scrubby_report_findings
Report review findings back to Scrubby for Hebbian learning — strengthens domain connections that matter, weakens ones that don't.
After reviewing a file (whether automatically or in collaboration with a human), call this tool to feed the findings back into Scrubby’s learning loop. Each finding strengthens the connection weights that produced useful signal; the absence of findings on a low-confidence connection weakens it.
This is what makes Scrubby get sharper over time. The phrase “Scrubby learns from usage” describes this tool’s behavior literally.
Parameters
| Name | Type | Required | Description |
|---|---|---|---|
file_path | string | Yes | Relative path to the file that was reviewed. |
findings | object[] | Yes | Array of findings. Each finding requires severity and message; domain and line are optional but recommended for accurate weight updates. |
repo_id | number | No | Scrubby repository ID. Auto-detected if omitted. |
repo_name | string | No | Repository name (e.g. api). Used to auto-resolve in remote mode when multiple repos are indexed. |
Finding shape
Each item in the findings array:
| Field | Type | Required | Description |
|---|---|---|---|
severity | "error" | "warning" | "info" | Yes | The severity of the finding. |
message | string | Yes | Human-readable description. |
domain | string | No | The domain that produced the finding. Recommended — weight updates target the connection between this domain and the file’s primary domain. |
line | number | No | The line number the finding applies to. |
Typical usage
After your AI editor completes a review:
"Report the findings you just produced to Scrubby via scrubby_report_findings."A well-prompted agent does this automatically as part of its review loop, attributing each finding to the domain that surfaced it.
What changes when you call this
- Reinforced. When a finding has a
domainset, the connection between that domain and the file’s primary domain gets stronger by a small delta. - Weakened. When a connected domain ran but produced no findings, the connection gets weaker by a smaller delta. This is asymmetric on purpose — useful signal counts more than its absence.
- Global vs repo-scoped. Global-domain weights (Ruby, React, Testing) update on a slower clock so a single repo’s signal doesn’t move them disproportionately.
See also
- Findings & the Learning Loop — the full mechanics.
- Reporting Findings — a task-oriented guide for getting your AI editor to report findings reliably.
Last updated