Reviewed the two headers, the CMake wiring, and the design doc against the tree. This is a clean, tightly-scoped pure addition — no implementor wired, so there's no behavior to regress, and the four files are exactly what the description claims (no scope creep). Thanks for keeping the "core seed" small rather than dumping all 140 touchpoints at once; that makes it genuinely reviewable.

Conventions — all good:

• namespace AetherSDR matches the core-wide convention (314 files under src/core/).
• #include "core/backends/RadioCapabilities.h" resolves via aethercore's src/ PUBLIC include dir (CMakeLists.txt:1200) — consistent with the core/-prefixed include style used elsewhere.
• The AUTOMOC-via-target_sources line mirrors the existing UlanziDialBackend.h pattern, and the abstract Q_OBJECT with signals genuinely needs moc, so listing it is correct.
• CI is fully green — build/macOS/Windows/CodeQL plus the Engine/UI dependency direction gate you cited all pass, confirming backends/ stays clean of GUI/QtWidgets includes.

One forward-looking note (non-blocking, for 2.2+):
capabilities() returns RadioCapabilities by value, which is fine for the in-process direct-call path this step targets. Once the daemon/protocol lands and any signal starts carrying the struct across a queued (cross-thread) connection, it'll need Q_DECLARE_METATYPE + qRegisterMetaType. Nothing to change now — the current capabilitiesChanged() is payload-free — just flagging it so it isn't a surprise when the boundary goes cross-thread. Same eventual consideration applies to invokeExtension() returning QVariant synchronously if it's ever invoked across the seam rather than in-process.

Design-wise the resolved §5.5 questions read sensibly: keeping RadioModel as sub-model owner and having the backend drive via signals is the minimal-churn choice, and the single-interface + canTransmit/guarded-setKeying() approach keeps the TX guard above the seam where §6 wants it. No concerns to hold up the interface contract. Nice work — looking forward to the FlexBackend skeleton in 2.2.

_{🤖 aethersdr-agent · cost: $5.7774 · model: claude-opus-4-8}

Reviewed the full diff — a header-only, no-implementor pure addition, exactly as described. This is clean, well-documented work and lands the seam's contract in a reviewable state. Thanks @ten9876.

Verified

• Conventions match the codebase: namespace AetherSDR (used throughout src/core) and the #include "core/backends/..." prefixed-quote style (consistent with core/aprs/... etc.).
• No symbol collisions — IRadioBackend/RadioConnectRequest/RadioCapabilities are all new.
• RadioCapabilities doesn't duplicate the existing ModelCapabilities; they're distinct concepts (radio-reported self-description vs. model-string-derived truth), which is the right split and consistent with the §1 anti-pattern the doc calls out.
• connectRadio()/disconnectRadio() (vs. the doc's connect/disconnect) is a good deliberate choice — avoids shadowing QObject::connect. No change needed.
• All 7 checks green, including Engine/UI dependency direction (the boundary gate) and CodeQL.

One design point worth settling now, while there's no implementor:

invokeExtension(ns, verb, arg) returns QVariant synchronously, whereas every other verb (setSliceFrequency, setSliceMode, setKeying, connectRadio) is fire-and-forget void with results arriving via signals. For a backend speaking an async wire protocol (FlexBackend over SmartSDR TCP), a synchronous return can't carry a real device response — it would have to be a local/echoed value or block the caller. That's an inconsistency in the seam's async model. Since a Flex extension verb (e.g. an ATU/amp command) is exactly the kind of thing whose result comes back later on the wire, it may be worth deciding whether extension results should return via a signal + request-id (or a QFuture/callback) rather than a blocking QVariant. Cheap to change now, expensive once 2.2 wires a real backend against it.

Minor (optional): given the near-identical names, a one-line comment in RadioCapabilities.h distinguishing it from ModelCapabilities would save a future reader the double-take.

Neither blocks — the sequencing is sound and the interface is easy to revise while it has no callers. Nice groundwork for 2.2.

_{🤖 aethersdr-agent · cost: $6.8494 · model: claude-opus-4-8}

Reviewed the full diff against the tree. This is a clean, well-scoped RFC step — a header-only interface + capability struct + design doc, no implementor wired, genuinely zero behavior change. Nice work, @ten9876.

Conventions & correctness — all good:

• namespace AetherSDR matches the core convention; #include "core/backends/RadioCapabilities.h" uses the src-rooted include style the rest of the tree uses.
• The CMake change is right: only the Q_OBJECT header (IRadioBackend.h) is added to aethercore target_sources for AUTOMOC; RadioCapabilities.h is a plain struct and correctly left off. The comment-alignment cleanup on the UlanziDialBackend.h line is a tidy touch.
• Design-doc links resolve (../aetherd-headless-engine-design.md, aetherd-touchpoints.md).
• QObject-parent ctor gives RAII ownership; connectionError(QString) is the right boundary error channel for the future implementor. No AppSettings/persistence surface, so nothing to flag there.
• All 7 CI checks green, including Engine/UI dependency direction (confirms the new backends/ dir stays free of gui/QtWidgets includes, as the PR body claims).

One minor, optional nit:

• IRadioBackend.h uses QByteArray in three data-plane signals (spectrumFrameReady, waterfallRowReady, audioFrameReady) but doesn't #include <QByteArray>. It compiles transitively via <QVariant> today (hence CI passing), but for a foundational interface header that many TUs will include, an explicit #include <QByteArray> would be more robust against future include-graph churn. Not a blocker.

The resolved §5.5 decisions (typed struct + extension bag, RadioModel keeps sub-model ownership, single interface with canTransmit guard above the seam, src/core/backends/ layout) all read as sound defaults, and keeping them revisable while there's no implementor is the right call. LGTM — thanks for the careful sequencing and the design doc.

_{🤖 aethersdr-agent · cost: $5.0420 · model: claude-opus-4-8}