Feedback and learnings
Classify review feedback and manage reusable agent guidance.
Tasuku records explicit reviewer feedback on findings and turns trusted review-thread conversations into reusable learnings. A learning created this way is enabled immediately; administrators control it afterwards from the Learnings page.
Finding feedback
Classify a finding by replying inside its Tasuku inline review thread:
@<app-handle> feedback helpful
@<app-handle> feedback not-helpful <reason>
@<app-handle> feedback not-helpful <reason> <note>The command must be a reply within the finding thread, not a top-level comment. The verdict is helpful or not-helpful. The reason is optional and must be one of:
| Reason | Use when |
|---|---|
FALSE_POSITIVE | The reported defect is not present. |
NOT_ACTIONABLE | The observation may be true but is not useful to change. |
DUPLICATE | Another finding already covers the problem. |
UNCLEAR | The wording or evidence is insufficient to act on. |
OTHER | The reason does not fit the standard categories. |
Feedback is retained with review and repository provenance. Analytics can use it to show acceptance and noise patterns without deleting the original finding.
Free text after the reason is stored as a note with the feedback and also runs the review-conversation workflow on the thread. When the note states an explicit durable preference from a trusted reviewer, the agent can record it as a learning for future reviews, exactly as with a plain conversation reply.
Learnings
A learning comes from a review-thread conversation. When a trusted human—a repository owner, member, or collaborator—replies under a Tasuku review finding thread, the review-conversation workflow runs and the agent can record an explicit durable preference from that exchange. The learning is stored approved and enabled immediately, and enabled learnings are supplied as conventions to later agent prompts.
Tasuku stores the source actor, repository, interaction, content, and state. It also records descriptive metadata from the review context—including language, file path, pull request number, and source comment—so the learning can be searched and presented in other views without losing its provenance.
The repository remains the applicability boundary for repository-scoped learnings. Language is descriptive metadata; it does not make a learning apply to other repositories automatically.
Administrative control
A user with the manage-learnings capability (ADMIN or SUPER_USER) can disable or re-enable a learning from the Learnings page. Disabled learnings remain available for audit but are not applied to new work. Re-enabling is explicit.
Related context controls
Repository context clusters and secondary-repository relationships remain under manual control. Suggested context is not trusted merely because an agent inferred it. Approve the relationship and keep secondary repositories read-only.