Partitioned-edit consistency debt¶

Definition¶

Partitioned-edit consistency debt is the residue produced when one global change is executed as many independent local edits: each unit is corrected correctly in isolation and every local check passes, but the invariants that span units — type signatures, call contracts, shared schemas, naming conventions — are owned by no worker, so the aggregate change lands locally valid and globally inconsistent.

Explanation¶

The debt is structural rather than a defect in any worker. Partitioning is what makes parallel agent work cheap — each worker sees one unit and nothing else — and that same narrowness is what makes cross-unit invariants invisible. The reported instance is unusually clean: a per-file sweep that replaced loose types with specific ones achieved a hundred per cent fix rate on the lint backlog it was given, and produced sixty-one downstream type errors in files it never touched, because a signature tightened in one file is a contract broken in its callers. The general shape is that any property whose truth depends on two units at once falls outside every worker's field of view, so a green per-unit result is not evidence of a green whole. Three mitigations follow directly, and each costs some of the parallelism the partition bought: run a global check after the fan-out and treat its output as a second backlog; partition along the invariant instead of along the file, grouping units that share a contract into one task; or reserve for a coordinating pass the classes of edit known to cross boundaries, letting workers handle only the local ones. The trap for the unwary is the metric — a completion rate measured per unit will look perfect at exactly the moment the debt is largest.

Key Properties¶

The debt arises from the partition itself, not from any worker doing its job badly
Only invariants spanning two or more units are affected; purely local corrections are unharmed
Per-unit success metrics stay green while the aggregate is broken, so the failure is silent at the dashboard
Detection needs a global check run after the fan-out, not during it
Every mitigation — global re-check, partitioning by invariant, or a coordinating pass — trades back some parallelism

Relationships¶

Stateless worker fan-out — is the price of that shape's scaling, since shared-nothing workers are by construction blind to anything that spans their units
Retry escalation ladder — partly answers it, because the exploration rung's licence to read neighbouring files is what lets a worker see a contract it would otherwise break

Applications¶

Planning any automated codebase-wide refactor — budget a global type or contract check after the parallel pass, and decide up front which edit classes are too cross-cutting to hand to per-file workers.

Sources¶

https://briandouglas.me/posts/2026/03/15/sweeper-why-agent-infrastructure-matters/