Imported from nvidia/skills (skills/jetson-video-capability/SKILL.md). Install upstream with npx skills add nvidia/skills --skill jetson-video-capability. Copyright stays with the author (Apache-2.0).

Jetson Video Capability

Purpose

Answer Video Codec SDK and PyNvVideoCodec support questions without confusing an API response, a successful operation, and product documentation. Query the live target first, retain exact operation evidence separately, and publish the final product-support verdict from applicable NVIDIA documentation.

Prerequisites

• Fresh evidence from jetson-video-setup is an optional authority, not a prerequisite. When it is supplied, authenticate and use it exactly. When it is absent, this skill authenticates only the selected installed surface: package-owned native sources and requested report targets, or the exact invoking PyNvVideoCodec interpreter and wheel. It never installs, repairs, scans for a venv, or imports setup code. A fresh environment artifact that the agent obtains from setup's public read-only probe counts as supplied evidence; it need not originate in the customer's prompt.
• Run live queries on the target Jetson with direct GPU access. Do not claim current availability from an x86 host or a result copied from another target.
• If selected-surface prerequisites are missing, stop without mutation and route that surface to jetson-video-setup. If that skill is not installed, tell the user to install it. Never infer codec or product support from a missing Python import or native build prerequisite.
• Keep the installed native and Python surfaces independent. Query only the requested surface unless the user explicitly selects both.
• Query execution requires this capability skill plus either valid supplied setup evidence or the selected local prerequisites described below. An exact operation check also requires jetson-video-recipe, which owns recipe resolution, and jetson-video-pipeline, which owns authenticated encode-to-independent-decode execution.
• Before a bounded operation check, when setup is installed read its shared video content policy. Query-only work is media-free; only the documented setup fixture is allowed for the capability-smoke exception. Setup is not required solely for this policy: without it, require one exact user-selected path or URL for any other operation, never substitute catalog or synthetic media, and preserve source URL, license, attribution, path, size, and SHA-256.

Resolve this skill from the installed skills root and set CAPABILITY_SKILL to its canonical absolute path. Invoke each owning script directly under isolated Python:

python3 -I "$CAPABILITY_SKILL/scripts/query_native_sample_reports.py" --help
python3 -I "$CAPABILITY_SKILL/scripts/query_encoder_caps.py" --help
python3 -I "$CAPABILITY_SKILL/scripts/query_decoder_caps.py" --help
python3 -I "$CAPABILITY_SKILL/scripts/validate_appenc_av1_ivf.py" --help

If an owning script is missing, stop with dependency_required. A setup artifact is optional; a supplied one is never optional to validate. Do not copy modules from another skill or add a fallback import path.

Compose requested sibling stages

Capability queries and documentation reconciliation require no sibling when the selected SDK prerequisites already exist. Use jetson-video-setup for installation, repair, or one read-only readiness handoff when registered PyNvVideoCodec authority is required; use jetson-video-recipe plus jetson-video-pipeline for an exact requested operation, and jetson-video-benchmark for requested throughput. Check the agent's installed skill catalog before each such stage. If the sibling is present, read its SKILL.md and invoke its documented public entry point; pass artifacts as data and never import sibling code. If it is absent, preserve every completed query result and say, using the actual names: I can run <stage>, but it requires <skill>, which is not installed. Install <skill> and retry this stage. Never require a sibling for an unrequested or optional refinement.

