chore(release): v0.23.0 — 8 bug fixes, cdp_native_errors tool, Android speed parity

Plugin: 0.22.0 → 0.23.0
MCP server (cdp-bridge): 0.17.0 → 0.18.0

Eight bugs fixed (B110-B118), one new MCP tool (cdp_native_errors), major
performance wins on Android. All validated live through a full 8-phase
rn-feature-dev run in workspace (Priority filter row on Tasks screen).

Highlights
- B118/D637: Android `cdp_interact` p95 went from 5343ms (timeout) to 45ms —
  120× faster, zero false-negative timeouts. Whole Android flow: 16.11s →
  7.2s (−55%).
- B114/D642: new `cdp_native_errors` MCP tool reads `simctl log` / `adb
  logcat`. Fills the gap when native-module errors fire before __RN_AGENT
  injects. `cdp_status` auto-hints at it on suspicious state.
- B111/D635: `cdp_connect` now accepts `targetId` + `bundleId` filters to
  disambiguate zombie Expo Go host pages from real app targets.
- B112/D641: `device_snapshot action=open attachOnly=true` skips app launch
  when app is already running — saves ~12s and avoids bundle-race cascades.
- B116/D639: `inferPlatforms` cross-checks `simctl listapps` + `adb pm list`;
  flags same-bundle-on-both with `ambiguousPlatform: true`.
- B117/D638: `device_screenshot` accepts `platform` + inherits from CDP
  target. Session-bound dispatch routes correctly to the right device.
- B115/D640: `cdp_nav_graph` flat-call fix for nested tab dispatch.
- B113/D636: `device_screenshot` stopped sending the unsupported `--format`
  flag; uses `--out <path>` explicitly.
- B110/D630: MCP server version read from package.json at module load;
  sync-versions.sh guards against hardcoded literals in src/.

Refactors
- cdp/connect.ts, cdp/helper-expr.ts, cdp/recovery.ts extracted — CDPClient
  facade shrunk further; every module now has a typed Context interface.
- cdp/state.ts setter-based ResettableState interface replaces unsafe cast.
- utils.ts freshness probe cached 2s per connectionGeneration.
- RingBuffer gains optional indexKey for O(1) getByKey.
- ResultEnvelope has optional structured ToolErrorCode.

Test count: 158 → 249 (+91 this release cycle).

Docs updated: README.md, CHANGELOG.md, docs-site/src/content/docs/changelog.md.
Decisions D630-D642 logged in workspace docs/DECISIONS.md.

Author: Lykhoyda

.claude-plugin/marketplace.json

@@ -9,7 +9,7 @@
     {
       "name": "rn-dev-agent",
       "description": "AI agent that fully tests React Native features on simulator/emulator — navigates the app, verifies UI, walks user flows, and confirms internal state.",
-      "version": "0.22.0",
+      "version": "0.23.0",
       "source": "./",
       "category": "mobile-development",
       "homepage": "https://github.com/Lykhoyda/rn-dev-agent"

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "rn-dev-agent",
-  "version": "0.22.0",
+  "version": "0.23.0",
   "description": "AI agent that fully tests React Native features on simulator/emulator — navigates the app, verifies UI, walks user flows, and confirms internal state.",
   "author": {
     "name": "Anton Lykhoyda",

CHANGELOG.md

@@ -4,6 +4,44 @@ All notable changes to rn-dev-agent will be documented in this file.
 
 Format follows [Keep a Changelog](https://keepachangelog.com/).
 
