A generator given the wrong context does not stop
Scoping a context pack by project type saved 12.6% per subagent, not the 40% I guessed. The real risk is a thin pack: a missing surface makes a generator invent.
All posts
Two places that both think they are the truth
A durable source of truth and an in-memory one give the same answers until a process dies. Deciding which is authoritative is a design act, not a default.
Checkbox, radio, switch: three controls, three different questions
Checkboxes, radio buttons and switches encode different semantics — multiple choice, exclusive choice, and immediate effect. Choosing by appearance produces interfaces that mislead, and the accessibility story differs too.
The export engine that exported nothing
A stub is not an empty function. It returns plausible values, satisfies its interface and does no work, which is why emptiness checks never find it.
A retry queue with nothing to run it is a graveyard
A failed-sync flag is not a retry mechanism. Three parts are needed: a connectivity signal, an enumeration of what failed, and backoff with a ceiling.
The corpus your generator reads is production code
The corpus a generator reads is as load-bearing as the code it emits. A hand-edit to it has no provenance and is dropped by the next re-sync.
An agent drove a KMP app to the Play Store. The codegen was the easy half.
Autonomous codegen produces an app that compiles; what makes it one that ships is a gate stack refusing any claim that carries no evidence.
Five buttons, one hierarchy, and the rule that decides which
Material 3 ships five button styles plus the FAB. They form an emphasis hierarchy, and the selection rule is about how many actions compete on a screen rather than about how each one looks.
Syndicate to a 404 and you donate the SEO
A canonical that 404s when the syndicating platform fetches it does not transfer authority. Their copy becomes the original, and the order is the whole fix.
It compiles, the test is green, and the screen is a grey box
A structural shell passes an AST scan and a unit test while rendering nothing the mockup specified. Only a rendered check can tell the difference.
Last write wins is a decision, not a default
When two writes reach one record and no code decides between them, the policy is whichever ran last. Name the resolver so the choice becomes reviewable.
A screen nothing navigates to still compiles
A route with no caller compiles, registers and never runs. Reachability is a graph property, so no per-file check can see it and nothing fails.