| name | locomo-segmentation-debugger |
| description | Run and inspect the LoCoMo event segmentation debugger in `crates/event_segmentation/examples/locomo_segmenter.rs`. Use when debugging `EventSegmenter`, checking segment boundaries on LoCoMo samples, inspecting the emitted event-segment JSON array, slicing or rewriting derived JSON through shell pipes or a REPL, focusing on `conv-47`, or sweeping every sample to judge overall behavior instead of a single conversation. |
LoCoMo Segmentation Debugger
Use crates/event_segmentation/examples/locomo_segmenter.rs as the primary debugger for segmentation work. It emits exactly one Vec<EventSegment> JSON document on stdout; warnings and errors go to stderr.
Environment Note
locomo_segmenter calls the embedding backend through plastmem_ai::embed_many. In Codex's sandbox, requests to the local provider (for example http://localhost:11434) may fail even when the service is healthy on the host. If the run errors at the embedding request step, re-run with escalated permissions instead of assuming the segmenter itself is broken.
Workflow
- Start with a targeted run on
conv-47 when iterating on segmentation logic. It is a strong long-context sample and is usually worth checking first.
- If the question is about event flattening rather than boundary decisions, add
--print-events and inspect stderr.
- If the question is about JSON shape or a suspicious segment, save
stdout to a file and inspect or transform it with the commands in references/commands.md.
- If a change looks good on
conv-47, run every sample before concluding anything about quality. Single-sample wins are not enough.
Rules
- Treat as machine-readable output only. Do not parse diagnostics from it.
