Flutter Gemma PR Review

Run comprehensive PR review with 10 parallel agents — 4 platform-specific + 6 general-purpose.

Usage

/review-pr 198        # Review PR by number
/review-pr            # Review current branch vs main

Process

Step 1: Get the diff

If PR number provided:

gh pr view {number} --json title,body,files --jq '.title'
gh pr diff {number} > /tmp/pr-{number}.diff

If no PR number:

BRANCH=$(git branch --show-current)
PR_NUMBER=$(gh pr list --head "$BRANCH" --json number -q '.[0].number')
if [ -n "$PR_NUMBER" ]; then
  gh pr diff "$PR_NUMBER" > /tmp/pr-branch.diff
else
  git diff main...HEAD > /tmp/pr-branch.diff
fi

Step 2: Identify changed areas

This is a Dart pub workspace monorepo — all code lives under packages/<pkg>/. From the diff, detect which package(s)/area(s) are affected:

Packages (packages/<pkg>/):

• flutter_gemma/ — core: registry, contracts, shells, ModelSource, slim native plugin (its android/ ios/ host only the bundled channel)
• flutter_gemma_litertlm/ — .litertlm FFI engine; native/litert_lm/ build scripts, lib/src/ffi/, hook/build.dart
• flutter_gemma_embeddings/ — LiteRT C API embeddings (isolate worker)
• flutter_gemma_mediapipe/ — .task MediaPipe; owns pigeon (lib/pigeon.g.dart) + Kotlin/Swift + web JS
• flutter_gemma_rag_qdrant/ — native RAG over the official qdrant_edge UniFFI SDK. Since 2.0.0 the package owns NO native code and NO hook: the engine and its Native Assets hook live in the SDK. lib/src/ only.
• flutter_gemma_rag_sqlite/ — sqlite-vec vec0 KNN on all six platforms; native via package:sqlite3 FFI, web via package:sqlite3/wasm.dart (wa-sqlite was dropped in 1.1.0)
• flutter_gemma_speech/ — opt-in STT (moonshine / Whisper / Parakeet) + TTS (Matcha / Qwen3 / Inflect) over the LiteRT C API; shares the litertlm bundle
• flutter_gemma_agent/ — opt-in SKILL.md agent skills over the function-calling loop (no Web)
• flutter_gemma_builtin_ai/ — OS models: Gemini Nano via ML Kit GenAI (Android, minSdk 26), Apple Foundation Models (iOS/macOS, sharedDarwinSource). Owns its own pigeon.
• genkit_flutter_gemma/, genkit_hybrid/ — Genkit integration packages (Dart; no native)
• flutter_gemma/example/ — example app + integration_test/ E2E

Native (per package): packages/<pkg>/android/, ios/, windows/, native/, hook/build.dart Web: packages/*/lib/src/web/, packages/*/web/ (JS interop, WASM) Desktop: packages/*/lib/desktop/ + FFI in flutter_gemma_litertlm Site: website/ — Jaspr landing + docs (deployed to fluttergemma.dev; CI .github/workflows/firebase-hosting-merge.yml) Repo-level: .github/workflows/, .claude/skills/, root pubspec.yaml (workspace + melos)

Ground rules for every agent — REVIEWERS DO NOT WRITE

Put this in every agent prompt, verbatim. It is not boilerplate; it is the lesson from a review round that had to be redone from scratch.

READ-ONLY. Do not write, create, edit, delete, move or rename ANY file in the
repository — including test files, scratch files and probe files. Do not run
mutation testing. Do not run `flutter test`, `flutter analyze`, `dart format`
or `flutter pub get` (they write into .dart_tool and build outputs). Do not run
git add / restore / checkout / stash / clean / commit. Reading is unlimited:
cat, sed, grep, find, `git show`, `git diff`, `git log`, and ~/.pub-cache.
If you want something verified by execution, DESCRIBE THE EXPERIMENT precisely
— setup, the call, what to print — and stop. The caller will run it and hand
you the output.

What this prevents. Reviewers told to "verify by mutation and restore afterwards" did exactly that, in the shared working tree, on top of UNCOMMITTED fixes. Those fixes were silently overwritten twice; one agent's mid-mutation state got picked up by a git add, so git diff showed the agent's revert as if it were HEAD, and it was nearly committed. It was caught only because the file's md5 changed twice in twenty seconds while nothing was being edited. Every test result taken during that window was meaningless.