+## [0.23.0] — 2026-04-16
+
+Major session of correctness and performance fixes surfaced by end-to-end benchmarks and a live feature-dev run. MCP server bumped to 0.18.0.
+
+### Added
+- **`cdp_native_errors` MCP tool** — reads `xcrun simctl log show` on iOS / `adb logcat -d` on Android, parses known native-module / bundle-fetch / FATAL EXCEPTION patterns, dedupes by message body. Fills the gap when `cdp_error_log` / `cdp_console_log` stay empty because native errors fired before `__RN_AGENT` injected. `cdp_status` also emits a suspicion hint pointing at this tool when `connected && !helpersInjected && !hasRedBox && errorCount === 0` (B114/D642).
+- **`targetId` + `bundleId` filters on `cdp_connect`** — disambiguate zombie Expo Go host pages from real app targets (B111/D635).
+- **`attachOnly: true` on `device_snapshot`** — skip app launch when it's already running; verifies via `xcrun simctl spawn booted launchctl list` / `adb shell pidof`. Prevents the ~12s app-restart cascade. Exported `isAppRunning(platform, bundleId, probes?)` helper (B112/D641).
+- **Platform-aware CDP timeouts** — `defaultTimeout(platform)` and `timeoutForMethod(method, platform)` apply a 2× Android multiplier via a single constant `ANDROID_MULTIPLIER`. `CDPClient` routes `Runtime.evaluate` paths through it using `this._connectedTarget?.platform`. iOS unchanged (B118/D637).
+- **`platform` param on `device_screenshot`** — inherits from `client.connectedTarget?.platform` or accepts explicit override. When no active session is open, the wrapper appends `--platform <p>` to agent-device CLI args. Session-bound dispatch remains the canonical path (B117/D638, partial — upstream agent-device CLI ignores `--platform` without a session; workaround via open session).
+- **`simctl listapps` cross-check in platform inference** — `cdp/discovery.ts::inferPlatforms` now reads both `adb shell pm list packages` AND `xcrun simctl listapps booted`; targets installed on both platforms are flagged with `ambiguousPlatform: true`. Readers are injectable for unit testing (B116/D639).
+- **Tab-dispatch fix for `cdp_nav_graph`** — `buildTabNavigateArgs(tab, screen, params)` emits the flat `ref.navigate(tab, params)` when target === tab, nested form when they differ. Prevents self-referential `navigate('TasksTab', { screen: 'TasksTab' })` that left RN stuck on the old tab (B115/D640).
+
+### Fixed
+- **B110: MCP server reports stale version** — server version now read from `package.json` at module load; `sync-versions.sh` gained a regex guard against hardcoded `version:` literals in src/ (D630).
+- **B113: `device_screenshot --format` always rejected** — agent-device >= 0.8.0 doesn't accept `--format`. Refactored into `buildScreenshotArgs()` + thin delegate; now uses `--out <path>` explicitly, extension drives encoding (D636).
+- **Freshness probe caching** — 2s TTL per `connectionGeneration`, WeakMap-keyed. Saves 30-150ms per back-to-back tool call by skipping redundant `__RN_AGENT.__v` round-trips (D631).
+- **Structured error codes on `ResultEnvelope`** — `ToolErrorCode` union (`STALE_TARGET`, `HELPERS_STALE`, `RECONNECT_TIMEOUT`, `NOT_CONNECTED`, `HELPERS_NOT_INJECTED`). Agents can branch on `code` instead of regex on error text. Back-compat preserved (D634).
+- **Extracted `cdp/recovery.ts`** — `probeFreshness()` + `recoverFromStaleTarget()` moved out of `utils.ts`. Replaced error-string matching for stale detection with the `__RN_AGENT.__v` probe as the primary signal (D633).
+- **`RingBuffer` requestId index** — optional `indexKey` extractor builds a parallel `Map<key, item>`; `getByKey(id)` is O(1). Swapped 5 call sites (`event-handlers.ts` ×4, `tools/network-body.ts` ×1) from `findLast` → `getByKey` (D632).
+
+### Refactors
+- **CDP module extraction continued** — `cdp/connect.ts` (213 lines), `cdp/helper-expr.ts`, `cdp/recovery.ts` (99 lines). CDPClient facade shrunk further; every module now has a typed Context interface instead of reaching into the facade directly.
+- **`cdp/state.ts` setter-based `ResettableState` interface** — replaces `as unknown as CDPResettableState` cast. Renaming a private field on `CDPClient` now produces a real TypeScript error.
+
+### Benchmarks validated live
+
+Cross-platform benchmark (Task Power User flow + Priority Filter Row feature):
+- **iOS (iPhone 17 Pro, iOS 26.3)**: 3.37s / 29 calls / 0 failures (`cdp_interact` p50 7ms)
+- **Android (Pixel_9_Pro, API 37) pre-fix**: 16.11s / 32 calls / 3 failures (incl. 5.3s `typeText` timeout)
+- **Android post-fix**: ~7.2s / 24 calls / 0 failures — **55% faster, zero false-negative timeouts** (`cdp_interact` p50 16ms, p95 45ms)
+
+### Test count
+158 → **249** (+91 this release cycle).
+
+### Decisions logged
+D630 through D642 in `rn-dev-agent-workspace/docs/DECISIONS.md`.
+
 ## [0.22.0] — 2026-04-16
 
 ### Added

README.md

@@ -2,7 +2,7 @@
 
 A [Claude Code](https://docs.anthropic.com/en/docs/claude-code) plugin that turns Claude into a React Native development partner. It explores your codebase, designs architecture, implements features, then **verifies everything live on the simulator** — reading the component tree, store state, and navigation stack through Chrome DevTools Protocol.
 
