Imported from menstood/unity-pipeline-commands-skill (SKILL.md). Install upstream with npx skills add menstood/unity-pipeline-commands-skill. Copyright stays with the author.

Unity CLI (com.unity.pipeline) — drive the Unity Editor over HTTP

The com.unity.pipeline package (v0.3.x, experimental) runs a local HTTP server inside the Unity Editor (and optionally inside development Player builds). Clients — the unity CLI, CI, or this agent — execute registered commands against it. This is how you can actually compile, test, screenshot, and edit scenes without a human in the Editor.

Prerequisite: package installed?

Check Packages/manifest.json for "com.unity.pipeline". If missing, report it and ask before installing — installing edits the manifest and triggers a domain reload. (Options once approved: unity pipeline install with the CLI, or Package Manager → Install package by name → com.unity.pipeline, version blank = latest.) The Editor must be running with the project open for the server to exist.

Connecting (do this first, every session)

1. Read the port descriptor the Editor writes:
   • Editor: <projectPath>/Library/Pipeline/.unity-pipeline-port
   • Player: <workingDirectory>/.unity-pipeline-runtime-port
2. It's JSON — extract port and evalToken (also has pid, projectName, unityVersion, mode, lastHeartbeat).
3. Call http://127.0.0.1:<port>/... with header Authorization: Bearer <evalToken>. Always use 127.0.0.1, not localhost (IPv6 pitfalls). Missing/wrong token → 401.

Ports: Editor 7800–7849 (tests 7850–7899), Runtime 7900–7949 (tests 7950–7999). Token is 256-bit CSPRNG, base64, regenerated after every domain reload — re-read the descriptor after any recompile/package change.

Treat evalToken as a secret: never print, log, or commit it (descriptors live under Library/, which Unity projects git-ignore); read it only to build the auth header and re-read after reloads instead of storing it.

Project identity check (before any mutation)

A reachable server + valid token does NOT prove you're talking to the right project. Before the first mutating command:

1. Confirm the descriptor's projectPath matches the project root you're working in.
2. Confirm pid is alive and lastHeartbeat is recent — a stale descriptor means a dead Editor.
3. With multiple Editors open, target explicitly (unity command --project-path <path> ...); never mutate when identity is ambiguous.

Endpoints

PowerShell recipe (Windows)

$d = Get-Content "Library/Pipeline/.unity-pipeline-port" | ConvertFrom-Json
$h = @{ Authorization = "Bearer $($d.evalToken)" }
Invoke-RestMethod -Uri "http://127.0.0.1:$($d.port)/api/exec" -Method Post -Headers $h `
  -ContentType 'application/json' `
  -Body (@{ command = 'editor_status'; parameters = @{} } | ConvertTo-Json -Depth 10)

The unity CLI — prefer this when installed (verified with v1.0.0-beta.2)

Run unity --version at session start — the CLI is an experimental beta and behavior shifts between releases; unity --help and bare unity command are the source of truth for the installed version. The CLI does port/token discovery itself — no descriptor reading, no 401-after-domain-reload handling:

unity pipeline list                                 # projects + server port + reachability
unity command                                       # no name = list ALL commands with params (best discovery)
unity command <name> [--arg value ...]              # run any command, e.g.:
unity command find_assets --type SceneAsset --name Init
unity command --project-path <path> <name> ...      # explicit project
unity command --runtime MyGame.exe runtime_status   # attach to a dev Player
unity command --runtime-path "C:\Builds\MyGame" runtime_status

Verified CLI behaviors:

• ObjectRef args take inline JSON — in PowerShell wrap in single quotes with no backslash escapes: --target '{"hierarchyPath":"/Foo"}' (escaped \" inside single quotes passes literal backslashes and breaks resolution).
• A plain string for an ObjectRef param is treated as a hierarchyPath shorthand: --target '/Foo'.
• For machine parsing prefer --format json --no-banner over the default human/tab layout; treat any non-zero exit code as failure and read stderr. (Observed with v1.0.0-beta.2: default output is tab-separated Command Success Result Parameters; failures print Error: lines and exit 6.)

Fall back to the raw HTTP recipe above only when the CLI isn't installed.

