Jetson Video Pipeline

Pipeline workflow

Use direct, authenticated NVIDIA samples for codec work. The agent owns planning, execution, artifact handoffs, and the final report.

Before planning

1. Apply the scope and request-classification gates in SKILL.md. Architecture questions end with a direct route answer and stated assumptions. Dry runs use only user-supplied identity and metadata. The target and media steps below apply to execution; recipe-bearing dry runs still use step 4.

2. Preserve an explicit native, pynvc, or both; otherwise use auto. Obtain fresh direct readiness from setup for the selected local runtime. With auto, run the only eligible surface, block with zero, or return selection_required offering native, pynvc, and both when both are eligible.

3. Require full-samples for Python encode, advanced decode, segmentation, or encode-performance. The smaller pynvc-smoke profile is sufficient only for its documented setup smoke and Python decode-performance route.

4. For encode or transcode dry runs and execution, obtain a validated schema-2 recipe. Decode-only, segmentation, and container triage do not require one.

5. Canonicalize and hash the exact user media. Preserve URL or null, license, attribution, path, byte count, SHA-256, and media metadata established by the user or a media parser/sample. Never infer metadata from the filename or replace the input with setup's synthetic fixture. For a URL, disable redirects; bound connection time, total time, and accepted bytes using an explicit user ceiling or trusted local metadata. Resolve the host first and accept only HTTP(S) addresses outside loopback, private, link-local, multicast, and reserved ranges. Recheck the connected peer address on every connection and preserve hostname-based TLS validation.

Plan, then execute

For a dry run, emit a readable planned result using the user-supplied media identity and metadata. Show the intended argument arrays, inputs, outputs, expected markers, frame counts, producer-to-consumer handoffs, and applicable io_contract from buffer-sharing-and-synchronization.md. Do not open the media, inspect the target, authenticate launchers, or claim the commands were verified.

For execution:

When a native binary still needs compilation, return dependency_required with the exact build command and source/tool identities.

1. Create a new mode-0700 workspace. Require every output path to be absent.

2. Authenticate the selected sample and every required helper as specified in official-sample-contract.md. Record canonical path, size, and SHA-256. Recheck them immediately before launch.

3. Launch the exact argument array directly, without a shell or eval, under the selected interpreter when Python is used. Retain unedited stdout/stderr in separate stage logs and capture exit code, elapsed monotonic time, timeout state, and working directory. Use a finite per-process timeout: 300 seconds unless the user supplies a different finite bound. Record the bound and terminate the launched process group on expiry before returning failed.

4. Require exit zero, the allowed positive marker set with every expected cardinality, count, and path, no explicit failure line, and a fresh nonempty output. A timeout, launch error, stale output, contradictory marker, or exit-zero unusable file fails that operation.

5. Immediately before a handoff, reopen the original producer output and verify its path, size, and SHA-256. The independent consumer must open that same path. Rehash it again after the consumer completes. Preserve the canonical io_contract on successful and failed boundaries and in the final result.

6. Apply the reference's byte/layout table to the actual raw source and decoded file. Do not infer usability from a marker or byte count alone.

7. Retain a successful branch when another selected surface or segment fails; label the aggregate partial. Retry once at most, with fresh paths, and only after the input, executable/package/interpreter, readiness, or unavailable resource has demonstrably changed. Retain both attempts and stop after the retry.

Routes

Encode and independently decode

• Native: validated recipe → AppEncCuda → fresh elementary stream → AppDec.

• Python: validated recipe/config → wheel-owned basic/encode.py → fresh elementary stream → wheel-owned advanced/decode.py.

• Require producer and decoder counts to equal the requested frame count.

• Require exact raw-input and decoded-output byte sizes from the format, geometry, and frame count. Record the independently decoded layout.

• For both, keep each surface's identities, commands, and result separate.

Native H.264-to-HEVC transcode

Require an H.264 input and an exact native HEVC recipe projection. Do not infer the input frame count from its filename. Run AppTrans with recipe options except raw-only -s, -if, and -gpu; supply the selected GPU once. Accept exactly one legacy or current transcode count marker, never both. Rehash the HEVC output and independently decode it with AppDec for the same positive frame count. If the input frame count was known before launch, both counts must also equal it; otherwise the two authenticated sample counts establish the tested frame count.

PyNvVideoCodec segments

Authenticate the wheel-owned segment sample, segments.txt, transcode_config.json, and advanced decoder. Validate every schedule row as finite 0 <= start < end. The created-file markers must exactly match the declared output set and the summary must equal its size. Each segment must be fresh, nonempty, rehashed, and independently decoded. Report failed segments without discarding verified peers.

Container triage

First establish an eligible released Jetson route. Decode the exact local or retrieved container through AppDec's intrinsic libavformat demux or the wheel-owned advanced decoder. Record the same content provenance and require a fresh usable decoded output. This route reports the tested operation only; it does not generalize hardware support.

AV1 verification

Require an AV1 recipe with an exact native projection and positive frame count. Run both requested host-output and video-memory modes independently. For every fresh IVF output, rehash it and have AppDec consume that exact path. Require the expected positive decoded frame count; do not claim a separate container- structure verdict from header bytes alone.

Concise acceptance report

Acceptance is a human-readable reproducibility report. Include only requested stages, with one row per stage and one of complete, partial, blocked, failed, not_evaluated, or deferred. Retain:

• readiness result and selected target/runtime;

• raw API capability facts separately from documentation and operations;

• P4/P5 recipe identities and the exact input identity they share;

• every encode/decode producer/consumer command, marker, frame count, output identity, decoded layout, and byte-size validation;

• every benchmark warmup, measured repetition, timing scope, and recomputed statistics as required by jetson-video-benchmark;

• limitations, failed/retained branches, and exact retry actions.

One report, a command log, and compact JSON evidence are sufficient when they contain those facts. Write a checksum manifest last if packaging is requested. Keep media, raw frames, bitstreams, SDK trees, builds, venvs, and caches outside the small evidence package; bind them by canonical path, bytes, and SHA-256.

Status meaning

• planned: an architecture plan with conceptual steps or a dry run with concrete commands was provided; nothing was launched. State unresolved assumptions.

• complete: every requested branch passed operation and handoff validation.

• partial: at least one independent requested branch passed and another did not.

• blocked: no requested branch could safely run.

• failed: launch or observable validation failed.

• input_required, dependency_required, and selection_required: the named terminal gate stopped work before launch.

• operation_verified: only the exact tested route worked.

• unsupported: reserve for directly applicable product documentation; never derive it from an API field or failed operation.