-**52 MCP tools** | **5 agents** | **13 commands** | **148 tests** | **46 best-practice rules** | [Full documentation](https://lykhoyda.github.io/rn-dev-agent/)
+**53 MCP tools** | **5 agents** | **13 commands** | **249 tests** | **46 best-practice rules** | [Full documentation](https://lykhoyda.github.io/rn-dev-agent/)
 
 ---
 
@@ -103,14 +103,22 @@ if (__DEV__) {
 
 ## MCP Tools
 
-52 tools across three layers. [Full reference](https://lykhoyda.github.io/rn-dev-agent/tools/)
+53 tools across three layers. [Full reference](https://lykhoyda.github.io/rn-dev-agent/tools/)
 
 | Category | Count | Examples | Docs |
 |----------|-------|---------|------|
-| **CDP** (React internals) | 25 | `cdp_component_tree`, `cdp_store_state`, `cdp_evaluate`, `cdp_set_shared_value` | [CDP tools](https://lykhoyda.github.io/rn-dev-agent/tools/#cdp-tools) |
+| **CDP** (React internals) | 26 | `cdp_component_tree`, `cdp_store_state`, `cdp_evaluate`, `cdp_set_shared_value`, `cdp_native_errors` | [CDP tools](https://lykhoyda.github.io/rn-dev-agent/tools/#cdp-tools) |
 | **Device** (native interaction) | 14 | `device_find`, `device_press`, `device_fill`, `device_screenshot` | [Device tools](https://lykhoyda.github.io/rn-dev-agent/tools/#device-tools) |
 | **Testing** (E2E + proof) | 13 | `proof_step`, `cross_platform_verify`, `maestro_run` | [Testing tools](https://lykhoyda.github.io/rn-dev-agent/tools/#testing-tools) |
 
