github-security-agent

name	github-security-agent
description	Use when asked to reduce GitHub security alerts across all active repositories in one selected profile.

GitHub Security Agent

Use this skill as the canonical entrypoint for one selected profile.

Source Of Truth

Read these files before changing anything:

1. docs/runtime-contract.md
2. docs/operating-model.md
3. docs/code-scanning-policy.md
4. docs/secret-scanning-policy.md
5. docs/review-gate.md
6. docs/reporting-model.md
7. the selected profile.yaml

Scope

• one selected profile
• all active repository entries in that profile
• Dependabot remediation first
• then selected code scanning remediation
• then secret scanning response for enabled targets

Required Workflow

1. Run account discovery and validate the selected profile.
2. Acquire the profile lock before any mutation.
3. Run the Dependabot remediation pass and keep iterating until eligible Dependabot work converges.
4. Run the code scanning triage pass across active repositories.
5. For code-scanning findings:
   • if code scanning is disabled for the repository, report skipped with reason code_scanning_disabled
   • if GitHub reports no analysis, report skipped with reason no_analysis_found
   • if the rule is not allowlisted in the selected profile, report skipped with reason unsupported_rule
   • if the rule is allowlisted and deterministic, run the code-scanning remediator
6. Run the secret scanning response pass across active repositories.
7. For secret-scanning findings:
   • if secret scanning is disabled or unavailable for the repository, report skipped with reason secret_scanning_disabled
   • if deterministic repository cleanup is possible, run the secret-scanning response skill and open or update a review-required cleanup PR
   • if cleanup would require history rewrite or external provider mutation, report blocked with the correct reason code and manual follow-up actions
8. Run the review gate before any merge decision.
9. Reconcile stale or superseded agent-managed PRs that remain in scope. If a PR no longer matches the live alert state because it was superseded, the advisory was withdrawn, or the dependency was removed, close it with a comment that points to the replacement remediation unit when one exists, and record the result with the correct skipped reason code.
10. Write final reporting artifacts for all supported alert classes after reconcile is complete.
11. Release the profile lock.

Execution Rules

• The selected profile is the live source of truth.
• Dependabot, code scanning, and secret scanning are separate passes with separate policies.
• Exhaust eligible Dependabot work before starting code scanning or secret scanning.
• For code scanning, only auto-remediate allowlisted deterministic rules from the selected profile's defaults.code_scanning policy.
• For secret scanning, use the selected profile's defaults.secret_scanning policy and never auto-merge cleanup PRs.
• Final reporting must reflect any reconcile actions taken in the same locked run.