Two more rules that follow from it:

1. Commit before you review. Reviewers read a moving tree otherwise, and their findings are anchored to a state that no longer exists.
2. Run them ONE AT A TIME when acting on findings. Parallel is right for a first sweep of a frozen commit. It is wrong once you are editing between reports: read one report, decide what you accept, apply it, commit, then start the next. The agents are advisors — the decision to change code is yours, and it is not delegable to ten of them at once.

Acting on a finding — the part that goes wrong

A report is a list of HYPOTHESES, not a list of defects. Converting one straight into an edit is how a review round ends with more bugs than it started.

Measured on one PR, one day: reviewers found six real pre-existing defects, and acting on them introduced five new ones. None shipped — four were caught by the next round, one by its own test — but four of the five came from the same omission, and it is not "this was hard":

The loop, per finding, in this order:

1. Reproduce it. Print the wrong value. A finding you cannot reproduce is not yet a defect — and twice this day a confident report was wrong on measurement.
2. Read the neighbours. If you change state, grep every reader of that state. If you change a file, find what consumes it. If you call a tool, check what it does to its output.
3. Fix.
4. Mutate the test. Break the production line the test names, and watch it go red. If it stays green the test is decoration — delete it or fix it, but do not keep it.
5. Only then take the next finding.

Step 4 is not optional and not a formality. Tests written after a fix "obviously" pass; that is the problem, not the reassurance.

Two more that cost real time:

• A green narrow check is not a green wide one. flutter test in a package passed while flutter analyze packages/ had 17 errors, because the ambiguity was only visible from the wider scope. Run the wide one before saying "clean".
• A fix that is not observable to the consumer is not a fix. A corrected binary was still not reaching builds, because the staging guard compared file SIZE and the correction happened not to change it. Verify at the point the user consumes it, not at the point you edited it.

Step 3: Launch ALL agents in parallel

Launch all 10 agents simultaneously using the Agent tool. Each agent gets:

• The full diff (or list of changed files)
• The affected platform(s)
• Its specific review checklist (below)

CRITICAL: All agents MUST run in parallel via a single message with multiple Agent tool calls.

Agent Specifications

Every path named in an agent prompt below was verified to exist. If one stops resolving, that is the finding — fix the skill before running the review, because an agent pointed at a missing directory returns a clean report.

Agent 1: Android native

subagent_type: android-architect

You are reviewing the Android native layer of flutter_gemma, a Dart pub workspace
monorepo. Inference itself is NOT in Kotlin — `.litertlm` runs through Dart FFI.
Kotlin exists in exactly three packages; confirm with
`find packages -path '*/android/src/main/kotlin' -name '*.kt'` before you start.

- packages/flutter_gemma/android/.../FlutterGemmaPlugin.kt — SLIM. Hosts only the
  `flutter_gemma_bundled` channel: file ops plus the litertlm NPU
  `getNativeLibraryDir`. If a change adds inference logic here, that is the
  finding.
- packages/flutter_gemma_mediapipe/android/ — the only real engine layer:
  own pigeon (PigeonInterface.g.kt), PlatformServiceImpl, InferenceModel, and
  engines/{InferenceEngine,EngineFactory,InferenceSession,EngineConfig}.kt plus
  engines/mediapipe/. EngineFactory handles `.task`/`.bin`/`.tflite` ONLY — it
  throws on `.litertlm` with a message pointing at the Dart FFI client. That
  throw is correct behaviour, not a bug.
- packages/flutter_gemma_builtin_ai/android/ — ML Kit GenAI / AICore. Declares
  `minSdkVersion 26`; an app on a lower floor fails the manifest merger.

CHECKLIST
1. NO KGP IN THE PLUGINS: since #440 none of the THREE android/build.gradle
   files applies `kotlin-android`, declares `ext.kotlin_version`, or carries the
   KGP classpath. Flutter's own Gradle plugin applies KGP to any plugin
   subproject that does not (FlutterPluginUtils `detectApplyingKotlinGradlePlugin`),
   which is why `flutter: '>=3.44.0'` is load-bearing. Re-adding a guard like the
   old `agpMajor < 9 || !builtInKotlinOn` is a REGRESSION, not a fix. What still
   matters is the #360 lesson: the `kotlin { compilerOptions { jvmTarget } }` block
   and its comment must be byte-identical in all three — a fix in one is a half-fix
   that ships under its own version number.
2. MANIFEST: the `<uses-native-library>` entries (libOpenCL.so, -car, -pixel,
   libvndksupport.so, libcdsprpc.so) live in the CORE plugin manifest and merge
   into consumer apps. libvndksupport is load-bearing for the OpenCL ICD on
   Android 12+ (#324) — removing it degrades GPU to WebGPU and can hard-freeze
   Mali.
3. minSdk: `.litertlm` needs 30 (API-30-only Bionic symbols, #265); builtin_ai
   needs 26; core declares 24. A change to any floor must be consistent with what
   the code actually dlopens.
4. PIGEON: *.g.kt files are generated — never hand-edited. Regenerated from the
   package's own pigeon.dart.
5. NATIVE LIBS: arm64-v8a only for .litertlm/embeddings/vision. A change that
   implies other ABIs ship broken APKs.
6. Coroutine scope, cancellation and prompt accumulation in the MediaPipe session.

