning: option labels and their routed action text share a line, so this is a rewrite, not a relocation; and there are six menu families in the range, not five.
ba-plan's detail-level templates (~150 ln). Same defect: Step 4 is reached on every run. The slice-2 brainstorm rejected one design partly because it made these an always-fired read; the adopted design has the identical property here.
Revisit trigger: when a design exists that restores conditionality without splitting a region across files — i.e. when either target's always-reached branch point can move above the extraction boundary, the way --persist's PERSIST=false gate already sits above its load site. Until then, extracting either is a measured regression, not a shrink.
Not blocked — just never started
ba-propose Step 5f (~150 ln, reachable only on commit_push_create) and the Step 3 composition spec (~170 ln). 5f is genuinely branch-conditional, so it is the one remaining target the slice-2 mechanism fits as-is. It was excluded from slice 2 on scope, not on the always-reached test. This is the pick-up-now item.
What slice 2 measured, which should temper the expected win
Extraction is not free, and #59's framing assumed it was. Slice 2's final numbers, across session types:
session type
vs pre-slice
/ba-review run
−544 est tok
any other session
+29 est tok
--persist run
+1536 est tok
The first version of that slice was a net regression: the convention bullet it needed loads in every session, and no acceptance criterion looked there. It only became net-positive after moving the rationale to a load-on-demand doc.
Carry that lesson into any further extraction here: measure the resident surface the change touches, not the one file it edits. An acceptance criterion of the form "file X shrank" can pass while the repo gets heavier.
Carries the unfinished scope of #59, which closed after slices 1 and 2 shipped. Three targets remain, and two of them are blocked on a design problem, not on effort.
Blocked on design — both are reached on every run
Slice 2 established the mechanism (skill-local reference, bare-relative citation, load site gated on a flag) and proved it at runtime. It does not help these two, because relocating an always-reached block makes the body heavier: shrunken body, plus the whole file, plus load overhead.
ba-reviewStep 5 resolvers (~308 ln today, was ~290 when [roadmap] Shrink the always-resident prompt surface in review.md and propose.md #59 was filed).## Step 5: Resolutionis reached on every run past theNO_CHANGESexit. Restoring conditionality means splitting by branch, which the slice-2 brainstorm's locked design forbids ("moves whole — never split"), and which duplicates the ~51-line post-apply guard reachable from both. Two further findings from slice-2 planning: option labels and their routed action text share a line, so this is a rewrite, not a relocation; and there are six menu families in the range, not five.ba-plan's detail-level templates (~150 ln). Same defect: Step 4 is reached on every run. The slice-2 brainstorm rejected one design partly because it made these an always-fired read; the adopted design has the identical property here.Revisit trigger: when a design exists that restores conditionality without splitting a region across files — i.e. when either target's always-reached branch point can move above the extraction boundary, the way
--persist'sPERSIST=falsegate already sits above its load site. Until then, extracting either is a measured regression, not a shrink.Not blocked — just never started
ba-proposeStep 5f (~150 ln, reachable only oncommit_push_create) and the Step 3 composition spec (~170 ln). 5f is genuinely branch-conditional, so it is the one remaining target the slice-2 mechanism fits as-is. It was excluded from slice 2 on scope, not on the always-reached test. This is the pick-up-now item.What slice 2 measured, which should temper the expected win
Extraction is not free, and #59's framing assumed it was. Slice 2's final numbers, across session types:
/ba-reviewrun--persistrunThe first version of that slice was a net regression: the convention bullet it needed loads in every session, and no acceptance criterion looked there. It only became net-positive after moving the rationale to a load-on-demand doc.
Carry that lesson into any further extraction here: measure the resident surface the change touches, not the one file it edits. An acceptance criterion of the form "file X shrank" can pass while the repo gets heavier.