adversarial-reviewRun the cross-model adversarial review gate on a pull request: a *different* model from the one that wrote the code reviews the open PR adversarially and posts its findings as PR comments, before the change is declared done. Covers reviewer selection (the fallback chain), the auditable PR-comment pr
agent-tiered-planningPlan and dispatch work across a mixed-capability pool of coding agents (humans, frontier models, cheaper models) by labeling every issue with an agent-capability tier orthogonal to its effort estimate, publishing a playbook cold agents read before picking up work, and enforcing an escalation rule wh
design-direction-lockAfter a design direction has converged through mockup iteration, freeze it into a short normative artifact (docs/design-direction.md) carrying the references, visual language, motion rules, and product truths — then stamp every design-facing issue with the lock revision and only the rules that apply
implementation-disciplineHow a coding agent conducts itself while implementing a change — the four disciplines that stop plausible-but-wrong work before it is written: surface assumptions and competing interpretations before coding (present them and stop rather than pick one silently); build the minimum that solves the stat
iterative-acceptanceDevelop a system on top of nondeterministic tools (LLMs, OCR, vision models) when the expectations cannot be fully specified up front — by iterating real test runs against a frozen baseline with predictions registered BEFORE each run and scored honestly after it. Covers the frozen-baseline invariant
silent-failure-designDesign and audit controls so that a control which never ran is distinguishable from one that ran and passed. Covers the failure class where absence wears the appearance of success — CI that was never scheduled, a watchdog whose exception is swallowed, a guard that cannot start and therefore denies n