synthesis-code-integration
Installation
SKILL.md
Code Integration
When multiple people contribute to a project built through synthesis coding, integration is fundamentally different from a standard open source merge workflow. This skill defines how it works.
Binding rules
- Adopt and adapt by default; never blind-merge. Branch fresh from current
main, bring changes in selectively, and fix issues before they reachmain, because the contributor built against a codebase that has since moved. Lighter phases are earned by a contributor's record and lost on a regression. - Zero accepted failures. Every test passes and every lint is clean, including failures you did not cause, fixed in the same branch; a suite with known failures cannot tell you whether a change is safe.
- Every contribution passes Gates 1 to 4 (completeness, security, architecture, project-specific standards) before integration. An undocumented standard is the lead's to document, not the contributor's fault.
- Never merge another long-lived branch into a PR branch. Put the commit on that branch as a separate operation; only a promotion PR (such as
developtomain) carries another branch's work. - After every cherry-pick, look for silent reversions and run the full suite. A conflict-free cherry-pick can bring back an old version of a file; never trust the GitHub diff for an old branch.
- Never force-push
mainto the staging branch. Check for divergence and merge instead, so contributors' unintegrated work survives. - Run the pre-squash PR manifest check before every squash merge, and write no changelog entry for a feature you have not found in the code on the target branch.
- Guard every deliberate upgrade of a critical config value (effort level, token budget, model, retries) with a test that explains why the value matters.
- Ground every technical reply in code. Separate what you observed from what you conclude, and say "I don't know" when you don't.
- Credit contributors with
Co-authored-bytrailers that match the integration intensity, because contribution graphs feed careers. - Fix first, talk later when the fix is quick and delay is risky.
- Production deploys always require explicit human approval.