Agent edits reintroduce null-safety errors in Flutter (Needs Verification) · Greybook

Agent edits reintroduce null-safety errors in Flutter

Hannah Becker · 2026-07-20T01:40:08.129Z

The diff pattern: when the analyzer flags a nullable value, the agent's dominant fix is the bang operator — `user!.profile!.avatarUrl!` — or a late keyword on a field it can't prove gets initialized. The analyzer goes green, the diff looks fine at review speed, and then Crashlytics fills up with `_CastError: Null check operator used on a null value` from paths where the value legitimately can be null (deep links, cold-start restore, race between provider init and first frame). Nine crash groups in two weeks traced to agent-authored bangs. It never reaches for ?., ??, or a proper early return unless prompted.

Needs verificationClaude Code2.4.1Claude Sonnet 5Medium severity1 affectedlast repro 2d ago

Filed by hbecker · Jul 20, 2026

Symptoms: The diff pattern: when the analyzer flags a nullable value, the agent's dominant fix is the bang operator — `user!.profile!.avatarUrl!` — or a late keyword on a field it can't prove gets initialized. The analyzer goes green, the diff looks fine at review speed, and then Crashlytics fills up with `_CastError: Null check operator used on a null value` from paths where the value legitimately can be null (deep links, cold-start restore, race between provider init and first frame). Nine crash groups in two weeks traced to agent-authored bangs. It never reaches for ?., ??, or a proper early return unless prompted.
Likely cause: The bang is the smallest edit that satisfies the analyzer, same shape as the 'use client' problem in Next: minimal-local-edit beats correct-global-reasoning. Flutter makes it worse because so much state is legitimately null during init, so the crash only appears under timing conditions no agent test run hits. Needs verification whether this is materially worse on some models than others — my data is mostly Sonnet.
Affected: observed with claude-code 2.3.5 – 2.4.1 and cursor 1.18.3; suspected agent-general
Environments: Flutter 3.32+ apps with legacy nullable models; worst in codebases mid-migration from pre-null-safety Dart · macOS, Windows (dev); crashes on Android/iOS
Confidence: Medium confidence

✓ What worked

Ban the bang in analysis_options + brief rulepartial
×4
Turned on `unnecessary_null_checks` and a custom lint denying `!` outside tests (avoid_non_null_assertion from custom_lint). Agent now writes ?? fallbacks and early returns because the bang literally won't pass analysis. Crash groups from new code: zero in the two weeks since. Small sample, hence needs-verification.
hbecker · 1w ago

✕ What did not work

Asking for 'proper null handling' in the task prompt
×3
Works for that one task, gone by the next session. Also produced defensive ?. chains on values that can never be null, which is its own review noise. Prompt-level fixes don't stick; lint-level ones do.
renzo.t · 2w ago

Agent	Version	Model	OS	Framework	Outcome	×
Claude Code	2.4.1	Claude Opus 5	macOS 15.5	Flutter	Not reproduced	2
Claude Code	2.4.1	Claude Sonnet 5	macOS 15.5	Flutter	Reproduced	4
Cursor	1.18.3	GPT-5.5	Windows 11	Flutter	Reproduced	2

Regression Testing Review workflows

Agent edits reintroduce null-safety errors in Flutter

Ban the bang in analysis_options + brief rulepartial

Asking for 'proper null handling' in the task prompt

Discussion · 3 comments

Fable 5 vs Opus 5 in Claude Code: early numbers and vibes

Claude Code beginners: CLI, desktop app, web, or IDE extension?

Claude Code keeps rewriting package-lock.json

Best AGENTS.md for SwiftUI production apps

Use Claude to review Codex output, step by step