Jetson Video Capability

Capability query guidance

Classification

Keep missing, malformed, unavailable, or failed authority unknown. Use unsupported as a product verdict only for applicable documentation No; a raw decoder API false remains explicitly an API result.

PyNvVideoCodec API query

Use the exact interpreter returned by the setup readiness check. Before and after querying, record the interpreter path, PyNvVideoCodec distribution version, imported module path, and requested GPU. Require the module path to be inside that interpreter's environment. Do not set import paths, scan for a venv, or switch to system Python.

Create a short task-local query from the installed public API and this reference. Inspect the installed API signature when necessary rather than guessing enum or argument names. For a broad encoder catalog, call GetEncoderCaps for h264, hevc, and av1; a request naming specific members of that set queries only those. Call GetDecoderCaps for the scoped tuple set below. Preserve the raw return values and per-call exceptions. Keep the program task-local and apply the user's normal workspace cleanup policy.

A successful GetEncoderCaps record is capability_reported, supported=null, and operation_status=not_tested. Fields absent from the installed extension are unknown rather than unsupported.

A broad decoder catalog covers these 120 unique tuples:

• codecs: mpeg1, mpeg2, mpeg4, vc1, h264, hevc, vp8, vp9, av1, jpeg;

• chroma: monochrome, 420, 422, 444;

• bit depths: 8, 10, 12.

Require exactly 120 unique attempted tuples for a broad catalog. A request naming specific decoder families queries only those families across the same four chroma formats and three bit depths, states the attempted count, and skips nothing within that scope. bIsSupported=1 makes the raw record capability_reported, supported=true, and its limits applicable. Zero makes the raw record supported=false; this is a namespaced API result, never the product verdict, and the remaining zeroed output fields are inapplicable. A missing enum, call, or flag is unknown.

If the public helper selects only GPU 0, retain every requested nonzero-GPU record as unknown rather than relabeling GPU-0 facts.

Native official-sample reports

For a query-only request, do not run setup's encode/decode smoke merely to obtain reports. Reuse current authenticated package-owned report binaries when available, or create a fresh user-owned report build from the single package-owned nvidia-video-codec-sdk 13.0.x Samples tree. Follow the pipeline native build contract, building only AppEncCuda and/or AppDec; building report binaries is the only mutation and does not authorize package installation, encode, or decode. Recheck the binary SHA-256 before and after, require real ldd resolution of libcuda.so.1 plus libnvidia-encode.so.1 for AppEncCuda or libnvcuvid.so.1 for AppDec, reject stub paths, use a clean bounded environment, and invoke exactly:

"$APPENC" -ec
"$APPDEC" -dc

If package, source, tool, or library authentication fails, report unknown and request setup repair. Never execute an encode/decode operation under query-only authorization.

Accept AppEncCuda -ec only when exit is zero, timeout is false, no explicit CUDA/NVENC/error/failure marker appears, and one recognized grammar is exact:

• legacy: one Encoder Capability Summary, one detail hint, unique GPU <ordinal> - <name> blocks including the selected GPU, one codec-support and capability-summary section per GPU, and exactly one H264/HEVC/AV1 row;

• SDK-13 compact: one Encoder Capability, unique GPU blocks including the selected GPU, and one basic H264/HEVC/AV1 yes|no row per GPU.

Accept AppDec -dc under the same process rules and either:

• legacy: unique GPU blocks including the selected GPU, one GPU Decoder Capabilities and Codec Support Summary per GPU, and one detail hint;

• SDK-13 compact: one Decoder Capability, one GPU in use: <name>, at least one Codec ... BitDepth ... ChromaFormat ... Supported ... row, and every Codec-prefixed line matching that strict grammar. Because this form has no ordinal, it applies only to requested GPU 0.

Preserve raw row values and recognized format variant. Native reports are independent: one failure must not erase the other. Never synthesize the Py decoder matrix from -dc output or translate a sample yes, no, or numeric Supported field into a product verdict.

Documentation reconciliation

Identify the live product independently from capability values. Prefer exact immutable device-tree/product evidence; never infer a SKU from GPU name, memory, engine count, queried limits, operation behavior, or similarity. A generic Thor identity remains a family identity.

Read the NUL-terminated device-tree nodes with tr '\0' '\n'. /proc/device-tree/compatible, or its /sys/firmware/devicetree/base/compatible mirror, supplies exact board/module identifiers when present; /proc/device-tree/model supplies only the family label. Record the exact path and literal values beside the documentation row.

Open each applicable official URL at execution time with the agent's web retrieval tool. Use the live page, not a cached excerpt or model memory. If a page cannot be retrieved or its row-to-column binding cannot be verified, record the attempted URL, UTC time, failure reason, and affected fields as unknown; do not reconstruct the values.

Use the official links in SKILL.md; do not preserve a copied product table in the skill. Record retrieval date, URL, table title, all header levels, exact row label, exact field label, and literal cell. If extraction loses the row-to-column binding, the field is unknown. Never use an application note from another SDK release.

Select an HTML table by its own exact caption or section anchor, not the first raw-text occurrence of its title, which may be a navigation link. Parse only that closed table element and require one ordered header binding for each cell. Repeated headers, duplicate matching tables, a row-width mismatch, extraction that crosses table boundaries, or conflicting cells makes the field unknown; never choose one value. For codec-specific tables, do not reuse rows from a preceding H.264 or HEVC table when evaluating AV1.

Capture documentation evidence manually as part of the agent workflow; there is no scraper or copied table to maintain. Retain one compact record per claimed field with: UTC retrieval time, source URL and title, SDK release, ordered header path, exact row label, exact field label, literal cell text, the live product-identity evidence used for applicability, and applicability as exact, candidate_consensus, or unknown. A missing value makes that field unknown rather than permission to reconstruct it from memory.

An authenticated candidate row is a current-document row whose product label is consistent with every fresh immutable device-tree or product identifier observed from the live target. Record those identifiers and their source paths beside the row. GPU name, memory, engine count, API fields, and test outcomes cannot authenticate or exclude a row. When immutable evidence identifies one exact product, use only its exact row and do not invoke consensus.

For a generic NVIDIA Jetson Thor Developer Kit or NVIDIA Thor identity, authenticate every matching Jetson/IGX Thor row. Publish candidate-row consensus only when all candidates agree for the exact field; do not narrow candidates using the query or test outcome.

The support matrix and application note may differ or omit an exact tuple. Preserve disagreement instead of choosing silently, and never expand a family-level statement into an undocumented chroma/bit-depth/profile claim.

Exact operation handoff

Documentation No ends the normal availability check. When documentation is positive or unknown and the user asks for live availability, obtain a schema-2 recipe and follow the pipeline official-sample contract. Require exact input geometry/format/frame count, fresh nonempty output, anchored success markers, and independent authenticated decode of the same path/size/SHA-256. Container structure, exit zero, and an encode marker alone are insufficient.

If the user explicitly requests a diagnostic despite documentation No, run only the exact requested tuple and label its result separately as operation_verified or operation_failed. The documentation-derived product verdict remains unsupported regardless of that diagnostic result.

For AppEncCuda AV1 output, independently decode the exact produced path with the package-owned decoder as the pipeline reference requires. Container structure alone deliberately carries operation_verified=false; only the independent decode, with a positive frame count matching the request, supports operation_verified=true.