Report CRITICAL / IMPORTANT / MINOR with file:line. Skip style nits.

Agent 2: Apple native (iOS + macOS)

subagent_type: swift-reviewer

You are reviewing the Apple native layer of flutter_gemma. Note the layouts
differ per package — verify with `find packages -name '*.swift' | grep -v example`
rather than assuming a single convention:

- packages/flutter_gemma/{ios,macos}/flutter_gemma/Sources/flutter_gemma/ —
  SwiftPM layout (NOT ios/Classes/). Slim plugin: bundled channel only.
- packages/flutter_gemma_mediapipe/ios/Classes/ — classic layout, real engine:
  FlutterGemmaMediaPipePlugin, PlatformServiceImpl, InferenceModel, pigeon .g.swift
- packages/flutter_gemma_builtin_ai/darwin/ — one source tree for iOS + macOS via
  `sharedDarwinSource: true`. Apple Foundation Models.

CHECKLIST
1. Swift 6 concurrency: Sendable conformance, actor isolation, and no
   captured-mutable-state across the pigeon boundary.
2. Podspec versions: four first-party podspecs exist and drift independently —
   core ios, core macos, mediapipe ios, builtin_ai darwin. Each must match its
   OWN package version.
3. iOS floor since #441: core and builtin_ai declare 15.0; only
   flutter_gemma_mediapipe declares 16.0. A litertlm-only app is NOT held to 16 —
   that was the bug. Both the podspec and the Package.swift must carry the same
   number for a package that has both (mediapipe has no Package.swift).
4. Entitlements: extended-virtual-addressing and increased-memory-limit for large
   models; on macOS `cs.disable-library-validation` is required for dlopen of the
   ad-hoc-signed companion frameworks — and it must be in BOTH DebugProfile and
   Release entitlements.
5. No `Podfile post_install` symlink step should be reintroduced: those lib*.dylib
   symlinks caused App Store rejection ITMS-90432 (#245). Any bundled dylib needs
   `vtool` minos 13.0 or the upload is rejected (ITMS-90208).
6. Generated pigeon .g.swift is never hand-edited.

Report CRITICAL / IMPORTANT / MINOR with file:line.

Agent 3: Web

subagent_type: general-purpose

You are reviewing the web layer of flutter_gemma. Two independent engines have
web arms, plus embeddings and RAG:

- packages/flutter_gemma_mediapipe/lib/src/web/ — `.task` via @mediapipe/tasks-genai
- packages/flutter_gemma_litertlm/lib/src/web/ — `.litertlm` via @litert-lm/core.
  EARLY PREVIEW: text only. No vision, audio, thinking, function calling or LoRA.
- packages/flutter_gemma_embeddings/ web arm — LiteRT.js, not the C API
- packages/flutter_gemma_rag_sqlite/ — package:sqlite3/wasm.dart + a custom
  sqlite3.wasm with vec0 linked in, which the APP copies into its own web/ dir
- packages/flutter_gemma/lib/web/ — the shared web shells and model source

CHECKLIST
1. CONDITIONAL IMPORT / STUB DRIFT — the highest-value check here. FFI clients
   have `*_stub.dart` counterparts that `flutter analyze` and `flutter test` do
   NOT type-check on the host. Signature drift only surfaces at
   `flutter build web`. If a real signature changed, its stub must change too.
   This is how a web break shipped in 0.15.0 with green analyze and green tests.
2. dart:io / dart:ffi must not reach the web graph. Check the conditional export.
3. The three required web/ assets are not auto-injected — the app copies them:
   cache_api.js (default cacheApi storage), opfs_helper.js (streaming), and
   litert_embeddings.js (web embeddings). A change that needs a new global must
   document the script tag.
4. Storage modes: cacheApi (default, <2GB), streaming (OPFS, large models),
   none. Web is GPU-only — MediaPipe has no web CPU backend.
5. CDN pins: @mediapipe/tasks-genai and @litert-lm/core versions must agree
   between the code and any documented script tag.