+### What's new in v0.23.0 (2026-04-16)
+
+- **`cdp_native_errors`** — surfaces `simctl log`/`adb logcat` entries for errors that fire before `__RN_AGENT` injects (missing native modules, bundle-fetch failures, FATAL EXCEPTIONs). `cdp_status` hints at it automatically when JS-layer signals are blank.
+- **Multi-device correctness** — `cdp_connect` accepts `targetId`/`bundleId` filters to disambiguate zombie Expo Go pages; `inferPlatforms` cross-checks `simctl listapps` and `adb pm list`; `device_screenshot` accepts `platform` and inherits from the connected CDP target.
+- **Android speed parity** — platform-aware CDP timeouts (2× on Android emulator) eliminate `typeText` false-negatives; real Android flow went from **16.1s/3 fails to 7.2s/0 fails** in the same benchmark.
+- **`device_snapshot attachOnly: true`** — skip app launch when already running, saving the ~12s restart cascade.
+- **91 new unit tests** (158 → 249) locking in the extracted `cdp/*` modules, platform filters, and native-error parsers.
+
 ## Architecture
 
 ```

docs-site/src/content/docs/changelog.md

@@ -9,6 +9,54 @@ All notable changes to rn-dev-agent will be documented in this file.
 
 Format follows [Keep a Changelog](https://keepachangelog.com/).
 
+## [0.23.0] — 2026-04-16
+
+Major session of correctness and performance fixes surfaced by end-to-end benchmarks and a live feature-dev run. MCP server bumped to 0.18.0.
+
+### Added
+- **`cdp_native_errors` MCP tool** — reads `xcrun simctl log show` on iOS / `adb logcat -d` on Android, parses known native-module / bundle-fetch / FATAL EXCEPTION patterns, dedupes by message body. Fills the gap when `cdp_error_log` / `cdp_console_log` stay empty because native errors fired before `__RN_AGENT` injected. `cdp_status` also emits a suspicion hint pointing at this tool when `connected && !helpersInjected && !hasRedBox && errorCount === 0` (B114/D642).
+- **`targetId` + `bundleId` filters on `cdp_connect`** — disambiguate zombie Expo Go host pages from real app targets (B111/D635).
+- **`attachOnly: true` on `device_snapshot`** — skip app launch when it's already running; verifies via `xcrun simctl spawn booted launchctl list` / `adb shell pidof`. Prevents the ~12s app-restart cascade (B112/D641).
+- **Platform-aware CDP timeouts** — `defaultTimeout(platform)` and `timeoutForMethod(method, platform)` apply a 2× Android multiplier. iOS unchanged (B118/D637).
+- **`platform` param on `device_screenshot`** — inherits from `client.connectedTarget?.platform` or accepts explicit override. When a device session is open, session-bound dispatch routes correctly (B117/D638).
+- **`simctl listapps` cross-check in platform inference** — `inferPlatforms` reads both `adb shell pm list packages` AND `xcrun simctl listapps booted`; targets on both platforms are flagged with `ambiguousPlatform: true` (B116/D639).
+- **Tab-dispatch fix for `cdp_nav_graph`** — `buildTabNavigateArgs` emits the flat `ref.navigate(tab, params)` when target === tab, nested form when they differ (B115/D640).
+
+### Fixed
+- **B110: MCP server reports stale version** — server version now read from `package.json` at module load; `sync-versions.sh` guards against hardcoded `version:` literals (D630).
+- **B113: `device_screenshot --format` always rejected** — now uses `--out <path>` explicitly; extension drives encoding (D636).
+- **Freshness probe caching** — 2s TTL per `connectionGeneration`, WeakMap-keyed. Saves 30-150ms per back-to-back tool call (D631).
+- **Structured error codes on `ResultEnvelope`** — `ToolErrorCode` union lets agents branch on `code` instead of regex on error text (D634).
+- **Extracted `cdp/recovery.ts`** — `probeFreshness()` + `recoverFromStaleTarget()` moved out of `utils.ts`. Replaced error-string matching with the `__RN_AGENT.__v` probe (D633).
+- **`RingBuffer` requestId index** — `getByKey(id)` is O(1); swapped 5 call sites from `findLast` → `getByKey` (D632).
+
+### Benchmarks validated live
+
+Cross-platform benchmark (Task Power User flow + Priority Filter Row feature):
+- **iOS**: 3.37s / 29 calls / 0 failures (`cdp_interact` p50 7ms)
+- **Android pre-fix**: 16.11s / 32 calls / 3 failures (incl. 5.3s `typeText` timeout)
+- **Android post-fix**: ~7.2s / 24 calls / 0 failures — **55% faster, zero false-negative timeouts** (`cdp_interact` p50 16ms, p95 45ms)
+
+### Test count
+158 → **249** (+91 this release cycle).
+
+## [0.22.0] — 2026-04-16
+
+Proof-capture infrastructure, fast-runner robustness, and the `cdp_set_shared_value` tool for Reanimated animation proofs.
+
+### Added
+- **`cdp_set_shared_value` tool** — Drive Reanimated SharedValue animations by testID for proof captures.
+- **Fast-runner auto-restart** — When fast-runner dies mid-session, `tryFastRunner` automatically attempts one restart.
+- **Reload counter + NativeWind corruption warning** — `cdp_status` warns after 5+ `cdp_reload` calls in a session.
+- **Auto-open device session in Phase 5.5** — Pipeline mandates opening a device session at verification start.
+
+### Fixed
+- **B103**: `cdp_navigate` false success — fallback now verifies target screen exists.
+- **B106**: `device_scroll`/`device_swipe` deadlock on Reanimated screens.
+- **B107**: `deviceId` parsing for agent-device v0.8.0.
+- **R2**: `device_screenshot` ignores requested path.
+- **R5**: Scroll amount semantics diverge.
+
 ## [0.19.2] — 2026-04-13
 
 ### Fixed

scripts/cdp-bridge/package-lock.json

@@ -1,6 +1,6 @@
 {
   "name": "rn-dev-agent-cdp",
-  "version": "0.17.0",
+  "version": "0.18.0",
   "type": "module",
   "main": "dist/index.js",
   "scripts": {