Teaching AI to read Xcode builds (tuist.dev)

Developers built a small probe into Xcode’s build plumbing and discovered that, unlike the flat text stream you see from xcodebuild, the real build engine (SWBBuildService) speaks a MessagePack-based, strongly typed protocol that emits discrete messages like BuildOperationTaskStarted/Ended, BuildOperationDiagnosticEmitted, BuildOperationTaskUpToDate and DependencyGraphResponse. Tools such as xcsift already try to make xcodebuild’s stdout AI-friendly by converting it to JSON or a token‑efficient TOON format, but that approach is fundamentally limited: the terminal stream flattens a rich dependency graph, per‑task timing, memory usage, cache hits, and structured diagnostics. By intercepting SWBProtocol messages (the team’s XCBLoggingBuildService) and leveraging Apple’s open‑sourced swift‑build message definitions, you can capture exact task IDs, parent/child relationships, start/end times, resource metrics, and explicit adjacency lists instead of guessing from logs. For AI/ML, that matters because agents would stop pattern‑matching on noisy text and start reasoning over precise, correlated build state and history. With structured messages an agent can reconstruct the dependency graph, compute critical paths and parallelism, pinpoint OOM or cache‑miss culprits, correlate recent edits with failures, and surface repeatable remediation steps (e.g., which target to clean or why DerivedData helped before). Realizing this end‑to‑end requires broader platform support from Apple, but the technical takeaway is clear: exposing build internals as structured telemetry transforms build debugging and performance analysis from heuristics into actionable, automatable insights for AI assistants.