Instructions

1. Classify the request first, before any target probe, capability query, or other workflow step. For a request solely for objective quality metrics, including PSNR or SSIM, state only that this skill does not provide them and that a separately authorized quality workflow is required, then stop. Do not name or recommend an external tool, and do not offer to configure or run the comparison; do not request media, probe, install anything, or launch an operation. For a request solely about Netflix, Widevine, or other DRM-protected streaming-service playback, state only that this skill covers hardware encode/decode of user-supplied non-DRM bitstreams and does not cover, enable, or verify streaming-service or content-DRM playback. Do not claim whether the service will work; do not describe Jetson content-DRM certification; do not recommend or offer to install a browser, a content-DRM module (Widevine, PlayReady), a playback tool, a workaround, or a bypass; do not probe the target or launch an operation. NVDEC decode of supported user-supplied non-DRM bitstreams stays fully in scope for this skill and is never discouraged by this boundary, so you may say so. Then stop. A mixed request that also asks an in-scope codec question is not refused wholesale: answer the in-scope part normally and apply this boundary only to the streaming-service part. An unqualified “DRM” does not by itself mean content protection: on Jetson it commonly means the Linux Direct Rendering Manager (DRM/KMS, /dev/dri, modesetting, display connectors), which this boundary does not cover. Apply this boundary only when the request identifies Netflix, Widevine, PlayReady, streaming-service protection, or otherwise clearly means content Digital Rights Management; if the request says only “DRM” and the context does not resolve which is meant, ask the user which before answering. A local DRM-free MP4 shown on a display is likewise not a content-DRM request, and its codec portion stays in scope. Otherwise handle capability discovery, exact support questions, and interpretation of saved capability evidence. Route package installation to jetson-video-setup, recipe construction to jetson-video-recipe, throughput measurement to jetson-video-benchmark, and multi-stage media work to jetson-video-pipeline. Within a capability request, a genuinely bare “video SDK” phrase with no product qualifier is ambiguous: ask whether the customer means native Video Codec SDK, PyNvVideoCodec, or both, then stop before probing either surface. Report-only intent or “probe the target” does not authorize --runtime both. A capability support/catalog request that names no SDK surface or product phrase uses native-preferred fallback selection, whether it is broad, exact, or a bounded subset. Select native when its route is eligible. Only when native is ineligible, evaluate the PyNvVideoCodec candidate in the authority order defined in step 3; select Py when that candidate is eligible. Do not ask the user to choose merely because this fallback was used. When native is selected, do not evaluate Py; if the response displays that unselected peer, report it as not_evaluated with reason surface_not_selected. If neither route is eligible, preserve both typed reasons and provide the applicable setup remediation. Serialize a successful fallback as an explicit native or pynvc request before applying the shared routing truth table. Carry that resolved surface explicitly into any authorized downstream operation so the operation controller does not reclassify it as auto. This capability-only unnamed policy is not auto: only explicit “auto”, “whichever”, “best available”, “choose for me”, or equivalent wording that expressly delegates the SDK choice is genuine auto; it is never both. Naming Python or PyNvVideoCodec is explicit pynvc. Naming Video Codec SDK, native, AppEncCuda, or AppDec is explicit native and never falls back.

2. Apply the authorization gate. A discovery or report-only request is query-only. Query and reconcile documentation before deciding whether an operation is useful. A directly applicable documentation No is the final product verdict and ends the normal support/availability check without a codec operation; a request to “check live availability” or “operational support” does not by itself require an experiment that cannot change that verdict. Preserve conflicting raw inventory as diagnostic evidence. Run a bounded matching official operation only when documentation is supported or unknown and live availability is requested, or when the user separately and explicitly requests a diagnostic experiment despite the unsupported product verdict. Such an experiment never promotes product support. No package, repository, signing-key, or credential change is authorized. If an otherwise required operation controller is unavailable, report not_tested and the required next action rather than inventing live availability.

3. Authenticate only the selected surface. Use matching fresh setup evidence when the caller or the agent supplied it. A malformed, stale, or mismatched supplied artifact fails closed and never falls back. For an explicit pynvc or both request, a genuine auto candidate gate, or the unnamed fallback after native is ineligible, with neither setup evidence nor an exact interpreter, check the installed skill catalog. When jetson-video-setup is present, invoke its public read-only probe_nvcodec.py with --runtime pynvc for explicit pynvc or the unnamed fallback, or --runtime both for both/auto, and a fresh --output; never pass --setup-candidate. Inspect the fresh output before using it. A Py candidate is eligible only when the artifact has mode=live, the requested GPU, pynvc.installed=true, and pynvc.identity.status=verified. If routing selects Py, invoke this skill's Py query only under the artifact's lexical pynvc.identity.interpreter with -I and pass the raw artifact path through --environment. If setup is absent, use reason setup_probe_unavailable; if it reports an absent, stale, unreadable, invalid-binding, or launch-failure result, preserve that exact typed reason. For explicit pynvc/both, ask for the exact canonical absolute interpreter. For auto, report PyNvVideoCodec as not_evaluated with that reason and continue only an eligible native branch. For the unnamed fallback, preserve the typed Py reason beside the ineligible native reason and provide setup remediation; never silently return a native-only unavailable result while a healthy registered Py candidate exists. Never scan or guess. Otherwise use this skill's local query path: a fresh explicit native build workspace, or the exact PyNvVideoCodec interpreter invoking the query under -I. Missing prerequisites are unknown/not_ready; route them to setup and, if setup is absent, instruct its installation.

4. Run the selected capability route. Before invoking it, read the matching command and authentication contract in native official-sample reports, encoder API query, or decoder API query. Native uses a supplied verification or a fresh explicit report workspace; Py uses the validated lexical interpreter and includes --environment when setup evidence was supplied. Default encoder scope is H.264, HEVC, and AV1. An unconstrained explicit Py decoder-catalog request, or a broad unnamed decoder-catalog request whose fallback selects Py, uses the complete 120-tuple matrix. A bounded subset queries only the named codec families. Keep an unnamed fallback answer at the requested family-summary scope; do not dump tuple-level claims unless the user requested them. For both, run the branches independently. Preserve every result as raw report/API evidence; never promote it to product support or operation proof. A failed or unavailable live branch never suppresses the documentation answer.

