feedback/
September 4, 2026 · View on GitHub
English | 中文
Summary
The feedback group collects human opinions about the harness's work: users can submit a free-text remark about a whole session, and rate or annotate individual assistant messages. Neither kind of feedback reaches the model — these are signals about the output, never input to it. Users record a session remark with the /feedback command; product surfaces read and change per-message ratings through the messageFeedback service. The two packages are independent: session remarks and per-message ratings do not interact. This page maps the group; the package READMEs and the feedback subsystem page own the per-package contracts.
Table of Contents
Packages
| Package | Role |
|---|---|
command-feedback | A /feedback command that records a free-text session remark with one command, without a model turn |
message-feedback | Per-message ratings and notes, served to product surfaces through the messageFeedback service |
Session remarks are a one-way signal: recording one is safe at any point in a conversation and never changes what the model sees. With a feedback-gated sharing policy, recording a session remark is what releases the session for sharing.
Per-message ratings and notes are stored with the session, survive restarts, and never appear in model history or telemetry.
Related documentation
- Feedback subsystem — the message-feedback types, service contract, and Web consumer.
- Session telemetry subsystem — the sharing policy disclosed by the
/feedbackacknowledgement. - Anonymous user identity — the per-harness-home id embedded in the feedback acknowledgement.
Dev Note
None.