Unreal Bridge

UnrealBridge

Execute Python directly inside a running UE 5.3+ editor. Auto-discovers via UDP multicast (239.255.42.99:9876); the TCP data port is OS-assigned per editor.

Preconditions

If bridge.py returns discovery: no UnrealBridge editors found, walk these in order — don't troubleshoot Python or firewalls first:

1. Plugin installed in <UEProject>/Plugins/UnrealBridge/ (use sync_plugin.bat from this repo; user must edit its DST= once). Check: <UEProject>/Plugins/UnrealBridge/UnrealBridge.uplugin exists.
2. Plugin enabled — check <UEProject>/<Project>.uproject "Plugins" block for {"Name":"UnrealBridge", "Enabled":false} and flip if present.
3. Editor up and ready — bridge.py ping returns "ready": true. false means MainFrame still loading; wait 10–60s.

Last resort if multicast is blocked (corp VPN, virtual NIC): --endpoint=127.0.0.1:<port> from the editor log line LogUnrealBridge: Listening on 127.0.0.1:<port>. Python 3.7+ stdlib only.

Waiting for the editor to become ready (post-launch / post-relaunch)

After launching/relaunching the editor, poll readiness with Monitor (streams progress events, recommended) or Bash with run_in_background: true (one completion event). Don't write a foreground for/sleep countdown — the harness blocks long leading sleeps and you'll miss readiness.

Paste this verbatim as the command for either tool:

end=$(( $(date +%s) + 300 ))
i=0
until python .claude/skills/unreal-bridge/scripts/bridge.py --json --timeout 3 ping 2>/dev/null | grep -q '"ready": *true'; do
  now=$(date +%s)
  [ "$now" -ge "$end" ] && { echo "[wait] TIMEOUT after 300s"; exit 1; }
  i=$((i + 1))
  [ $((i % 3)) -eq 0 ] && echo "[wait] still booting ($((end - now))s left)"
  sleep 10
done
echo "[wait] READY"

Then call one of:

Monitor(description="UE editor ready-poll", command=<block above>,
        timeout_ms=360000, persistent=false)

Bash(description="Wait for UE editor ready", command=<block above>,
     run_in_background=true, timeout=360000)

Locks (don't change without reading):

• Grep "ready": *true — TCP-up ≠ MainFrame-ready. success:true, ready:false means exec calls will be rejected; ping success alone is not enough.
• Ping FIRST then sleep — no leading sleep (harness blocks long ones).
• i % 3 echo gate — caps notifications at ~10 over 5 min so Monitor doesn't trip its event-flood auto-stop.
• end deadline must stay below the tool timeout_ms/timeout so you get [wait] TIMEOUT rather than a silent harness kill.
• Path is relative to repo root; ${CLAUDE_SKILL_DIR} is not reliably set in Monitor/Bash subshells — don't substitute it here.
• After the loop returns, do one foreground bridge.py ping before real work — bg success is past tense.
• Multi-editor host: append --project=<name> to the ping inside the loop.
• Skip this entirely when invoking rebuild_relaunch.py — it polls internally and prints [rebuild] bridge is ready. on success.

Bridge CLI

python "${CLAUDE_SKILL_DIR}/scripts/bridge.py" [options] <command> [args]

Optional flags: --project=<name|path> (disambiguate when >1 editors run; or env UNREAL_BRIDGE_PROJECT), --endpoint=host:port, --token=<secret>, --timeout=<s>, --json, --no-preflight.

Workflow

1. Always ping first.
2. Default to exec --stdin heredoc, NOT exec-file. A heredoc is the right mode for ~95% of one-shot work — no temp file to name, no cleanup, prompt stays self-contained, no risk of dangling scripts in $TEMP / .tmp / project root. Only reach for exec-file when you genuinely intend to re-run the same script multiple times (iterating on a fix, comparing runs). One-shots like "find X, list Y, build a report" → heredoc. If you find yourself writing with open("/tmp/foo.py", "w") followed by bridge.py exec-file /tmp/foo.py, stop and rewrite as a heredoc.
3. --json for parseable output.
4. Exit codes: 0 success · 1 runtime/transport · 2 bad CLI args · 3 AST preflight rejected.
5. If exec hangs: from a separate terminal try gamethread-ping (high latency = GT mid-exec, queue will drain) or resume (BP breakpoint).

Multi-line example:

python "${CLAUDE_SKILL_DIR}/scripts/bridge.py" exec --stdin <<'EOF'
from unreal_bridge import Asset
paths, _ = Asset.search_assets_in_all_content(query="Hero", max_results=5)
for p in paths:
    print(p.export_text())
EOF

Windows path gotcha (Bash tool only): unquoted backslashes get eaten by bash word-splitting — python G:\Claude\UnrealBridge\.claude\skills\unreal-bridge\scripts\bridge.py exec ... arrives as G:ClaudeUnrealBridge.claudeskillsunreal-bridgescriptsbridge.py and fails with "No such file or directory". Use forward slashes (G:/Claude/...) or single-quote the path ('G:\Claude\...'). The body inside <<'EOF' ... EOF is literal and unaffected — only the arg before the heredoc matters. PowerShell tool is unaffected.

