Scaffold-to-Production Promotion Gate (Delivery)

Scaffold-to-Production Promotion Gate (Delivery)

1. 01

   Freeze the generated scaffold and tag the commit before any human edits land

   Record the agent session ID, the prompt that produced the scaffold, and the model version so the baseline is reproducible when a client disputes scope six weeks later.

2. 02

   Run the dependency and license inventory against the client's approved stack

   Generated projects routinely pull transitive packages the client's security team has not cleared; flag anything outside the agreed allowlist before the first deploy.

3. 03

   Assign a named human owner to every architectural boundary the agent touched

   Auth flows, payment paths, data schemas, and third-party integrations each get one accountable engineer, because no agent holds context across a retainer's full duration.

4. 04

   Re-run the agent's own test suite on a clean checkout, not the working directory

   Passing tests in a dirty tree hide missing migrations and uncommitted config; a fresh clone is the only honest signal of what the client actually receives.

5. 05

   Instrument the build with error and cost telemetry before handoff to the delivery pod

   Track token spend per feature branch and capture runtime exceptions in staging; agencies billing fixed-fee retainers need this data to price the next statement of work.

6. 06

   Document the regeneration path for every file the agent produced

   Note which modules can be safely regenerated from prompts and which are now hand-maintained, so a future model upgrade does not silently overwrite bespoke client logic.

7. 07

   Hold a 30-minute promotion review with the delivery lead and one engineer who did not build the feature

   The reviewer's job is to find the one thing the builder rationalized away; a second set of eyes catches the silent CSS and race-condition class of defects that never throw errors.