v0.22.0 — Server self-identification, command rename, Windows scene-folder fix

No wire/protocol change (1.0.0, 102 catalog commands, 0 drift).

Added

• Server self-identification. list_unity_instances now returns server: { package, version } alongside
  registryDir, and the startup log (stderr) states the package, version and registry directory. This tells a client
  which server package answered — an unrelated package shares the unity-editor-mcp bin name.

Changed

• The installed command is now unity-editor-mcp-server (was unity-editor-mcp, a name also installed by the
  unrelated unity-editor-mcp npm package). npx @burakaydinofficial/unity-editor-mcp is unaffected. A global install
  now provides unity-editor-mcp-server; configs that invoke the bare unity-editor-mcp command must switch to it (or
  to the npx form).
• One shared HostPlatformInfo OS detector for both the registry directory (InstanceRegistry) and path-containment
  case handling (PathContainment), detected once per process.

Fixed

• create_scene / save_scene failed on Windows for a folder that didn't exist yet ("Failed to save scene").
  Path.GetDirectoryName returns \-separated paths on Windows and the folder-creation loop split only on /, so
  no folder was created. Paths are now normalized first (one shared helper for both commands). Regression tests
  added; scene-test assertions now print the handler's error message on failure.
• Registry-directory fallback parity: with USERPROFILE/HOME missing or empty, the C# side fell back to "."
  (and treated an empty value as set) while Node uses env.X || homedir() — so in such environments the editor would
  publish to a different directory than the server reads. C# now falls back to the OS user-profile directory too.
  Parity tests on both sides.
• The registry-directory log line added in 0.21.5 fired on every domain reload; it now logs once per editor session
  (and again only if the directory changes).

Removed

• Dead Node derivePort / DEFAULT_PORT_RANGE (the Node mirror of the editor's port derivation — only reachable
  through the removed UNITY_PROJECT_PATH branch). The editor keeps its own DerivePort for its listen port.

Notes

• Breaking for global installs that run the bare unity-editor-mcp command — use unity-editor-mcp-server or the
  npx @burakaydinofficial/unity-editor-mcp form. Both packages ship at 0.22.0. The README install URL is now pinned to
  the release tag.

v0.21.5 — Registry-directory parity (Windows editor visibility)

Fixes a discovery bug where a Unity editor could be invisible to the server on the same machine — with no error — and removes a dead configuration knob. No wire/protocol change (1.0.0, 102 catalog commands, 0 drift). Both packages ship at 0.21.5.

Fixed

• The editor and server could resolve DIFFERENT registry directories. Same rules, but different OS detection — C# via RuntimeInformation.IsOSPlatform, Node via process.platform. On a disagreement (e.g. RuntimeInformation misreporting under Unity's Mono while HOME is set), the editor published its descriptor to a directory the server never read → invisible editor, no error. C# now detects the OS with primitives stable on every Unity Mono down to the 2019.4 floor (Path.DirectorySeparatorChar / Environment.OSVersion.Platform + a macOS probe), and both sides are pinned to the SAME per-platform path by parity tests.
• The editor now logs its resolved registry directory once at startup, so a mismatch is diagnosable instead of silent.

Removed

• UNITY_PROJECT_PATH env resolution — vestigial under the 3-tool surface (every editor is targeted per-call by an explicit instance; list_unity_instances enumerates the whole registry). Pinning it only risked collapsing the multi-editor design to one project. UNITY_PORT and the discovery registry are unchanged.
• The README's "alternative (if globally installed)" bare-unity-editor-mcp client config (bin-name collision that can launch the wrong server). Use npx @burakaydinofficial/unity-editor-mcp.
• Dead internal helper isSameHost.

Gates: Node 304/0, dotnet Core 157/0, compat-lint / meta-check / drift all green.

v0.21.4 - test-runner: refuse-while-compiling + self-identifying results

More production feedback on the test runner. No wire/protocol change (1.0.0, 102 catalog commands, 0 drift).

Fixed

• run_tests refuses while the editor is compiling/importing (COMPILING) instead of silently dropping the run — Unity fires no RunFinished for a run started mid-compile, which then wedged the state. Wait with get_compilation_state waitForIdle:true, then retry.

Added

• Self-identifying test results — run_tests returns a runId; get_test_results returns the stored run's runId + testMode, and (given expectRunId) a runIdMismatch flag, so a caller can no longer be silently served a different run's or mode's results. runId + mode are journaled, surviving domain reloads and editor restarts. testCount now reports the real matched count.

Most of the feedback was already fixed in 0.21.1–0.21.3 (stale runner, .meta pollution, compile-error visibility, results-across-reloads) or false positives (save_scene exists; get_editor_info exposes editor state). The deeper timeout item — calls landing during compile/reload — needs a wire-contract change (an editorState envelope + busy-fast-return), scoped in the roadmap and deferred as a deliberate design.

