Annotated, sanitized terminal moments — a guard firing, a loopback, an acceptance gate holding. Behavior only: no rule contents, no step files, no client code.
Blocked by its own product's blueprint guard, the model refused to fake the lockfile — then diagnosed and shipped the real fix.
A misfired autonomous-loop heartbeat told the wrong session to take over a live campaign. It declined — repeatedly.
Blocked from generating an auth magic link, the model explained why its own classification had been wrong.
One shell command would have flipped a stale status field — through a guard. The model flagged it to the human instead.
Blocked mid-campaign by a guard with a real bug, the model routed around the path — not the rule — and filed the lesson for every future run.
The model deviated from an approved spec for a defensible reason — and led its final report with the disclosure.
A campaign session hit the context limit twice and kept going — closing with a versioned release shipped to every install.
A single campaign command took a product from research through five built-and-verified phases across 19 wall-clock hours — surviving a context limit on the way.
A release-hook bug surfaced late in a session — so the model wrote the complete handoff into memory before closing.
A single campaign session ran a 52-instance fleet to deliver a full back office — with dedicated reviewer agents in the loop.
One session audited and corrected the package dependency direction in six separate multi-tenant SaaS codebases.
The April build sprint, as recorded by its own session metadata.
Before touching a repo other sessions work on, there was a record of what had been said — a cue at the start, searchable afterward, surviving restarts.