Report CRITICAL / IMPORTANT / MINOR with file:line.

Agent 4: Desktop / FFI and native assets

subagent_type: general-purpose

You are reviewing desktop inference and the native-asset pipeline of
flutter_gemma. Desktop is Dart FFI directly into the LiteRT-LM C API — there is
NO JVM, NO gRPC, NO separate server process and no proto layer. If the diff or a
task description mentions any of those, it predates 0.14.0.

- packages/flutter_gemma_litertlm/lib/src/ffi/ — the FFI client, generated
  bindings, and the inference model
- packages/flutter_gemma_litertlm/hook/build.dart — the SOLE hook that owns the
  shared libLiteRtLm bundle. embeddings and speech consume it transitively and
  have no hook of their own.
- packages/flutter_gemma_litertlm/native/litert_lm/ — build_*.sh, patch_c_api.sh,
  stream_proxy.c
- packages/flutter_gemma/{linux,windows}/ — thin plugin registration C++
- packages/flutter_gemma/lib/desktop/ — the registry-dispatch shell

CHECKLIST
1. HOOK CHECKSUMS: a bundle `version:` bump requires all seven per-platform
   SHA256 entries updated, and the same value must appear in the released
   tarball, in checksums_litertlm.txt, and in the hook. A stale txt sent a user
   down the wrong path in #316.
2. NEVER re-upload an existing native-v* tag. tar is not reproducible, so the
   published SHA256 can never be recovered — it breaks every user already on a
   plugin version referencing that tag.
3. windowsExtraLibs / androidExtraLibs and the CI allow-list are ONE SET IN TWO
   PLACES. A name staged by CI but absent from the hook extracts to the cache and
   is never bundled — that is how 78.8 MB of the Intel NPU stack went missing at
   v0.16.0 while every CI assertion passed.
4. Bazel `--define` names rot silently: Bazel accepts an unknown define and
   builds the default. Grep the tree being built before trusting one.
5. stream_proxy.c probes the callback ABI at runtime via dlsym/GetProcAddress
   because v0.15.0 changed the shape with no version symbol. Do not replace that
   with a compile-time branch.
6. `stage()` in the hook is Apple-only on purpose (an Xcode directoryTreeSignature
   cycle); staging on Windows splits companion DLLs and hangs cancel/close.
7. Any claim that a native check "passed" must name the artifact it inspected —
   nm/otool on a missing path prints nothing and reads as success.

Report CRITICAL / IMPORTANT / MINOR with file:line.

Agent 5: Flutter Architect

subagent_type: flutter-architect

Prompt: Review the PR diff for flutter_gemma — a multi-platform Flutter plugin for on-device AI inference. Focus on: plugin architecture (platform channels via Pigeon), SOLID principles, ModelSource sealed class design, handler chain pattern (NetworkSourceHandler, AssetSourceHandler), dependency injection (ServiceRegistry), platform abstraction layer. Check separation of concerns between install-time identity (modelType, fileType) and runtime configuration (maxTokens, preferredBackend). Read CLAUDE.md for project conventions.

Agent 6: Flutter Coder

subagent_type: flutter-coder

Prompt: Review the changed Dart files in flutter_gemma for code quality. Check: null safety, proper async/await patterns, Stream handling (no leaks, proper cancellation), Message class usage (isUser: true for user messages), PreferencesKeys constants (no inline string keys), proper close()/dispose() in finally blocks, type safety with ModelSource sealed classes. Read CLAUDE.md for coding standards — especially "No Inline String Keys" rule.

Agent 7: Codex Review (Second Opinion)

subagent_type: codex:codex-rescue

Not the copilot CLI. That binary is GitHub Copilot, billed against a different quota entirely, and it answers 402 You have exceeded your monthly quota on every model regardless of the ChatGPT plan — an earlier revision of this skill called it and lost a whole review round to that.

Prompt:

Second-opinion review of PR #{number} for flutter_gemma. Read the diff with
`gh pr diff {number}` (or `git diff main...HEAD` if no PR), and read CLAUDE.md
for the project's conventions.

READ-ONLY. Do not write, create, edit, delete or move any file in the
repository, and do not run `flutter test` / `flutter analyze` / `dart format` /
`flutter pub get` — they write into .dart_tool and build outputs. Do not run
git add / restore / checkout / stash / clean / commit. Reading is unlimited.
If you want something verified by execution, DESCRIBE the experiment precisely
and stop; the caller will run it and hand you the output.