Full details in CHANGELOG.md.

v0.21.3 - test-runner reliability: self-healing run state + run-and-wait

Acts on production feedback that the test runner could get stuck reporting "running" when no run was active, with no way to stop or clear it. No wire/protocol change (1.0.0, 102 catalog commands, 0 drift).

Fixed

• Self-healing run state — the IsRunningTests latch could stick true (a missed RunFinished, an empty run, or an interrupted run) and persist across domain reloads, so a later agent saw "running" having started nothing. It's now reconciled against a callback heartbeat: believed only while set AND a Test Runner callback fired within ~5 min (progress, not duration — a long suite is never falsely cleared). Stale state self-heals.
• run_tests won't latch a no-match run (Unity may fire no RunFinished for an empty run — the likely wedge trigger); returns "nothing to run".
• cancel_tests clears a stuck EditMode state — a live run is still refused (poll get_test_results), but a stale latch is reset so run_tests is unblocked; force:true resets regardless. (Was: EditMode cancel refused unconditionally → unrecoverable wedge.)
• Honest get_test_results messaging; real testCount.

Added

• get_test_results waitForCompletion — server-side poll until the run finishes, tolerating the domain-reload reconnect, returns the final results + summary (mirrors 0.21.2's get_compilation_state waitForIdle).
• force on run_tests / cancel_tests.

Aside (feedback): save_scene + save_assets are the dedicated save tools — File/Save via execute_menu_item is a valid fallback, not the only way. Full details in CHANGELOG.md.

v0.21.2 - production-feedback polish: compile-wait, menu logs, composite writes

Acts on feedback from an agent using the tool in production. No wire/protocol change (1.0.0, 102 catalog commands, 0 drift).

• get_compilation_state waitForIdle mode — a server-side poll that blocks until compilation + asset import finish, reconnecting across the domain reload, so an agent can refresh_assets then wait instead of sleep-and-hope. Compile errors are in messages (type "Error") + errorCount — the reliable compile-error channel.
• execute_menu_item now returns the console output emitted during the invoke (logs + errorCount). Menu methods are void — no return value — but their logs are captured, so an agent needn't a separate read to see what a menu did.
• SerializedValue.Write recurses into a nested [Serializable] struct/class, so set_serialized_properties and modify_serialized_array insert can set a composite element/field in one call (was "Generic is read-only").
• CI meta-check — fails if a package .cs/.asmdef lacks its .meta (a missing .meta warns in every consumer's editor).

Deferred with reasons (see CHANGELOG.md): compile-error console mapping (use get_compilation_state), batch asset creation (a generic batch deserves its own cycle), asset snapshot/rollback (git + Undo cover it).

v0.21.1 - SDK modernization, fresh-dimensions hardening, and Unity 6 CI

Maintenance release — no wire/protocol change (still 1.0.0, 102 catalog commands, 0 drift) and no new tool surface.

• Dependency modernization: the Node server's @modelcontextprotocol/sdk goes 0.5.0 → 1.x (25 versions of drift). This clears the high-severity DNS-rebinding advisory that surfaced in consumers' npm audit — not reachable here (the server is stdio-only; that advisory only affects the HTTP/SSE transport), but it's real hygiene. No server.js change was needed (the low-level Server + StdioServerTransport API is stable across the jump). npm audit now reports 0 vulnerabilities (runtime + dev). Node floor unchanged (>=18).
• Security: update_script / delete_script are now contained to Assets/ (were gated on the whole project root — a raw local TCP client, which bypasses the Node layer, could otherwise overwrite .git/config, Packages/manifest.json, or any existing file). read_script → Assets/ + Packages/.
• Robustness (fresh-dimensions audit — scale / DoS / fuzzing): caps on modify_serialized_array resize and capture_screenshot dimensions (OOM), analyze_scene_contents triangle counting without allocating index buffers, Roslyn fetch timeouts, an O(1) framing-recovery scan, and a reapStale TOCTOU re-verify.
• manage_packages: explicit update action + packageId validation.
• CI: Unity 6.0 (6000.0.78f1) added to the floor-matrix — now cold-compiles + EditMode-tests 2019.4 / 2020.3 / 2021.3 / 2022.3 / 6000.0 (297/297 green on Unity 6).

Full details in CHANGELOG.md.

v0.21.0 - live-editor E2E harness, reload-recovery, and an 8-round hardening campaign

The largest correctness pass in the fork's history.

