Changelog

All notable changes to bagx are documented here. The format follows Keep a Changelog, and versions follow the policy in Versioning and compatibility.

Unreleased

Added

• bagx tf: TF tree validation (multiple parents, cycles, disconnected trees, transforms on both /tf and /tf_static, pauses and early stops, zero or regressing stamps, clock mismatches between topics, sensor frame_ids missing from the tree, --require-frames, --fail-on). The same checks run inside bagx eval (tf.* findings, TF: recommendations, report.tf) and as the MCP tool check_tf. They do not change scores.
• Message content checks in bagx eval (msg.* findings, Data: recommendations): scans with no or few valid ranges, invalid scan config, beam count changes, uncalibrated or zero-size camera_info, camera_info vs image size mismatch, truncated images, resolution or encoding changes, joint_states shared by publishers with different joint sets, name/position length mismatch, NaN joints, and empty frame_id. These do not change scores.
• CDR fallback parsers (no ROS needed) for sensor_msgs LaserScan, CameraInfo, Image, CompressedImage and JointState.
• Header-stamp matched pipeline latency: when a stage's output carries its input's stamp, latency is measured exactly per message and reported with (header-stamp matched).
• Scoreboard manifest integrity: provenance, availability, bag fingerprints and scripts/generate_scoreboard.py --check.
• CI runs the test suite on ROS 2 Humble and Jazzy, including rosbag2_py vs SQLite backend parity tests.

Changed

• Report schema_version 1.3.0 → 1.4.0. This is additive: eval JSON gains tf and new finding categories tf_tree and message_content.
• eval.py is split into eval_topics, eval_workflow, eval_latency and eval_domains. This is a code move with no behavior change, and the old bagx.eval imports still work.
• Scoreboard: TurtleBot3 Walker and Franka Panda rows are marked source_unavailable; r2b rows point at NVIDIA NGC. Livox MID-360 is pinned to the AIST Zenodo bag (the same bag the 0.4.0 score came from) and NTU VIRAL to eee_03 on DR-NTU. Newer College (download needs a request form) is unscored until a sequence is pinned. Ouster OS0-32 is source_unavailable, because its recording was never tied to a public file. Every available public bag is rescored after the sync fix below, and rows scored on unreleased code record bagx_version as X.Y.Z+dev.

Fixed

• Sync score (bagx eval, all SLAM / Autoware / perception bags with two or more sensor streams): each pair walked whichever topic sorted first by name and took the nearest message on the other. For a dense/sparse pair such as 200 Hz IMU ↔ 10 Hz LiDAR this measured a quarter of the LiDAR period (~25 ms) however well the streams were aligned, and raised a false "enable per-point deskew" warning. The pair now walks the sparser topic, so sync scores rise on bags with mixed rates (Livox MID-360 70 → 100, Autoware bag4 67.7 → 87.4) and the tuned FAST-LIO comment about deskew disappears where it was spurious. bagx scenario sync_degraded uses the same rule.
• sensor_msgs Temperature, MagneticField, FluidPressure, RelativeHumidity, Illuminance and BatteryState topics no longer count as sync candidates because their name contains imu or camera (e.g. /imu/temperature was paired with a camera on NTU VIRAL).
• NavSatFix CDR parser misread status / service unless len(frame_id) + 1 was a multiple of 4 (e.g. gnss_link), which skewed the GNSS fix rate on the SQLite fallback.
• On ROS 2 Jazzy every rclpy-decoded message failed to convert (_check_fields slot), so the ROS backend silently fell back to the CDR parser and custom message types lost their content.
• pip install bagx[mcp] pulled mcp 2.x, which broke bagx mcp; mcp is now <2.
• pyarrow 26 needs NumPy 2 at import time, so bagx export failed on ROS 2 systems with NumPy 1.x; pyarrow is now <26.
• Ruff rule selection is pinned, so CI no longer breaks when ruff changes its defaults.

0.6.0 - 2026-06-10

Added

• ROS 1 .bag support via pip install bagx[ros1] (verified on TUM VI and EuRoC).
• bagx mcp: stdio MCP server (eval_bag, list_topics, detect_anomalies, compare_bags, query_messages).
• Public scoreboard: 26 datasets scored, with reproduction commands.
• bagx-action composite GitHub Action and CI dogfooding.

0.5.0 - 2026-06-10

Added

• bagx eval --tune generates FAST-LIO / LIO-SAM / GLIM / KISS-ICP configs from a bag.
• bagx eval --html self-contained visual reports; bagx report converts saved JSON.

0.4.0 - 2026-06-10

Added

• Finding.time_range for temporal findings across eval / anomaly / diff / benchmark.
• Per-check custom rule findings with optional severity.
• Domain plugin SDK for stack detection (bagx.domains entry points).
• bagx eval --badge shields.io readiness badges.

0.3.0 - 2026-05-21

Added

• bagx diff: compare two eval reports by stable finding id (text / markdown / json, --exit-on).
• bagx eval --findings-only and --severity-min.
• Benchmark forbidden_findings, per-category max_severity, and bagx benchmark --exit-on.
• bagx/schema/findings.schema.json ships with the wheel.
• GitHub Actions PR-check example workflow.

Changed

• Eval and benchmark JSON schema_version 1.1.0 → 1.2.0. Benchmark reports add worst_severity.

0.2.0 - 2026-05-21

Added

• Structured findings in eval --json with stable <domain>.<area>.<qualifier> ids.
• Benchmark expected_findings.
• Custom rules from JSON files, BAGX_RULE_PLUGIN_PATH or bagx.rules entry points; bagx rules list.
• Action / service workflow outcome observability and a generic control-loop latency check.
• Built-in perception and robotarm domains.
• Versioned report contract (schema_version, report_type, bagx_version).

0.1.3 - 2026-03-24

Added

• Pipeline latency between Nav2 / Autoware / MoveIt stages.
• Actionable recommendations at the end of bagx eval.

Fixed

• IMU frequency with bursty timestamps (28306 Hz → 388 Hz).
• Anomaly output summarised instead of listing thousands of events.
• No false sync_degraded from /rosout; clean errors on corrupt or empty bags.

0.1.2 - 2026-03-23

Fixed

• Meta topics (/rosout, /parameter_events, /tf_static, …) are excluded from scenario sync and dropout detection.

0.1.1 - 2026-03-18

Changed

• IMU noise via first-order differencing (= Allan deviation at τ = dt); multi-IMU evaluation; full Allan variance on static data.

Fixed

• Anomaly false positives on real data (31,972 → 640 on Newer College).

0.1.0 - 2026-03-18

Initial release: eval, compare, sync, export, anomaly, scenario, ask, scene, batch, info. Works without ROS 2 via the SQLite + CDR fallback; supports .mcap, multi-segment bags and both CDR endiannesses.