check-feature-support

Answers "can CARLA do X here, and is there a vetted procedure for it" before you promise anything, and detects client/server/tree version skew — a support matrix over the whole collection (covered by

SKILL.md

GitHub

Is this feature usable, and is there a procedure for it?

Paths. scripts/… and references/… below are relative to the directory holding this SKILL.md. Your working directory is the user's project, not that directory, so prefix them with its absolute path or the command is not found.

The failure this prevents is specific: an agent asked to use an uncovered CARLA feature reconstructs a plausible procedure from UE4-era documentation, and it is wrong. Every row in this skill's [broken] list used to be a working, well documented UE4 workflow. Reciting those steps on ue58-dev produces a silent no-op, an empty result, or a crashed server.

So answer from this skill instead of from memory, and when the answer is "no vetted procedure exists", say that and offer to investigate.

Instructions

Progress:
- [ ] Step 1: bash scripts/check_env.sh   (no blockers — it reports probe depth)
- [ ] Step 2: matrix — find the feature's status
- [ ] Step 3: broken — read this BEFORE promising anything on 0.10.0
- [ ] Step 4: probe — confirm against this machine's build and server
- [ ] Step 5: version — only if a measurement looks impossible (skew check)
source scripts/env.sh

bash scripts/support.sh matrix     # the table (works with no CARLA)
bash scripts/support.sh broken     # verified-broken index, with evidence
bash scripts/support.sh probe      # this machine: flags, tree, live server
bash scripts/support.sh version    # the identity tuple + skew warnings

version: am I reasoning about the thing I am talking to?

"Which CARLA is this" has five answers that can disagree — installed client, server, source tree, content branch, engine — and a disagreement is silently wrong rather than loudly wrong. version prints all five plus the build flags, and warns on the two skews that cause wasted debugging:

  • client ≠ server. CARLA usually still connects, then misbehaves in ways that look like your own bug.
  • installed client older than the tree HEAD, or than a wheel built from it. You read a source change that the client you are calling does not contain. This is not hypothetical: it happened while these skills were being written, and it is why one API was briefly reported as removed when it had only moved class.

Note what version deliberately does not claim: that matching versions mean matching behaviour. ENABLE_ROS2, ENABLE_RSS and ENABLE_OSM2ODR are compile-time, and the map-layer breakage is a property of the content branch — two servers reporting 0.10.0 can differ completely. Use matrix for behaviour; use version to establish that your measurements are even about the right target.

Engine lines, and the version number

UE 5.8 (ue58-dev) is the line that continues, and it is the one released as CARLA 1.0. UE 5.5 (ue5-dev) is an earlier revision of the same line, not a fork: PythonAPI/carla/src/Sensor.cpp is byte-identical between them. So the verdicts below apply to both engines unless a row says otherwise, and the five features 5.5 lacks are catalogued in [[check-ue5-limitations]].

On the version string. Every measurement here was taken on a pre-release ue58 build, which reports 0.10.0 — that is the number version will print on those, and the number 5.5 reports today. Read 0.10.0 below as naming the UE5 line: 1.0 renamed it, it did not change the behaviour underneath. Where a verdict is genuinely engine-specific the row says so.

The four verdicts, and what each obliges you to do

Verdict Meaning What to do
[skill] a vetted procedure exists here invoke that skill; do not improvise
[works] verified working, no skill use the API directly, say it is unvetted
[untested] in the build, never exercised here say so, offer to investigate, do not write steps
[broken] verified broken or removed on 0.10.0 say so, name the alternative if one exists

What probe adds that the matrix cannot

Two features are gated by compile-time flags that no runtime check reveals, and every -D option must be repeated on each re-configure — so the CMake cache is the only honest source:

  • ENABLE_RSS — sensor.other.rss appears in the blueprint library either way, and cannot function without it. On the tree measured here it is OFF.
  • ENABLE_OSM2ODR — the OSM→OpenDRIVE half of digital twins. OFF here.
  • ENABLE_ROS2 — no native ROS 2 interface at all without it, and it cannot be switched on at runtime ([[run-autoware-ue58]], [[build-carla-ue58]]).

probe also reports tree-visible facts that decide whether a documented procedure can work at all: whether Co-Simulation/ exists (it does not on ue58-dev), whether RecastBuilder sits in Util/DockerUtils/dist (it does not, which is why map-import navmesh generation silently produces nothing), whether the Autoware integration is shipped, and whether the multi-GPU sources are there.

Examples

Example 1: "grab the GBuffer normals from the camera"

broken first. On 0.10.0 listen_to_gbuffer() asserts on Stream.has_value() and takes the server down, while the whole API surface still exists. Tell the user it crashes the simulator; offer sensor.camera.normals instead, which is a real sensor.

Example 2: "hide the buildings so I can see the road network"

matrix shows map layers as [broken]. enable_environment_objects is the mechanism that works — that is [[toggle-env-objects]], a [skill] row.

Example 3: "set up SUMO co-simulation"

[broken]: ue58-dev has no Co-Simulation/ directory at all, while UE4 ships Sumo, PTV-Vissim, Chrono and Carsim bridges. Do not port the UE4 instructions blind — say the tooling is absent and ask whether they want it investigated.

Example 4: "can I use the RSS sensor?"

probe. The blueprint exists; if ENABLE_RSS=OFF the honest answer is "not in this build — it needs a rebuild with -DENABLE_RSS=ON, and nobody here has exercised it afterwards".

Example 5: "add a vehicle to ue58"

[untested] — to be done; there is no skill for it in this version. The pipeline is known and gets as far as a vehicle that registers, spawns and reports four wheels in get_physics_control(); it does not move, because the physics asset does not survive an editor restart on the skeletal mesh. Deferred to a later version rather than shipped half-working.

So: say it is not supported yet, offer to investigate, and do not adapt [[import-carla-ue58-walker]]'s procedure. Walkers bind to a shared skeleton; vehicles need a per-vehicle skeleton plus a persisted physics asset, which is the one step still unsolved.

Troubleshooting

Problem: the matrix disagrees with the CARLA docs Cause: the public docs describe the UE4 tree. Solution: trust probe and the evidence in references/matrix.md — they were taken from the tree in front of you. Docs are the weaker source here.

Problem: probe reports nothing about build flags Cause: no root variable set, or the tree is unconfigured, or CARLA_PRESET names a preset with no cache. Solution: export CARLA_UE58_ROOT (or CARLA_UE4_ROOT), and set CARLA_PRESET to a preset you have actually configured.

Problem: a feature is missing from the matrix Cause: the matrix covers what has been examined; CARLA is larger than that. Solution: say it is unexamined — which is not the same as unsupported — and offer to check the tree. Do not extrapolate from a neighbouring row.

Problem: a measurement contradicts the source you just read Cause: version skew — most often an installed client older than the tree. Solution: version. On the machine these skills were written on it caught a wheel built one day after the installed client, which is exactly how a moved API gets misreported as a removed one.

Problem: this skill is stale after an upgrade Cause: every verdict is stamped to ue58-dev HEAD 718efd7cc / 0.10.0. Solution: re-run the measurements in references/matrix.md; each row records how it was established so it can be re-checked cheaply.

Outputs

Nothing is written and nothing is started: all four modes are read-only. matrix and broken print text; probe prints what this machine's variables, CMake cache, tree layout and running server report; version prints the identity tuple and exits non-zero for nothing — read its MISMATCH lines, they are advisory but they invalidate measurements.

Per-feature evidence — file paths, line numbers, the measurements behind each [broken] row, and what a future skill would have to verify to promote a row — is in references/matrix.md.