5. Cross-check NVIDIA documentation before deciding whether to run an operation. Follow the complete documentation cross-check and read the versioned R39.2/SDK 13.0 Thor baseline before transcribing a dynamic table. Record URLs, retrieval date, literal live identity, exact field label, and the exact row or complete candidate set. Never infer a cell from flattened or ordinal table text. Use only the version-matched SDK 13.0 note for this release; a field it does not establish remains unknown.

6. Run the smallest matching official operation only when steps 2 and 5 require it. Resolve and validate one exact minimal recipe with jetson-video-recipe, then hand its sealed recipe, environment, and input identities to the pipeline-owned controller. Set PIPELINE_SKILL to that installed skill's canonical path:

   python3 -I "$PIPELINE_SKILL/scripts/encode_controller.py" \
     --request "$OPERATION_REQUEST" --workspace "$FRESH_WORKSPACE" \
     --output "$OPERATION_RESULT"

   A bounded capability smoke operation may use the setup workflow's documented deterministic one-frame raw fixture; it is never representative media and cannot support performance, quality, or pipeline claims. For encode, operation proof requires the independent authenticated decoder to consume the exact output path and SHA-256 and produce the exact decoded frame count. An API response, exit zero, encode marker, or output file alone is insufficient. If jetson-video-recipe or jetson-video-pipeline is absent, report not_tested and name every missing dependency as the next action. AV1 IVF structure may be checked with:

   python3 -I "$CAPABILITY_SKILL/scripts/validate_appenc_av1_ivf.py" \
     --input "$BITSTREAM" --width "$WIDTH" --height "$HEIGHT" \
     --expected-frames "$FRAMES" --output "$IVF_REPORT"

   structure_verified proves container structure only; it never establishes operation_verified.

7. Publish the reconciled result. Keep these signals separate:

   • caps_query: raw Py API fields and status;
   • official_sample_report: raw native -ec/-dc inventory, never an API or operation verdict;
   • operation_evidence: operation_verified, operation_failed, or not_tested;
   • documentation_crosscheck: the product-support authority.

   Publish customer-facing support from documentation_crosscheck, and report live availability only from a successful exact operation. A directly applicable documentation No, including unanimous authenticated candidate rows when exact row identity is unresolved, is the final unsupported verdict even if the API advertises fields or an operation succeeds.

8. Enumerate ambiguous product rows completely. When one live identity maps to several authenticated documentation candidates, list every authenticated candidate row by its documentation label and exact queried value and state the candidate count. A subset cannot establish consensus. If candidate values disagree, keep the documentation verdict unknown; never choose a nearby product row. Never shrink the candidate set using live API fields such as engine count, dimensions, or format flags; GUID enumeration; an operation outcome; performance; or similarity to a marketing specification. Those are evidence being reconciled, not independent product identity. For a generic NVIDIA Jetson Thor Developer Kit / NVIDIA Thor identity, the word Jetson alone is not an authenticated row discriminator: unless an NVIDIA one-to-one product mapping narrows it, include every Thor row in the combined Jetson/IGX table. Apply this rule to any queried field, not only one codec.

Read capability-queries.md for exact evidence semantics and surface-selection-contract.md for native, pynvc, auto, and both behavior.

Available Scripts

The four owning scripts are listed in the prerequisite help commands above. Invoke them directly, and read capability-queries.md for route-specific purposes, arguments, and contracts.

Published artifacts

Capability artifacts are optional refinements; authenticated schema-1.2 nvcodec-environment capabilities remain sufficient for sibling workflows. Read capability-queries.md for their exact contracts. Never promote sample/API evidence to support or operation proof.

Troubleshooting

• Preserve capability_reported, operation_verified, operation_failed, raw API unsupported, and unknown as distinct internal states.
• Treat missing query authority, failed registry authentication, absent API fields, nonzero-GPU PyNv queries, and unavailable exact operations as unknown or not_tested with a concrete next action.
• Report a launched exact operation failure as operation_failed, not global product unsupported. A native -ec/-dc report failure remains raw unknown.
• Retry at most once and only after an evidenced condition changes. Use a fresh work directory and output path for the retry.

Limitations

• PyNvVideoCodec 2.1 encoder and decoder capability helpers select GPU 0; nonzero-GPU results remain unknown.
• Capability fields do not measure throughput, quality, latency, camera count, or successful concurrent sessions.
• Objective quality measurement, including PSNR and SSIM, is outside this skill.
• Results apply to the exact target, software versions, GPU, codec tuple, and operation that produced the evidence.