description: Review, fix, verify, land KKG PRs together
argument-hint: "[PR numbers]"
Land open KKG PRs: $ARGUMENTS.
You do drudge work; I decide exceptions.
No PRs specified → propose ordered batch; await selection.
Sort by: dependabot > small PRs > easy PRs > rest.
Only consider PRs without reviewers or those where I am reviewer.
Once we decide to work on PRs -> set me as reviewer.
Use relevant skills for mechanics.
Review
Analyze → show important diffs + explanations (visual, pseudo code, diagrams, mockups, ephemlets) → fix → verify → land.
Be granular; unless there is learned taste → routine summaries, exceptions surfaced.
Mission-critical code always granular: migrator, entities, foundational APIs/interfaces, …
Infer boundaries from repo + callers + persistence impact; uncertain → ask.
During edits → scoped tests, compiles, app verification (manual, scripts, fxdriver, grpcurl); best effort.
Review complete → full tests + app build once before final push; ¬ endless full-gate loops.
Deps → verify consuming app: GUI | gRPC CLI | equivalent.
Prefer adapting existing tests; add fast, useful coverage to PR.
Runtime loading: reflection | ServiceLoader | … → regression test hitting app-used library API.
Unable to verify → simplest proposal to empower you; I choose enable now | defer | skip.
Explicit skip → named check + current PR revision only.
Land
Bare repo layout → trunk + PR worktrees beside bare Git dir.
Independent PRs → isolated worktrees, parallel work; dependent PRs → ordered.
¬ shared mutable build/app state during concurrent verification.
Ordinary commits/pushes → autonomous; force-push → explicit approval.
After push → monitor CI; meanwhile work next PRs.
Autonomous merge within learned preferences only if required checks pass, reviews resolved, ¬ conflicts, ¬ unexplained changes.
Missing evidence | uncertainty → ask; explicit human override wins.
Confirmed merge → remove only agent-created, clean worktrees/branches with no unpushed work.
Human-facing text
Proposed messages/edits → my voice (technical german, dev speak) + exact-text approval before publishing.
Existing PR text → usable unchanged; intermediate commit messages → no review.
Dependabot HTML garbage → rewrite, simplify, no approval.
Squash → approved PR title/body; unchanged existing text needs no new approval.
Learn
Consult memory before work; auto-update abstracted coding aesthetics, principles, taste from my decisions.
Apply across review | debugging | implementation | tests; infer global vs KKG scope.
¬ generalize single-PR exceptions.
Current instruction → repo rules → local architecture → remembered taste.
Consequential conflict → surface.
---
description: Review, fix, verify, land KKG PRs together
argument-hint: "[PR numbers]"
---
Land open KKG PRs: $ARGUMENTS.
You do drudge work; I decide exceptions.
No PRs specified → propose ordered batch; await selection.
Sort by: dependabot > small PRs > easy PRs > rest.
Only consider PRs without reviewers or those where I am reviewer.
Once we decide to work on PRs -> set me as reviewer.
Use relevant skills for mechanics.
## Review
Analyze → show important diffs + explanations (visual, pseudo code, diagrams, mockups, ephemlets) → fix → verify → land.
Be granular; unless there is learned taste → routine summaries, exceptions surfaced.
Mission-critical code always granular: migrator, entities, foundational APIs/interfaces, …
Infer boundaries from repo + callers + persistence impact; uncertain → ask.
## Scope
Goal: merge PRs, ¬ fix world.
Fix PR-caused issues + necessary regression coverage.
Touched code → formatting + IntelliJ inspections welcome.
Unrelated edits → approval; future improvements/bugs → Ephemlet.
Dependency adjustments, major upgrades, new deps → follow POM rules; report before push, ¬ pre-edit approval.
Combine reviews, ¬ branches without approval.
## Verify
During edits → scoped tests, compiles, app verification (manual, scripts, fxdriver, grpcurl); best effort.
Review complete → full tests + app build __once__ before final push; ¬ endless full-gate loops.
Deps → verify consuming app: GUI | gRPC CLI | equivalent.
Prefer adapting existing tests; add fast, useful coverage to PR.
Runtime loading: reflection | ServiceLoader | … → regression test hitting app-used library API.
Unable to verify → simplest proposal to empower you; I choose enable now | defer | skip.
Explicit skip → named check + current PR revision only.
## Land
Bare repo layout → trunk + PR worktrees beside bare Git dir.
Independent PRs → isolated worktrees, parallel work; dependent PRs → ordered.
¬ shared mutable build/app state during concurrent verification.
Ordinary commits/pushes → autonomous; force-push → explicit approval.
After push → monitor CI; meanwhile work next PRs.
Autonomous merge within learned preferences only if required checks pass, reviews resolved, ¬ conflicts, ¬ unexplained changes.
Missing evidence | uncertainty → ask; explicit human override wins.
Confirmed merge → remove only agent-created, clean worktrees/branches with no unpushed work.
## Human-facing text
Proposed messages/edits → my voice (technical german, dev speak) + exact-text approval before publishing.
Existing PR text → usable unchanged; intermediate commit messages → no review.
Dependabot HTML garbage → rewrite, simplify, no approval.
Squash → approved PR title/body; unchanged existing text needs no new approval.
## Learn
Consult memory before work; auto-update abstracted coding aesthetics, principles, taste from my decisions.
Apply across review | debugging | implementation | tests; infer global vs KKG scope.
¬ generalize single-PR exceptions.
Current instruction → repo rules → local architecture → remembered taste.
Consequential conflict → surface.