check-work

Work Validation

Overview

Run all quality gates after implementation is complete. Each gate is tracked as a task. If any gate fails, fix it — using subagents for parallel resolution when possible.

Announce at start: "I'm using the zenflow:check-work skill to verify this work."

Command Discovery

Before running gates, detect the project's toolchain by reading package.json, Makefile, pyproject.toml, or equivalent. Look for:

• Lint command — lint script in package.json, make lint, ruff check, etc.
• Lint fix command — lint:fix, make lint-fix, ruff check --fix, etc.
• Format command — format script, make format, black --check, etc.
• Format fix command — format:fix, make format-fix, black ., etc.
• Test command — test script, make test, pytest, etc.

If CLAUDE.md documents the project's commands, use those. Otherwise discover from config files.

The Process

Create all 5 tasks via TaskCreate upfront, then execute them in order.

Gate 1: Pass Lint

• Mark in_progress via TaskUpdate
• Run the project's lint command
• If failures: run the lint fix command first. If errors remain, launch a subagent to fix them manually.
• Confirm zero errors before marking completed

Gate 2: Pass Format

• Mark in_progress via TaskUpdate
• Run the project's format check command
• If failures: run the format fix command. If errors remain, launch a subagent to fix them.
• Confirm zero errors before marking completed

Gate 3: Pass Tests

• Mark in_progress via TaskUpdate
• Run the project's full test suite with no filters or specific file paths. Running a subset is NOT acceptable — partial test runs hide regressions.
• If failures: launch subagent(s) to fix failing tests. Each failing test file can be a separate subagent.
• After fixes, re-run the full test suite (not just the fixed files)
• Confirm zero failures before marking completed
• Note the test count — report if it changed (new tests added or tests removed)

Gate 4: Update Documentation

• Mark in_progress via TaskUpdate
• Review what changed in this session (use git diff against the starting state)
• Update relevant documentation:
  • README.md — if new features, commands, or setup steps were added
  • CLAUDE.md — if new conventions, patterns, or project structure changed
  • ARCHITECTURE.md — if architectural changes were made
  • Inline doc comments — if exported symbols were added or changed
• If no documentation updates needed, note why and mark completed
• If updates needed: make them, then mark completed

Gate 5: Write Journal Entry

• Mark in_progress via TaskUpdate
• Invoke field-notes:write using the Skill tool
• The journal plugin handles the format and file location (.claude/journal.jsonl)
• If field-notes is not installed, append directly to .claude/journal.jsonl using the full schema:

{
  "timestamp": "YYYY-MM-DDTHH:MM:SS.000Z",
  "workedOn": "plan: resources/plans/NNN-name.md | issue: #123 | description",
  "branch": "current git branch name",
  "type": "work",
  "origin": "primary",
  "skill": "zenflow:check-work",
  "outcome": "completed",
  "summary": "One sentence: what was accomplished",
  "details": {
    "filesModified": ["path/to/file1.ts", "path/to/file2.ts"],
    "filesCreated": ["path/to/new-file.ts"],
    "testsAdded": 0,
    "testsModified": 0
  },
  "issues": "Problems encountered (empty string if none)",
  "blockers": "What stopped progress (empty string if none)",
  "workarounds": "Temporary solutions used (empty string if none)",
  "insights": "What worked well, surprising discoveries, useful patterns",
  "wouldDoDifferently": "Hindsight improvements for next time",
  "feedback": "Observations about the workflow or process",
  "duration": "approximate time spent"
}

• Mark completed

Failure Resolution

When a gate fails:

1. Try auto-fix first — lint fix, format fix for gates 1-2
2. Launch subagent for manual fixes — if auto-fix doesn't resolve it
3. Re-run the gate — confirm it passes after the fix
4. If stuck after 2 attempts — stop and report the failure to the user instead of looping

Parallel Execution

Gates 1 and 2 (lint + format) can run in parallel since they're independent. Gate 3 (tests) should run after 1 and 2 are clean. Gates 4 and 5 (docs + journal) can run in parallel after tests pass.

Remember

• Every gate must pass — no skipping
• Auto-fix before manual fix
• Use subagents for independent failures
• Report test count changes
• Stop after 2 failed attempts on the same gate