Summarize a data source without returning its messages. Includes: a brief summary, basic metadata (start time, message count, config parameters), and a list of available topics. Excludes: detailed topic definitions or actual messages.

Generate a structured summary of a topic without returning its messages. Includes: short summary, DuckDB schema, original IDL definition, and guidelines for SQL queries. Excludes: actual topic data.

Run a DuckDB SQL query on messages from a single topic in a data source. Returns the query results as structured dictionaries. Use this tool to answer user questions about message data, including filtering, aggregation, and downsampling.

Extract INFO, WARN, and ERROR messages from a data source. Supports optional time filtering. Use for debugging or diagnostics.

Use this tool to inspect a live data stream and list the topics that can be subscribed to. Helpful before starting a subscription.

Use this tool to connect to a live data stream and subscribe to one or more topics. Messages are written to a local sink directory, which can be used later as input for other tools (via the path argument in SourceFactory). Optionally attach a pipeline config to create a STANDING pipeline that runs on incoming messages -- e.g. an on_event cadence that captures and uploads a window around every anomaly.

Use this tool to run a predefined capability described in a .poml file. Discover available capabilities and their paths with list_agent_capabilities. The file specifies task instructions and output formats. Optional context values can be injected to customize its behavior.

List the predefined POML capabilities shipped with Bagel: each entry has a name, a path to pass to run_poml_capability, and a one-line summary. Use this to discover available capabilities instead of guessing file paths.

List the tasks and gates available to compose a data pipeline, including each one's module path, kind (task or gate), constructor parameters, and a short summary. Use this before authoring a pipeline so the correct module and args are chosen instead of guessed.

Dry-run an event-windowed reduction WITHOUT writing any files. Detects the rising-edge events where a SQL predicate becomes true on a topic, builds pre/post windows around them, merges overlaps, and reports how much data would be kept. Use this to audit a reduce/snippet pipeline before running it.

Persist a pipeline configuration to a YAML file so it can be reused, edited, or run later with run.py. Returns the path to the written file.

Build and run a pipeline from a configuration and return the artifact paths it produced. Prefer running preview_pipeline first for event-driven reductions so the effect is audited before anything is written.

Run one pipeline configuration against many data sources -- explicit paths or glob patterns like 'logs/*'. Each source is processed independently; a failure on one source is reported but does not stop the batch. Returns per-source results and a summary. For an event reduction, preview a representative source first.

Export a time window of topic data as a PlotJuggler session: a flattened CSV (one scalar column per signal) plus a layout file with the curves pre-added and the window pre-framed. Opening the returned command shows the event already plotted and zoomed. Use after preview_pipeline to hand an event to a human for visual inspection.

Export a time window of topic data as a Rerun recording (.rrd): every scalar signal becomes a Rerun time series, so rerun <file> opens the event in the Rerun viewer. Use after preview_pipeline to hand an event to a human for visual inspection. Needs the optional rerun-sdk dependency (uv sync --group viz).

Export a time window of topic data as a Lichtblick session: an MCAP file with JSON-encoded channels plus a layout with the plot series and time/value ranges pre-set. Works in Lichtblick (open source) and Foxglove, which share the layout format. Use after preview_pipeline to hand an event to a human.

Export time windows as a LeRobotDataset v3.0 for robot-learning training: each window becomes an episode, resampled to a uniform fps, with the given signals composing feature vectors like observation.state and action. Use after preview_pipeline to turn detected events into a curated dataset. Beta: load-tests clean with the lerobot package; awaiting validation by real training runs.

Auto-detect the robot's current hardware, firmware, and software using waffle-iron and return the resulting hardware state. Requires the waffle CLI on PATH (cargo install waffle-iron). The WaffleForm it writes is immediately queryable as a data source.