| name | rerun-nbv-inspector |
| description | Use when creating, reviewing, or fixing ARIA-NBV Rerun integrations for immutable VIN offline-store inspection, NBV candidate/frustum visualization, RRI/validity diagnostics, depth/RGB/keyframe layers, OBB/mesh/trajectory logging, `.rrd` smoke artifacts, or Rerun frame-coordinate issues involving PoseTW, CameraTW, PyTorch3D cameras, and display-only CW90 handling. |
| metadata | {"mode":"implementation","not_when":["generic Streamlit UI behavior with no Rerun artifact or visual logging path","offline-store data validity before Rerun receives a compatible sample","geometry-frame changes outside the Rerun logging or inspection surface"],"handoff_to":["diagnose-aria for concrete Rerun launch failures or suspicious visual output","dataset-cache-ops for incompatible or invalid offline stores","nbv-geometry-contracts for PoseTW, CameraTW, projection, or CW90 frame issues","aria-litkg-memory for official-doc/API lookup when Rerun behavior is uncertain"],"evidence_required":["focused inspector test, saved .rrd smoke artifact, or exact store incompatibility","entity-path, frame-coordinate, and display-only downsampling evidence","official Rerun/API evidence when changing SDK usage patterns"],"applies_to":["aria_nbv/aria_nbv/app/**","aria_nbv/aria_nbv/rerun_inspector/**","aria_nbv/aria_nbv/**/rerun*.py","docs/reference/**",".agents/skills/rerun-nbv-inspector/**"],"triggers":["Rerun","offline inspector",".rrd","candidate frustum"],"must_read":["AGENTS.md","aria_nbv/AGENTS.md",".agents/skills/rerun-nbv-inspector/references/nbv-inspector-contract.md",".agents/skills/rerun-nbv-inspector/references/context7-queries.md"],"canonical_sources":["aria_nbv/AGENTS.md#core-rules",".agents/skills/rerun-nbv-inspector/references/nbv-inspector-contract.md",".agents/skills/rerun-nbv-inspector/references/context7-queries.md",".agents/skills/rerun-nbv-inspector/references/rerun-python-patterns.md",".agents/references/external_stack_contracts.md"],"context7_refs":["/rerun-io/rerun","/websites/streamlit_io","/facebookresearch/pytorch3d"],"literature_refs":["quality-driven-rri","egocentric-aria-substrate"],"tool_refs":["mcp__MCP_DOCKER.get_library_docs","mcp__MCP_DOCKER.browser_run_code","mcp__code_index.search_code_advanced"],"verification":["focused Rerun inspector tests or smoke command for changed surface"]} |
Rerun NBV Inspector
OMX Integration
OMX owns task orchestration; this skill owns ARIA Rerun evidence production.
Return the exact inspector command, artifact path, frame/entity evidence, and
handoffs for data or geometry blockers.
Overview
Use this skill for ARIA-NBV Rerun work where visual diagnostics must protect
data, geometry, and RRI contracts. Treat Rerun as an observer sink for offline
samples and rollout diagnostics, not as the owner of scientific semantics.
Start from current official Rerun docs and examples before changing API usage.
Use the PRML rerun-slam-integration skill as a pattern source, but keep
ARIA-NBV-specific decisions grounded in immutable VIN stores and NBV geometry.
Workflow
- Load current repo guidance first:
AGENTS.md, aria_nbv/AGENTS.md, and
aria_nbv/aria_nbv/data_handling/AGENTS.md when the offline store is
touched.
- Query Rerun docs with Context7 library id
/rerun-io/rerun; start from
references/context7-queries.md and use the smallest relevant query.
- Read
references/rerun-python-patterns.md for recording, entity-tree,
camera/depth, transform, and timeline guardrails.
- Read
references/nbv-inspector-contract.md before editing or reviewing
ARIA-NBV inspector code, tests, docs, or .configs/rerun_offline.toml.
- Read
references/official-examples-map.md when choosing a Rerun example to
compare against.
- Inspect every touched
rr. call site along the logging path, not just the
failing sink.
- For real smoke, prefer a saved
.rrd artifact from one sample:
cd aria_nbv
uv run nbv-rerun-inspect --config-path ../.configs/rerun_offline.toml \
--split val --index 0 --save ../.artifacts/rerun/sample.rrd
If the local store is version-blocked or partial, report the exact store error
and run fixture/fake-Rerun tests instead of weakening validation.
Guardrails
- Keep Rerun read-only: never mutate offline-store poses, RRI labels, candidate
ordering, validity masks, cached geometry, or training payloads for display.
- Declare one scene basis at the root, normally
rr.ViewCoordinates.RIGHT_HAND_Z_UP
for ARIA-NBV world-space diagnostics unless the code documents another choice.
- Treat
PoseTW candidate poses as T_world_cam and reference rig poses as
T_world_rig; do not pass raw matrices across public Rerun helper boundaries.
- Prefer native
rr.Transform3D + rr.Pinhole camera entities for candidate
cameras when transform direction and camera convention are covered by tests.
Keep rr.LineStrips3D for trajectories and selected-view paths.
- Put
Pinhole, RGB, and metric DepthImage(..., meter=1.0) on matching
camera/image entities when depth is meant for 3D interpretation.
- Keep camera-local points under posed camera entities; keep world-space
semidense, fused, mesh, OBB, and trajectory geometry under world-space paths.
- Preserve zero-candidate and all-invalid candidate cases as empty/no-top
visual layers; never let padded rows become real candidates.
- Use stable, low-cardinality entity paths for candidate sets; batch frusta,
centers, OBBs, and point clouds rather than logging one entity per candidate
unless a specific selected candidate needs isolation.
- Treat blueprints as viewer layout only. They must not encode candidate
validity, RRI, frame, or dataset semantics.
- Keep downsampling deterministic and display-only.
Review Checklist
- Confirm config-as-factory usage and
.configs/rerun_offline.toml stay aligned
with RerunOfflineInspectorConfig.
- Confirm required visual inventory failures happen before
rr.init, rr.save,
rr.spawn, or rr.connect_grpc.
- Confirm candidate count, validity mask, oracle RRI, accuracy/completeness
deltas, and frustum labels use the same prefix/order.
- Confirm
candidate_count=0 logs no candidates and an all-invalid mask logs no
top-oracle frustum.
- Confirm
Pinhole.resolution is [width, height], while PyTorch3D
image_size remains (height, width).
- Confirm
DepthImage layers are either explicit 2D diagnostics or paired with
camera pose/intrinsics for 3D interpretation.
- Confirm generated
.rrd outputs are smoke artifacts under .artifacts/ or
another non-training path.
Verification
Use the narrowest checks that cover the touched surface:
cd aria_nbv
uv run ruff check aria_nbv/rerun_inspector tests/rerun_inspector \
aria_nbv/data_handling/_offline_visual_inventory.py
uv run pytest tests/rerun_inspector tests/data_handling/test_offline_visual_inventory.py -q
When the store is compatible:
cd aria_nbv
uv run nbv-rerun-inspect --config-path ../.configs/rerun_offline.toml \
--split val --index 0 --save ../.artifacts/rerun/sample.rrd
