Open Fusion Quality Gates

Open Fusion Quality Gates

Required Context

Before finalizing a change, inspect:

• modified files;
• relevant specs and ADRs;
• existing package scripts once a package manifest exists;
• test output from the current task.
• recent review/audit reports when the task is addressing review feedback or preventing known regressions.

Use this skill together with open-fusion-code-quality when reviewing design risks.

Minimum Gate

A change should not be considered done until these are true or explicitly reported as not runnable:

• targeted tests pass;
• relevant broader tests pass;
• typecheck passes;
• lint passes;
• formatting is consistent;
• docs/specs/ADRs are updated when behavior or decisions changed;
• public API compatibility is preserved or the breaking change is documented;
• secrets are not logged, returned, or committed;
• public error envelopes do not expose internal configuration identifiers such as internal model ids, provider model ids, provider config keys, env var names, stack traces, or registry details;
• new provider/network behavior is mocked unless a live integration run was intentional.

Suggested Commands

Discover commands from the repo before inventing them:

• inspect package.json;
• inspect workspace tooling files;
• use package manager already present in lockfiles.

Typical commands once available:

npm test
npm run test:e2e
npm run lint
npm run typecheck

Adapt to pnpm, yarn, or npm based on the repository.

Review Checklist

Check:

• Does a failing test protect the changed behavior?
• Is the smallest responsible module changed?
• Are controllers thin?
• Are provider details isolated behind adapters?
• Are config and secrets handled only by config/secret services?
• Are OpenAI-compatible envelopes stable?
• Are streaming chunks valid and terminated?
• Are timeouts and limits enforced by backend code?
• Is duplicate logic removed or intentionally tolerated?
• Are errors useful without exposing sensitive data?

Known regression checks from PR review audit:

• numeric request validators reject non-finite numbers;
• payload, message-count, and message-content limits are enforced where required;
• shared client/request domain shapes are not duplicated across services and Express typings;
• provider-supplied tool calls and delegate messages are runtime-validated before typed use;
• internal gateway misconfiguration maps to internal_error, not provider_error;
• public internal_error messages are generic and do not include unresolved internal model ids or other internal configuration identifiers;
• validation/model/tool-policy failures produce structured failure logs;
• depends_on: [] does not misclassify final delegate calls;
• parallel delegation failures cancel in-flight work or document/test ignored late results;
• implemented specs and PRD links do not still say "next", "draft", or "upcoming";
• package runtime requirements such as engines.node are explicit when dependencies require them;
• tests restore process.env by mutating the existing object.

Risk Reporting

When finishing, report:

• checks run;
• checks not run and why;
• any residual risk;
• files changed at a high level.

Do not claim full verification if only static edits or documentation checks ran.