Key project context:
- A Dart pub workspace monorepo: core flutter_gemma plus opt-in packages under
  packages/. Core registers no engine; engines and backends are passed to
  FlutterGemma.initialize().
- .litertlm inference is Dart FFI into the LiteRT-LM C API on every native
  platform, desktop included. There is NO JVM, NO gRPC, no separate server
  process and no proto layer — removed at 0.14.0. Kotlin exists only in the
  MediaPipe and builtin_ai packages plus a slim core plugin; MediaPipe never
  handles .litertlm.
- Engine selection is by the DECLARED ModelFileType via canHandle(spec), NOT by
  sniffing the file name. installModel defaults fileType to .task, so a
  .litertlm model must declare it explicitly or it is routed to MediaPipe.
- ModelSource sealed class: NetworkSource, AssetSource, BundledSource, FileSource.
- Installation stores identity (modelType, fileType); runtime accepts config
  (maxTokens, preferredBackend).
- maxTokens is the CONTEXT WINDOW, not the reply length. Below 1024 the
  .litertlm KV cache allocation fails; the engine clamps up with a warning. To
  cap a reply use maxOutputTokens on the session.
- Error handling: NO silent fallbacks. Throw or return an error; never swallow
  in a catch. gemmaLog is debug-only, so anything reported only through it is
  unreported in release builds.
- No inline string keys — use PreferencesKeys constants.
- Generated files are never hand-edited: pigeon *.g.dart / *.g.kt / *.g.swift,
  and the ffigen bindings.

Cross-check specifically: engine routing by declared fileType; conditional-import
stub drift against the real signatures (analyze does not catch it, only
`flutter build web` does); and platform-specific limitations.

Focus on bugs, logic errors, security, dead code, silent error swallowing, race
conditions in async/streaming code, and memory leaks (unclosed sessions/models).
Be concise — only real issues, each with file:line and a concrete failure
scenario. Skip style nits. Categorize CRITICAL / IMPORTANT / MINOR.

If no PR number, detect via: gh pr list --head $(git branch --show-current) --json number -q '.[0].number'

Agent 8: Code Reviewer

subagent_type: pr-review-toolkit:code-reviewer

Prompt: Review the PR for adherence to project guidelines in CLAUDE.md. Focus on recently changed files. Key rules: no inline string keys, proper PreferencesKeys usage, no AI attribution in commits, sessions/models always closed, Message(isUser: true) for user messages.

Agent 9: Type Design Analyzer

subagent_type: pr-review-toolkit:type-design-analyzer

Prompt: Analyze any new or modified types/interfaces in the PR for encapsulation, invariant expression, and design quality. Focus on: ModelSource sealed class hierarchy, EngineConfig/SessionConfig, InferenceEngine/InferenceSession interfaces, EngineCapabilities, TestModelConfig.

Agent 10: Silent Failure Hunter

subagent_type: pr-review-toolkit:silent-failure-hunter

Prompt: Check all changed files for silent failures, inadequate error handling, catch blocks that swallow errors, and inappropriate fallback behavior. Key concern areas: image conversion fallbacks (should throw, not return original bytes), tryEmit() that drops errors silently, cancelGeneration() that swallows exceptions, download/checksum verification gaps.

Step 4: Collect and Deduplicate

After all agents complete:

1. Collect all findings from all agents
2. Deduplicate — if two agents report the same issue (same file, same line, same concern), keep the one with better description
3. Categorize by severity: CRITICAL > IMPORTANT > MINOR
4. Group by file path within each severity

Step 5: Generate Report

Save to test_reports/pr-reviews/pr-{number}-review.md (or pr-branch-{branch}-review.md):

# PR Review: #{number} — {title}

**Branch:** {branch}
**Date:** {date}
**Reviewers:** 10 agents (4 platform-specific + 6 general)
**Platforms affected:** {platforms}

## Critical Issues
{blocking issues that must be fixed before merge}

## Important Issues
{should be fixed, but not blocking}

## Minor Issues
{nice-to-have improvements}

## Passed Checks
{list of checks that passed cleanly}

## Summary
- Critical: X
- Important: Y
- Minor: Z
- Recommendation: APPROVE / REQUEST CHANGES / NEEDS DISCUSSION

Step 6: Output Summary

Print a concise summary to the user:

• Total findings by severity
• Top 3 most important issues
• Recommendation (approve/changes requested)
• Path to full report file