HFlow’s latest release bases robot-data curation on each check’s latest run. Hippo adds VS Code hooks that capture new agent-chat turns and keep a task handoff for each chat.
HFlow bases curation on each check’s latest run
Robot recordings combine camera video with movement and action streams. A frozen camera or a stream that falls out of sync can make an episode unsuitable for training, and the problem gets harder to track across a large dataset. HFlow organizes ingestion, checks and curation, recording each step’s versions and measurements in a queryable catalog. Its current data boundary is MCAP, a format for synchronized streams, with an importer for part of LeRobot Dataset v3.
HFlow’s version 0.2.21 release, dated Oct. 8 on GitHub, changes curation to use the latest run of each check. When a check’s latest run errors or its version changes, HFlow stops using evidence from earlier runs to decide dataset admission and coverage. A crash in a critical check also can no longer clear an existing quarantine. The update addresses stale evidence; it does not report an improvement in the accuracy of models trained on the data.
The project describes itself as pre-v1. Teams can run its quickstart in-process; scheduled workflows use Airflow 3 or the included Docker Compose runtime. Its code is Apache 2.0 licensed. Version 0.2.21 also requires an explicit base_url when constructing HFlowHostedExecution; omitting it now raises an error. The repository says the hosted control plane is not implemented there, and hosted API access requires contacting the team.
Hippo adds memory hooks for VS Code chats
Long coding-agent chats can be shortened when they approach a model’s context limit, which can drop earlier details from the active conversation. Hippo’s version 1.69.0 release, dated Oct. 7 on GitHub, adds two hooks for VS Code agent mode. Before compaction, VS Code’s PreCompact hook saves a task snapshot to Hippo’s local .hippo store. After each reply, VS Code’s Stop hook reads only new user and assistant turns from that chat’s transcript, extracts candidate memories into the store’s SQLite database with Markdown mirrors, and rewrites one handoff per chat in the same local store. Hippo documents a handoff as a summary and next step that a later session can reload as context.
The hooks require VS Code 1.109.3 or later with its chat.useHooks setting enabled; older versions receive the server and instructions without the hooks. Hippo requires Node.js 22.16 or later. The pinned Copilot integration guide says VS Code agent mode is built and covered by fixture tests, but has not yet been observed end to end in a live VS Code session. The available evidence therefore does not establish live reliability or improved coding-task success.
Sources
HFlow project and setup notes ·
HFlow 0.2.21 release notes ·
HFlow license ·
Hippo project and setup notes ·
Hippo 1.69.0 release notes ·
Hippo’s VS Code integration and capture limits ·
Hippo capture implementation ·
Hippo license
Image: HFlow pipeline demonstration from its Apache 2.0-licensed repository at the cited commit.