IT & Security/First-party starter agent
Dependabot evidence review
Review a bounded set of dependency alerts against available source evidence and recommend what needs human attention.
// what it does
- 01Select the review batch. Read open Dependabot alerts for the named repository, prioritizing the highest severity within the supplied limit. Report how many alerts were returned, selected, and left outside the batch; do not claim a repository-wide review when pagination is incomplete.
- 02Establish the source snapshot. Read repository metadata and resolve the current default-branch commit. Inspect the affected manifests, dependency files and supplied production entry points at that commit, within the file limit.
- 03Assess the available evidence. For each alert, identify the affected dependency and vulnerable behavior. Cite file paths and the pinned commit for any relevant usage, configuration or mitigation. A dependency's location in a development manifest, a missing search result, or a limited tree read alone does not prove it cannot affect production.
- 04Recommend a disposition. Classify each reviewed alert as needs attention, candidate for human review with mitigating evidence, or insufficient evidence. Explain the exact facts and uncertainty. Do not declare exploitability disproved unless the inspected evidence supports that narrow conclusion; even then leave the alert open for the owner.
- 05Return a review report. Include the alert links, severity, inspected commit, evidence, recommendation, and unreviewed scope. Identify which candidates need a fresh check if the code or deployment changes. Report selected, assessed and uncertain counts without presenting them as a native outcome dashboard.
// guardrails in the draft
Never dismiss, reopen, close, comment on, or edit an alert; never open a pull request, upgrade a package, or merge code.
Never run an exploit or probe a deployed service to validate a finding.
Configure read-only grants; written instructions are not a substitute for runtime authority limits.
Do not equate severity with reachable risk or claim complete source analysis after a bounded read.
Keep inaccessible files and truncated responses visible as missing evidence.
Do not send source code or secret values outside the configured workspace report.