Byte-Interval Mediation: Read-Receipt Write Admission for Concurrent LLM Coding Agents

Concurrent coding tools need to distinguish a still-matching edit target from a still-current observation of the context used to propose that edit. We study a public implementation that persists scope-bound byte observations, validates them before admitting a patch, rebases disjoint observations after length-changing edits, and journals single-file publication. The mechanism has explicit cross-resource and target-resource modes; its default whole-file write API permits blind replacement and must not be confused with its read-covered patch API. A frozen functional corpus contains 156 schedules, executed under three admission policies and two validation scopes. All 936 executions match an independently implemented finite oracle, with physical files and 1,300 committed journal operations reconciled. In the 144-schedule narrow-read grid, byte observations and an exact-match admission ablation both accept 66 second writes per scope, versus 36 for whole-file observations. Thus this grid demonstrates reduced coarse-grained rejection but no additional accepted-outcome benefit over exact matching. Other cases distinguish validation of previously observed context from exact matching and exhibit semantic failures when relevant dependencies were never observed. Fifteen additional behavioral checks cover selected point boundaries, repeated edits, repair, coverage, expiry and retirement. We position the mechanism relative to STORM, CoAgent, AgentRoom and ByteFence without claiming priority or superiority. A separate 135-batch shared-host study finds little accepted-throughput scaling from one to eight workers under its synchronized-round boundary, while five abrupt-process probes verify selected recovery outcomes. A retained scripted-client integration case illustrates end-to-end tool use and repair. No autonomous-agent performance experiment is reported.