Dev Loop

The Dev Loop executes the implementation plan phase by phase. It repeats until all phases are complete, then validates the output and hands off to the Validate Loop.

What it does

Each iteration:

1. Identifies the next unchecked phase in tasks.md
2. Executes every task in that phase, checking each off as it completes
3. Stops after the phase and waits for your review
4. Commits staged changes with a Conventional Commits message
5. Checks whether all phases are complete; if not, repeats
6. Runs validators when all phases are done
7. Chains to the Validate Loop on success

How to start it

The Dev Loop fires automatically when the Plan Loop completes. You can also run it manually:

/dmx/run-loop dev

The human gate

After each phase, the loop pauses. Review the code changes. When you're satisfied:

/dmx/loop-continue

The loop checks whether all phases are complete. If not, it starts the next phase immediately. When all phases are done, validators run.

repeat_until: all_phases_complete

This property tells the orchestrator to iterate the skill sequence until the condition is met. After each iteration, dmx reads tasks.md and checks for unchecked - [ ] items. If any remain, the loop repeats. When all items are - [x], validators run.

Skills

implement-next-phase executes all tasks in the next unchecked phase of tasks.md. The AI implements the code, checks off each task, and stops. It never continues to the next phase automatically.

commit stages changes, updates the memory bank with light progress notes, and creates a Conventional Commits message referencing the current ticket. It never auto-pushes.

For finer control, you can customize the Dev Loop to use implement-next-task instead of implement-next-phase. This implements one task at a time, which is useful when individual tasks are large or risky.

Validators

run_tests

Detects and runs your project's test suite. Detection is automatic:

• Python + pyproject.toml + uv.lock → uv run pytest -q
• Python + pyproject.toml → pytest -q
• Node.js + package.json test script → npm test --silent
• Makefile with test: target → make test

Check	Required
tests_pass	Yes
coverage_threshold	No

If your test command isn't detected, place a custom validators/run_tests.py at your repo root. See Writing Validators.

spec_adherence

Grades the structured validation-report.json that the validate skill writes from its own diff analysis — not free-text skill output. See Validators for the full check logic.

Check	Required
scope_matches_spec	Yes
edge_cases_addressed	No

If the report is missing or stale, every check fails and the loop pauses with a message to re-run /dmx/validate first.

Loop config

name: dev
skills:
  - implement-next-phase
  - commit
trigger:
  type: on_complete
goal_state: "All phases implemented and committed, no unchecked tasks remaining"
repeat_until: all_phases_complete
validators:
  - tool: run_tests
    checks:
      - name: tests_pass
        required: true
      - name: coverage_threshold
        required: false
  - tool: spec_adherence
    checks:
      - name: scope_matches_spec
        required: true
      - name: edge_cases_addressed
        required: false
on_optional_failure: warn
failure_handling: pause
human_gate: true
on_complete:
  on_success:
    trigger_loop: validate

Tips

Review each phase before continuing. The AI implements the full phase in one go. Read the diff before calling /dmx/loop-continue. Fix anything wrong before the next phase runs.

Let tests fail naturally. If run_tests fails, the loop pauses. Fix the failing tests, then continue. Don't bypass the validator. Test failures are exactly what the harness is designed to catch.