API surface — use the wrapper module first

from unreal_bridge import Asset, Level, Blueprint, Editor, Anim, Material, PoseSearch, Chooser, ...
paths, _ = Asset.search_assets_in_all_content(query="Hero", max_results=20)

The wrapper has 21 classes (one per UnrealBridge*Library) with kwargs-only signatures — positional args raise TypeError immediately, no UE round-trip. This is the structural fix for positional-arg-order hallucinations. Regenerate after C++ header changes via python tools/gen_manifest.py.

Fallback: raw unreal.UnrealBridge*Library.foo(...) works (preflight catches errors), but prefer the wrapper.

AST preflight (automatic — safety net only)

Every exec* call runs an AST preflight in the bridge client BEFORE sending. Reads scripts/bridge_manifest.json and rejects locally (exit 3, no UE round-trip) when:

• Function/library name doesn't exist (with did-you-mean)
• Required args missing / unknown kwarg / param given twice
• unreal.BridgeXxx.YYY enum member doesn't exist

Type validation is NOT done by preflight. Wrong asset paths / wrong scope still surface at runtime — references prose still matters for those.

Treat preflight as a backstop, not a planning tool. If you write code expecting preflight to autocorrect your guesses, you'll produce noisy output and still hit runtime tracebacks for the half it can't catch (USTRUCT field names, Python attribute access, EditDefaultsOnly violations). See "Verify before you call" below.

Bypass with --no-preflight (rare). Preview with bridge.py preflight <path>.

Common semantic traps (preflight does NOT catch)

Verify before you call — don't lean on preflight

Preflight is a safety net, not a planning tool. Writing code from imagined API / field / attribute names and waiting for preflight to tell you the right one produces noisy output (4-8 [WARN] per script, often still ending in a runtime traceback) and wastes round-trips. Verify FIRST, then write the script. Four axes:

1. Bridge function names — grep scripts/bridge_manifest.json or read the relevant references/bridge-*-api.md. Don't translate the C++ name into snake_case in your head and assume.
2. Bridge USTRUCT return-value fields — return types like BridgeGameplayEffectInfo keep a minimum set of fields (not exhaustive); UE 5.x evolves struct shape between versions. Look up the fields in bridge_manifest.json structs section, OR run a one-off probe print([a for a in dir(obj) if not a.startswith('_')]) BEFORE writing the consumer code.
3. UE object attributes — never <obj>.<attr>. Always get_editor_property (and respect protected: / EditDefaultsOnly rejection). Use Level.list_actor_properties(actor_name=...) or Level.list_class_properties(class_path=...) to list first.
4. Function signatures — when unsure: unreal.UnrealBridgeXxx.fn.__doc__ first line is the signature. Cheaper than a guess + correction round-trip.

Symptom you're in the bad loop: your last few exec attempts produced [WARN] lines from preflight followed by AttributeError tracebacks. Stop, run one heredoc with dir() / manifest grep, then write the real script.

Never silently bail to raw unreal.*

If a bridge call returns empty / errors, suspect your parameters (wrong scope / path / filter), not the wrapper. Falling back to unreal.AssetRegistryHelpers etc. bypasses preflight, audit logging, and often walks 100k–2M registry entries.

Preflight emits [WARN] on known raw-fallback patterns and shows the bridge equivalent inline. To look up before writing: bridge.py suggest [pattern]. Patterns currently mapped:

If you genuinely believe no bridge function covers your case, ask the user before reaching for raw unreal.*.

Reference index — read these for usage / traps / workflow patterns

Signatures are now mechanically enforced (preflight). References carry semantic traps, scope warnings, and workflow patterns preflight can't catch.

Pawn control is non-obvious — read bridge-gameplay-api.md first. Driving the player (sticky IA_Move, apply_look_input, navigating to a moving target, holding an input over time) has hard constraints a fresh API read won't reveal: bridge.exec runs on GameThread so time.sleep inside one exec freezes the engine and stops the sticky ticker; continuous steering must run from a reactive register_runtime_timer callback, not a Python while loop; IA_Move is camera-relative and the forward-axis convention varies per project. The "Pattern: chase a (possibly moving) target and stop on arrival" section has the working template.

Asset lookup by name defaults to search_assets_in_all_content(name, max_results). When the user names an asset without a path, do not call unreal.AssetRegistryHelpers.get_assets_by_path('/Game', recursive=True) and filter — that walks 100k–2M+ entries and times out. The full search_assets form needs BridgeAssetSearchScope.ALL_ASSETS (not PROJECT) when the asset might live in a plugin mount (/PluginName/...). PROJECT covers /Game only; using it for a plugin asset returns [] silently. Valid scope members: ALL_ASSETS, PROJECT, CUSTOM_PACKAGE_PATH — there is no GAME_FOLDER.

