Cli build sync pipeline

cli-build-sync-pipeline

A repeatable consistency pass that brings a working tree to a known-good state before a commit or PR. The pipeline itself — regenerate → format → lint → build — is universal and uncontroversial (the pre-commit framework codifies the same chain). The non-obvious, load-bearing step is the last one: read the IDE's diagnostics through its MCP server as a second opinion, because xcodebuild / ./gradlew assemble / swift build routinely miss issues the editor's Issue Navigator shows.

When to invoke

• Before declaring a change ready or opening a PR.
• On /sync.
• When wiring format-on-write or pre-commit automation for a repo.

Announce on invoke: "Using cli-build-sync-pipeline to regenerate, format, lint, build, and take an IDE second-opinion diagnostics pass."

The five steps

#	Step	iOS / Xcode	Android / Gradle	Node	Go
1	Regenerate project	xcodegen generate (or Tuist)	gradle wrapper sync	—	—
2	Format	swiftformat .	./gradlew ktlintFormat	prettier --write	gofmt -w .
3	Lint	swiftlint --strict	./gradlew detekt	eslint	golangci-lint run
4	Build	xcodebuild … build	./gradlew assembleDebug	pnpm build	go build ./...
5	IDE second opinion	XcodeListNavigatorIssues (native Xcode MCP)	Android Studio / JetBrains MCP inspections	—	—

#!/usr/bin/env bash
# scripts/sync.sh — invoked by /sync (step 5 is a separate MCP tool call, see below)
set -euo pipefail
case "${1:-auto}" in
  ios)     xcodegen generate; swiftformat .; swiftlint --strict; xcodebuild -scheme "$SCHEME" -destination "$DEST" build -quiet ;;
  android) ./gradlew ktlintFormat detekt assembleDebug lintDebug ;;
  node)    pnpm install --frozen-lockfile && pnpm format && pnpm lint && pnpm build ;;
  go)      go mod tidy && gofmt -w . && golangci-lint run && go build ./... ;;
esac

Step 5 — the IDE second opinion (the core insight)

CLI build output ≠ Issue Navigator content. xcodebuild runs without the editor's live SourceKit-LSP indexing, so warnings like unused let/var, "Will never be executed," and static analyzer findings ("Potential leak") are frequently emitted only in the IDE. After step 4 build succeeds, query the editor's MCP server for issues it sees and the build log didn't surface.

For Apple platforms there are two complementary MCP servers — call them by their real tool names:

• Native Xcode MCP (Xcode 26.3+). mcp__xcode__XcodeListNavigatorIssues returns the Issue Navigator state (compiler diagnostics from SourceKit-LSP, the static analyzer, package-resolution warnings). mcp__xcode__XcodeRefreshCodeIssuesInFile re-runs diagnostics for one file.
• XcodeBuildMCP. Wraps xcodebuild into structured JSON build/test errors — agent-friendly where the raw log is megabytes. Verified tools include build_sim, build_run_sim, test_sim, list_schemes, show_build_settings, discover_projs, clean, and the session defaults pair session_show_defaults / session_set_defaults (set scheme + simulator once instead of re-passing them every call).

A robust pass calls both: XcodeBuildMCP for the build verdict, the native Xcode MCP for the Navigator state. For Android the closest analog is the JetBrains / Android Studio MCP exposing inspections; if that server isn't installed, document the gap rather than pretending the CLI lintDebug covers it.

Format-on-write via PostToolUse (verified)

To keep files formatted as Claude writes them, use a PostToolUse hook — not PreToolUse. Verified against the hooks reference: PreToolUse fires before the tool and can block it; PostToolUse fires after a tool call succeeds — which is what you want when formatting what was just written. Hooks receive their input as stdin JSON; the edited path is tool_input.file_path (there is no CLAUDE_TOOL_INPUT_FILE_PATH env var — read it from stdin):

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          { "type": "command",
            "command": "f=$(jq -r '.tool_input.file_path // empty'); case \"$f\" in *.swift) swiftformat \"$f\" ;; *.kt) ktlint -F \"$f\" ;; *.ts|*.tsx) npx prettier --write \"$f\" ;; esac" }
        ]
      }
    ]
  }
}

The matcher is matched against tool names (Write|Edit), confirmed in the hooks docs.

Pitfalls (verified)

• Formatting generated files breaks codegen output. Exclude *.generated.swift, build/, .build/, DerivedData/, Gradle plugin output. Enforce via .swiftformat / ktlint config so the hook and the script both respect it.
• PreToolUse for format-on-write is wrong. It fires before the write and can block the edit; PostToolUse formats what was actually written.
• XcodeBuildMCP wants session defaults first. Per its own instructions, call session_show_defaults before the first build/run/test; only call discover_projs when defaults are missing or wrong — calling it speculatively wastes a tool turn.
• Project regeneration (step 1) can be lossy. xcodegen generate rewrites *.xcodeproj; if the team hand-edits the project, make step 1 opt-in.

Related skills

• global-skills/claude-code-workflow/android-build-sync/SKILL.md — the Android-specific build-sync (Gradle + Compose + Android Studio inspections). This skill is the generic cross-platform version.
• global-skills/claude-code-workflow/claude-mcp-lesson-capture-pipeline/SKILL.md — captures the lessons that build/lint failures here tend to produce.
• global-skills/claude-code-workflow/mcp-orchestration-pattern/SKILL.md — decides which MCP server owns the build/diagnostics domain so step 5 routes to the right tool.

Sources

• Automate workflows with hooks (PostToolUse timing, matcher on tool names, stdin tool_input)
• XcodeBuildMCP (structured build output, session defaults)
• Xcode 26.3 agentic coding — native MCP Issue Navigator tools

---

Last verified: 2026-06-03. Confirmed live: PostToolUse fires after a tool succeeds and the edited path arrives via stdin tool_input.file_path (no CLAUDE_TOOL_INPUT_FILE_PATH env var); matcher filters on tool names; XcodeBuildMCP tools (build_sim, build_run_sim, test_sim, session_show_defaults) and native Xcode MCP tools (XcodeListNavigatorIssues, XcodeRefreshCodeIssuesInFile) confirmed present in the live tool surface. Flagged: the IDE-second-opinion step is emerging practice (well-evidenced for Xcode 26.3, not a named Anthropic feature); the Android inspection MCP is a documented gap, not a guarantee. Re-check after: any Claude Code minor release or new IDE-MCP tool surface, or by 2026-09-03. Decay risk: high. Found a drift? Run /skill-pattern-freshness-audit claude-code-workflow.