AI Agent Contribution Directives (Draft)

Scope

• General directives for any AI-assisted contribution (patch generation, PR drafting, review assistance).

Mandatory Rules

1) No unverified claims - Do not state "tested", "validated", "fixed", "works on X" without evidence.

2) No scope creep - Do not add unrelated changes (formatting, refactors, opportunistic cleanups). - One issue per PR; if multiple issues are detected, split the work.

3) Preserve behavior unless explicitly requested - Do not change defaults, semantics, or feature gates beyond what is required.

4) Minimize diff and surface area - Prefer local changes. - Avoid new files/layers/frameworks unless explicitly required.

5) Maintain reviewer ergonomics - Changes must be easy to review, easy to revert, and logically grouped. - Avoid mechanical churn.

Required Output Structure (for any proposed change)

• Context:
• Problem:
• Root cause:
• Change:
• Risk:
• Validation commands:
• Evidence required (outputs/logs to paste):

Validation Protocol When the AI Cannot Run Tests

• Provide exact commands for:
• configure/build
• relevant runtime checks
• Request concrete artifacts:
• configure output excerpt
• build output tail (or error)
• OS/compiler/dependency versions
• runtime log lines demonstrating behavior

Build / Detection Constraints

If proposing any detection change, the AI must specify:

• detection criteria and ordering
• flag/macro names produced
• where those flags are consumed
• how consistency is ensured (headers/cflags/libs/call sites)

Prohibited Behaviors

• Inventing test results, environment details, or tool outputs.
• Quietly changing behavior without explicit request.
• Introducing a new compatibility layer because it "feels cleaner".
• Large stylistic rewrites that increase diff noise.

Review Interaction Protocol

• Apply maintainer feedback with minimal churn.
• Do not introduce additional improvements in the same review iteration.
• If the approach is rejected, stop and propose the smallest alternative.