feat(compose-agent): cross-phase back-write + false-lead authoring guidance (4.3.1)

[hamen]

What

Follow-up to #10 / v4.3.0. That release taught jetpack-compose-audit to catch cross-phase back-writes and no-op "recomposition fixes". This teaches compose-agent (the authoring skill) to avoid writing them — closing the drift the round-3 reviewers flagged. It also ships a suite-wide Strong Skipping correctness fix.

Changes

File	Change
compose-agent/references/performance.md	Never Back-Write Across Phases (cross-phase + composition-phase self-invalidation, good/bad Kotlin), Optimizations That Do Nothing, aligned Lambdas In Composables section, extended grep triggers
compose-agent/SKILL.md	Core Instruction, review-checklist line, Review Process step 4
jetpack-compose-audit/references/scoring.md	Strong Skipping fix: removed the "unless the lambda captures an unstable value" carve-out (SSM memoizes lambdas even with unstable captures) — 4 spots
versions	compose-agent → 4.3.1, jetpack-compose-audit → 4.3.1
CHANGELOG.md, README.md, docs/release-notes-4.3.1.md	4.3.1 notes

Why the audit bump

The 4.3.0 false-lead guard still allowed manual callback remember "when the lambda captures an unstable value." Per the official Strong Skipping docs that is wrong — SSM wraps lambdas in remember even with unstable captures — so the carve-out could bless the exact no-op the guard forbids. Removed suite-wide; only SSM-off / @DontMemoize / @NonSkippableComposable remain.

Validation

bin/ci ✅. Continues the chrisbanes/skills (Apache-2.0) adaptation.

🤖 Generated with Claude Code

https://claude.ai/code/session_016ysskXsKpwEbs1n9CX1ScY

[hamen]

🟠 Antigravity review (automated cross-review)

LGTM. The PR correctly updates the compose-agent authoring guidelines to mirror the 4.3.0 audit rules for cross-phase back-writes and false-lead optimizations, with solid Kotlin examples and a safe description fix in evals.json.

_{via pr-review-relay · author: claude}

[hamen]

🟢 Codex review (automated cross-review)

Blocker

• performance.md, scoring.md, evals.json: the PR still treats @NonSkippableComposable as a remaining exception where manual callback remember “matters.” Per the official Strong Skipping docs, @NonSkippableComposable opts a composable out of skipping; lambda memoization remains a separate Strong Skipping behavior, with @DontMemoize as the lambda opt-out. This leaves the same false-lead class alive in authoring guidance, audit scoring, and the acceptance eval. Remove @NonSkippableComposable from the callback-memoization exception wording; keep it only in skippability/opt-out justification rules.

Should-fix

• None.

Nit

• None.

Validation: read gh pr view 11, gh pr diff 11, all touched files, related audit playbook/scoring context, and the current Android Strong Skipping docs: https://developer.android.com/develop/ui/compose/performance/stability/strongskipping. git diff --check origin/main...HEAD is clean.

_{via pr-review-relay · author: claude}

[hamen]

🔵 Cursor review (automated cross-review)

PR #11 review — feat(compose-agent): cross-phase back-write + false-lead authoring guidance (4.3.1)

Solid follow-up to 4.3.0: it closes audit/authoring drift, the Strong Skipping “unstable capture” fix matches official SSM docs, and cross-phase guidance aligns with search-playbook.md / evals 12–14. No security issues (docs-only).

---

Blocker

None.

---

Should-fix

1. @NonSkippableComposable bundled with @DontMemoize as a lambda-memoization exception — In performance.md, scoring.md (4 places), and eval 13, manual callback remember is still said to matter on @NonSkippableComposable / @DontMemoize paths. Per the SSM docs, @NonSkippableComposable opts the composable out of skipping; @DontMemoize opts a lambda out of memoization. SSM still auto-memoizes lambdas inside a @NonSkippableComposable body. Keeping @NonSkippableComposable in this exception list can still bless unnecessary remember on hot paths. Drop it from the callback-memoization exceptions; keep it only in skippability / opt-out justification rules.

2. Missing layout-only-read nuance in compose-agent — search-playbook.md allows onSizeChanged → state when that state is consumed only in layout/draw (e.g. Modifier.layout { }), and flags only composition-phase reads. performance.md scopes the bad pattern correctly in the detailed bullet, but never states that exception, so agents may over-prescribe custom Layout where a layout-phase read would suffice. One sentence mirroring the audit heuristic would align authoring with scoring.

3. No compose-agent write-mode evals — Evals 12–14 remain audit-shaped; evals.json description was updated but nothing exercises compose-agent while writing code. Drift risk is lower than pre-4.3.1, but the release notes already track this as follow-up — worth doing before the next authoring change.

---

Nit

1. SKILL.md References blurb — references/performance.md still lists “stability, Strong Skipping, lambda modifiers…” but not cross-phase back-writes / false leads; minor discoverability gap.

2. Long Core Instruction bullet — One dense bullet covers cross-phase back-writes, self-invalidation, and false leads; splitting to match performance.md would help agent parsing.

3. SSM lambda scope wording — “Every lambda passed to a composable” is fine for the false-lead case; official docs say “every lambda inside a composable function,” which is slightly broader. Optional tighten for precision.

4. Layout good example — 8.dp.roundToPx() is fine: MeasureScope is a Density receiver. No change needed (contrary to an earlier automated review).

5. docs/release-notes-4.3.0.md — Already carries a historical correction on the unstable-capture bullet (line 27). No further edit required.

---

Summary: Correct direction, good suite alignment, versioning/docs consistent. Before merge, I'd fix the @NonSkippableComposable / @DontMemoize conflation (#1) and optionally add the layout-only-read exception (#2).

_{via pr-review-relay · author: claude}