Remove the stdio transport and the legacy TCP bridge

[vbucc]

Removes the stdio transport and the legacy framed TCP bridge it used to reach Unity. This fork only uses HTTP, and stdio was still the Python default — Server/Dockerfile and manifest.json both omitted --transport and inherited it, while Unity already defaulted UseHttpTransport = true. Net −9,000 lines across 29 deleted files.

Python

• Delete Server/src/transport/legacy/ (unity_connection, port_discovery, stdio_port_registry). _extract_response_reason / _is_reloading_response are rescued into models/unity_response.py — the HTTP path still uses them.
• Drop the send_fn escape hatch from send_with_unity_instance; it is now a single HTTP path. 60 files / 72 call sites.
• Convert the four sites that bypassed the shim entirely (sync_tool_visibility_from_unity, script_apply_edits get_sha/read, the manage_script reload sentinel) rather than only dropping their imports — deleting those would have silently broken them.
• --transport is retained, accepting only http, so client configs written by older packages keep working on upgrade.

Unity

• Delete StdioBridgeHost (1134 lines), StdioTransportClient, StdioBridgeReloadHandler, PortManager. TransportManager collapses to one client; TransportMode goes.
• Remove UseHttpTransport. With stdio gone, a stale false from a prior install would have silently disabled auto-start and reload-resume.
• Remove ConfiguredTransport.Stdio. A leftover stdio config now reports MissingConfig so the UI prompts re-Configure, and StdIoVersionMigration → LegacyClientConfigMigration rewrites it on the next version bump. (Kept, not deleted — it is the only thing that migrates upgrading users off stdio.)

Test harness

tools/local_harness.py drove Unity entirely over the TCP bridge, so it is ported to HTTP: it starts its own MCP server, boots Unity against it via McpCiBoot.StartHttpForCi, discovers the instance through /api/instances, and sends through /api/command.

Port isolation — 8080 is the everyday local server port, and Unity's StartLocalHttpServer terminates whatever process owns the port it is about to bind. The harness therefore:

• never binds 8080 or $UNITY_MCP_HTTP_PORT (--http-port 8080 is a hard exit 2);
• points its Editor at its own server with the UNITY_MCP_HTTP_URL env override rather than EditorPrefs, which are shared machine-wide per Unity version;
• no longer sets UNITY_MCP_ALLOW_BATCH, the flag that lets a batch Editor reach StopLocalHttpServerInternal at all;
• starts and stops only the server process it spawned.

Verified across seven runs against a live 8080 server — including one killed mid-flight and one where another Unity project connected to 8080 concurrently. The listener PID never changed and /health stayed 200.

Four bugs in the port were found only by running it end-to-end: /api/instances field names, a blocked main thread in McpCiBoot (Unity registered but could not execute), the /api/command envelope colliding with the UTF job status in _poll_utf, and missing reload retry (/api/command routes to PluginHub.send_command, which has no retry wrapper, while the UTF legs trigger domain reloads).

Also

• Delete e2e-bridge.yml and claude-nl-suite.yml — both pinned to a 2021.3 container that cannot open this 6.5 project, so both were already unrunnable. claude-nl-suite.yml was workflow_dispatch-only and gated nothing.
• Delete tools/stress_*.py — they implement the deleted framed socket protocol directly.
• Docs: CLAUDE.md, Server/README.md, the transports page, install/clients/guides, zh translations, the homepage cards, the bug-report template.
• Regenerate website/docs/reference/ — this drift is pre-existing on main (verified in a clean worktree); CI's docs-generate check runs because this PR touches Server/src/services/**, so it had to be fixed here.

Gate

• Python: 1502 passed, 3 skipped.

• Harness unit tests: 139 passed (including 18 for port isolation, 7 for reload retry, 6 for envelope handling).

• Unity compile (6000.5.5f1): pass.

• Unity EditMode, native -runTests: pass.

• Harness smoke leg: 7/7, exit 0.

• Harness EditMode leg: 3 failures, all pre-existing —

  EditorStateCache_IsChanging_TracksTransitionWindowNotWholeSession
  ReadConsoleTests.HandleCommand_Get_PreservesMultilineMessageBody
  ReadConsoleTests.HandleCommand_Get_Works
  

  Proven two independent ways: Unity's native runner (no MCP transport at all) passes all three, and running unmodified origin/main with its original stdio harness in a separate worktree produces a character-identical failure set. They are artifacts of driving tests through an MCP run_tests job — the activity phase reads running_tests, and the console carries session traffic — and they predate this change. Not fixed here; they are unrelated to stdio and would block any Unity-touching PR.

🤖 Generated with Claude Code

[vbucc]

Correction: the gate note above is wrong where it says Unity's native runner passes all three pre-existing failures. It passes only EditorStateCache_IsChanging.

The two ReadConsoleTests failures were never MCP-specific — they reproduce under the plain -batchmode -nographics -runTests runner with no MCP involvement at all (native full suite at that commit: 1039 tests, 972 passed, 2 failed, exactly those two). Unity does not record a log emitted inside a Test Framework test into the console's LogEntries buffer when headless, so those tests only ever passed in a GUI editor. HandleCommand_Clear_Works was passing only by luck, on log entries left by unrelated tests; it failed as soon as the class was run alone.

All three are fixed in #41 (b3ece1b0). The native suite is now 0 failures and the harness gate is green with no documented exception.