Skip to content

Workflow

HELIX organizes software delivery into six phases, governed by an authority hierarchy that ensures upstream artifacts always take precedence.

Phases

PhasePurposeKey Commands
FrameDefine the problem — vision, PRD, feature specs, user storieshelix frame
DesignArchitect the solution — ADRs, solution designs, technical designshelix design
TestWrite failing tests that define behavior(implicit in build)
BuildImplement code to make tests passhelix build, helix run
DeployRelease to production with monitoring(future)
IterateLearn, review, align, and improvehelix review, helix align

Authority Order

When artifacts disagree, HELIX uses this hierarchy:

  1. Product Vision — what the product should become
  2. Product Requirements — what must be built
  3. Feature Specs / User Stories — detailed behavior
  4. Architecture / ADRs — structural decisions
  5. Solution / Technical Designs — how to build it
  6. Test Plans / Tests — verification criteria
  7. Implementation Plans — build sequencing
  8. Source Code — current state (not source of truth)

Higher layers govern lower layers. Source code reflects what exists, not what should exist. If a higher layer is missing, HELIX does not infer intent from lower layers.

The Supervisory Loop

helix run is the supervisory autopilot. It continuously:

  1. Checks the tracker for ready work
  2. Selects the highest-leverage bounded next action
  3. Executes one pass (build, design, polish, align, etc.)
  4. Reviews the result with cross-model verification
  5. Repeats until human input is needed or no work remains

The autopilot uses the principle of least power — prefer refining existing artifacts over creating new ones, prefer polishing issues over implementing from incomplete specs.

When HELIX Stops

HELIX stops and asks for human input when:

  • Authority is missing (no governing spec for the work)
  • Ambiguity cannot be resolved from existing artifacts
  • Product judgment or prioritization is needed
  • A decision requires stakeholder approval
  • No actionable work remains in the tracker

Trigger Rules

ConditionHELIX Action
Ready governed work existsbuild — bounded implementation
Specs changed with open issuespolish — refine issues first
Requirement change detectedevolve or design before build
Queue drainedcheck — decide next action
Periodic intervalalign — reconciliation audit
Build completereview — fresh-eyes cross-model review

Tracker-First Execution

The tracker (ddx bead) is the steering wheel. Users direct agents by:

  • Creating issues with priorities and acceptance criteria
  • Setting dependencies between work items
  • Approving or blocking work through tracker state
  • Closing completed work

Agents read tracker state and execute. The mental model:

User <-> Tracker <-> helix run (background)

Epic Focus Mode

When helix run picks an epic, it stays focused:

  1. Decompose the epic into child tasks
  2. Implement children one at a time
  3. Review the epic when children complete
  4. Only move to the next scope after the epic finishes or is blocked

Cross-Model Verification

When configured, HELIX alternates AI models for adversarial review:

  • Primary agent handles reasoning-heavy work (frame, design, evolve, align, review)
  • Alternate agent handles mechanical work (build, polish, check)

This ensures different models review each other’s output, catching blind spots that self-review misses.