Blueprint authoring policy (MUST read before BP write ops)

非必要不写蓝图 Node 和 Graph. Agent-authored BPs are visibly weaker than C++ on readability/maintainability/perf. Default response to "用蓝图实现 X":

1. State the tradeoff. Bridge has full graph-write primitives but agent-Graphs are weaker than C++.
2. Offer alternative: a C++ UFUNCTION(BlueprintCallable) with thin BP wrapper, or another bridge surface (Material / Anim / UMG / DataTable / Level).
3. Re-confirm explicitly: "你仍然希望我用蓝图 Node/Graph 实现，还是改走 C++？"
4. Only proceed on explicit insistence ("就用蓝图写"). Otherwise redirect.

Exceptions (skip the dance):

• Pure data writes (set_blueprint_variable_default etc.) without new nodes
• Bulk automation explicitly framed as such ("扫 100 个 BP 改默认值")
• Standing authorization in CLAUDE.md or current conversation

Hard line: any op spawning/connecting/removing nodes, adding events, or creating functions needs the confirmation dance.

Blueprint review loop (mandatory after BP authoring)

After authoring/modifying a BP graph (and the user confirmed they want it):

1. plan        — list events / functions / local vars in text first
2. build       — add_*_node / connect_graph_pins / add_blueprint_variable …
3. auto_layout — MANDATORY after ANY graph mutation; use 'pin_aligned' (3-exec flow) or 'exec_flow' (bulk).
                 See bridge-blueprint-api.md for the exact open→sleep→layout sequence.
4. lint        — lint_blueprint(...) and resolve every warning/error
5. collapse    — collapse_nodes_to_function on any LongExecChain finding
6. straighten  — straighten_exec_chain on main rails (after 'exec_flow' only)
7. comment     — add_comment_box + set_comment_box_color per section (>10 nodes = mandatory)
8. compile     — compile_blueprints clean

• Lint is non-optional. Fix every warning/error or comment-box-justify it.
• Name things. Rename UnnamedCustomEvent / CustomEvent_0 to intent (OnHealthChanged).
• Comment boxes are section titles. Any graph >10 nodes needs at least one comment box with a meaningful title ("1. Validate inputs", "2. Apply damage") and an appropriate preset color (Section, Validation, Danger, Network, UI, Debug, Setup).
• Pick the right layout strategy. auto_layout_graph(... 'pin_aligned' ...) uses DFS-ordered leaves + downstream-driven Y alignment + live Slate geometry — matches hand-laid BP shape (.then rails horizontal, siblings stacked in one column). Requires the three-exec flow (open graph → sleep → layout) to read live widgets. Use 'exec_flow' (Sugiyama-lite, layer-center Y) for bulk tidy on large graphs where pixel-accuracy matters less, then follow with straighten_exec_chain on the main rail.
• Compact by spacing, not by node. pin_aligned treats h_spacing as a loose upper bound and internally uses DataHSpace = max(15, h/3) and ExecGap = max(30, h/2) — so h_spacing=100 renders tighter than the raw number suggests. Pass smaller values (40, 32) for max density.
• Size predict before spawning. If placing several nodes in a row by hand, call predict_node_size for each kind first so X offsets don't overlap.
• Post-layout geometry reads. After auto_layout_graph runs, Slate widgets don't refresh NodePosX/Y until they tick — so get_rendered_node_info returns pre-layout pin coords in the same exec. For crossing-detection / wire-length audits right after layout, read get_node_layout(bp, fn, guid).pos_* (authoritative from node model) and estimate pin Y as pos_y + 40 + 22 × dir_index. Only pay the open-graph + sleep cost when you specifically need Slate-accurate coords.
• auto_insert_reroutes is intentionally NOT in the loop — empirically it produces many-knot routing that reads worse than the original. Opt in per-graph if a specific case needs it.

Non-ASCII output (CJK / Greek / emoji)

Bridge is UTF-8 byte-perfect end-to-end since 2026-05-04. Mojibake = display, not data. Confirm with hex:

print(got.encode('utf-8').hex(), '==', '测试'.encode('utf-8').hex())

If hex matches, it's a Windows cp936/cp1252 console issue — fix the display, not the bridge:

Verified: identifiers + literals in script source, exception messages/tracebacks, DataTable FName keys, asset paths with Chinese folders, save→restart→read persistence. Korean/Japanese use the same UTF-8 path — verify with hex if anything looks off.

**Never write temp files to C:\ root** — use $env:TEMP\… or /tmp/….

Safety Rules

• NEVER delete assets / actors / files without explicit confirmation.
• NEVER modify or overwrite assets without describing the change first.
• Wrap state-changing ops in unreal.ScopedEditorTransaction.
• On failure: show full traceback. Don't silently retry destructive ops.

Notes

• Python interpreter is persistent — variables carry between exec calls.
• print() returns to bridge; unreal.log() only to UE Output Log.
• Structured data: import json; print(json.dumps(...)) with --json.
• If ping fails: editor not running, plugin not enabled, or wrong project (preconditions above).
• Timeout: retry with --timeout 120.

Source: TornLux/UnrealBridge — distributed by TomeVault.