build-sweep

name	build-sweep
model	sonnet
description	Produces a verified build/test status report and auto-fixed modules by sweeping all Go modules in a workspace. Use when: "run a sweep", "check all builds", "nightly sweep", "fix all failing modules", "workspace health check".
category	system-health
triggers	["run a sweep","check all builds","nightly sweep","fix all failing modules","workspace health check"]
tier	1
agents	["primary"]
tool_dependencies	["file_system"]
inputs	[{"name":"workspace_path","type":"string","description":"Root directory containing Go modules to sweep","required":true}]
outputs	[{"name":"sweep_report","type":"text","format":"markdown-table","description":"Verified status table of all modules with build/test results, fixes applied, and commit hashes"}]

Build Sweep Skill

Philosophy

A healthy workspace is one where every module compiles and every test passes. The build sweep is not a debugging session — it is a systematic sweep that treats each module as an independent unit, diagnoses failures mechanically, and applies minimal fixes. The sweep must never trust self-reports: every fix is verified by re-running the exact build and test commands that originally failed.

When to Use

• As a scheduled nightly task to catch drift across a multi-module workspace
• Before a release to verify all modules are green
• After a large refactor or dependency update that may have broken multiple modules
• When onboarding to a workspace to establish a health baseline

Workflow

Phase 1: Detect

1. Find all directories containing go.mod under the workspace path (skip vendor/, .git/, node_modules/, .claude/worktrees/)
2. For each module, run go build ./... and go test ./...
3. Collect results into a table: module | build status | test status | error summary

Phase 2: Diagnose and Fix

For each failing module, dispatch a parallel agent (max 5 concurrent) to:

1. Read the failing build/test output
2. Read the relevant source files to diagnose the root cause
3. Implement the minimal fix — do not refactor, do not add features
4. Re-run that module's build and tests to confirm green

Phase 3: Verify

After ALL agents complete, independently verify each fix:

1. Run the tests yourself — do NOT trust agent self-reports of completion
2. Check actual file diffs with git diff
3. Confirm no duplicate imports, wrong formats, or regressions
4. Run the full suite one final time to catch cross-module integration issues

Phase 4: Commit

If all fixes pass verification:

1. Stage only the fixed files (not unrelated changes)
2. Commit each module separately: fix(sweep): [module] — [one-line description]

Phase 5: Report

Output a verified summary table: module | failure type | fix applied | verification status | commit hash

Output

The skill produces:

1. A markdown table summarizing all module statuses
2. Git commits for each fix applied (one per module)
3. A final pass/fail verdict for the entire workspace

Examples

Clean workspace:

/sweep
> All 29 modules pass build and test. No action needed.

Two failures detected and fixed:

/sweep
> Phase 1: 27 pass, 2 fail (DIP: missing interface method, MCP: stale test count)
> Phase 2: Dispatched 2 agents
> Phase 3: Both fixes verified independently
> Phase 4: 2 commits created
> RESULT: 29/29 modules green

Edge Cases

• Module with no test files: Report as SKIP, not FAIL — [no test files] is not an error
• Flaky tests: If a test fails then passes on rerun, report as FLAKY with the original error
• Circular dependency between modules: Run builds in topological order if dependency info is available
• Module requires external service: Skip with a note if go test fails due to missing database/service connection
• Agent fix breaks another module: Resolve the cross-module conflict before committing either fix

Anti-Patterns

• Trusting agent self-reports: Agents may report "all tests pass" without actually running them. Always verify.
• Committing untested fixes: Never commit a fix without running the full build and test suite for that module.
• Fixing unrelated issues: The sweep fixes build/test failures only. Do not refactor, add features, or clean up code.
• Running the full workspace test suite after each individual fix: Too slow. Verify each module independently, then run one final integration pass.
• Force-pushing sweep commits: Sweep commits are always local. Never push without explicit user confirmation.