Method boundary: evaluate contribution quality and maintainer cost. Never tune timing, wording, account fields, branch names, metadata, or repository selection to influence an AgentScan score.
Before every public GitHub write
- Define the exact external action.
Separate local research, patching, and testing from the issue, pull request, review, or comment that would be visible to maintainers.
- Read the repository's current rules.
Inspect contribution guides, AI policy, code of conduct, security policy, templates, and relevant workflow files.
- Study local precedent.
Search recent issues, pull requests, discussions, and labels for the contribution type, AI, agent, bot, automation, and slop. Learn what previously cost maintainers time or trust.
- Inspect automated-contributor controls.
Run the read-only inspector and record direct configuration, workflow use, labels, comments, or machine-targeted hidden content.
- Build an evidence packet.
Use one coherent scope, a minimal reproduction or motivation, appropriate tests, provenance, limitations, and a truthful disclosure when required.
- Brief the owner and wait for approval.
Present exact sources, uncertainty, likely repository-local consequences, and the concrete learning delta from any prior failure.
Repository attention levels
Assess live state immediately before contributing. A public adopter list is a discovery lead, not proof of current enforcement.
| Level | Observed condition | Practical response |
|---|---|---|
| High | Public auto-close or honeypot behavior, a prohibition on the proposed AI-generated content, or recent rejection of similar work. | Require explicit owner review and maintainer alignment before any write. |
| Elevated | AgentScan label, comment, or check mode; related labels; or a strict disclosure policy. | Prepare complete evidence and ask the owner before every write. |
| Policy | No visible AgentScan signal, but contribution or AI rules exist. | Follow the rule and preserve the ordinary human approval gate. |
| Unknown | No public signal observed. | Do not infer safety; apps, silent modes, custom labels, and browser tools may be invisible. |
A healthy contribution pattern
- The owner can explain why the problem matters and how it was discovered.
- The contribution follows the target's policy and does not claim disclosure as a universal exception.
- The change has one independently reviewable scope and project-appropriate validation.
- The owner has read the diff, understands the design, and accepts responsibility.
- The public text accurately states tests, limitations, provenance, and assistance when required.
- Review communication remains human-owned wherever the project requires it.
- Concurrent work stays within what maintainers invited and can reasonably review.
Behavior that warrants a stop and review
- Many unsolicited pull requests across unrelated repositories or multiple active contributions that outrun review.
- Fast issue-to-patch transitions without credible investigation evidence.
- Repeated rejected or closed contributions with no learning or maintainer agreement.
- Generic, templated, promotional, coercive, or confrontational public messages.
- A patch without reproduction, tests, project-specific rationale, provenance, or owner understanding.
- Targeting beginner, bounty, or security work at scale.
- Concealing AI assistance where disclosure is requested.
- Changing activity patterns solely to influence a detector.
These are accountability and review-cost concerns. They are not a recipe for staying below a score.
When a contribution is flagged
Hard stop for autonomous communication: freeze every queued reply to the thread, preserve the check, label, closing comment, report link, current diff, tests, and prior approvals, then notify the owner.
- Distinguish an algorithmic classification, repository-local action, community report, and GitHub platform enforcement.
- Find the documented appeal, correction, or reopening path.
- Let the owner choose silence, withdrawal, invited maintainer contact, or one evidence-based correction.
- Do not generate repeated denials, apologies, parallel issues, or social pressure.
Local help continues. The agent can preserve state, reproduce behavior, improve tests, review owner-authored text, and research policy within the user's authorization.
Legitimate repetitive work
Mechanical migrations can be valuable. Agree on a public plan before execution, define batch size and review cadence with maintainers, use a clearly identified bot or GitHub App if work is autonomous, keep runs idempotent, and stop on rejection, changed policy, or failed validation.
Read the complete repository reference: contribution-risk-methodology.md.