Problem to Fix
Code Cannon's non-negotiable workflow rules — work only on an approved feature branch, one change per ticket, always open a pull request, pass review before merging — currently live as written instructions that each developer's AI agent is asked to follow. Different agents on different tools honor them to different degrees, and a rule written as a suggestion can simply be skipped. Nothing actually stops a rule from being broken.
Why it Matters
Consistent enforcement of business rules across a team using different AI tools is the entire point of Code Cannon. Instructions that agents "should" follow drift — that is the exact problem Code Cannon was built to escape. For rules that must never be broken (especially in compliance-sensitive work), enforcement has to sit outside the agent so it binds everyone identically, no matter which tool they use. This backbone is also a prerequisite for extending Code Cannon to new domains: it becomes the reusable pattern for how any future skill group enforces its own rules.
General Approach
Move the hard, non-negotiable rules out of written skill instructions and into automated gates that live in the code repository and on the hosting platform — checks that run automatically and block the action when a rule is violated, regardless of which agent or person triggered it. Written skills continue to guide the normal flow; the automated gates become the backstop that makes compliance mandatory rather than voluntary. Document the setup as a repeatable pattern a future skill group can adopt.
Complexity
Verification / QA effort: significant
Requires standing up repository-side and platform-side checks and confirming they block violations when driven by different agent tools, while ensuring they don't obstruct the legitimate happy path. Part of the effort is design: deciding exactly which invariants belong in local repository checks versus platform-side merge gates.
Acceptance Criteria
- Committing or pushing outside an approved feature branch is automatically rejected, regardless of which agent or tool is used.
- Merging a change that has not passed the required review is blocked at the hosting-platform level, not merely discouraged in instructions.
- The normal, compliant workflow proceeds without new friction for developers who follow the rules.
- The enforcement setup is documented as a repeatable pattern that a future skill group can adopt.
Problem to Fix
Code Cannon's non-negotiable workflow rules — work only on an approved feature branch, one change per ticket, always open a pull request, pass review before merging — currently live as written instructions that each developer's AI agent is asked to follow. Different agents on different tools honor them to different degrees, and a rule written as a suggestion can simply be skipped. Nothing actually stops a rule from being broken.
Why it Matters
Consistent enforcement of business rules across a team using different AI tools is the entire point of Code Cannon. Instructions that agents "should" follow drift — that is the exact problem Code Cannon was built to escape. For rules that must never be broken (especially in compliance-sensitive work), enforcement has to sit outside the agent so it binds everyone identically, no matter which tool they use. This backbone is also a prerequisite for extending Code Cannon to new domains: it becomes the reusable pattern for how any future skill group enforces its own rules.
General Approach
Move the hard, non-negotiable rules out of written skill instructions and into automated gates that live in the code repository and on the hosting platform — checks that run automatically and block the action when a rule is violated, regardless of which agent or person triggered it. Written skills continue to guide the normal flow; the automated gates become the backstop that makes compliance mandatory rather than voluntary. Document the setup as a repeatable pattern a future skill group can adopt.
Complexity
Verification / QA effort: significant
Requires standing up repository-side and platform-side checks and confirming they block violations when driven by different agent tools, while ensuring they don't obstruct the legitimate happy path. Part of the effort is design: deciding exactly which invariants belong in local repository checks versus platform-side merge gates.
Acceptance Criteria