When the server recycled, the agent deleted the feature to fix the build

The repair was correct for the world the agent could see. An import pointed at `src/generated/api_client.ts`, nothing on disk resolved it, and nothing in the repository recorded that the file was ever meant to exist ^[5]. Removing the import made the build green again ^[3].

Look at what the session logs actually certified. They showed a generation step that succeeded, a test run that passed, and a commit that referenced the generated file by name ^[4]. All three entries were true. None of them was evidence that the file was inside the commit, because the agent had written it to the workspace and a fresh clone had no way to learn it had existed ^[5].

The reproduction is the only step in the postmortem that produced new information. A clean checkout failed at once while the same tree, left dirty, still passed ^[6], which means pass and fail were a property of the local disk rather than of the history ^[2]. Between the last signal that looked like success and the first signal that was true, twelve hours of unattended running elapsed ^[1]. The author's reading is that free servers are ephemeral by design and the workflow was the defect, because the agent was allowed to treat local disk as durable storage and given no way to verify that its world had been reset ^[7]. He likens the workspace to a cache and the restart to an invalidation event the agent never noticed ^[8].

The remedy is smaller than the diagnosis. A manifest at `.agent/artifacts.json` declares required paths; a preflight script refuses to run on a dirty tree, checks that each declared artifact exists, prints the regeneration command when one is missing, clears the build directory and runs the build ^[11]. The example manifest carries three entries and `pnpm generate` as the regeneration command ^[12]. Notice that the existence check asks whether the path is present on disk, which is the same question that would have answered yes the night before the restart ^[3]. The value is not detection at authoring time. It is that the drill is bound to the restart, so the reset produces a hard failure with a clear message rather than a silent deletion ^[10].

The author concedes the obvious alternative: for small projects, generated files should simply be committed ^[13]. That is the cheaper fix, and the manifest is for the cases where it does not apply. What makes this worth reading twice is the economics underneath. When tokens are free, agents run long loops and retry endlessly, which makes them less likely to checkpoint state, and the bottleneck moves from compute cost to state persistence ^[9]. Retries are billed to nobody. Untracked artifacts are billed to whoever clones next.