• Live-editor E2E harness — drives the real client → server → editor chain to cover the ~two dozen tools untestable in the EditMode floor matrix (play mode, script recompile, refresh), verifying each effect through both the bridge and an independent channel.
• Bridge reload-recovery (ADR 0007) — the editor now force-closes accepted client sockets on domain reload, so the server reconnects in ~1s instead of hanging. This was latent for every client after any reload (incl. script recompile), not just the harness.
• ~125 verified faults from a pre-release bug hunt + an eight-round loop-until-dry adversarial audit (convergence 31 → 27 → 13 → 7 → 5 → 2 → 0). Classes: silent data corruption/loss (Flags-enum ordinal, material shape-vs-type, composite zero-fill, ambiguous-scenePath wrong-object), false success (ignored API results, no-ops, partial applies, play-mode-evaporated mutations), security (capture-screenshot write-anywhere, create_script into Library/, File/Exit menu, invoke-policy over-exposure), reads-with-side-effects (material/mesh getters instancing on a read), prefab-stage scene-scoping, partial-mutation-under-error, and Node connection/wire robustness (project-path verification, oversize-command handling, reconnect-loop resurrection, Roslyn crash-safety).

No protocol change (wire stays at 1.0.0, 102 catalog commands, 0 drift). CI-verified across 2019.4 / 2020.3 / 2021.3 / 2022.3 (298 EditMode + 152 dotnet + 314 server unit tests) with the full live suite green.

Full details in CHANGELOG.md.

v0.20.6 - second-review fixes + docs refresh

A deeper second code review (a 5-lens panel with adversarial verification, over the released 0.20.5) plus a full documentation audit. CI-green on Unity 2019.4 / 2020.3 / 2021.3 / 2022.3 (296 EditMode tests).

Fixed:

• create_gameobject components no longer reports a rejected component as added — adding e.g. a 2nd Rigidbody (a [DisallowMultipleComponent] violation, where Unity returns null rather than throwing) now fails all-or-nothing (INVALID_STATE), with a regression test. (A real false-success the first review missed.)
• get_hierarchy now declares its emitted truncated / maxNodes result fields (the strict schema omitted them — the pagination signal was undiscoverable).
• set_active_scene's sceneName selector is now captured by the H5 audit log; manage_tools lost a dead cache branch.
• Four aftermath tests that re-read through the same path as the handler now assert against a genuinely independent source.

Documentation: corrected the compatibility floor across CONTRIBUTING / the setup + floor-testing guides / the roadmap — they still claimed 2020.3 / C# 8; the floor is 2019.4 / C# 7.3, CI-verified on all four LTS versions. Plus refreshed command/test counts, marked completed roadmap/ADR items done, added the Roslyn code-intelligence sidecar to the roadmap, and documented the deferred round-7 friction (FR1/FR3/FR5) + known placeholders.

Same commit on npm (@burakaydinofficial/unity-editor-mcp) and OpenUPM (com.burakk.unity-editor-mcp). See the CHANGELOG for the full list.

v0.20.5 - dogfood rounds 6-8, reviewed, + aftermath test suite

Dogfood rounds 6–8, an independent code review (no Critical/High), and a rigorous aftermath sweep (34 mutating tools driven read→act→independent-re-read→assert, zero false-successes) — plus a deterministic aftermath EditMode test suite covering the whole testable-in-EditMode tool surface. 295 EditMode tests, CI-green on Unity 2019.4 / 2020.3 / 2021.3 / 2022.3.

New / improved: set_active_scene; create_gameobject components array (now declared in the catalog, so it's discoverable); multi-scene get_hierarchy / analyze_scene_contents (loadedScenes + per-root scene); inactive GameObjects reachable by path (no longer stranded when deactivated); serialization tools (inspect_serialized_object / set_serialized_properties / modify_serialized_array) see the open prefab stage.

Fixed: play_game refuses in -batchmode (was bricking the bridge); create_gameobject rejects an empty name; delete_gameobject returns NOT_FOUND on no match (was a 0-count "success"); manage_tools reports real installed package versions; set_active_scene verifies the effect (Unity's bool return is unreliable); set_serialized_properties / modify_serialized_array — componentType match scoping + force/CAS bypass.

Same commit on npm (@burakaydinofficial/unity-editor-mcp) and OpenUPM (com.burakk.unity-editor-mcp). See the CHANGELOG for the full list.

v0.20.4 — prefab-stage shadowing fix (npm + OpenUPM re-sync)

Re-sync release — brings npm and OpenUPM back in lockstep on the corrected code.

Fixes the prefab-stage same-name shadowing bug present in 0.20.3 (the Unity/C# half): the by-path GameObject
resolvers now resolve the open prefab stage first, so a same-named main-scene object can't shadow the stage
object during an in-stage add_component / modify_gameobject / delete_gameobject. Found by code review; guarded
by a [UnityTest]. CI-green on Unity 2019.4 / 2020.3 / 2021.3 / 2022.3.

(0.20.3 reached OpenUPM with this bug, and npm without a 0.20.3 at all; 0.20.4 is the same corrected commit on both
registries — 0.20.3 = b182bba, 0.20.4 = 056a99a everywhere.)