Response envelope

Every command returns a CommandExecutionResponse: { success: bool, command: string, result: <payload>, executionTimeMs: long?, error: string? }. Authoring commands put an AuthoringResult in result — assets get assetPath/guid/fileId/type/globalId; scene objects get instanceId/hierarchyPath/type/globalId. Chain those identities into the next command's ObjectRef args.

ObjectRef — how you point at things

Any param typed as a reference accepts ONE of: {"globalId": "..."}, {"path": "Assets/Foo.mat"}, {"guid": "...", "fileId": ...}, {"instanceId": 12345}, {"hierarchyPath": "/Root/Child"}.

Safety conventions

confirm/dry_run are command-level conventions, not a global guarantee — non-destructive commands and custom [CliCommand]s may implement neither. Check the command's schema (unity command / GET /api/commands) before relying on them.

• dry_run: true → validate + preview, mutate nothing (wins even if confirm is also true).
• confirm: false (default) on destructive/overwriting commands → command refuses; pass confirm: true to apply.
• Scene/GameObject mutations are Undo-able (grouped as one Ctrl+Z step). AssetDatabase writes, file writes, package ops, and settings are NOT undoable.
• Bare paths (Materials/Stone) resolve under the authoring root (default Assets; read/change via get_authoring_root / set_authoring_root). .. traversal is rejected. Explicit Assets/... paths work as-is.
• Most scene/asset authoring commands are blocked during Play mode.

Async pattern (fire → poll)

Long operations return immediately; poll their status command until completed:

A domain reload (recompile, package add/remove, scripting-backend change) can drop the in-flight HTTP reply — treat a dropped connection as "poll the status command with a fresh token".

If the Editor is unfocused/minimized it may stop ticking — call set_autotick (enable:true) first; it auto-disables after a domain reload.

Common recipes

• Verify a code change compiles: recompile → poll recompile_status → get_console_logs --severity error.
• Run play-mode smoke test: editor_play → get_console_logs / screenshot --view game → editor_stop.
• Run the test suite: run_tests --mode editor --filter <name> (or async_tests:true + poll test_status).
• See what the scene looks like: screenshot (game/scene view) or capture_game_view (returns base64 PNG, params width/height/camera/save_path).
• Inspect a scene: get_scene_hierarchy → nodes carry instanceId+hierarchyPath → feed into get_component_properties / set_component_properties.
• Escape hatch: eval --code "<C#>" runs arbitrary C# via Roslyn (5s default timeout); eval_file for a .cs file; menu --path "Assets/Reimport All" runs any menu item (omit path to list all). Policy: prefer a registered command when one exists; use eval primarily for read-only inspection; ask before eval/menu calls that write files, start processes, or touch the network; verify the effect afterwards.

Field-tested gotchas (verified on this project, 2026-07)

• /api/exec body key is parameters, not params. The wrong key doesn't error on commands whose args are all optional — it silently binds nothing. If a command behaves like you passed no args, check this first.
• Entering play mode rotates the evalToken (domain reload). Never reuse a token read before editor_play — split scripts at every play/recompile/package boundary and re-read the descriptor after.
• get_console_logs can return total: 0 even when the game is running. Don't treat an empty result as "no errors" — cross-check by tailing %LOCALAPPDATA%\Unity\Editor\Editor.log.
• eval is the best truth source for runtime state. list_open_scenes reflects the edit-mode setup; in play mode query SceneManager via eval instead.
• Interpolated strings inside eval code work fine; return a single string built with StringBuilder for multi-part dumps.

Full reference (read as needed)

• references/commands-content.md — assets & files, scenes, GameObjects & components, prefabs, scripts, animation/animator/timeline, materials & shaders (every command + every parameter).
• references/commands-editor.md — baking (lighting/navmesh/occlusion), search & selection, capture, build/compile/tests, project settings, package manager, editor lifecycle & observability, runtime/player commands, hot reload.
• references/connectivity-and-setup.md — descriptors, ports, auth, runtime (Player) server setup, log file locations.
• references/extending.md — writing your own [CliCommand]s, the authoring contract (ProjectPaths/ObjectResolver/AuthoringUndoScope), safety conventions, hot-reload authoring ([HotReload]/[HotReloadWithOverrides]), test architecture.